An online appearance defect identification method and system after hot pressing

Abstract

This invention relates to the field of industrial online optical inspection and machine vision inspection technology, specifically disclosing a method for layer-based attribution identification of online appearance defects after hot-press bonding. The method first performs initial screening and image acquisition on the entire area of ​​the hot-pressed product to extract candidate abnormal regions. Then, for the same candidate abnormal region, surface-sensitive images, interlayer-sensitive images, first focal plane images, and second focal plane images are acquired separately. After image registration, differences in surface response, interlayer response, and focal plane sharpness are extracted, and based on these features, the defect is determined to be a surface defect, an interlayer bonding defect, or an anomaly requiring further verification. This method can distinguish between surface contamination, protective film defects, and interlayer bonding defects, reducing the risk of misjudgment and missed detection, and improving the targeted nature of online detection and handling.

Description

Technical Field

[0001] This invention relates to the field of industrial online optical inspection and machine vision inspection technology, and more specifically, to an online method and system for identifying appearance defects after hot-press bonding. Background Technology

[0002] Hot-press bonding is widely used in the manufacturing of display panels, touch modules, functional films, and electronic components. This process applies predetermined temperature and pressure to two or more layers of functional materials, causing them to bond tightly at the interface, thus creating a composite product with specific optical, electrical, or mechanical properties. After hot-press bonding, the composite product typically undergoes online visual inspection to identify and reject defective products with appearance abnormalities, preventing them from flowing into downstream processes.

[0003] Existing online appearance inspection solutions after hot pressing typically use an industrial camera with fixed-mode lighting to scan and image the entire surface of the product. Then, image processing algorithms mark areas with brightness deviations in the image and classify suspected anomalies based on the brightness, area, or shape characteristics of these areas.

[0004] However, in composite products after hot-press bonding, the suspected anomalies appearing in the appearance images may originate from either the outermost surface of the product (e.g., particulate contamination or localized damage to the protective film) or from within the bonding layers (e.g., interlayer bonding defects such as bubbles, delamination, or inclusions between the bonding layers). In images acquired under conventional single-illumination conditions, these two types of suspected anomalies from different layers exhibit a high degree of similarity in brightness distribution, morphological contours, and contrast characteristics. The technical basis for this similarity lies in the fact that image signals acquired under conventional single-illumination conditions lack layer resolution capabilities, making it impossible to distinguish between optical response signals from the outermost surface and those from within the bonding layers at the image level.

[0005] Because existing solutions cannot distinguish the actual location of suspected anomalies at the image level, two opposing misjudgments are likely to occur in subsequent judgment and handling stages: On the one hand, outermost surface contamination, protective film damage, or foreign matter that can be eliminated or confirmed through cleaning or film peeling operations are mistakenly judged as irreparable interlayer bonding defects and excessively rejected, causing unnecessary losses to the production line yield; on the other hand, real interlayer bonding defects are mistakenly judged as cleanable surface anomalies and missed, posing potential risks to subsequent processes and finished product quality. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for identifying the layer location of appearance defects in online hot-press bonding, which solves the problem that image signals acquired under single illumination conditions do not have layer location resolution capabilities and cannot distinguish between optical response signals from the outermost surface and optical response signals from the interior of the bonding layer at the image level.

[0007] This invention is achieved through the following technical solution: Firstly, a layer-based attribution method for identifying online appearance defects after hot-press bonding is proposed, including the following steps: The entire area of ​​the hot-pressed product is initially screened and imaged. Image preprocessing and candidate region extraction are performed on the initial screened image to obtain a candidate abnormal region queue. Each candidate abnormal region in the candidate abnormal region queue contains the region center coordinate information. Based on the regional center coordinates of the candidate anomaly region, align the candidate anomaly region with the field of view of the local verification imaging unit; For the same candidate abnormal region, a first verification image is acquired under surface-sensitive lighting conditions, and a second verification image is acquired under interlayer-sensitive lighting conditions; wherein, the surface-sensitive lighting conditions are used to make the optical response of defects at the outermost surface of the product stronger than the optical response of defects inside the product's bonding layer, and the interlayer-sensitive lighting conditions are used to make the optical response of defects inside the product's bonding layer stronger than the optical response of defects at the outermost surface of the product. For the same candidate abnormal area, the focus position is switched sequentially to the first focal plane close to the outermost surface of the product and the second focal plane close to the bonding interface using the focal plane switching component, and the first focal plane image and the second focal plane image are obtained respectively. Image registration aligns the image coordinates of the first verification image, the second verification image, the first focal plane image, and the second focal plane image in the candidate anomaly region. The differential features of candidate anomalous regions are extracted from the registered images. The differential features include at least the surface response contrast features of the candidate anomalous regions in the first verification image, the interlayer response contrast features in the second verification image, the first focal plane sharpness features in the first focal plane image, and the second focal plane sharpness features in the second focal plane image. Based on the differences in characteristics, the layer attribution is determined, and the candidate abnormal areas are classified into surface defects, interlayer bonding defects, or anomalies to be reviewed. The corresponding handling instructions are output according to the classification results. Surface defects include surface contamination defects and protective film defects.

[0008] Secondly, a layer-based attribution and identification system for online appearance defects after hot-press bonding is proposed, including: The conveying and positioning unit is used to carry the hot-pressed products, pass the products through the initial screening station and the verification station in sequence, and provide the conveying displacement information of the products. The initial screening image acquisition unit is set at the initial screening station and is used to perform initial screening image acquisition on the entire area of ​​the product to obtain the initial screening image. The candidate anomaly extraction unit is used to perform image preprocessing and candidate region extraction on the initial screening image, and output a queue of candidate anomaly regions containing the center coordinate information of the regions. A local verification imaging unit, located at the verification station, includes: a surface-sensitive illumination component for providing surface illumination conditions that make the optical response of defects on the outermost surface of the product stronger than the optical response of defects inside the bonding layer; a second industrial camera for acquiring a first verification image under the surface illumination conditions; an interlayer-sensitive illumination component for providing interlayer illumination conditions that make the optical response of defects inside the bonding layer of the product stronger than the optical response of defects on the outermost surface; a second industrial camera for acquiring a second verification image under the interlayer illumination conditions; and a focal plane switching component for switching the focus position of the second industrial camera to a first focal plane close to the outermost surface of the product and a second focal plane close to the bonding interface. The image registration and candidate region management unit is used to align the candidate anomaly region with the field of view of the local verification imaging unit according to the regional center coordinate information of the candidate anomaly region, and to align the image coordinates of the first verification image, the second verification image, the first focal plane image and the second focal plane image in the candidate anomaly region through image registration; The layer attribution determination unit is used to extract the differential features of candidate abnormal regions from the registered image and perform layer attribution determination; the differential features include surface response contrast features, interlayer response contrast features, first focal plane sharpness features, and second focal plane sharpness features; the layer attribution determination is used to classify candidate abnormal regions into surface defects, interlayer bonding defects, or anomalies to be verified; surface defects include: surface contamination defects and protective film defects. The result output and execution unit is used to output the abnormal level attributes and corresponding handling instructions based on the classification results. The control unit is used to uniformly control the coordinated operation of all units.

[0009] Compared with existing technologies, this invention has the following advantages and beneficial effects: Addressing the problem that surface anomalies and interlayer anomalies are easily confused in a single image during online appearance inspection after hot-press bonding, this technical solution first uses a full-image initial screening method to quickly locate suspected abnormal areas in the product, and then performs local verification only on the candidate abnormal areas. This ensures that the online inspection process retains full-image coverage while avoiding repeated multi-condition fine imaging of the entire product, thus balancing inspection efficiency and layer analysis accuracy. In the local verification stage, the same candidate abnormal area is imaged under surface-sensitive illumination, interlayer-sensitive illumination, and different focal plane conditions, so that its appearance anomaly is no longer simply manifested as a single bright or dark spot, but is transformed into multi-dimensional difference information such as surface response, interlayer response, surface focal plane sharpness, and interface focal plane sharpness. Since the above multiple images are registered and aligned to the image coordinates of the same candidate abnormal area before feature extraction, subsequent contrast and sharpness features can be compared based on the same physical location, reducing feature errors caused by the displacement of multiple imaging steps. Therefore, the layer location attribution determination no longer relies on a single illumination image or a single focal plane image, but instead comprehensively considers the response differences of candidate anomaly regions under different illumination conditions and focal plane positions, classifying them into surface defects, interlayer bonding defects, or anomalies to be verified; among them, surface defects further cover surface contamination defects and protective film defects. The detection result is upgraded from "detecting anomalies" to "identifying the layer where the anomaly is located and matching the treatment path," which can reduce the risk of over-rejection caused by surface contamination or protective film defects being misjudged as interlayer defects, and also reduce the risk of missed detection of real interlayer bonding defects being treated as surface problems. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of the overall module structure of an online layer attribution identification system for appearance defects after hot pressing, provided in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the structure of the conveying and positioning unit provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the structure of the initial screening imaging unit provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the structure of the local verification imaging unit provided in Embodiment 1 of the present invention; The attached diagram shows the markings and corresponding component names: 100 - Conveying and positioning unit; 110 - Conveying mechanism; 120 - Side guide rails; 130 - Adsorption-type bearing platform; 140 - Entrance detection photoelectric sensor; 150 - Conveying encoder; 160 - Verification station stop mechanism; 200 - Initial screening image acquisition unit; 210 - First industrial camera; 220 - First lens; 230 - First illumination assembly; 240 - Second illumination assembly; 300 - Candidate anomaly extraction unit; 400 - Local verification imaging unit; 410 - Second industrial camera; 420 - Second lens Head; 430-Surface sensitive illumination component; 440-Interlayer sensitive illumination component; 450-Polarization switching component; 451-Incident polarizer; 452-Analysis polarizer; 453-Polarizer switching frame; 454-Polarizer rotation drive mechanism; 460-Focal plane switching component; 500-Image registration and candidate area management unit; 600-Layer attribution determination unit; 700-Result output and execution unit; 800-Control unit; 900-Product formula management unit; 910-Abnormal review priority scheduling unit. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0013] Example: On a hot-press lamination production line, after the product completes the hot-press lamination process, it enters the online appearance inspection stage. The drawback of the existing online appearance inspection solution after hot-press lamination is that the suspected anomalies presented in the image may originate from either the outermost surface of the product or from within the product's lamination layer. The image representations of the two are highly similar under normal single lighting conditions, making it impossible to distinguish them based on a single image. This results in subsequent processing instructions being difficult to match with the actual layer where the defect is located.

[0014] To address the aforementioned issues, this embodiment provides an online appearance defect identification solution after hot-press bonding. The overall logical architecture of this solution is as follows: using candidate abnormal regions as the smallest processing unit, a two-level detection architecture is employed. At the initial screening station, the entire product area is rapidly imaged and a candidate abnormal region queue is extracted. At the verification station, different lighting conditions and focus positions are switched multiple times for each candidate abnormal region to acquire multiple verification images with layer-specific characteristics. After image registration of the multiple verification images, multi-dimensional difference features, including at least surface response contrast features, interlayer response contrast features, first focal plane sharpness features, and second focal plane sharpness features, are extracted. Then, based on preset multi-condition judgment rules, the candidate abnormal regions are classified as surface defects, interlayer bonding defects, or abnormalities to be verified, and processing instructions matching the layer attributes are output accordingly.

[0015] Based on the above overall logical architecture, this embodiment provides an online appearance defect recognition system after hot-press bonding, which executes the following technical route step by step: Step 1, performing initial screening and image acquisition on the entire area of ​​the hot-pressed product, performing image preprocessing and candidate region extraction on the initial screening image to obtain a candidate abnormal region queue containing regional center coordinate information; Step 2, aligning the candidate abnormal regions with the field of view of the local verification imaging unit according to the regional center coordinate information of the candidate abnormal regions; Step 3, acquiring a first verification image for the same candidate abnormal region under surface sensitive illumination conditions, and under interlayer sensitive illumination conditions... Step 4: For the same candidate anomaly region, use the focal plane switching component to sequentially switch the focus position to the first focal plane and the second focal plane, and acquire the first focal plane image and the second focal plane image respectively; Step 5: Perform image registration on the first verification image, the second verification image, the first focal plane image and the second focal plane image; Step 6: Extract the difference features of the candidate anomaly region from the registered image; Step 7: Perform layer attribution determination based on the difference features, classify the candidate anomaly region into surface defects, interlayer bonding defects or anomalies to be verified, and output the corresponding processing instructions based on the classification results.

[0016] Accordingly, to implement the above technical approach, the online appearance defect identification system after hot-press bonding includes: Figure 1 The system includes a transport and positioning unit 100, a preliminary screening and imaging unit 200, a candidate anomaly extraction unit 300, a local verification imaging unit 400, an image registration and candidate area management unit 500, a layer attribution determination unit 600, a result output and execution unit 700, and a control unit 800. All components are uniformly scheduled by the control unit 800. The system also includes a product formula management unit 900 and an anomaly verification priority scheduling unit 910. The control unit 800 is connected to each of the above units and is used to sequentially control each unit to complete product transport and positioning, full-frame preliminary screening, candidate area extraction, local verification imaging, image registration, difference feature extraction, layer attribution determination, and processing instruction output according to the detection formula stored in the product formula management unit 900.

[0017] In this embodiment, each step in the technical route corresponds one-to-one with the function of each unit in the system. The processing actions recorded in each step are the processing actions executed by the corresponding units in the system under the scheduling of the control unit 800. The purpose, specific implementation method and relationship between each unit will be explained in detail below.

[0018] 1. Conveying and positioning unit 100 The conveying and positioning unit 100 is used to carry the hot-pressed products, allowing them to pass sequentially through the initial screening station and the verification station, and providing the control unit 800 with the product's conveying displacement information. Figure 2As shown, the conveying and positioning unit 100 includes: a conveying mechanism 110, two side guide rails 120, an adsorption-type bearing platform 130, an entrance detection photoelectric sensor 140, a conveying encoder 150, and a verification station stop mechanism 160.

[0019] The conveying mechanism 110 is used to drive the product to move along the inspection line. As a product conveying actuator, the conveying mechanism 110 can be a belt conveyor or a roller conveyor.

[0020] The two guide rails 120 are respectively set on both sides of the conveying mechanism 110 to limit and guide the position of the product in the direction perpendicular to the conveying direction, so that the product maintains a stable lateral position during the conveying process.

[0021] The adsorption-type support platform 130 is set at the product support position in the verification station corresponding to the field of view of the local verification imaging unit 400. When the product arrives at the verification station, the adsorption-type support platform 130 applies negative pressure adsorption to the product, so that the product remains flat during the local verification imaging stage and avoids product position drift caused by mechanical vibration.

[0022] An entrance detection photoelectric sensor 140 is installed at the entrance of the detection line. When the front edge of the product blocks the entrance detection photoelectric sensor 140, a trigger signal is output. After receiving the trigger signal, the control unit 800 determines that the product has entered the detection line and uses the trigger signal as the starting time reference for the displacement counting of the conveyor encoder 150.

[0023] The conveying encoder 150 is mounted on the drive shaft or driven shaft of the conveying mechanism 110 and is used to output pulse signals proportional to the conveying displacement in real time; the control unit 800 counts the pulse signals and converts them into the actual displacement of the product from the entrance of the detection line.

[0024] The verification station stop mechanism 160 can be a pneumatic stop pin or an electric baffle. During the partial verification imaging stage, it extends to prevent the product from continuing to move forward, stopping the product at the preset stop position of the verification station. After the partial verification imaging is completed, the verification station stop mechanism 160 retracts to release the product for continued transport.

[0025] The control unit 800 is electrically connected to the inlet detection photoelectric sensor 140, the conveying encoder 150, and the verification station stop mechanism 160. Based on the product entry signal from the inlet detection photoelectric sensor 140 and the conveying displacement feedback provided by the conveying encoder 150, the control unit 800 controls the movement and positioning of the product between the initial screening station and the verification station, and uses the conveying displacement information as the basis for subsequent coordinate transformation of candidate abnormal areas.

