An image recognition-based composite substrate defect detection method, device and medium

Abstract

The application discloses a composite substrate defect detection method and device based on image recognition and a medium, relates to the technical field of image recognition, and comprises the following steps: feeding a defect candidate image segment into an image recognition model, separating out a layer mark residual image and a gray mutation edge image, removing the layer mark residual image, closing the gray mutation edge image into a defect boundary image, extracting a transmission abnormal image from a composite substrate image under the same detection coordinate, and recombining the defect boundary image and the transmission abnormal image to generate a candidate defect parameter group; performing physical forward simulation on candidate defect parameters in the candidate defect parameter group respectively, converting multiple candidate defect parameters into multiple simulated defect images, and performing competitive verification on the multiple simulated defect images through the defect candidate image segment to generate a parameter verification sequence. Through the physical forward simulation and the competitive verification of the simulated defect image, the accuracy, stability and interpretability of target defect parameter determination are improved.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a method, device and medium for detecting defects in composite substrates based on image recognition. Background Technology

[0002] As the application of composite substrates in electronic packaging, functional films, laminates, and industrial protective materials continues to expand, the interlayer state, surface texture continuity, and local defect morphology of composite substrates have a significant impact on product reliability. Image recognition technology is maturing in industrial visual inspection, enabling automated identification of material surfaces and interlayer images through image acquisition, texture analysis, boundary extraction, feature encoding, and image comparison. For composite substrate inspection scenarios, existing image recognition methods typically combine high-resolution imaging, grayscale feature extraction, edge recognition, and region segmentation to locate abnormal textures, local grayscale abrupt changes, and suspected defect areas in material images, providing a technological foundation for online inspection, quality grading, and defect traceability in composite substrate production processes.

[0003] Existing composite substrate defect detection technologies typically focus on directly identifying grayscale anomalies when dealing with complex layer texture backgrounds. They lack the ability to construct a co-positional reference relationship between normal layer texture images, interlayer boundary images, and defect candidate regions. This can easily lead to the mixed identification of layer texture residues, texture undulations, and real defect edges. Especially when composite substrates have multi-layer structures, texture direction variations, and differences in transmission brightness, relying solely on image recognition results makes it difficult to accurately distinguish defect boundaries from the layer texture background, resulting in insufficient reliability in matching candidate defect layers and candidate defect boundaries. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a composite substrate defect detection method based on image recognition to solve the problem of insufficient matching reliability between candidate defect layers and candidate defect boundaries.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for detecting defects in composite substrates based on image recognition, comprising, Acquire composite substrate images and record the detection coordinates of the composite substrate images. Perform texture benchmarking on the surface texture and interlayer boundary images in the composite substrate images to form a layer texture reference. Cut the image areas that deviate from the layer texture reference into defect candidate image fragments. The candidate image fragments of defects are fed into the image recognition model to separate the residual image of layer texture and the gray-scale abrupt edge image. The residual image of layer texture is removed and the gray-scale abrupt edge image is closed into the defect boundary image. The transmission abnormal image is extracted from the composite substrate image under the same detection coordinate and the defect boundary image and the transmission abnormal image are recombined to generate the candidate defect parameter group. Physical forward simulation is performed on the candidate defect parameters in the candidate defect parameter group, and multiple candidate defect parameters are converted into multiple simulated defect images. The multiple simulated defect images are competitively verified by defect candidate image fragments to generate a parameter verification sequence. The candidate defect layer and candidate defect boundary of the candidate defect parameter group are corrected according to the parameter verification sequence, the target defect parameter is determined, and the target defect parameter is mapped to the detection coordinates in the composite substrate image to generate the defect detection result.

[0007] In a preferred embodiment of the image recognition-based composite substrate defect detection method of the present invention, the layer texture is referenced, specifically, The composite substrate image is divided into multiple composite substrate image segments according to the detection coordinate order, and arranged sequentially to generate a continuous image band. The interlayer boundary images that span multiple composite substrate image segments are traced within a continuous image band. The discontinuous interlayer boundary images are connected into an interlayer boundary trajectory line, and the surface texture images on both sides of the interlayer boundary trajectory line are preserved to generate an interlayer boundary trajectory map. Establish a texture coordinate zone with the interlayer boundary trajectory line in the interlayer boundary trajectory map as the center line, and arrange the surface texture images on both sides of the interlayer boundary trajectory line along the extension direction of the interlayer boundary trajectory line to the texture coordinate zone to generate a texture unfolding map. Full-resolution texture flattening is performed on the surface texture image in the layer texture unfolding image to reduce the fine gray-level fluctuations in the surface texture image, while preserving the width variation and brightness transition on both sides of the interlayer boundary trajectory line, thus generating a layer texture flattening image. The lamination flattening map is mapped to the detection coordinates of the composite substrate image and superimposed with the interlayer boundary trajectory map to form a lamination reference map.