[0026] 2. Initial screening imaging unit 200 The initial screening imaging unit 200 is located at the initial screening station and is used, triggered by the control unit 800, to perform initial screening imaging on the entire effective detection area of ​​the hot-pressed product, acquiring initial screening images for candidate abnormal area extraction. For example... Figure 3 As shown, the initial screening imaging unit 200 includes a first industrial camera 210, a first lens 220, a first illumination component 230, and a second illumination component 240. The first industrial camera 210 and the first lens 220 constitute the initial screening imaging mechanism. The first illumination component 230 provides coaxial light illumination, and the second illumination component 240 provides low-angle dark-field illumination. A control unit 800 is connected to the first industrial camera 210, the first illumination component 230, and the second illumination component 240, respectively, and is used to control time-division imaging under the two illumination conditions.

[0027] (1) First industrial camera 210 The first industrial camera 210 is used to receive product images imaged by the first lens 220, and, triggered by the control unit 800, acquires a first preliminary screening image under coaxial illumination and a second preliminary screening image under low-angle dark field illumination. The first industrial camera 210 can be an area array industrial camera or a line array industrial camera, and its resolution is determined based on the effective inspection area size of the product and the minimum defect size to be detected in the preliminary screening stage. Specifically, the imaging field of view of the first industrial camera 210 in the object direction is first determined based on the effective inspection width W and effective inspection length L that the product needs to cover in one imaging session at the preliminary screening station; then, the minimum defect size d to be detected in the preliminary screening stage is determined based on... min Determine the maximum permissible object size P corresponding to a single pixel on the product surface. To ensure that the minimum defect can be stably identified in the initial screening image, the minimum defect size d is determined. min The image must cover at least N pixels, where N can be set to 3–8 depending on the detection requirements. When only candidate anomalous regions need to be detected, N can be 3–5; when a preliminary assessment of defect morphology is also required, N can be 5–8. Therefore, the object size P corresponding to a single pixel should satisfy P ≤ d. min / N.

[0028] When using an area array industrial camera, the horizontal pixel count X of the first industrial camera 210 should be no less than W / P, and the vertical pixel count Y should be no less than L / P, so that the effective detection area of ​​the product has a number of sampling points that meet the minimum defect detection requirements in both the width and length directions.

[0029] When using a linear scan industrial camera, the number of pixels per row of the first industrial camera 210 should not be less than the effective detection width W / P of the product. At the same time, the control unit 800 controls the line frequency f of the linear scan industrial camera according to the product conveying speed v, so that the line spacing v / f in the product conveying direction is not greater than P, thereby ensuring that the sampling accuracy in the conveying direction also meets the minimum defect detection requirements.

[0030] For example, when the effective detection width W of a product is 200 mm and the effective detection length L is 120 mm, the minimum point defect size d that needs to be detected in the initial screening stage. min When the minimum defect size is 0.20mm and the minimum defect is required to cover at least 5 pixels, the object size P corresponding to a single pixel should not exceed 0.04mm. In this case, if an area array industrial camera is used, the horizontal pixel count X should not be less than 5000 and the vertical pixel count Y should not be less than 3000. Therefore, an area array industrial camera with a resolution of not less than approximately 5000×3000 pixels can be selected.

[0031] (2) First shot 220 The first lens 220 is mounted on the side of the first industrial camera 210 facing the product, and is used to image the effective detection area of ​​the product onto the sensor target surface of the first industrial camera 210. The first lens 220 can be an industrial-grade fixed-focus lens or a telecentric lens that matches the pixel size and target surface size of the first industrial camera 210. The telecentric lens can reduce the impact of slight height fluctuations of the product or imaging distortion in edge areas on the initial screening image; the industrial-grade fixed-focus lens can achieve full-area imaging through matching working distance and focal length.

[0032] The magnification of the first lens 220 is determined based on the sensor target size of the first industrial camera 210 and the effective detection area size of the product. Specifically, if the effective detection width of the product is W and the effective detection length is L, the effective imaging width of the sensor of the first industrial camera 210 is S. w The effective imaging height is S h If the preset imaging margin coefficient is m, then the magnification β of the first lens 220 is set according to the following condition: β≤S w / [W×(1+m)], and β≤S h / [L×(1+m)]. Where m is used to reserve imaging margin for product placement deviation, conveying offset and edge positioning error, and m can be 0.05 to 0.15. The control unit 800 or the system debugging personnel determine the target magnification of the first lens 220 according to the smaller upper limit value of the above two conditions, so that the effective detection area of ​​the product and its surrounding margin area fall within the effective imaging range of the sensor of the first industrial camera 210.

[0033] For example, when the effective detection width W of the product is 200mm and the effective detection length L is 120mm, the effective imaging width S of the sensor of the first industrial camera 210 is... w 11mm, effective imaging height S hWith a thickness of 7mm and a preset imaging margin coefficient m of 0.1, β≤11 / [200×1.1] and β≤7 / [120×1.1], then β should not be greater than approximately 0.05. In this case, a first lens 220 with a magnification of approximately 0.05 can be selected to ensure that the effective detection area of ​​the product and its surrounding margin area are completely imaged within the sensor range of the first industrial camera 210.

[0034] In terms of installation location, the first industrial camera 210 is positioned above the product conveying path, and the first lens 220 is mounted on the side of the first industrial camera 210 facing the product. The imaging optical axis of the first industrial camera 210 coincides with the optical axis of the first lens 220 and is arranged in a direction perpendicular to the surface of the product to be inspected, so that the effective detection area of ​​the product can be imaged onto the sensor target surface of the first industrial camera 210 through the first lens 220. Both the first industrial camera 210 and the first lens 220 are mounted on a fixed bracket at the initial screening station, and their relative positions are calibrated and confirmed during the system installation and commissioning phase.

[0035] (3) First lighting component 230 The first illumination component 230 is a coaxial illumination component used to provide coaxial illumination conditions, enabling the first industrial camera 210 to acquire a first preliminary screening image under coaxial illumination conditions. The first illumination component 230 can be positioned in the coaxial illumination path between the first lens 220 and the surface of the product to be inspected, or it can be positioned to one side of the first lens 220 and guided into the optical axis direction of the first lens 220 via a beam splitter. Specifically, the illumination light emitted by the first illumination component 230 is reflected by the beam splitter and then illuminates the surface of the product to be inspected along the optical axis direction of the first lens 220; the light signal reflected back from the surface of the product to be inspected then passes through the beam splitter along the original optical axis direction and enters the first lens 220, where it is received and imaged by the first industrial camera 210. Thus, the first illumination component 230 forms a positive bright field illumination coaxial with the first industrial camera 210.

[0036] Under coaxial illumination, areas of the product surface that are flat and without obvious abnormalities present a uniform bright field background. Areas with surface bumps, depressions, obvious contaminants, or optical discontinuities will appear as relatively dark areas, bright spots, or areas with sudden changes in local brightness in the image due to changes in the reflected light path. Therefore, the first illumination component 230 is mainly used to acquire an overall reflection image of the product surface and to detect positive bright field anomalies such as large-area stains, large-area damage to the protective film, and localized reflection abnormalities.

[0037] When the first lighting component 230 is lit under the control of the control unit 800, the control unit 800 synchronously triggers the first industrial camera 210 to complete the exposure and obtain the first preliminary screening image.

[0038] (4) Second lighting component 240 The second illumination component 240 is a low-angle dark field illumination component, used to provide low-angle dark field illumination conditions so that the first industrial camera 210 can acquire a second preliminary screening image under low-angle dark field illumination conditions. The second illumination component 240 is disposed to the side or circumferentially outside the effective detection area of ​​the product, and its light emission direction is obliquely incident relative to the surface of the product to be detected, and intersects with the imaging optical axis of the first industrial camera 210 on the surface of the product to be detected.

[0039] In one embodiment, the second illumination component 240 employs a ring-shaped low-angle light source, which is arranged around the optical axis of the first lens 220 and located circumferentially below the first lens 220 or above the product inspection area, so that the low-angle illumination light obliquely illuminates the product surface to be inspected from all sides. In another embodiment, the second illumination component 240 employs a strip-shaped low-angle light source, which is disposed on one side or opposite sides of the product conveying path, so that the illumination light obliquely incident on the product surface to be inspected along a direction forming a preset angle with the normal to the product surface.

[0040] The preset incident angle is determined based on the product surface's reflectivity, roughness, and the protective film's haze, transmittance, and surface scattering capability. Specifically, during the system debugging phase, the control unit 800 controls the second illumination component 240 to sequentially capture images of the standard sample and the sample with known surface defects at multiple candidate incident angles. The candidate incident angles can be set to 60 degrees, 65 degrees, 70 degrees, 75 degrees, and 80 degrees from the product surface normal. For the images corresponding to each candidate incident angle, the brightness contrast of the known surface defect area relative to the surrounding defect-free background area, the background brightness uniformity, and the overexposed area area are calculated. The candidate incident angle that ensures the surface defect area contrast meets the preset detection requirements, the background area is not overexposed, and the scattering noise is below the preset noise threshold is selected as the low-angle dark field illumination angle for the current product model and stored in the product formula management unit 900.

[0041] Under low-angle dark-field illumination conditions, the first industrial camera 210 primarily receives scattered and diffused light from the product surface. Areas with strong scattering characteristics, such as fine scratches, particulate contaminants, and localized blistering of the protective film on the product surface, appear as bright features in the second preliminary screening image; while smooth, defect-free areas retain a darker image background due to less scattered light entering the first industrial camera 210. Therefore, the second illumination assembly 240 is mainly used to enhance the preliminary screening detection capability of scattering-type surface anomalies such as fine particles, fine scratches, and localized blistering of the protective film.

[0042] When the second lighting component 240 is lit under the control of the control unit 800, the control unit 800 synchronously triggers the first industrial camera 210 to complete the exposure and obtain the second preliminary screening image.

[0043] (5) Initial screening image acquisition sequence and image output The first lighting component 230, the second lighting component 240, and the first industrial camera 210 are respectively connected to the control unit 800. The control unit 800 controls the first lighting component 230, the second lighting component 240, and the first industrial camera 210 to operate in a coordinated manner according to a preset image acquisition sequence.

[0044] In one embodiment, the control unit 800 first controls the first illumination component 230 to light up and triggers the first industrial camera 210 to complete one exposure under coaxial illumination to obtain a first preliminary screening image. Then, the control unit 800 turns off the first illumination component 230, controls the second illumination component 240 to light up, and triggers the first industrial camera 210 to complete another exposure under low-angle dark field illumination to obtain a second preliminary screening image. The first illumination component 230 and the second illumination component 240 are illuminated in a time-sharing manner to avoid interference between the coaxial light signal and the low-angle dark field scattering signal. In another embodiment, when fused illumination is required, the control unit 800 can also control the first illumination component 230 and the second illumination component 240 to light up simultaneously according to a preset brightness ratio, allowing the first industrial camera 210 to acquire a preliminary screening image under fused illumination conditions. However, in a conventional preliminary screening process, a time-sharing illumination method is preferred to obtain the first preliminary screening image under coaxial illumination conditions and the second preliminary screening image under low-angle dark field illumination conditions, respectively.

[0045] After completing two initial screening images, the control unit 800 sends the first and second initial screening images to the candidate anomaly extraction unit 300. The candidate anomaly extraction unit 300 performs image preprocessing and candidate region extraction on the first and second initial screening images respectively, and merges the extracted candidate regions from the two images to form a candidate anomaly region queue. This initial screening method, through the complementarity of coaxial illumination and low-angle dark-field illumination, enables the initial screening stage to simultaneously detect both overall reflection anomalies and scattering-type fine surface anomalies.

[0046] It should be noted that: 1. The purpose of the initial screening is to scan the entire area of ​​the hot-pressed product at a high imaging coverage speed, mark all candidate abnormal areas that need to be further analyzed at the review station without missing any potential abnormalities, and record the location information of each candidate abnormal area for subsequent review station to use.

[0047] 2. Initial screening images are acquired using two illumination modes, one for each image. Specifically, a first initial screening image is acquired under coaxial illumination, and a second initial screening image is acquired under low-angle dark-field illumination. Coaxial illumination provides excellent detection capabilities for bright-field defects (such as large-area stains, extensive damage to the protective film, etc.), while low-angle dark-field illumination offers strong detection capabilities for smaller surface micro-defects with strong scattering signals (such as fine particles, minor scratches, localized blistering of the protective film, etc.). These two modes complement each other in terms of defect morphology.

[0048] 3. Candidate anomaly extraction unit 300 The candidate anomaly extraction unit 300 is data-connected to the initial screening image acquisition unit 200. It is used to perform image preprocessing and candidate region extraction on the first initial screening image and the second screening image under the call of the control unit 800, and output a candidate anomaly region queue containing the center coordinate information of each candidate anomaly region.

[0049] (1) Image preprocessing The control unit 800 calls the candidate anomaly extraction unit 300 to perform image preprocessing on the first and second preliminary screening images respectively. The purpose of image preprocessing is to eliminate non-defect image interference such as the uneven illumination of the imaging system itself, camera dark current, and background texture, and to shield product edges and known non-detection areas, so as to provide processed images with higher signal-to-noise ratio and more stable judgment criteria for subsequent candidate region extraction steps.

[0050] In this embodiment, image preprocessing includes: flat field correction, background normalization, noise suppression, and invalid region masking.

[0051] 1) Leveling Correction The flat-field correction is implemented as follows: during the system calibration stage, a defect-free standard reference image is pre-acquired as the flat-field reference image, and an image under off-light conditions is acquired as the dark-field reference image; for each pixel in the current preliminary screening image to be preprocessed, pixel-by-pixel gain compensation is performed based on the flat-field reference image and the dark-field reference image. Specifically, let the gray value of the pixel with coordinates (x,y) in the current preliminary screening image to be preprocessed be I(x,y), the gray value of the corresponding pixel in the flat-field reference image be R(x,y), and the gray value of the corresponding pixel in the dark-field reference image be D(x,y). Then, the formula for calculating the gray value I'(x,y) of the pixel after flat-field correction is: I'(x,y)=[I(x,y)-D(x,y)] / [R(x,y)-D(x,y)]×M. Wherein, I(x,y) represents the original pixel grayscale value at coordinates (x,y) in the current initial screening image, used to reflect the total light intensity response of the object point corresponding to the pixel under the current illumination conditions; D(x,y) represents the pixel grayscale value at the same coordinate in the dark field reference image, used to reflect the background response of camera dark current and fixed pattern noise; R(x,y) represents the pixel grayscale value at the same coordinate in the flat field reference image, used to reflect the total light intensity response at the same position on the defect-free standard reference product, combining the effects of illumination non-uniformity and inconsistent camera pixel response; M is the preset normalized target mean grayscale, used to normalize the overall brightness of the corrected image to a predetermined reference level, which can be determined according to the image bit depth and the input requirements of subsequent processing algorithms (e.g., 128 or 255). The mathematical logic of the formula is as follows: first, the dark current background is removed by the numerator [I(x,y)-D(x,y)]; then, the denominator [R(x,y)-D(x,y)] is used as a joint reference for illumination intensity and pixel response for normalization; finally, it is multiplied by M to restore a reasonable gray level. The logic of the formula for calculating the pixel gray value I'(x,y) after flat-field correction is as follows: through pixel-by-pixel normalization, the overall brightness gradient caused by uneven illumination space and inconsistent pixel response is separated from the original image, so that the gray values ​​of each pixel in the corrected image only reflect the optical differences of the object point itself, thereby eliminating non-defective overall brightness differences. Flat-field reference images and dark-field reference images can be re-acquired and updated each time the system starts or at preset time intervals to cope with illumination aging and camera response characteristic drift.

[0052] 2) Background normalization Background normalization is implemented as follows: For the flat-field corrected image, a Gaussian filter kernel or a mean filter kernel with a kernel size larger than the expected maximum defect size (e.g., 5 to 10 times the expected maximum defect size) is used to estimate the low-frequency background brightness distribution of the image. Then, the low-frequency background is subtracted from the flat-field corrected image to obtain the residual image after removing macroscopic background brightness variations. The purpose of background normalization is to eliminate slowly varying backgrounds such as macroscopic gloss gradients and coating thickness gradients inherent in the product itself, so that subsequent thresholding steps can apply a unified judgment benchmark to the entire image.

[0053] 3) Noise suppression The noise suppression is achieved by smoothing the residual image after background normalization using Gaussian filtering or median filtering to suppress random noise from the image sensor and camera readout noise, while preserving the local brightness deviation information of the real defect area.

[0054] The Gaussian or median filter kernels are determined through sample calibration. Specifically, during the system debugging phase, the control unit 800 uses multiple candidate filter kernel sizes to perform noise suppression processing on the reference image. These candidate kernel sizes include 3×3, 5×5, 7×7, and 9×9. For each candidate kernel size, the grayscale standard deviation of the defect-free background region and the gradient magnitude of the known fine defect edge region are calculated. The grayscale standard deviation of the defect-free background region characterizes the noise suppression effect; a lower grayscale standard deviation indicates less background random noise. The gradient magnitude of the known fine defect edge region characterizes the defect edge preservation effect; a higher gradient magnitude indicates a clearer defect edge. The control unit 800 selects the smallest candidate filter kernel size that ensures the grayscale standard deviation of the defect-free background region is below a preset background noise threshold and the gradient magnitude of the known fine defect edge region is not less than a preset proportion of the corresponding gradient magnitude in the unfiltered image. This is used as the filter kernel size for the current product model, and the filter kernel size is written into the product formula management unit 900. This avoids weakening the fine defect edges due to an excessively large filter kernel size while meeting noise suppression requirements.

[0055] 4) Invalid area shielding The invalid area shielding is implemented as follows: based on the preset invalid shielding area range in the product formula management unit 900, the pixels corresponding to the areas in the noise-suppressed image that do not participate in the candidate abnormal area extraction are set as background reference values, or the pixels in this area are ignored in the subsequent candidate area extraction process. The invalid shielding area range refers to the image area in the initial screening image that is not considered as an effective appearance inspection object, including the outer edge area of ​​the product, the product border area, the positioning hole area, the alignment mark area, the QR code area, the silk screen area, the clamping and obstructing area, the fixture pressing area, and other preset non-detection areas. The invalid shielding area range is preset and stored in the product formula management unit 900 during the system debugging phase; specifically, the non-detection area in the product coordinate system is determined according to the product design drawings, product dimensions, effective detection area requirements, and fixture obstruction position. Then, the standard sample image is acquired by the first industrial camera 210, and the image coordinates of the product edge, positioning hole, QR code, alignment mark, silk screen area, and clamping and obstructing area are calibrated. The coordinates of the corresponding rectangular, circular, or polygonal area are used as the invalid shielding area range of the current product model. During the production process, the control unit 800 reads the invalid shielding area range corresponding to the current product model and performs shielding processing on the range during the image preprocessing stage, so that fixed structures or occluded areas within the range will not be extracted as candidate abnormal areas.

[0056] (2) Candidate region extraction The control unit 800 calls the candidate anomaly extraction unit 300 to perform candidate region extraction on the preprocessed image. The purpose of candidate region extraction is to extract candidate anomaly regions from the preprocessed image whose local brightness deviation exceeds the initial screening deviation threshold, whose area is not less than the preset area threshold, and whose area ratio and location meet the candidate region filtering conditions, and to calculate the corresponding region center coordinate information for each candidate anomaly region.

[0057] The initial screening deviation threshold is used to determine whether the brightness deviation of a pixel or local area is sufficient to constitute a candidate anomaly. It is determined during the system debugging phase through statistical analysis of background grayscale fluctuations in defect-free standard sample images. Specifically, the control unit 800 acquires multiple defect-free standard sample images, calculates the absolute value of the deviation of each pixel within the effective detection area relative to the average brightness of its neighborhood, and selects the deviation value corresponding to the 99th, 99.5th, or 99.9th percentile of this absolute deviation distribution as the initial value of the initial screening deviation threshold.

[0058] A preset area threshold is used to remove randomly noisy connected regions with excessively small areas. It is determined based on the minimum defect size to be detected in the initial screening stage and the object-side pixel size of the initial screening image. Specifically, the control unit 800 first determines the object-side pixel size P corresponding to a single pixel on the product surface in the initial screening image based on the imaging parameters of the first industrial camera 210 and the first lens 220. When the minimum defect to be detected in the initial screening stage is a point defect, the minimum point defect diameter d is used as the threshold. min Calculate the theoretical pixel area A of the defect in the initial screening image. min A min =π×(d min / 2P) 2 Then set the preset area threshold A th Set as theoretical pixel area A min The preset ratio, i.e., A th =k×A min Where k is the area retention coefficient, which can take values ​​from 0.3 to 0.6. The area is less than a preset area threshold A. th Connected regions are identified as random noise or invalid candidate regions and are removed during the candidate region filtering step. For elongated defects, the minimum scratch length L can be used as a reference. min and minimum scratch width W min Calculate the theoretical pixel area A min =(L min / P)×(W min / P), and determine the corresponding preset area threshold in the same way.

[0059] A preset slenderness ratio threshold is used to eliminate interference areas such as conveyor-direction stripes caused by conveyor vibration, guide rail scratches, camera noise, or fixed stripes. Specifically, during the system debugging phase, the control unit 800 acquires multiple standard sample images and sample images with known real defects. After extracting candidate regions, it calculates the length of the long side *a*, the length of the short side *b*, and the direction of the major axis of the circumscribed rectangle for each candidate connected region, and defines the slenderness ratio K as K = a / b. Real defect areas typically appear as dot-like, patchy, nearly circular, irregular block-like, or non-fixed-direction scratches, and their slenderness ratio K is usually low, or although they are slender, their major axis direction is not fixed. Conveyor-direction stripe interference areas typically extend continuously along the product conveying direction, with the long side of the circumscribed rectangle significantly longer than the short side, and the major axis direction is basically consistent with the product conveying direction. Based on the slenderness ratio distribution of known real defect areas and known stripe interference areas, the control unit 800 selects the boundary value between the two distributions as the preset slenderness ratio threshold and stores this preset slenderness ratio threshold in the product formula management unit 900. During candidate area filtering, if the elongation ratio K of a candidate area exceeds the preset elongation ratio threshold, and the angle θ between the major axis of the candidate area and the product conveying direction is less than the preset direction angle threshold, then the candidate area is determined to be a conveying direction stripe interference area and is removed; wherein, the preset direction angle threshold can be set to 10 degrees to 20 degrees.

[0060] The initial screening deviation threshold, preset area threshold, and preset slenderness ratio threshold are all stored in the product formula management unit 900 and can be adjusted based on the false alarm rate and missed detection rate during production line operation.

[0061] The input for candidate region extraction is the preprocessed image, the processing object is the brightness deviation and connectivity information of each local location in the image, and the output object is a queue of candidate abnormal regions with the coordinate information of the region center.

[0062] In this embodiment, candidate region extraction includes: local contrast enhancement, pixel-by-pixel deviation calculation and threshold segmentation, connected region clustering, and candidate region filtering.

[0063] 1) Local contrast enhancement Local contrast enhancement is a process that performs local contrast enhancement on a preprocessed image. Conventional image enhancement methods in the field, such as histogram equalization or local contrast stretching, can be used to make local brightness deviations more clearly presented in the image, thereby improving the sensitivity of subsequent threshold segmentation in detecting low-contrast defects.

[0064] 2) Pixel-by-pixel deviation calculation and threshold segmentation Pixel-by-pixel deviation calculation and threshold segmentation, for each pixel in the locally contrast-enhanced image, calculates the deviation between its grayscale value and the average brightness of its neighborhood. The average brightness of the neighborhood is represented by the average of all pixels within the neighborhood window of that pixel. The neighborhood window size is used to determine the range of pixels around the current pixel that participate in the calculation of the average background brightness, and it is determined based on the minimum defect size to be detected in the initial screening stage, the object-side pixel size of the initial screening image, and the scale of the product background brightness change. Specifically, the control unit 800 calculates the minimum defect size d to be detected in the initial screening stage. min Given the object-side pixel size P of a single pixel in the initial screening image on the product surface, calculate the pixel width n of the minimum defect in the image. min =d min / P; then set the side length of the neighborhood window to n min The size of the neighborhood window should be an integer multiple (e.g., 4 to 10 times) and an odd number of pixels to ensure that the neighborhood window covers the background area surrounding the defect, while preventing the neighborhood window from being dominated by the defect area itself. If the background brightness of the product surface changes gradually, the neighborhood window can be larger (e.g., setting the side length of the neighborhood window to n). min (7 to 10 times larger); if the product surface has borders, silkscreen printing, patterns, or areas with rapid changes in brightness, the neighboring window should be smaller (e.g., set the side length of the neighboring window to n). min (4 to 7 times the size of the current product), and in conjunction with the invalid shielding area range, avoid crossing different background areas. The neighborhood window size is stored as the neighborhood window size parameter for the current product model in the product formula management unit 900.

[0065] When the absolute value of the grayscale value of a pixel deviates from the average brightness of its neighborhood, exceeding a preset deviation threshold, the pixel is marked as a candidate abnormal pixel. The initial screening deviation threshold is used to determine whether the brightness deviation of the current pixel or local area relative to the surrounding background meets the candidate abnormality triggering condition. It is determined during the system debugging phase through statistical analysis of background brightness fluctuations in defect-free standard sample images and stored in the product formula management unit 900. Specifically, the control unit 800 acquires multiple defect-free standard sample images and performs the same image preprocessing as the formal inspection on each image. Subsequently, the control unit 800 calculates the absolute value of the grayscale value of each pixel within the effective detection area relative to the average brightness of its neighborhood, sorts all absolute deviation values ​​from smallest to largest, and selects the deviation value corresponding to the preset sorting position closest to the maximum value as the initial value of the initial screening deviation threshold. The preset sorting position can correspond to the 99th percentile, 99.5th percentile, or 99.9th percentile of all absolute deviation values, ensuring that the brightness fluctuations of the vast majority of normal backgrounds are below the initial screening deviation threshold. The initial screening deviation threshold is stored as a test formula parameter for the current product model in the product formula management unit 900.

[0066] 3) Connected region clustering Connected region clustering is a process that performs connected region labeling on the binarized candidate anomalous pixel image after threshold segmentation. Spatially adjacent candidate anomalous pixels (using four-neighbor or eight-neighbor connectivity rules) are grouped into the same candidate region. The pixel area, bounding box, region center coordinates (taking the centroid coordinates of the connected region pixels as the region center coordinates), major axis direction, and slenderness ratio (defined as the ratio of the long side to the short side of the bounding box) of each candidate region are calculated.

[0067] 4) Candidate region filtering Candidate region filtering involves filtering the clustering results of connected regions item by item, eliminating the following three types of candidate regions: First, candidate regions with an area below a preset area threshold. These regions typically correspond to random impulse noise or sub-pixel-level fine particles in the image sensor and are not worth verifying. Second, candidate regions with a length-to-width ratio exceeding a preset length-to-width ratio threshold and whose long axis is aligned with the product's transport direction. These regions typically correspond to regular scratches or fixed-pattern stripes formed by the product's contact with the guide rail during transport, rather than randomly distributed real defects. Third, candidate regions located at the edge of the invalid shielding area. These regions may be generated by image processing boundary effects at the shielding area boundary and are not within the effective detection range. The remaining candidate regions after candidate region filtering, along with their respective region center coordinates, constitute a candidate abnormal region queue.

[0068] In a further preferred implementation of this embodiment, the candidate anomaly extraction unit 300 independently executes the aforementioned candidate region extraction process on the first preliminary screening image and the second preliminary screening image, respectively, and then merges the candidate regions extracted from the two images to form the final candidate anomaly region queue. During fusion, a preset candidate region merging distance threshold is used to determine whether the candidate regions from the two images overlap or are adjacent in spatial position. Specifically, after the candidate anomaly extraction unit 300 extracts candidate regions from the first preliminary screening image and the second preliminary screening image, the control unit 800 calculates the center coordinate distance and the minimum boundary distance between the candidate regions from the two preliminary screening images; when the center coordinate distance between the two candidate regions is not greater than the candidate region merging distance threshold, or when the minimum boundary distance between the two candidate regions is not greater than the candidate region merging distance threshold, the control unit 800 determines that the two are imaging results of the same physical anomaly region under different lighting conditions, and merges the two into one candidate anomaly region; when both the center coordinate distance and the minimum boundary distance between the two candidate regions are greater than the candidate region merging distance threshold, the two are retained as two independent candidate anomaly regions.

[0069] The candidate region merging distance threshold is determined and stored in the product formula management unit 900 during the system debugging phase. Specifically, the control unit 800 acquires standard sample images containing known defects, extracts candidate regions corresponding to the same known defect under coaxial illumination and low-angle dark field illumination conditions, and calculates the center coordinate offset and boundary offset of the same known defect in the two initial screening images. The maximum offset obtained is added to a preset safety margin and used as the candidate region merging distance threshold for the current product model. Different product models have different object pixel sizes, surface reflection characteristics, protective film scattering characteristics, and differences in candidate region boundaries between the two types of initial screening images. Therefore, the candidate region merging distance threshold is configured separately for different product models.

[0070] By employing a dual-illumination complementary imaging and candidate region fusion strategy, the initial screening stage can achieve a reasonable detection rate for both surface protrusion defects and scattering defects, while reducing the risk of missing certain types of defects due to a single illumination mode.

[0071] 4. Local verification imaging unit 400 like Figure 4 As shown, the local verification imaging unit 400 is disposed at the verification station and is used to perform multi-condition local verification imaging on the same candidate abnormal area under the control of the control unit 800. In this embodiment, the local verification imaging unit 400 includes: a second industrial camera 410, a second lens 420, a surface sensitive illumination component 430, an interlayer sensitive illumination component 440, a polarization switching component 450, and a focal plane switching component 460.

[0072] (1) Local verification image acquisition mechanism The second industrial camera 410 and the second lens 420 constitute a local verification image acquisition mechanism.

[0073] The second industrial camera 410 can be an area array industrial camera, the resolution of which is determined based on the size of the local verification field of view and the minimum layer discrimination feature size required for identification in the local verification stage. Specifically, the control unit 800 or the system debugging personnel first determine the width W of the local verification field of view based on the maximum expected size of the candidate anomaly region, the width of the surrounding background band, the required expansion range for scattering halo evaluation, and the verification positioning error margin. r and height H r Then, based on the minimum layer size d that needs to be identified in the local verification stage, the feature size d is determined. r Determine the maximum allowable object size P corresponding to a single pixel on the product surface. r Minimum layer discrimination feature size d rThis includes at least one of the following: surface defect edge width, interlayer bubble bright ring width, scattering halo ring width, or minimum texture scale used for sharpness evaluation. To ensure stable extraction of the minimum layer-level discriminant feature, the minimum layer-level discriminant feature size d... r Cover at least N in the image r Pixels, N r A value of 8 to 15 can be used; when it is necessary to quantify the bright ring dark core structure, the area of ​​the scattering halo ring, or the edge sharpness, N should be used. r A value of 10 to 20 can be taken. Therefore, P... r ≤d r / N r When using an area array industrial camera, the horizontal pixel count X of the second industrial camera 410 is... r It should be no less than W r / P r Vertical pixel count Y r It should be no less than H r / P r The 410 resolution of the second industrial camera determined in this way enables the candidate anomaly region and its surrounding background region to have sufficient pixel sampling density in the local verification image, so as to support the subsequent extraction of differential features such as surface response contrast, interlayer response contrast, focal plane sharpness, edge sharpness, scattering halo rings and bright ring dark core structures.

[0074] The second lens 420 can be an industrial-grade fixed-focus lens, telecentric lens, or motorized focusing lens that matches the second industrial camera 410, and the magnification of the local verification imaging unit 400 is higher than that of the initial screening imaging unit 200, so as to achieve fine local imaging of the candidate anomaly area. The magnification of the second lens 420 is determined based on the local verification field of view size, the minimum layer discrimination feature size required to be identified in the local verification stage, and the pixel size of the second industrial camera 410. Specifically, let the width of the local verification field of view be W. r Height is H r The effective imaging width of the sensor in the second industrial camera 410 is S. rw The effective imaging height is S rh The second lens has a magnification of 420 volts, which is β. r , then β r β should be satisfied r ≤S rw / W r And β r ≤S rh / H r This ensures that the candidate anomaly region, its surrounding background band, and the extended region of the scattering halo evaluation are completely imaged within the sensor range of the second industrial camera 410; simultaneously, the minimum layer discrimination feature size required for identification in the local verification stage is set to d. r The pixel size of the second industrial camera 410 is ps The minimum level discriminative feature needs to cover N pixels in the image. r , then β r β should also be satisfied r ×d r / p s ≥N r The control unit 800 or system debugging personnel determine the target magnification of the second lens 420 within the range that simultaneously meets the above-mentioned field of view coverage and sampling accuracy conditions. The magnification determined in the above manner can ensure complete imaging of the local verification field of view, and also ensure that the edges of surface defects, bright rings of interlayer bubbles, scattering halos, or fine textures used for sharpness evaluation have sufficient pixel sampling density in the verification image.