[0008] In a preferred embodiment of the image recognition-based composite substrate defect detection method of the present invention, the step of cropping the image region deviating from the layer texture reference into defect candidate image segments specifically involves: The composite substrate image and the layer texture reference image are aligned segment by segment under the detection coordinates to form a reference image pair; The reference image pair is split and converted by a dual-path feature encoding method. The composite substrate image in the reference image pair is converted into a measured feature field, and the layer texture reference in the reference image pair is converted into a reference feature field, thus generating a dual-path feature field. Within the dual-path feature field, the reference feature field and the measured feature field are assigned to the same feature coordinates. The deviation positions of the layer texture in the measured feature field are repositioned to the reference position, and the unrepositioned deviation positions of the layer texture are connected to form a deviation feature map. Based on the non-positioned layer texture deviation position within the deviation feature map, the co-position image range is delineated in the composite substrate image, and adjacent co-position image ranges are merged into the deviation image region. Apply layer texture to the deviated image region and refer to the clipping constraint of the interlayer boundary trajectory line in the image. Deviated image regions that cross the interlayer boundary trajectory line are divided into interlayer clipping windows, and deviated image regions that do not cross the interlayer boundary trajectory line are divided into surface clipping windows, generating a candidate clipping window sequence. The composite substrate image is cropped using interlayer cropping windows and surface cropping windows in the candidate cropping window sequence. The original image grayscale and detection coordinates within the cropped area are preserved and identified as candidate image fragments for defects.

[0009] As a preferred embodiment of the image recognition-based composite substrate defect detection method of the present invention, wherein: the generation of candidate defect parameter sets specifically includes, Based on the detection coordinates carried by the defect candidate image fragments, the corresponding layer texture reference fragments are extracted from the layer texture reference image, and the defect candidate image fragments and the corresponding layer texture reference fragments are arranged into a candidate reference fragment set. Image recognition models are used to perform isotopic texture resolution on the candidate reference fragment set, separating the residual texture image that extends with the isotopic texture reference fragment and the gray-scale abrupt edge image that deviates from the extension direction of the isotopic texture reference fragment in the defect candidate image fragment. The pixel occupancy of the residual layer image in the defect candidate image segment is expanded into a residual coverage area. The residual coverage area is used as a counter-evidence masking area. The counter-evidence masking area is masked in the defect candidate image segment while retaining the gray-scale abrupt change edge image to obtain the edge candidate segment. Within the edge candidate segments, the broken endpoints of the gray-scale abrupt edge image are connected and arranged to form a closed contour by connecting the broken endpoints belonging to the same deviated image region, thus forming a defect boundary image. Based on the detection coordinates of the defect boundary image, the in-situ transmission image region is obtained from the composite substrate image. The transmission background image that changes with the in-situ layer texture reference segment in the in-situ transmission image region is removed to form a transmission abnormality image. The defect boundary image is organized into candidate defect boundaries. The position of the brightness and darkness distribution of the transmission anomaly image is aligned with the interlayer boundary image in the co-position layer reference segment to obtain the candidate defect layer. The candidate defect boundary and candidate defect layer are then recombined to generate a candidate defect parameter group.

[0010] In a preferred embodiment of the image recognition-based composite substrate defect detection method of the present invention, the step of converting multiple candidate defect parameters into multiple simulated defect images specifically involves: The candidate defect boundaries and candidate defect layers in the candidate defect parameter group are paired one by one to form a candidate defect parameter sequence; The candidate defect boundary in the candidate defect parameter sequence is projected onto the substrate layer corresponding to the candidate defect layer, and the image range enclosed by the candidate defect boundary is set as the defect occupancy area to form a layer perturbation sheet. Within the layer perturbation sheet, the edge region of the defect-occupying area is assigned to the reflection imaging path, and the inner region of the defect-occupying area is assigned to the transmission imaging path, thus forming an optical perturbation sheet. By physically forward simulation, the reflection imaging path and the transmission imaging path in the optical perturbation sheet are respectively derived, and the reflection imaging response simulation of the edge region and the transmission imaging response simulation of the inner region are completed, thus forming the reflection simulation image and the transmission simulation image. The simulated reflection image and the simulated transmission image are mapped to the detection coordinates of the composite substrate image and superimposed with the normal layer texture image in the layer texture reference to form a simulated defect image; After the candidate defect parameters in the candidate defect parameter sequence are physically simulated, multiple simulated defect images are formed.