[0075] (2) Surface-sensitive lighting component 430 The surface-sensitive illumination component 430 is used to provide surface-sensitive illumination conditions. These conditions ensure that the optical response of defects on the outermost surface of the product is stronger than that of defects within the bonding layers, thereby causing the first verification image acquired by the second industrial camera 410 under these conditions to preferentially display the image features of surface defects and suppress the response of interlayer defects in the image. Alternatively, the surface-sensitive illumination component 430 may be at least one of a low-angle annular dark-field light component, a side-strip oblique illumination light component, a highly directional surface scattering enhanced light component, or a surface reflection enhanced polarization illumination component. The low-angle ring-shaped dark-field light component is uniformly incident from all around the product at a low incident elevation angle (e.g., 60 to 80 degrees), causing particles, bumps, or foreign matter in the protective film on the outermost surface to generate omnidirectional scattering enhancement signals. The side-facing strip-shaped oblique illumination light component is set on one side of the effective detection area of ​​the product, and its light emission direction is obliquely incident on the surface to be detected at an angle of 60 to 80 degrees with the normal to the product surface. It is used to enhance the scattering response of directional scratches, unidirectional surface abrasions, and strip-shaped protective film damage from one side, and to detect directional scratches or unidirectional surface defects. The trap has a good detection effect; the high-directional surface scattering enhancement light component uses a highly collimated unidirectional beam shaped by a collimating lens or slit aperture to illuminate the surface of the product to be tested. The incident direction of the highly collimated unidirectional beam forms an angle of 2 to 15 degrees with the surface of the product to be tested, or an angle of 75 to 88 degrees with the normal of the product surface, to enhance the scattering response of the incident light to the fine protrusions, shallow scratches, particulate contaminants and local blistering of the protective film on the outermost surface of the product; the surface reflection enhancement polarization illumination component uses the surface reflection characteristics of polarized light to enhance the surface signal response.

[0076] (3) Interlayer sensitive lighting component 440 Interlayer sensitive illumination component 440 is used to provide interlayer sensitive illumination conditions. These conditions are designed to make the optical response of defects within the bonding layers of the product stronger than the optical response of defects on the outermost surface of the product. This causes the second verification image acquired by the second industrial camera 410 under these conditions to preferentially display the image features of interlayer defects, suppressing the response of surface defects in the image. Alternatively, the interlayer sensitive illumination component 440 may be at least one of a collimated backlight, a semi-transparent illumination component, a polarized transmission illumination component, or a reflective illumination component that suppresses surface specular reflection. Collimated backlights transmit illumination from the back of the product in a direction perpendicular to the product plane, causing interlayer defects with transmission attenuation characteristics, such as interlayer bubbles, to appear as dark areas in transmission imaging, creating high contrast with the bright background. Semi-transmissive illumination components introduce partial scattering on top of collimated backlighting, making them suitable for detecting interlayer scattering defects in products bonded to transparent materials. Polarized transmission illumination components utilize the transmission characteristics of polarized light, enhancing the contrast of interlayer defects by changing the polarization state as it passes through different interlayer defects. Reflective illumination components that suppress surface specular reflection suppress the specular reflection component of the outermost surface in reflective imaging mode through illumination direction design or polarization configuration, relatively enhancing the diffuse signal from between layers.

[0077] (4) Polarization switching component 450 The polarization switching component 450, under the control of the control unit 800, switches between different polarization states when acquiring a first verification image and when acquiring a second verification image, thereby further improving the layer differentiation capability between the first and second verification images. In this embodiment, the polarization switching component 450 includes an incident polarizer 451, an analytical polarizer 452, a polarizer switching frame 453, and a polarizer rotation drive mechanism 454. The incident polarizer 451 is disposed in the light path of the light source, so that the illumination beam first passes through the incident polarizer 451 to form polarized illumination light with a preset polarization direction before reaching the product; the analytical polarizer 452 is disposed in the receiving light path of the second industrial camera 410, at the front or rear end of the second lens 420, so that the imaging beam before reaching the sensor of the second industrial camera 410 first passes through the analytical polarizer 452. The polarizer switching frame 453 is used to carry the incident polarizer 451 and the analytical polarizer 452. The polarizer rotation drive mechanism 454 is connected to the polarizer switching frame 453 for transmission and is used to drive the polarizer switching frame 453 under the control of the control unit 800 to realize the relative adjustment of the polarization axis direction between the incident polarizer 451 and the analytical polarizer 452.

[0078] (5) Focal plane switching component 460 The focal plane switching assembly 460 is connected to the second industrial camera 410 and is used, under the control of the control unit 800, to switch the focus position of the second industrial camera 410 to a first focal plane close to the outermost surface of the product to obtain a first focal plane image, and to switch to a second focal plane close to the bonding interface to obtain a second focal plane image. Alternatively, the focal plane switching assembly 460 may be at least one of an electrically adjustable focusing lens, a camera precision displacement platform for driving the second industrial camera 410 to move along its imaging optical axis, or a lens precision displacement platform for driving the second lens 420 to move along its imaging optical axis.

[0079] The motorized focusing lens moves the focusing group inside the lens along the imaging optical axis through a built-in drive mechanism to change the focusing distance. It can switch focal planes without moving the second industrial camera 410 body or the product body. The focal plane switching response time is short, making it suitable for online rapid switching scenarios.

[0080] The camera precision displacement platform drives the second industrial camera 410 to generate a precise displacement along its imaging optical axis, thereby changing the imaging conjugate relationship between the sensor of the second industrial camera 410 and the surface of the measured product, thus achieving the switching of the focus surface position; its displacement accuracy can reach the micrometer level, and it is suitable for scenarios with high requirements for the repeatability accuracy of the focus surface position.

[0081] The lens precision displacement platform drives the second lens 420 to generate a precision displacement along its imaging optical axis, changing the distance between the second lens 420 and the sensor of the second industrial camera 410, thereby achieving the switching of the focal plane position. Its principle is the same as that of the camera precision displacement platform, which achieves focal plane switching by changing the conjugate relationship of the imaging system. It is suitable for configurations where the mechanical structure of the second lens 420 and the second industrial camera 410 allows for separate installation.

[0082] In an optional extended embodiment, the focal plane switching component 460 is further configured with a third focal plane, the focusing position of which is located between the first and second focal planes, for detecting defects located at a depth in the middle of the outer layer material of the product; or the focusing position of the third focal plane is located at an interlayer position below the bonding interface corresponding to the second focal plane, for further analysis of the depth distribution of interlayer defects. The second industrial camera 410 completes exposure at the focusing position corresponding to the third focal plane to obtain a third focal plane image, which can serve as evidence of supplementary layer information and participate in subsequent layer attribution determination.

[0083] Furthermore, based on the aforementioned second industrial camera 410, second lens 420, surface-sensitive illumination component 430, interlayer-sensitive illumination component 440, polarization switching component 450, focal plane switching component 460, and control unit 800, in this embodiment, the specific implementation methods for acquiring the first verification image under surface-sensitive illumination conditions, acquiring the second verification image under interlayer-sensitive illumination conditions, and acquiring the first focal plane image and the second focal plane image using the focal plane switching component are as follows: 1) Acquire the first verification image under surface-sensitive illumination conditions. The control unit 800 controls the surface-sensitive illumination component 430, the polarization switching component 450, and the second industrial camera 410 to work together. When a candidate anomaly region enters the field of view of the local verification imaging unit 400, the control unit 800 controls the surface-sensitive illumination component 430, the polarization switching component 450, and the second industrial camera 410 to work together to image the candidate anomaly region under surface-sensitive illumination conditions, thereby acquiring a first verification image. The purpose is to ensure that, through a layer-selective illumination configuration, defects at the outermost surface of the product have a stronger image contrast response in the imaging results of the second industrial camera 410 compared to defects inside the product's bonding layer. This allows the first verification image to preferentially present the image features of surface defects and suppress the response of interlayer defects.

[0084] The optical principle behind surface-sensitive illumination for layer selectivity lies in the following: contaminants, foreign objects, or localized damage to the protective film located on the outermost surface of the product, being directly exposed at the interface between the air and the product surface, will generate significantly enhanced scattering signals compared to the surrounding smooth, defect-free areas under oblique incident light or surface scattering-enhanced light. Conversely, interlayer bubbles or delamination defects located within the bonding layer require their scattering signals to penetrate multiple layers of material before reaching the second industrial camera 410. After interface refraction and internal scattering attenuation, the scattering signal reaching the second industrial camera 410 is relatively weak. By selecting an illumination angle and light source shape that primarily feature surface scattering, a signal contrast relationship favorable for imaging the outermost surface defects can be established between these two types of defects.

[0085] As described above, the surface-sensitive illumination component 430 can be at least one of a low-angle annular dark-field light component, a side-strip oblique illumination component, a highly directional surface scattering enhanced light component, or a surface reflection enhanced polarized illumination component. In this embodiment, a low-angle annular dark-field light component is used, with the light source uniformly incident from all around the product at an angle of incidence ranging from 60 to 80 degrees with the surface normal of the product. The specific incident angle is referenced to the configuration of the preset angle in the second illumination component 240 described above.

[0086] When acquiring the first verification image, the polarization switching component 450 switches to a first polarization state that enhances the surface scattering signal. The first polarization state is specifically implemented as follows: the polarizer rotation drive mechanism 454 drives the polarizer switching frame 453 to rotate the incident polarizer 451 and the analysis polarizer 452 to an angle where their polarization axes are parallel (i.e., parallel polarization configuration). In the parallel polarization configuration, the polarization component in the light scattered from the outermost surface of the product, which has the same polarization direction as the incident polarization, can smoothly pass through the analysis polarizer 452 and reach the sensor of the second industrial camera 410, thereby enhancing the surface scattering signal. The second industrial camera 410 completes the exposure under the condition that the surface sensitive illumination component 430 is lit and the polarization switching component 450 is in the first polarization state, thus obtaining the first verification image.

[0087] 2) Acquire the second verification image under interlayer sensitive illumination conditions. The control unit 800 controls the interlayer sensitive illumination component 440, the polarization switching component 450, and the second industrial camera 410 to work together to acquire a second verification image for the same candidate abnormal area under interlayer sensitive illumination conditions. The purpose is to make the internal defects of the product bonding layer have a stronger image contrast response in the imaging result of the second industrial camera 410 than the outermost surface defects of the product by using an alternative illumination configuration that contrasts with the surface sensitive illumination conditions. This allows the second verification image to preferentially present the image features of the interlayer defects and suppress the response of the surface defects.

[0088] The optical principle behind layer-sensitive illumination for layer selectivity lies in the fact that bubbles, delaminations, or inclusions within the bonding layer have a different refractive index than the surrounding bonding material. When light penetrates the product along the transmission direction, these interlayer defects refract, scatter, or absorb the transmitted light, resulting in significant contrast changes at corresponding positions in the transmission image. Conversely, contaminants or protective film damage on the outermost surface of the product primarily affect the light path near the outermost interface. Compared to the overall impact of interlayer defects on the transmitted light path, their contribution to contrast in the image under interlayer-sensitive illumination is relatively weak. Furthermore, when an illumination configuration that suppresses surface specular reflection is used (e.g., by offsetting the incident light from the specular reflection direction, or by using polarization suppression to prevent the surface specular reflection component from reaching the sensor of the second industrial camera 410), the specular reflection signal from the outermost surface is effectively suppressed, further highlighting the scattering signal from between layers in the image.

[0089] As described above, the interlayer sensitive illumination component 440 can be at least one of a collimated backlight, a semi-transmissive illumination component, a polarized transmissive illumination component, or a reflective illumination component that suppresses surface specular reflection. In this embodiment, a collimated backlight is used to provide transmissive illumination from the back of the product in a direction perpendicular to the product plane, so that interlayer defects such as interlayer bubbles appear as dark areas in the transmissive imaging, creating high contrast with the bright background. The intensity and direction parameters of the collimated backlight are configured in the product formulation management unit 900 in the form of interlayer sensitive illumination mode parameters.

[0090] When acquiring the second verification image, the polarization switching component 450 switches to a second polarization state that suppresses surface specular reflection and enhances interlayer transmission and scattering signals. The second polarization state is specifically implemented as follows: the polarizer rotation drive mechanism 454 drives the polarizer switching frame 453 to rotate the analytical polarizer 452 to an angle where its polarization axis is orthogonal to the polarization axis of the incident polarizer 451 (i.e., orthogonal polarization configuration). Under the orthogonal polarization configuration, specular reflection light generated from the outermost surface of the product cannot pass through the analytical polarizer 452, which has an orthogonal polarization axis, because its polarization direction remains consistent with the incident polarization direction, and is therefore effectively suppressed. Meanwhile, scattering light generated from defects within the product's bonding layer undergoes depolarization after multiple scatterings, allowing the scattering component with a polarization component consistent with the transmission axis of the analytical polarizer 452 to pass smoothly through the analytical polarizer 452 and reach the sensor of the second industrial camera 410, thereby enhancing the interlayer scattering signal. The second industrial camera 410 completes the exposure under the condition that the interlayer sensitive illumination component 440 is lit and the polarization switching component 450 is in the second polarization state, and obtains the second verification image.

[0091] It should be noted that both surface-sensitive illumination and interlayer-sensitive illumination follow the principle of relative contrast differentiation. That is, surface-sensitive illumination does not require the complete elimination of the signal response of interlayer defects, nor does interlayer-sensitive illumination require the complete elimination of the signal response of surface defects; rather, through a reasonable differential configuration between the two illumination conditions, a measurable change in the relative signal intensity relationship between the two types of defects is achieved between the first and second verification images. This relative difference forms the information basis for subsequent differential feature extraction steps.

[0092] 3) Use the focal plane switching component to acquire the first focal plane image and the second focal plane image. The control unit 800 controls the focal plane switching component 460 and the second industrial camera 410 to work together. For the same candidate abnormal area, the focus position of the second industrial camera 410 is switched sequentially to the first focal plane near the outermost surface of the product and the second focal plane near the bonding interface, respectively acquiring the first focal plane image and the second focal plane image. The purpose is to use the spatial selectivity of optical depth of focus by switching the focus position of the second industrial camera 410 to different depth layers, so that defects located at the focus layer present the highest clarity in the image, while defects located at the out-of-focus layer present lower clarity or even obvious blur, thereby obtaining clarity difference information with layer depth resolution capability.

[0093] The principle of optical depth-of-focus spatial selectivity is as follows: When the imaging system focuses on a certain focal plane, the blur spot formed on the sensor of the second industrial camera 410 at the object point located at that focal plane is the smallest (ideally close to an ideal point), and the image sharpness is the highest. For object points located in front of or behind the focal plane, the blur spot increases with the increase of defocus, and the corresponding image sharpness decreases. For hot-pressed products with a certain thickness, there is a certain physical distance between the outermost surface and the bonding interface. When the focus position is switched to the first focal plane (closer to the outermost surface of the product), defects located on the outermost surface of the product appear clear, while defects located at the bonding interface appear blurry; when the focus position is switched to the second focal plane (closer to the bonding interface), defects located at the bonding interface appear clear, while defects located on the outermost surface of the product appear blurry.

[0094] It should be further noted that the physical distance between the outermost surface and the bonding interface is equal to the thickness of the outer functional material, and this value varies depending on the product model. This physical distance can be configured in the product formula management unit 900 as a first focal plane position parameter and a second focal plane position parameter, respectively. Specifically, the first focal plane position parameter and the second focal plane position parameter refer to the driving parameters invoked by the focal plane switching component 460 when switching the focus position of the second industrial camera 410 to the first focal plane and the second focal plane, respectively. The first focal plane is the focal plane closest to the outermost surface of the product, and the second focal plane is the focal plane closest to the bonding interface. When the focal plane switching component 460 uses a motorized focusing lens, the first focal plane position parameter and the second focal plane position parameter can be the number of driving steps, the number of motor pulses, or the driving voltage value of the motorized focusing lens. When the focal plane switching component 460 uses a camera precision displacement platform for driving the second industrial camera 410 to move along its imaging optical axis, or a lens precision displacement platform for driving the second lens 420 to move along its imaging optical axis, the first focal plane position parameter and the second focal plane position parameter can be the displacement of the camera or lens along the imaging optical axis. During the system debugging phase, the control unit 800 controls the focal plane switching component 460 to gradually change the focus position within a preset search range, and controls the second industrial camera 410 to acquire images at each candidate focus position. Subsequently, the control unit 800 calculates the sharpness value in the image corresponding to each candidate focus position, determines the driving parameter corresponding to the peak sharpness of the outermost surface feature of the product as the first focal plane position parameter, and determines the driving parameter corresponding to the peak sharpness of the bonding interface feature, interlayer reference mark, or known interlayer defect feature as the second focal plane position parameter. If the bonding interface features are not easily observed directly in the standard sample, the initial driving position of the second focal plane is calculated based on the design thickness between the outermost surface of the product and the bonding interface, the material refractive index, and the position parameters of the first focal plane. This initial driving position is then corrected using the peak image sharpness of a sample with known interlayer reference marks or known interlayer defects. The control unit 800 writes the finally determined first and second focal plane position parameters as the test formula parameters for the current product model into the product formula management unit 900. When subsequently testing the same model of product, these parameters are used to control the focal plane switching component 460 to complete the focal plane switching.

[0095] As mentioned above, the focal plane switching component 460 employs at least one of the following: an electrically adjustable focusing lens, a camera precision displacement platform for driving the second industrial camera 410 to move along its imaging optical axis, or a lens precision displacement platform for driving the second lens 420 to move along its imaging optical axis. In this embodiment, the focal plane switching component 460 employs an electrically adjustable focusing lens. The specific process of focal plane switching is as follows: In the product formula management unit 900, a first focal plane position parameter and a second focal plane position parameter are pre-calibrated and stored for each product model; wherein, the first focal plane position parameter is the number of electrically adjustable focusing lens driving steps, the number of motor pulses, or the driving voltage value corresponding to switching the focus position to the vicinity of the outermost surface of the product, and the second focal plane position parameter is the number of electrically adjustable focusing lens driving steps, the number of motor pulses, or the driving voltage value corresponding to switching the focus position to the vicinity of the bonding interface. When performing focal plane switching, the control unit 800 first sends a drive command corresponding to the first focal plane position parameter to the focal plane switching component 460. The focal plane switching component 460 performs a focusing action and feeds back a positioning signal. After the positioning signal is fed back, the second industrial camera 410 completes the exposure and obtains the first focal plane image. Subsequently, the control unit 800 sends a drive command corresponding to the second focal plane position parameter to the focal plane switching component 460. The focal plane switching component 460 again performs a focusing action and feeds back a positioning signal. After the positioning signal is fed back, the second industrial camera 410 completes the exposure again and obtains the second focal plane image. The imaging illumination conditions of the first and second focal plane images can use the same illumination configuration as either the first or second verification image, or they can use a neutral illumination configuration specifically for focal plane image acquisition. The imaging illumination conditions, as focal plane image acquisition parameters, are stored in the product formula management unit 900 for different product models and are called by the control unit 800 during focal plane image acquisition.

[0096] Furthermore, the focal plane switching component 460 is also equipped with a third focal plane. In an optional extended embodiment, after switching to the focus position corresponding to the third focal plane, the control unit 800 controls the second industrial camera 410 to complete the exposure and obtain the third focal plane image. The third focal plane image can participate in the corresponding sharpness feature calculation in the difference feature extraction step, serving as supplementary torsional evidence for assisting torsional attribution determination.

[0097] 5. Image registration and candidate region management unit 500 The image registration and candidate region management unit 500 is connected to the candidate anomaly extraction unit 300 and the local verification imaging unit 400, respectively. On one hand, the image registration and candidate region management unit 500 is used to align the candidate anomaly region with the field of view of the local verification imaging unit 400 according to the regional center coordinate information of the candidate anomaly region; on the other hand, the image registration and candidate region management unit 500 is used to perform image registration on the first verification image, the second verification image, the first focal plane image and the second focal plane image, so that the image coordinates of the four images are aligned in the candidate anomaly region.

[0098] (1) Field of view alignment Field alignment is a preliminary step in acquiring the first verification image, the second verification image, the first focal plane image, and the second focal plane image. Its purpose is to align the candidate anomaly region with the field of view of the local verification imaging unit based on the regional center coordinate information of the candidate anomaly region.

[0099] The control unit 800, the transport and positioning unit 100, and the image registration and candidate area management unit 500 work together to align the candidate abnormal areas with the field of view of the local verification imaging unit 400. The purpose is to ensure that when each candidate abnormal area enters the field of view of the local verification imaging unit 400, the product is in a stable and stationary state, and the image coordinate position of the candidate abnormal area meets the field of view coverage requirements of subsequent multi-condition fine imaging.

[0100] The specific implementation of field alignment is as follows: The image registration and candidate area management unit 500 reads the region center coordinate information of the candidate abnormal area to be reviewed from the candidate abnormal area queue. The region center coordinate information is based on the initial screening imaging coordinate system (i.e., the image coordinate system of the first industrial camera 210, combined with the product coordinate system obtained by the calibration parameters of the initial screening station). The control unit 800 controls the conveying mechanism 110 in the conveying and positioning unit 100 to convey the product to the review station, and controls the review station stop mechanism 160 to extend so that the product remains stationary at the preset stop position of the review station; at the same time, the adsorption-type bearing platform 130 applies negative pressure adsorption to the product to further suppress the slight vibration of the product during the local review imaging stage.

[0101] After the product stops, the image registration and candidate area management unit 500 first converts the coordinates of the center of the candidate abnormal area in the initial screening image to its physical position in the product coordinate system. Then, based on the conveying displacement of the product when it moves from the initial screening station to the stopping position of the verification station, it converts this physical position to the verification station coordinate system and determines whether the candidate abnormal area falls within the field of view of the local verification imaging unit 400. Specifically, let the coordinates of the center of the candidate abnormal area in the initial screening image coordinate system be (u... i ,v i In the initial screening image, the object-space pixel size corresponding to a single pixel on the product surface is P.x and P y Then the center coordinates of the region can be converted to physical coordinates (x, y) in the product coordinate system. p ,y p ), where x p =u i ×P x y p =v i ×P y If there is a difference in rotation, translation, or scale between the initial screening image coordinate system and the product coordinate system, the coordinate transformation matrix from the initial screening coordinate system to the product coordinate system obtained during the system calibration phase is used for conversion. During the process of the product moving from the initial screening station to the verification station, the conveyor encoder 150 outputs the conveying displacement in real time. The control unit 800 calculates the displacement compensation amount Δs between the product stopping time and the initial screening image acquisition time based on the pulse count of the conveyor encoder 150. The image registration and candidate area management unit 500, based on the displacement compensation amount Δs and the fixed installation distance between the initial screening station and the verification station, determines the physical coordinates (x, y, x) of the candidate abnormal area in the product coordinate system. p ,y p Convert to coordinates (x) in the verification workstation coordinate system. r ,y r Verify the coordinates (x) in the workstation coordinate system. r ,y r This is used to indicate the positional offset of the candidate abnormal region relative to the center of the field of view of the local verification imaging unit 400. If the positional offset is within the effective field of view width and effective field of view height of the local verification imaging unit 400, it is determined that the candidate abnormal region has entered the field of view of the local verification imaging unit 400; if the positional offset exceeds the effective field of view range, the control unit 800 adjusts the product stop position according to the offset, or drives the displacement mechanism of the local verification imaging unit 400 to perform position compensation until the candidate abnormal region enters the verification field of view. Among them, the coordinate information of the center of the candidate abnormal region is based on the initial screening coordinate system; the conveying displacement of the product from the initial screening station to the verification station stop time is obtained by cumulative calculation of the pulse signal provided by the conveying encoder 150; the pre-calibrated coordinate transformation relationship between the initial screening coordinate system and the verification coordinate system is determined by imaging through the calibration plate during the system installation and debugging stage; the current field of view parameters of the local verification imaging unit 400 include: the magnification of the second lens 420 and the center position of the field of view of the second industrial camera 410.

[0102] Based on the above calculation results, the image registration and candidate region management unit 500 obtains the lateral offset Δx and vertical offset Δy of the candidate anomaly region relative to the center of the field of view of the second industrial camera 410, and determines whether the candidate anomaly region completely falls within the effective field of view of the second industrial camera 410. Specifically, let W be the effective field of view width of the second industrial camera 410 on the product surface. r The effective field of view height is H r The bounding dimensions of the candidate anomaly region in the horizontal and vertical directions are w, respectively. c and h c The width of the surrounding background band is b; when |Δx|+w c / 2+b≤W r / 2 and |Δy|+h c / 2+b≤H r When the deviation is Δx, it is determined that the candidate abnormal area and its surrounding background have entered the effective field of view of the second industrial camera 410; otherwise, it is determined that the candidate abnormal area has not completely entered the effective field of view of the second industrial camera 410. The control unit 800 adjusts the product stop position according to Δx and Δy, or drives the local verification imaging unit 400 to perform position compensation, until the above field of view judgment conditions are met. If it has entered the field of view, multi-condition imaging is performed directly; if it exceeds the field of view, the control unit 800 adjusts the product stop position according to the deviation, or performs field of view position compensation through the displacement mechanism of the movable local verification imaging unit 400, until it is confirmed that the candidate abnormal area has entered the field of view of the second industrial camera 410.

[0103] (2) Image registration Image registration occurs after the first verification image, the second verification image, the first focal plane image, and the second focal plane image have been obtained.

[0104] The control unit 800 invokes the image registration and candidate area management unit 500 to perform image registration on the first verification image, the second verification image, the first focal plane image, and the second focal plane image. The purpose of image registration is to eliminate the relative positional deviations caused by factors such as illumination switching, focal plane switching, minor residual movement of the product, and lens breathing effect during the separate acquisition of the four images, so as to align the image coordinates of the four images in the candidate abnormal area and establish a unified spatial reference for subsequent pixel-by-pixel difference feature extraction.

[0105] It should be noted that during the acquisition of four images, each switch in illumination mode or focal plane position requires corresponding mechanical actions (including polarizer rotation, focus adjustment, etc.), which inevitably introduce minor vibrations. Even with the adsorption platform 130 applying an adsorption effect to the product, there may still be sub-pixel-level residual product movement. Furthermore, the motorized focusing lens, while changing the focusing distance, usually experiences a slight change in the field of view (i.e., lens breathing), resulting in slight differences in magnification between images at different focal planes, manifesting as an overall scaling deviation between images. Without image registration, these positional and scaling deviations will introduce systematic errors during feature extraction, causing the corresponding coordinates of the same defect to shift in different images, leading to inaccurate results in contrast calculation, sharpness calculation, and edge sharpness calculation.

[0106] The specific implementation of image registration is as follows: Any one of the four images is used as the registration reference image (e.g., the first verification image is used as the registration reference image), and the remaining three images are registered separately. Registration processing can employ methods based on image cross-correlation or local feature point matching to calculate the two-dimensional transformation parameters of the image to be registered relative to the reference image. These two-dimensional transformation parameters include translation, rotation angle, and scaling. Considering that the main types of deviation between images of different focal planes are overall translation and scaling, and that the deviation is usually within a few pixels, a sub-pixel-level translation estimation method based on phase correlation can be used, combined with pre-calibrated lens breathing scaling compensation parameters (pre-calibrated during the system calibration phase for each product model and focal plane position combination). Image transformation is then performed on each image to be registered, aligning it with the coordinates of the reference image.

[0107] After image registration is completed, the image registration and candidate region management unit 500 calculate the registration error. The registration error can be represented by the root mean square value of the feature point coordinate residuals in the registered image. The preset registration error threshold is determined based on the object-space pixel size of the verification image and the minimum difference feature size required for subsequent difference feature extraction. Specifically, the control unit 800 determines the object-space pixel size P corresponding to a single pixel on the product surface in the verification image based on the imaging parameters of the second industrial camera 410 and the second lens 420. r Then, based on the minimum difference feature size d required for extraction in the local verification stage... r Determine the pixel width d of the minimum difference feature in the verification image. r / P r Then, the preset registration error threshold E is set. th Set to a preset ratio for that pixel width, for example, E. th ≤0.2×d r / P r Minimum difference feature size d rThis can be the minimum value among the width of the surface defect edge, the width of the bright ring of interlayer bubbles, or the width of the scattering halo ring. Preset registration error threshold E th Stored in the product formula management unit 900.

[0108] When the registration error does not exceed the preset registration error threshold E th If the residual misalignment of the four images at the candidate anomaly region is less than the maximum allowable deviation for subsequent differential feature extraction, then proceed to subsequent differential feature extraction; if the registration error exceeds the preset registration error threshold E, then proceed to subsequent differential feature extraction. th If the residual misalignment between the four images is not clear, it indicates that the accuracy of the calculation of surface response contrast, interlayer response contrast, focal plane sharpness, edge sharpness, or scattering halo features may be affected. Candidate abnormal areas are identified as abnormalities to be verified, or low-speed reshooting is triggered.

[0109] 6. Hierarchical Attribution Determination Unit 600 The layer attribution determination unit 600 is connected to the image registration and candidate region management unit 500, and is used to extract the differential features of candidate abnormal regions from the registered image and perform layer attribution determination.

[0110] (1) Extracting differential features The control unit 800 calls the tonal attribution determination unit 600 to extract the difference features of candidate anomalous regions from the four registered images. The purpose of difference feature extraction is to extract multi-dimensional quantitative features from the four registered images with aligned coordinates, for the pixels of the candidate anomalous regions, to reflect the signal differences between different lighting conditions and different focal plane positions, and to provide input basis for subsequent tonal attribution determination.

[0111] The differential features extracted in this embodiment include at least the following four basic features, and optionally further include four supplementary features. The definitions, calculation methods, and layer orientations of each feature are explained in detail below.

[0112] 1) Surface response contrast characteristics.

[0113] The surface response contrast feature uses the local contrast of the candidate anomaly region in the first verification image as a quantitative indicator of the surface response intensity. Specifically, in the first verification image, the average gray value of the candidate anomaly region is calculated within its pixel range, and the average gray value of the background is calculated within a ring-shaped background band formed by extending a preset width outside the candidate anomaly region. The absolute value of the difference between the average gray value of the candidate anomaly region and the average gray value of the background is divided by the average gray value of the background to obtain a normalized contrast value C1, which is used as the feature value of the surface response contrast feature. The layer orientation of the surface response contrast feature is as follows: In the first verification image, the control unit 800 calculates the first response contrast C1 of the candidate anomaly region relative to the surrounding background region; when C1 is not less than a preset surface response threshold T... C1 Furthermore, the difference between C1 and the second response contrast C2 satisfies C1−C2≥response difference threshold ΔT. C When C1 is less than the preset surface response threshold T, the candidate anomaly region is determined to meet the surface response characteristics; C1 Or the difference between C1 and C2 does not satisfy C1−C2≥ΔT C When the candidate anomaly region does not meet the surface response characteristics, it is determined that the anomaly region does not meet the surface response characteristics.

[0114] 2) Interlayer response contrast characteristics.

[0115] The interlayer response contrast feature uses the local contrast of the candidate anomaly region in the second verification image as a quantitative indicator of the interlayer response intensity. The calculation method is consistent with the surface response contrast feature, yielding a normalized contrast value C2, which is then used as the feature value of the interlayer response contrast feature. The layer-specificity of the interlayer response contrast feature is as follows: In the second verification image, the control unit 800 calculates the second response contrast C2 of the candidate anomaly region relative to the surrounding background region; when C2 is not less than a preset interlayer response threshold T... C2 Furthermore, the difference between C2 and the first response contrast C1 satisfies C2−C1≥ΔT C When C2 is less than the preset interlayer response threshold T, the candidate anomaly region is determined to meet the interlayer response characteristics; C2 Or the difference between C2 and C1 does not satisfy C2−C1≥ΔT C When the candidate anomaly region does not meet the interlayer response characteristics, it is determined that the anomaly region does not meet the interlayer response characteristics.