[0011] As a preferred embodiment of the composite substrate defect detection method based on image recognition described in this invention, the reflection imaging path is an imaging branch formed in the optical disturbance sheet after the edge region of the defect-occupying area enters the surface reflection direction, carrying the reflection grayscale change of the candidate defect boundary in the composite substrate image. The transmission imaging path is an imaging branch formed in the optical perturbation sheet after the internal region of the defect-occupying area enters the interlayer transmission direction, carrying the transmission brightness and darkness changes of the candidate defect layer in the composite substrate image.

[0012] As a preferred embodiment of the image recognition-based composite substrate defect detection method of the present invention, wherein: the generation of the parameter verification sequence specifically comprises, Based on the detection coordinates of the defect candidate image fragments, multiple simulated defect images are superimposed onto the same coordinate positions of the defect candidate image fragments to form a co-position verification piece; In the co-position verification film, the simulated defect outer edge of the simulated defect image is co-positioned with the gray-scale abrupt change edge image in the defect candidate image segment, and the bonding position, offset position and disconnection position are preserved and organized into a boundary verification record. The grayscale change direction between the candidate defect image fragments and the simulated defect image is arranged along the boundary verification record, and the position of the overlap between the transmission simulated image and the transmission abnormal image is marked to form a transmission verification record. The candidate defect parameters are constrained by the boundary verification records and the transmission verification records. Candidate defect parameters with continuous boundary fit and complete overlapping positions are placed at the front end, and candidate defect parameters with offset and broken positions are placed at the back end, thus generating a parameter verification sequence.

[0013] As a preferred embodiment of the image recognition-based composite substrate defect detection method of the present invention, wherein: the determination of the target defect parameter is achieved by correcting the candidate defect layer and candidate defect boundary in the candidate defect parameter group by the front-end candidate defect parameter in the parameter verification sequence, and synthesizing the corrected candidate defect layer and candidate defect boundary into the target defect parameter.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the image recognition-based composite substrate defect detection method described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the image recognition-based composite substrate defect detection method as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: By using physical forward simulation and competitive verification with simulated defect images, the imaging consistency screening of candidate defect parameters is achieved. After the candidate defect boundary and candidate defect layer enter the reflection imaging path and transmission imaging path respectively, they can form reflection simulation images and transmission simulation images, which are then verified with the defect candidate image fragments under the same detection coordinates. This ensures that the candidate defects obtained by image recognition no longer rely solely on a single gray-level anomaly. The boundary verification record can reflect the fitting state between the outer edge of the simulated defect and the gray-level abrupt change edge image, and the transmission verification record can reflect the overlap state between the transmission simulation image and the transmission anomaly image. The parameter verification sequence further places candidate defect parameters with continuous boundary fitting and complete overlap positions at the front end, thereby improving the accuracy, stability, and interpretability of the target defect parameter determination. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a composite substrate defect detection method based on image recognition.

[0019] Figure 2 The flowchart for generating layer texture references and defect candidate segments.

[0020] Figure 3 A flowchart for generating simulated defect images.

[0021] Figure 4 A flowchart for parameter verification and defect result generation.

[0022] Figure 5 This is a comparison chart of consistency rates for repeated detections.

[0023] Figure 6 This is a thermogram of the transmission overlap rate perturbation. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0027] Reference Figures 1-6 This is one embodiment of the present invention, which provides a method for detecting defects in composite substrates based on image recognition, comprising the following steps: S1. Acquire images of the composite substrate and record the detection coordinates of the composite substrate images. Perform texture benchmarking on the surface texture and interlayer boundary images in the composite substrate images to form a layer texture reference. Cut the image areas that deviate from the layer texture reference into defect candidate image fragments.