[0116] 3) Sharpness characteristics of the first focal plane The first focal plane sharpness feature uses the local image sharpness of the candidate anomaly region in the first focal plane image as a quantitative indicator. Sharpness calculation can be achieved using the Laplacian operator variance method: Apply the Laplacian operator to each pixel within the pixel range of the candidate anomaly region in the first focal plane image, and calculate the variance of the operator response values. A larger variance value indicates richer edge information and a sharper image within the region. The resulting sharpness value F1 is the feature value of the first focal plane sharpness feature. The layer orientation of the first focal plane sharpness feature is as follows: In the first focal plane image, the control unit 800 calculates the first focal plane sharpness index F1 of the candidate anomaly region; when F1 is not less than the preset first focal plane sharpness threshold T... F1 Furthermore, the difference between F1 and the second focal plane sharpness index F2 satisfies F1−F2≥focal plane sharpness difference threshold ΔT. F When the candidate abnormal region satisfies the surface focal plane sharpness feature, the region is determined to be: when F1 is less than the preset first focal plane sharpness threshold T. F1 Or the difference between F1 and F2 does not satisfy F1−F2≥ΔT F When the candidate abnormal region does not meet the surface focal plane clear feature, it is determined that the candidate abnormal region does not meet the feature.

[0117] 4) Second focal plane sharpness characteristics The second focal plane sharpness feature uses the local image sharpness of the candidate anomaly region in the second focal plane image as a quantification index. The calculation method is the same as the first focal plane sharpness feature, yielding a sharpness value F2, which is then used as the feature value of the second focal plane sharpness feature. The layer orientation of the second focal plane sharpness feature is as follows: In the second focal plane image, the control unit 800 calculates the second focal plane sharpness index F2 of the candidate anomaly region; when F2 is not less than the preset second focal plane sharpness threshold T... F2 Furthermore, the difference between F2 and the first focal plane sharpness index F1 satisfies F2−F1≥ΔT. F When the candidate abnormal region satisfies the interlayer focal plane sharpness feature, the condition is determined; when F2 is less than the preset second focal plane sharpness threshold T... F2 Or the difference between F2 and F1 does not satisfy F2−F1≥ΔT F When the candidate abnormal region does not meet the requirements for sharp interlayer focal plane, it is determined that the abnormal region does not meet the requirements for sharp interlayer focal plane.

[0118] 5) Surface edge sharpness characteristics As a supplementary feature, the surface edge sharpness feature uses the image gradient magnitude of the edge of the candidate anomaly region in the first verification image as a quantification index. Specifically, in the first verification image, within a strip-shaped region of a preset pixel width along the contour line of the candidate anomaly region, the Sobel operator or Canny operator is applied to each pixel to calculate the image gradient magnitude, and the mean or a preset high percentile (e.g., the 90th percentile) of the gradient magnitude within this strip-shaped region is taken as the gradient magnitude of the surface edge sharpness feature. The layer orientation of the surface edge sharpness feature is as follows: the control unit 800 calculates the average gradient magnitude G at the boundary of the candidate anomaly region in the first verification image. e When the average gradient magnitude G e Not less than the preset edge gradient threshold T G When the candidate anomaly region has a clear boundary and meets the surface boundary characteristics, the average gradient magnitude G is determined to be... e Less than the preset edge gradient threshold T G When the candidate anomaly region boundary does not meet the condition of a clear surface boundary, it is determined that the boundary of the candidate anomaly region does not meet the condition of a clear surface boundary. For interlayer bonding defects, since their boundaries are located inside the material and are affected by the optical blurring effect of the upper material layer, their corresponding average gradient magnitude is usually lower than the preset edge gradient threshold T. G .

[0119] 6) Interlayer scattering propagation characteristics As a supplementary feature, the interlayer scattering spread feature uses the ratio of the scattering halo area of ​​the edge of the candidate anomaly region in the second verification image as a quantitative indicator. The scattering halo area ratio is defined as follows: In the second verification image, a ring-shaped region is formed by extending outwards from the boundary of the candidate anomaly region by a preset width (e.g., 5 to 15 pixels). The sum of the pixel areas within this ring-shaped region whose brightness deviation exceeds a preset scattering brightness threshold is counted, and then divided by the area of ​​the candidate anomaly region itself. The resulting ratio is the scattering halo area ratio. The layer orientation of the interlayer scattering spread feature is as follows: The control unit 800 calculates the scattering halo area ratio A within the ring-shaped region outside the candidate anomaly region in the second verification image. halo When the area of ​​the scattering halo is greater than A halo Not less than the preset threshold T for the ratio of the area of ​​the scattering halo ring A When the candidate anomaly region satisfies the interlayer scattering halo feature, the area of ​​the scattering halo is compared to A. halo Less than the preset scattering halo area ratio threshold T A When the candidate anomalous region does not meet the interlayer scattering halo feature, it is determined that the anomalous region does not meet the feature.

[0120] The aforementioned thresholds were determined through statistical calibration of sample images during the system debugging phase and written into the product formula management unit 900 as the layer attribution judgment threshold group for the corresponding product model. Specifically, the control unit 800 acquires the first verification image, second verification image, first focal plane image, and second focal plane image of a defect-free standard sample, a sample with known surface defects, and a sample with known interlayer defects, respectively. The control unit performs the same image registration and difference feature extraction as in formal testing on these images to obtain surface response contrast C1, interlayer response contrast C2, first focal plane sharpness F1, second focal plane sharpness F2, and edge gradient magnitude G. e Scattering halo area ratio A halo The ratio of average brightness of the bright ring and the dark core, R ring and position offset D pos The control unit 800 determines the judgment thresholds that can distinguish between surface defects, interlayer bonding defects, and defect-free background fluctuations based on the value distribution of various differential characteristics of the defect-free standard sample, the known surface defect sample, and the known interlayer defect sample. For judgment conditions that require comparing the strength of two types of characteristics, the control unit 800 further calculates the corresponding difference and determines the difference threshold. After calibration, the control unit 800 stores the surface response threshold, interlayer response threshold, response difference threshold, focal plane sharpness threshold, focal plane sharpness difference threshold, edge gradient threshold, scattering halo area ratio threshold, bright ring dark core structure threshold, and position offset threshold as the layer attribution judgment threshold group for the current product model in the product formula management unit 900, and calls this threshold group to perform layer attribution judgment when subsequently testing the same model of product. 7) Bright ring dark core structure index As a supplementary feature, the bright ring-dark core structure index is used to characterize whether the candidate anomaly region has an interlayer bubble-like structure feature in the second verification image where the brightness of the edge region is higher than that of the center region. Specifically, in the second verification image, the control unit 800 first determines the boundary of the candidate anomaly region; then, inside the candidate anomaly region, a preset width w is extended from the boundary as the starting line towards the center of the candidate anomaly region. ring This forms a ring of pixels inside the boundary of the candidate anomaly region, and this ring region is defined as the annular region; the remaining pixels within the annular region are defined as the central region. Preset width w ring The value can be determined based on the equivalent diameter or circumscribed dimension of the candidate abnormal region. It can be set to 10% to 30% of the equivalent diameter of the candidate abnormal region, or to 2 to 10 pixels, to ensure that both the annular region and the central region have a sufficient number of pixels for average brightness calculation.

[0121] Control unit 800 calculates the average brightness I of the annular region respectively. ring and the average brightness I of the central regioncenter And calculate the bright ring dark core structure index R according to the following formula. ring Bright ring dark core structure index R ring The calculation formula is: R ring =I ring / I center .

[0122] For interlayer bubble defects, under interlayer sensitive illumination conditions, the central region of the bubble may appear as a region of reduced brightness due to the refraction and deflection of transmitted light, while the edge region of the bubble may appear as a region of increased brightness due to interface reflection and edge scattering. Therefore, its R ring It can typically reach or exceed the preset bright ring dark core structure threshold T. R The control unit 800 will R ring With the preset bright ring dark core structure threshold T R Comparison: When R ring ≥T R When R is selected, the candidate anomaly region is determined to have a bright ring-dark core structure; when R ring <T R When the candidate anomaly region is determined to not have a bright ring-dark core structure, it is determined that the anomaly region does not have the characteristics of a bright ring-dark core structure. A preset bright ring-dark core structure threshold T is used. R During the system debugging phase, the layer location is determined by calibrating samples with known interlayer bubble defects and samples with known surface contamination defects, and stored in the product formula management unit 900 as the layer location attribution judgment parameter for the current product model.

[0123] For example, if the equivalent diameter of a candidate anomaly region is 30 pixels, the control unit 800 can... ring Set to 6 pixels; in this case, a ring of pixels with a width of 6 pixels close to the boundary inside the candidate anomaly region is defined as a ring region, and the remaining pixels inside this ring region are defined as the central region. If the average brightness I of the ring region... ring The value is 150, and the average brightness of the central area is I. center If R is 100, then ring =1.5; if the preset bright ring dark core structure threshold T corresponding to the current product model R If the value is 1.3, then the candidate anomaly region is determined to have a bright ring dark core structure.

[0124] 8) Position stability characteristics As a supplementary feature, the positional stability feature uses the positional offset of the candidate anomaly region's center in four registered images (first verification image, second verification image, first focal plane image, and second focal plane image) as a quantification index. Specifically, the centroid coordinates of the candidate anomaly region are independently detected in each of the four registered images, and the maximum mutual distance between the four sets of centroid coordinates is used as the positional offset. The layer orientation of the positional stability feature is as follows: the control unit 800 extracts the region center coordinates of the candidate anomaly region in each of the four registered images and calculates the positional offset D between the region center coordinates. pos When the position offset D pos Not greater than the preset position offset threshold T D When the positional consistency of the candidate anomaly region in the four images meets the requirements, the positional offset D is determined to be within acceptable limits. pos Greater than the preset position offset threshold T D At that time, the positional consistency of the candidate anomaly region in the four images was determined to be insufficient. For the positional offset D... pos Greater than the preset position offset threshold T D For candidate anomaly regions, the control unit 800 can classify them as anomalies to be reviewed or reduce their confidence as real physical defects. Therefore, the positional stability feature itself does not directly distinguish between surface defects and interlayer defects, but rather serves as a basis for judging the confidence of the authenticity of candidate anomaly regions, assisting in identifying real defects and false positive candidate regions in the layer attribution determination.

[0125] Among the aforementioned differences, surface response contrast features, interlayer response contrast features, first focal plane sharpness features, and second focal plane sharpness features constitute the basic feature group; surface edge sharpness features, interlayer scattering spread features, bright ring dark core structure index, and positional stability features constitute the supplementary feature group.

[0126] (2) Attribution determination at the execution level The control unit 800 calls the layer attribution determination unit 600 to perform layer attribution determination based on the difference characteristics, classifying the candidate abnormal area into one of three categories: surface defects, interlayer bonding defects, or anomalies to be reviewed. The purpose is to: make an interpretable layer attribution conclusion for the candidate abnormal area based on the multi-dimensional quantitative information obtained by extracting the difference characteristics, according to the pre-set multi-condition determination rules, and output the corresponding handling instructions accordingly.

[0127] The logic behind the multi-condition judgment rule is as follows: Considering the diversity of defect morphology, size, and optical characteristics on real production lines, no single distinguishing feature can independently and reliably differentiate between surface defects and interlayer defects in all situations; any single feature carries a certain possibility of misjudgment. To improve the overall reliability of the judgment, candidate anomaly regions are required to simultaneously meet at least two of the multiple independent feature judgment conditions corresponding to a certain type of defect before being classified as that type of defect; when the judgment conditions for two types of defects are met simultaneously, it is judged as an anomaly awaiting review; when none of the features reach the corresponding threshold, it is also judged as an anomaly awaiting review. This judgment logic establishes the layer attribution conclusion on the basis of multi-feature mutual verification, effectively reducing the overall attribution error rate caused by misjudgment due to a single feature.

[0128] The stratum attribution determination in this embodiment is performed according to the following rules.

[0129] (1) Conditions for classifying a candidate anomaly as a surface defect (including surface contamination defects or protective film defects): A candidate anomaly is classified as a surface defect if it meets at least two of the following four conditions: Condition (A1): The eigenvalue C1 of the surface response contrast feature is higher than the eigenvalue C2 of the interlayer response contrast feature, i.e., C1>C2. This condition indicates that the contrast response of the candidate anomalous region under surface-sensitive lighting conditions is higher than that under interlayer-sensitive lighting conditions, which is consistent with the typical relative response pattern of surface defects under the two types of lighting conditions. Condition (A2): The feature value F1 of the first focal plane sharpness feature is higher than the feature value F2 of the second focal plane sharpness feature, i.e., F1>F2; This condition indicates that the sharpness of the candidate abnormal area in the first focal plane image focused on the outermost surface of the product is higher than that in the second focal plane image focused on the bonding interface, which is consistent with the focal plane sharpness feature of the outermost surface layer defect. Condition (A3): The gradient magnitude of the surface edge sharpness feature exceeds the preset edge sharpness threshold T. edge This condition indicates that the candidate anomaly region exhibits clear and sharp edge features in the first verification image, consistent with the edge features of solid contaminants or localized damage to the protective film located on the outermost surface. Condition (A4): The ratio of the scattering halo area of ​​the interlayer scattering spread characteristic is lower than the preset scattering spread threshold T. scatter This condition indicates that the scattering halo around the outer edge of the candidate anomaly region in the second verification image is small and does not have the typical interlayer bubble edge scattering extension characteristics.

[0130] (2) Conditions for determining interlayer bonding defects: When a candidate abnormal region simultaneously meets at least two of the following four conditions, the candidate abnormal region is determined to be an interlayer bonding defect: Condition (B1): The characteristic value C2 of the interlayer response contrast feature is not lower than the characteristic value C1 of the surface response contrast feature, i.e., C2≥C1. This condition indicates that the contrast response of the candidate anomalous region under interlayer sensitive lighting conditions is not weaker than that under surface sensitive lighting conditions, which is consistent with the typical relative response mode of interlayer defects under the two types of lighting conditions. Condition (B2): The feature value F2 of the second focal plane sharpness feature is higher than the feature value F1 of the first focal plane sharpness feature, i.e., F2>F1; This condition indicates that the sharpness of the candidate abnormal area in the second focal plane image focused on the bonding interface is higher than that in the first focal plane image focused on the outermost surface of the product, which is consistent with the focal plane sharpness feature of the bonding interface layer defect. Condition (B3): The ratio of the scattering halo area of ​​the interlayer scattering spread characteristic exceeds the preset scattering spread threshold T. scatter This condition indicates that the candidate anomaly region has a significant scattering halo on the outer edge of the second verification image, which is consistent with the edge scattering extension characteristics of typical interlayer bubbles. Condition (B4): The ratio of the average brightness of the annular region to the central region in the bright ring-dark core structure index. ring Exceeding the preset bright ring dark core threshold T ring This indicates that the candidate anomaly region exhibits a distinct optical structure in the second verification image where a central dark area and a surrounding bright ring appear simultaneously, consistent with the bright ring and dark core characteristics of interlayer bubbles; or the positional offset of the positional stability feature is lower than the preset stability threshold T. stable This indicates that the candidate anomaly region is highly stable in location across the four registered images, representing a genuine defect with a definite physical location. Combining this with other inter-layer feature evidence enhances the confidence of the inter-layer attribution. Condition (B4) is included in the satisfaction count of the condition group in a logical "OR" manner, i.e., as long as R... ring More than T ring Or the position offset is less than T stable If any one of them is true, then condition (B4) is satisfied.

[0131] (3) Conditions for determining an anomaly to be reviewed: When one of the following situations occurs, the candidate anomaly region is determined as an anomaly to be reviewed: 1) The judgment condition group for surface defects and the judgment condition group for interlayer bonding defects are simultaneously satisfied, that is, at least two conditions in both condition groups are met, resulting in attribution conflict; 2) The feature values ​​of each differential feature do not reach their corresponding preset thresholds, and there is a lack of discriminative evidence required for clear classification; 3) The registration error of image registration exceeds the preset registration error threshold, which cannot guarantee the accuracy of differential feature extraction. The conditions for anomalies to be reviewed are a mechanism for handling uncertainty, avoiding the forced conclusion of potentially erroneous attribution when information is insufficient.

[0132] The preset thresholds involved in the above-mentioned judgment rules include the preset edge sharpness threshold T.edge Preset scattering spread threshold T scatter Preset bright ring dark core threshold T ring Preset stability threshold T stable The preset registration error thresholds are stored as layer attribution judgment threshold groups in the product formula management unit 900 and configured separately for different product models. Different product models have different layer structures, protective film haze, bonding material transmittance, adhesive layer thickness, surface reflection characteristics, and verification imaging resolution. Therefore, the preset thresholds do not use uniform fixed values, but are calibrated for the current product model during the system debugging phase.

[0133] Specifically, during the system debugging phase, the control unit 800 acquires first verification images, second verification images, first focal plane images, and second focal plane images of defect-free standard samples, samples with known surface defects, and samples with known interlayer defects, respectively. Among these, samples with known surface defects include those with surface contaminants, scratches on the protective film surface, localized damage to the protective film, or residual adhesive on the surface; samples with known interlayer defects include those with interlayer bubbles, interlayer inclusions, or localized debonding. The control unit 800 performs the same image registration, boundary extraction, annular region segmentation, central region segmentation, and differential feature extraction on the above sample images as in the formal inspection, obtaining the surface edge sharpness feature value G for each sample. edge Scattering halo area ratio A halo Bright ring dark core structure index R ring Position offset D pos And registration error E reg .

[0134] For the preset edge sharpness threshold T edge The control unit 800 statistically analyzes the surface edge sharpness characteristic value G of samples with known surface defects and samples with known interlayer defects. edge Since surface defects typically exhibit more concentrated boundary grayscale changes in the first verification image, while interlayer defects are affected by optical blurring of material layers, their boundary grayscale changes are usually more gradual. Therefore, the control unit 800 is selected to ensure that most known surface defect samples satisfy G. edge ≥T edge Furthermore, most known interlayer defect samples do not satisfy G. edge ≥T edge The value is used as the preset edge sharpness threshold T edge The initial value.

[0135] For the preset scattering spread threshold T scatter The control unit 800 statistically analyzes the ratio A of the scattering halo area of ​​samples with known surface defects and samples with known interlayer defects. haloBecause interlayer bubbles, inclusions, or localized debonding typically form a scattering propagation region outside the defect body boundary under interlayer sensitive illumination conditions, while surface contaminants or defects on the protective film surface usually do not form the same degree of internal scattering halo around the material, the control unit 800 is selected to ensure that most known interlayer defect samples meet A. halo ≥T scatter Furthermore, most known surface defect samples do not meet A. halo ≥T scatter The value is used as the preset scattering spread threshold T scatter The initial value.

[0136] For the preset bright ring dark core threshold T ring The control unit 800 statistically analyzed the bright ring dark core structure index R of known interlayer bubble samples and known surface defect samples. ring R ring This represents the ratio of the average brightness of the annular region to the average brightness of the central region within the candidate anomalous region. Since interlayer bubbles tend to form structural features where the brightness of the edge region is higher than that of the central region under interlayer sensitive illumination conditions, and surface contaminants or protective film defects typically do not possess this interlayer bubble optical structure, the control unit 800 selects samples that satisfy R... ring ≥T ring Furthermore, most known surface defect samples do not satisfy R. ring ≥T ring The value is used as the preset bright ring dark core threshold T ring The initial value.

[0137] For the preset stability threshold T stable The control unit 800 statistically analyzes the offset D of the region center coordinates of the same real physical defect in the first verification image, the second verification image, the first focal plane image, and the second focal plane image. pos And calculate the offset D of the region center coordinates of the image noise, reflection artifacts, or segmentation mis-triggered regions in the above image. pos Since the positions of real physical defects in multiple registered images should remain near the same physical region, while the positional stability of artifacts or noise-triggered regions is poor, the control unit 800 is selected to ensure that most real physical defects satisfy D. pos ≤T stable Furthermore, most artifacts or noise triggering regions do not satisfy D. pos ≤T stable The value is used as the preset stability threshold T stable The initial value.

[0138] For the preset registration error threshold, the control unit 800 determines it based on the object-side pixel size of the verification image and the minimum difference feature size required for subsequent difference feature extraction. Specifically, the control unit 800 determines the object-side pixel size P of a single pixel in the verification image corresponding to the product surface based on the imaging parameters of the second industrial camera 410 and the second lens 420. r And extract the minimum differential feature size d as needed. r Determine the pixel width d of this feature in the verification image. r / P r Then, the preset registration error threshold is set to a preset ratio of the pixel width, for example, to no more than 0.2×d. r / P r Therefore, the residual registration error between the four verification images will not significantly affect the calculation of surface response contrast, interlayer response contrast, focal plane sharpness, edge sharpness, scattering halo area ratio, and bright ring dark core structure index.

[0139] After determining the initial threshold as described above, the control unit 800 will... edge T scatter T ring T stable The preset registration error threshold is written into the product formula management unit 900 as the layer attribution judgment threshold group for the current product model. During subsequent trial operation, the control unit 800 reviews the difference characteristic values ​​corresponding to misjudged and missed samples based on manual verification results or confirmation results from subsequent processes. When a surface defect is misjudged as an interlayer bonding defect, the control unit 800 corrects the thresholds related to interlayer judgment; when an interlayer bonding defect is misjudged as a surface defect, the control unit 800 corrects the thresholds related to surface judgment or interlayer judgment. The corrected thresholds are rewritten into the product formula management unit 900 as the layer attribution judgment threshold group to be used in subsequent testing of this product model.

[0140] In an optional extended implementation, the layer attribution determination unit 600 can also be configured with a classification model based on the above-mentioned multi-condition rule determination to assist in verifying the rule determination conclusion. Specifically, the difference features extracted from the registered image are input into the classification model, and the classification model outputs layer attribution probability values ​​corresponding to the two categories of "surface defects" and "interlayer bonding defects".

[0141] The classification model can be implemented using conventional classifier structures in the field, such as support vector machines, gradient boosting decision trees, or lightweight convolutional neural networks. The input to the classification model is the extracted multidimensional differential feature vector, or preprocessed candidate anomaly region image patches; the output is the probability value of each attribution category. The classification model is trained offline using a large number of historical defect samples with labeled hierarchical attribution conclusions. During training, conventional loss functions in the field, such as cross-entropy loss or hinge loss, can be used, and the model's hyperparameters are fine-tuned using a validation set. After training, the classification model is deployed in inference mode within the runtime environment of the hierarchical attribution determination unit 600.

[0142] The output of the classification model serves as an auxiliary verification basis for the above rule-based judgment conclusions: when the rule judgment gives a clear classification conclusion and the attribution probability output by the classification model is consistent with the conclusion, the confidence of the attribution conclusion is further enhanced; when there is a discrepancy between the rule judgment and the output of the classification model, the candidate abnormal region can be transferred to the anomaly processing flow to be reviewed, and further confirmed by manual review or low-speed re-shooting mechanism.

[0143] 7. Result Output and Execution Unit 700 The result output and execution unit 700 is connected to the stratum attribution determination unit 600, and is used to output abnormal stratum attributes and corresponding handling instructions based on the classification results.

[0144] When a candidate abnormal area is classified as a surface defect, the result output and execution unit 700 output a cleaning re-inspection command or a film peeling re-inspection command. The cleaning re-inspection command indicates that the defect may be a cleanable surface contaminant; the product is cleaned at the cleaning station and then returned to the inspection line for re-inspection. The film peeling re-inspection command indicates that the defect may be damage to the protective film itself or foreign matter beneath the protective film; the protective film is peeled off at the film peeling station, and the substrate is re-inspected to confirm its integrity. The specific selection of the cleaning re-inspection command and the film peeling re-inspection command can be further subdivided based on the morphological characteristics in the initial screening or review images. For example, a cleaning re-inspection command is tended to be output for contaminants appearing as scattered dots, while a film peeling re-inspection command is tended to be output for damage appearing as sheets or associated with the edge of the protective film.

[0145] When a candidate abnormal area is classified as an interlayer bonding defect, the result output and execution unit 700 output either a defect rejection instruction or a rework transfer instruction. The defect rejection instruction is applicable when the defect area is large, or the defect location is in a critical functional area of ​​the product, and rework is not feasible. In this case, the defect rejection mechanism diverts the defective product from the production line. The rework transfer instruction is applicable when the defect area is small, located in a non-critical area, and rework is possible. In this case, the rework station processes the product for rework. The specific selection of the defect rejection instruction and the rework transfer instruction can be based on a comprehensive judgment of the defect's position in the product coordinate system (different sub-regions within the effective detection area have different criticality levels) and parameters such as the defect's area and shape.

[0146] When a candidate abnormal area is classified as an abnormality requiring verification, the result output and execution unit 700 output either a manual verification instruction or a low-speed re-image instruction. The manual verification instruction indicates that the current automatic detection conclusion has uncertainty and needs to be confirmed by manual visual verification; the low-speed re-image instruction instructs the product to re-pass through the initial screening station and the verification station for image acquisition in a low-speed operation state to reduce information loss caused by factors such as vibration or insufficient imaging quality.

[0147] 8. Control Unit 800 The control unit 800 is connected to the aforementioned units and is used to uniformly control the coordinated operation of the units. Specifically, the control unit 800 can be implemented using an industrial computer, a dedicated image processing server, or a combination of a field-programmable gate array (FPGA) and an industrial computer. The image processing algorithms (including image preprocessing, candidate region extraction, image registration, differential feature extraction, and layer attribution determination) run in the industrial computer or image processing server. The motion control of the conveying and positioning unit 100, the initial screening and image acquisition unit 200, and the local verification imaging unit 400 can be executed by a PLC or a field controller, with the control unit 800 handling the instruction coordination and data interaction between the two.

[0148] 9. Product Formula Management Unit 900 To enable the testing system of this invention to adapt to flexible production scenarios with multiple product models, the system may also include a product formula management unit 900. The product formula management unit 900 stores testing formulas for different product models, with each product model corresponding to a complete set of testing formula parameters.

[0149] In this embodiment, the detection formula includes the following parameters: effective detection area range (used to define the effective detection coordinate range within the product area), ineffective shielding area range (used to define the coordinate range of the product edge area, alignment mark area, and other areas that do not need to be detected that need to be shielded), initial screening deviation threshold (used to determine the local brightness deviation threshold for candidate area extraction in the initial screening stage), surface sensitive lighting mode parameters (including the light source power, lighting angle, and lighting trigger timing parameters of the surface sensitive lighting component 430), interlayer sensitive lighting mode parameters (including the light source power, lighting angle or backlight intensity, and lighting trigger timing parameters of the interlayer sensitive lighting component 440), first focal plane position parameters (driving parameters of the focal plane switching component 460 corresponding to the outermost surface focal plane of the product), second focal plane position parameters (driving parameters of the focal plane switching component 460 corresponding to the bonding interface focal plane), and layer attribution determination threshold group (including the aforementioned preset thresholds).

[0150] When a product enters the testing line, the control unit 800 reads the testing formula corresponding to the current product model from the product formula management unit 900 and automatically configures each parameter in the testing formula to the corresponding unit. Specifically, the control unit 800 sends the effective detection area range and the invalid shielding area range to the candidate anomaly extraction unit 300, so that it processes the image preprocessing and candidate region extraction according to the effective area and shielding area of ​​the current product model; sends the initial screening deviation threshold to the candidate anomaly extraction unit 300 for threshold segmentation; sends the surface sensitive illumination mode parameters and interlayer sensitive illumination mode parameters to the drive interfaces of the surface sensitive illumination component 430 and the interlayer sensitive illumination component 440 in the local verification imaging unit 400, respectively; sends the first focal plane position parameters and the second focal plane position parameters to the drive interface of the focal plane switching component 460; and sends the layer attribution determination threshold group to the layer attribution determination unit 600.

[0151] When the production line switches product models, the corresponding testing formula only needs to be switched through the product formula management unit 900. The control unit 800 can then run the entire system according to the testing formula of the new product model after the formula switch is completed, without the need for physical adjustments to the system hardware.

[0152] 10. Anomaly Review Priority Scheduling Unit 910 In actual production line operation, the candidate abnormal area queue may contain multiple candidate abnormal areas. If multi-condition local review imaging is performed on all candidate abnormal areas one by one in the reading order, the total review time for a single product may exceed the production line cycle time limit. To address this, this system further includes an abnormal review priority scheduling unit 910, which prioritizes the review order of multiple candidate abnormal areas.

[0153] The anomaly review priority scheduling unit 910 assigns a review priority to each candidate anomaly region based on at least one of the following: area, brightness deviation, and morphological category. Specifically, the larger the area of ​​a candidate anomaly region, the more significant the potential impact of the corresponding defect, and the higher the corresponding review priority; the greater the brightness deviation of a candidate anomaly region in the initial screening image, the higher the probability that the candidate anomaly region belongs to a real significant defect, and the higher the corresponding review priority; the morphological category of the candidate anomaly region can be pre-classified in the initial screening stage. For example, circular or elliptical candidate anomaly regions may initially point to bubble-like interlayer defects, and irregularly shaped candidate anomaly regions may initially point to contaminant-like surface defects. A higher review priority is assigned to candidate anomaly regions that may correspond to interlayer bonding defects based on their morphological category.

[0154] The anomaly review priority scheduling unit 910 sends scheduling instructions to the local review imaging unit 400 according to the set review priority. The local review imaging unit 400 performs multi-condition local reviews on each candidate anomaly region in order of priority. For candidate anomaly regions with lower priority, if the production line cycle time does not allow all reviews to be completed within the current cycle, an anomaly mark to be reviewed is output for the remaining candidate anomaly regions, and subsequent processing is carried out by the supplementary review mechanism outside the production line cycle time.

[0155] In summary, the online appearance defect recognition system after hot-press bonding provided in this embodiment has a complete closed-loop operation as follows: the control unit 800 uniformly manages the coordinated operation sequence of the conveying and positioning unit 100, the initial screening and imaging unit 200, the candidate anomaly extraction unit 300, the local verification imaging unit 400, the image registration and candidate area management unit 500, the layer attribution determination unit 600, the result output and execution unit 700, the product formula management unit 900, and the anomaly verification priority scheduling unit 910, ensuring that the data transmission and action triggering between each step of initial screening and imaging, candidate area extraction, candidate area alignment, multi-condition local verification imaging, image registration, difference feature extraction, layer attribution determination, and disposal instruction output are executed in the correct order.

[0156] In a preferred embodiment of this system, the transport positioning unit 100, the initial screening image acquisition unit 200, the candidate anomaly extraction unit 300, the local verification imaging unit 400, the image registration and candidate area management unit 500, the layer attribution determination unit 600, and the result output and execution unit 700 sequentially form a collaborative link of "transport-initial screening-extraction-alignment-verification-registration-determination-output": the transport positioning unit 100 provides a unified product pose reference for the initial screening station and the verification station; the initial screening image acquisition unit 200 and the candidate anomaly extraction unit 300 cooperate to generate a candidate anomaly area queue; the local verification imaging unit 400 performs multi-condition imaging on the candidate anomaly areas according to the coordinate information provided by the image registration and candidate area management unit 500; the layer attribution determination unit 600 extracts difference features based on the multiple registered images and makes a layer attribution conclusion; the result output and execution unit 700 outputs corresponding handling instructions in cleaning re-inspection, film peeling re-inspection, defect rejection, rework transfer, manual verification, or low-speed re-shooting according to the attribution results.

[0157] There are two typical operating conditions during the operation of this system: When the difference characteristics of the candidate abnormal area meet the clear layer attribution judgment conditions, the system enters the normal output condition, and the result output and execution unit 700 directly outputs the processing instructions corresponding to the attribution conclusion; when the registration error of image registration exceeds the preset registration error threshold, each difference characteristic does not reach the corresponding preset threshold, or the judgment conditions of surface defects and interlayer bonding defects are met at the same time, the system enters the pending review output condition, and outputs manual review instructions or low-speed re-shooting instructions through the result output and execution unit 700 to ensure the reliability of the judgment conclusion.

[0158] Example 2: Corresponding to Example 1, this example provides a layer location attribution identification for online appearance defects after hot pressing, including the following steps: The entire area of ​​the hot-pressed product is initially screened and imaged. Image preprocessing and candidate region extraction are performed on the initial screened image to obtain a candidate abnormal region queue. Each candidate abnormal region in the candidate abnormal region queue contains the region center coordinate information. Based on the regional center coordinates of the candidate anomaly region, align the candidate anomaly region with the field of view of the local verification imaging unit; For the same candidate abnormal region, a first verification image is acquired under surface-sensitive lighting conditions, and a second verification image is acquired under interlayer-sensitive lighting conditions; wherein, the surface-sensitive lighting conditions are used to make the optical response of defects at the outermost surface of the product stronger than the optical response of defects inside the product's bonding layer, and the interlayer-sensitive lighting conditions are used to make the optical response of defects inside the product's bonding layer stronger than the optical response of defects at the outermost surface of the product. For the same candidate abnormal area, the focus position is switched sequentially to the first focal plane close to the outermost surface of the product and the second focal plane close to the bonding interface using the focal plane switching component, and the first focal plane image and the second focal plane image are obtained respectively. Image registration aligns the image coordinates of the first verification image, the second verification image, the first focal plane image, and the second focal plane image in the candidate anomaly region. The differential features of candidate anomalous regions are extracted from the registered images. The differential features include at least the surface response contrast features of the candidate anomalous regions in the first verification image, the interlayer response contrast features in the second verification image, the first focal plane sharpness features in the first focal plane image, and the second focal plane sharpness features in the second focal plane image. Based on the differences in characteristics, the layer attribution is determined, and the candidate abnormal areas are classified into surface defects, interlayer bonding defects, or anomalies to be reviewed. The corresponding handling instructions are output according to the classification results. Surface defects include surface contamination defects and protective film defects.

[0159] Furthermore, the differential features also include: the surface edge sharpness of the candidate anomaly region in the first verification image, and the interlayer scattering spread of the candidate anomaly region's edge in the second verification image. Additionally, the presence of a central dark area and a bright-ring-dark-core structure index of the candidate anomaly region in the second verification image, as well as the positional stability features constituted by the positional offset of the candidate anomaly region's center in the first verification image, the second verification image, the first focal plane image, and the second focal plane image.

[0160] Furthermore, the hierarchical attribution determination includes: A candidate anomalous region is classified as a surface defect when it meets at least two of the following conditions: the feature value of the surface response contrast feature is higher than the feature value of the interlayer response contrast feature; the feature value of the first focal plane sharpness feature is higher than the feature value of the second focal plane sharpness feature; the gradient magnitude of the surface edge sharpness feature exceeds a preset edge sharpness threshold; and the scattering halo area ratio of the interlayer scattering spread feature is lower than a preset scattering spread threshold. A candidate abnormal region is determined to be an interlayer bonding defect when it meets at least two of the following conditions: the feature value of the interlayer response contrast feature is not lower than the feature value of the surface response contrast feature; the feature value of the second focal plane sharpness feature is higher than the feature value of the first focal plane sharpness feature; the area ratio of the scattering halo of the interlayer scattering spread feature exceeds the preset scattering spread threshold; the ratio of the average brightness of the annular region to the central region of the bright ring dark core structure index exceeds the preset bright ring dark core threshold, or the positional offset of the positional stability feature is lower than the preset stability threshold. When the criteria for determining surface defects and the criteria for determining interlayer bonding defects are met simultaneously, the feature values ​​of each difference feature do not reach the corresponding preset threshold, or the registration error of image registration exceeds the preset registration error threshold, it is determined to be an abnormality pending review.

[0161] Furthermore, preliminary imaging is performed on the entire area of ​​the hot-pressed product, including: The first preliminary screening image was acquired under coaxial light illumination, and the second preliminary screening image was acquired under low-angle dark field illumination. Candidate regions are extracted from the first and second preliminary screening images respectively, and the candidate regions extracted from the two images are merged to form a candidate anomaly region queue.

[0162] Furthermore, image preprocessing includes: flat field correction, background normalization, noise suppression, and invalid region masking; candidate region extraction includes performing local contrast enhancement on the preprocessed image, calculating the deviation of each pixel from the average brightness of its neighborhood and performing threshold segmentation with a preset deviation threshold, connected region clustering, and candidate region filtering. Candidate region filtering removes regions with an area lower than a preset area threshold, a slenderness ratio exceeding a preset slenderness ratio threshold and an orientation consistent with the transport direction, and regions located at the edge of the masking area.

[0163] Furthermore, aligning the candidate anomaly area with the field of view of the local verification imaging unit includes: controlling the conveying and positioning unit to stop the product at the verification station, and calculating the corresponding position of the candidate anomaly area in the verification station coordinate system based on the regional center coordinates of the candidate anomaly area, the product conveying displacement, and the field of view parameters of the verification camera, to confirm that the candidate anomaly area has entered the field of view of the verification camera.

[0164] Furthermore, acquiring the first verification image under surface-sensitive illumination conditions further includes: switching the polarization switching component to a first polarization state to enhance the surface scattering signal; acquiring the second verification image under interlayer-sensitive illumination conditions further includes: switching the polarization switching component to a second polarization state to suppress surface specular reflection and enhance the interlayer transmission scattering signal.

[0165] Furthermore, this method also includes: using a focal plane switching component to switch the focus position to a third focal plane, and obtaining a third focal plane image. The focus position of the third focal plane is located between the first focal plane and the second focal plane, or at an interlayer position below the corresponding bonding interface of the second focal plane.

[0166] Furthermore, based on the classification results, corresponding processing instructions are output, including: when classified as a surface defect, a cleaning re-inspection or film peeling re-inspection instruction is output; when classified as an interlayer bonding defect, a defect rejection or rework transfer instruction is output; when classified as an abnormality pending review, a manual review or low-speed re-shoot instruction is output.

[0167] Furthermore, when there are multiple candidate abnormal regions in the candidate abnormal region queue, the review priority is set according to at least one of the area, brightness deviation and morphology category of each candidate abnormal region, and multi-condition local review is performed in sequence according to the review priority.

[0168] Furthermore, the stratum attribution determination also includes: inputting the differential features into the classification model, outputting the stratum attribution probability from the classification model, and using the output of the classification model as an auxiliary verification basis for rule determination.

[0169] The above specific embodiments further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for layer location attribution and identification of online appearance defects after hot-press bonding, characterized in that, Includes the following steps: The entire area of ​​the hot-pressed product is initially screened and imaged. Image preprocessing and candidate region extraction are performed on the initial screened image to obtain a candidate abnormal region queue. Each candidate abnormal region in the candidate abnormal region queue contains the region center coordinate information. Based on the regional center coordinates of the candidate anomaly region, align the candidate anomaly region with the field of view of the local verification imaging unit; For the same candidate abnormal region, a first verification image is acquired under surface-sensitive lighting conditions, and a second verification image is acquired under interlayer-sensitive lighting conditions; wherein, the surface-sensitive lighting conditions are used to make the optical response of defects at the outermost surface of the product stronger than the optical response of defects inside the product's bonding layer, and the interlayer-sensitive lighting conditions are used to make the optical response of defects inside the product's bonding layer stronger than the optical response of defects at the outermost surface of the product. For the same candidate abnormal area, the focus position is switched sequentially to the first focal plane close to the outermost surface of the product and the second focal plane close to the bonding interface using the focal plane switching component, and the first focal plane image and the second focal plane image are obtained respectively. Image registration aligns the image coordinates of the first verification image, the second verification image, the first focal plane image, and the second focal plane image in the candidate anomaly region. The differential features of candidate anomalous regions are extracted from the registered images. The differential features include at least the surface response contrast features of the candidate anomalous regions in the first verification image, the interlayer response contrast features in the second verification image, the first focal plane sharpness features in the first focal plane image, and the second focal plane sharpness features in the second focal plane image. Based on the differences in characteristics, the layer attribution is determined, and the candidate abnormal areas are classified into surface defects, interlayer bonding defects, or anomalies to be reviewed. The corresponding handling instructions are output according to the classification results. Surface defects include surface contamination defects and protective film defects.

2. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 1, characterized in that, The differential features also include: the surface edge sharpness features of the candidate anomaly region in the first verification image, and the interlayer scattering spread features of the candidate anomaly region in the second verification image.

3. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 2, characterized in that, The differential features also include: whether the candidate anomaly region has a central dark area and a bright ring-dark core structure in the second verification image, and the positional stability features constituted by the positional offset of the center of the candidate anomaly region in the first verification image, the second verification image, the first focal plane image, and the second focal plane image.

4. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 3, characterized in that, Stratification attribution includes: A candidate anomalous region is classified as a surface defect when it meets at least two of the following conditions: the feature value of the surface response contrast feature is higher than the feature value of the interlayer response contrast feature; the feature value of the first focal plane sharpness feature is higher than the feature value of the second focal plane sharpness feature; the gradient magnitude of the surface edge sharpness feature exceeds a preset edge sharpness threshold; and the scattering halo area ratio of the interlayer scattering spread feature is lower than a preset scattering spread threshold. A candidate abnormal region is determined to be an interlayer bonding defect when it meets at least two of the following conditions: the feature value of the interlayer response contrast feature is not lower than the feature value of the surface response contrast feature; the feature value of the second focal plane sharpness feature is higher than the feature value of the first focal plane sharpness feature; the area ratio of the scattering halo of the interlayer scattering spread feature exceeds the preset scattering spread threshold; the ratio of the average brightness of the annular region to the central region of the bright ring dark core structure index exceeds the preset bright ring dark core threshold, or the positional offset of the positional stability feature is lower than the preset stability threshold. When the criteria for determining surface defects and the criteria for determining interlayer bonding defects are met simultaneously, the feature values ​​of each difference feature do not reach the corresponding preset threshold, or the registration error of image registration exceeds the preset registration error threshold, it is determined to be an abnormality pending review.

5. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 1, characterized in that, Initial screening and imaging of the entire area of ​​the hot-pressed product, including: The first preliminary screening image was acquired under coaxial light illumination, and the second preliminary screening image was acquired under low-angle dark field illumination. Candidate regions are extracted from the first and second preliminary screening images respectively, and the candidate regions extracted from the two images are merged to form a candidate anomaly region queue.

6. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 1, characterized in that, Image preprocessing includes: flat field correction, background normalization, noise suppression, and invalid region masking; candidate region extraction includes performing local contrast enhancement on the preprocessed image, calculating the deviation of each pixel from the average brightness of its neighborhood and performing threshold segmentation with a preset deviation threshold, connected region clustering, and candidate region filtering. Candidate region filtering removes regions with an area lower than a preset area threshold, a slenderness ratio exceeding a preset slenderness ratio threshold and an orientation consistent with the transport direction, and regions located at the edge of the masking area.

7. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 1, characterized in that, Aligning the candidate anomaly area with the field of view of the local verification imaging unit includes: controlling the conveying and positioning unit to stop the product at the verification station, and calculating the corresponding position of the candidate anomaly area in the coordinate system of the verification station based on the regional center coordinates of the candidate anomaly area, the product conveying displacement, and the field of view parameters of the verification camera, and confirming that the candidate anomaly area has entered the field of view of the verification camera.

8. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 1, characterized in that, Acquiring a first verification image under surface-sensitive illumination conditions further includes: switching a polarization switching component to a first polarization state to enhance the surface scattering signal; acquiring a second verification image under interlayer-sensitive illumination conditions further includes: switching a polarization switching component to a second polarization state to suppress surface specular reflection and enhance interlayer transmission scattering signal.

9. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 1, characterized in that, It also includes: using a focal plane switching component to switch the focus position to a third focal plane, and obtaining a third focal plane image. The focus position of the third focal plane is located between the first and second focal planes, or in the interlayer position below the interface corresponding to the second focal plane.

10. The method for layer location attribution identification of online appearance defects after hot pressing according to claim 1, characterized in that, Based on the classification results, corresponding processing instructions are output, including: when classified as a surface defect, a cleaning re-inspection or film peeling re-inspection instruction is output; when classified as an interlayer bonding defect, a defect rejection or rework transfer instruction is output; when classified as an abnormality pending review, a manual review or low-speed re-shoot instruction is output.

11. A layer location attribution and identification system for online appearance defects after hot-press bonding, characterized in that, include: The conveying and positioning unit (100) is used to carry the hot-pressed products, pass the products through the initial screening station and the verification station in sequence, and provide the conveying displacement information of the products. The primary screening image acquisition unit (200) is set at the primary screening station and is used to perform primary screening image acquisition on the entire area of ​​the product to obtain the primary screening image; The candidate anomaly extraction unit (300) is used to perform image preprocessing and candidate region extraction on the initial screening image and output a queue of candidate anomaly regions containing the center coordinate information of the regions. A local verification imaging unit (400), disposed at a verification station, includes: a surface-sensitive illumination component (430) for providing surface illumination conditions that make the optical response of defects on the outermost surface of the product stronger than the optical response of defects inside the bonding layer; a second industrial camera (410) for acquiring a first verification image under surface illumination conditions; an interlayer-sensitive illumination component (440) for providing interlayer illumination conditions that make the optical response of defects inside the bonding layer of the product stronger than the optical response of defects on the outermost surface; a second industrial camera (410) for acquiring a second verification image under interlayer illumination conditions; and a focal plane switching component (460) for switching the focus position of the second industrial camera (410) to a first focal plane close to the outermost surface of the product and a second focal plane close to the bonding interface. The image registration and candidate region management unit (500) is used to align the candidate anomaly region with the field of view of the local verification imaging unit (400) according to the regional center coordinate information of the candidate anomaly region, and to align the image coordinates of the first verification image, the second verification image, the first focal plane image and the second focal plane image in the candidate anomaly region through image registration; The layer attribution determination unit (600) is used to extract the differential features of candidate abnormal regions from the registered image and perform layer attribution determination; the differential features include surface response contrast features, interlayer response contrast features, first focal plane sharpness features, and second focal plane sharpness features; the layer attribution determination is used to classify candidate abnormal regions into surface defects, interlayer bonding defects, or anomalies to be verified; surface defects include: surface contamination defects and protective film defects; The result output and execution unit (700) is used to output the abnormal level attributes and corresponding handling instructions according to the classification results; The control unit (800) is used to uniformly control the coordinated operation of all units.

12. The layer location attribution identification system for online appearance defects after hot pressing according to claim 11, characterized in that, The local verification imaging unit (400) also includes a polarization switching component (450); the polarization switching component (450) includes an incident polarizer (451), an analytical polarizer (452), a polarizer switching frame (453), and a polarizer rotation drive mechanism (454); the polarization switching component (450) is used to switch to a first polarization state that enhances the surface scattering signal when acquiring a first verification image, and to switch to a second polarization state that suppresses surface specular reflection and enhances interlayer transmission scattering signal when acquiring a second verification image.

13. The layer location attribution identification system for online appearance defects after hot pressing according to claim 11, characterized in that, The initial screening image acquisition unit (200) includes a first industrial camera (210), a first lens (220), a first illumination component (230) for providing coaxial light illumination, and a second illumination component (240) for providing low-angle dark field illumination; the first illumination component (230) and the second illumination component (240) are used to acquire the first initial screening image and the second initial screening image, respectively; the candidate anomaly extraction unit (300) is used to perform candidate region extraction on the first initial screening image and the second initial screening image, respectively, and to fuse the candidate regions extracted from the two images to form a candidate anomaly region queue.

14. The layer location attribution identification system for online appearance defects after hot pressing according to claim 11, characterized in that, The surface-sensitive illumination component (430) is at least one of a low-angle ring dark field light component, a side strip oblique illumination component, a high-directional surface scattering enhancement light component, or a surface reflection enhancement polarization illumination component; the interlayer-sensitive illumination component (440) is at least one of a collimated backlight, a semi-transmissive illumination component, a polarized transmission illumination component, or a reflective illumination component that suppresses surface specular reflection.

15. The layer location attribution identification system for online appearance defects after hot pressing according to claim 11, characterized in that, The focal plane switching assembly (460) is at least one of an electric focusing lens, a camera Z-axis micro-displacement platform, or a lens Z-axis micro-displacement platform; the focal plane switching assembly (460) is also equipped with a third focal plane, the focusing position of which is located between the first focal plane and the second focal plane, or at an interlayer position below the corresponding bonding interface of the second focal plane.

16. The layer location attribution identification system for online appearance defects after hot pressing according to claim 11, characterized in that, The conveying and positioning unit (100) includes a conveying mechanism (110), two side guide rails (120), an adsorption-type bearing platform (130), an entrance detection photoelectric sensor (140), a conveying encoder (150), and a verification station stop mechanism (160). The verification station stop mechanism (160) is used to stop the product at the verification station during the local verification imaging stage, and the conveying encoder (150) is used to provide conveying displacement information.

Images