[0028] S1.1 After the composite substrate enters the image acquisition area, it is continuously transported along the longitudinal acquisition direction, and the composite substrate passes sequentially through the imaging field of view of the line scan industrial camera. A surface reflection light source is set on the imaging side of the line scan industrial camera, and a transmission backlight source is set on the back side of the composite substrate. The surface reflection light source and the transmission backlight source are illuminated in a time-division manner according to the same acquisition trigger cycle. When the surface reflection light source is illuminated, the line scan industrial camera reads the reflected grayscale of the composite substrate surface, forming a reflected image line, which carries the surface texture image of the composite substrate surface. When the transmission backlight source is illuminated, the transmission backlight source passes through the interlayer region of the composite substrate, and the line scan industrial camera reads the transmission grayscale at the interlayer boundary, forming a transmission image line, which carries the interlayer boundary image of the composite substrate. According to the acquisition trigger sequence of the line scan industrial camera, the reflected image lines and transmission image lines formed in the same acquisition trigger cycle are associated as a horizontal image line group, and the continuous... The continuously formed horizontal image line groups are sequentially stitched together along the vertical acquisition direction to obtain a composite substrate image with surface texture and interlayer boundary images. Simultaneously with the generation of the composite substrate image, the acquisition batch of the horizontal image line groups is determined as the image segment number, the pixel column position of the line scan industrial camera is determined as the horizontal pixel position, and the stitching order of the horizontal image line groups is determined as the vertical acquisition position. The image segment number, horizontal pixel position, and vertical acquisition position are written into the detection coordinates of the composite substrate image, ensuring that the surface texture, interlayer boundary images, and subsequent cropping areas in the composite substrate image can all be located according to the same detection coordinates. The detection coordinates of the composite substrate image are used to mark the coordinate information of the image segment number, horizontal pixel position, and vertical acquisition position in the composite substrate image. The photosensitive component of the line scan industrial camera is a single-row or multi-row linear pixel array, used to read the image grayscale at the horizontal pixel position row by row as the composite substrate moves continuously.

[0029] S1.2. The composite substrate image is divided into multiple composite substrate image segments according to the detection coordinate order, and these segments are arranged sequentially to generate a continuous image band. Specifically, Using the detection coordinates of the composite substrate image as the segmentation basis, the image segment number, horizontal pixel position, and vertical acquisition position of the composite substrate image are read. Image areas with continuous vertical acquisition positions and complete horizontal pixel positions are classified into the same composite substrate image segment. The position where the vertical acquisition position jumps and the image segment number changes is used as the segmentation boundary of adjacent composite substrate image segments. The surface texture, interlayer boundary image, and detection coordinates within each composite substrate image segment are retained from the composite substrate image. When the vertical acquisition positions of adjacent composite substrate image segments can be continuously connected, the composite substrate image segments are arranged from front to back according to the image segment number. When the vertical acquisition positions of adjacent composite substrate image segments are missing, repeated, or the horizontal pixel positions are incomplete, the corresponding composite substrate image segment retains the coordinate anomaly mark and continues to participate in the arrangement. The arranged composite substrate image segments are used as continuous connecting objects, and the boundaries of adjacent composite substrate image segments are stitched together according to the detection coordinates to keep the surface texture and interlayer boundary image continuous in the detection coordinates, generating a continuous image band.

[0030] S1.3. Within a continuous image band, trace the interlayer boundary images spanning multiple composite substrate image segments, connect the discontinuous interlayer boundary images into an interlayer boundary trajectory line, and retain the surface texture images on both sides of the interlayer boundary trajectory line to generate an interlayer boundary trajectory map. Specifically, Interlayer boundary images within each composite substrate image segment are read from the continuous image band according to the image segment numbering order, and the corresponding lateral pixel positions and longitudinal acquisition positions of the interlayer boundary images are recorded in the detection coordinates. Interlayer boundary images that connect the light and dark transition directions at the edges of adjacent composite substrate image segments and whose lateral pixel positions can be connected are connected to form the same interlayer boundary trajectory line. When the interlayer boundary images are discontinuous between adjacent composite substrate image segments, the connection relationship is completed along the lateral pixel positions, longitudinal acquisition positions, and light and dark transition directions before and after the discontinuity. Interlayer boundary images that cannot form a continuous connection relationship retain the boundary discontinuity mark. The surface texture images on both sides of the interlayer boundary trajectory line located in the detection coordinates are simultaneously retained, so that the surface texture images on both sides of the interlayer boundary trajectory line can continue to support the establishment of subsequent layer texture coordinate bands. Finally, the interlayer boundary trajectory line, the surface texture images on both sides of the interlayer boundary trajectory line, and the boundary discontinuity mark are organized into an interlayer boundary trajectory map.

[0031] S1.4. Establish a texture coordinate zone using the interlayer boundary trajectory line in the interlayer boundary trajectory map as the center line. Arrange the surface texture images on both sides of the interlayer boundary trajectory line along the extension direction of the interlayer boundary trajectory line to the texture coordinate zone to generate a texture unfolded map. Specifically, Using the interlayer boundary trajectory line in the interlayer boundary trajectory map as the center line of the texture coordinate zone, the texture coordinate zone is a coordinate arrangement area centered on the interlayer boundary trajectory line, used to receive and expand the surface texture images on both sides of the interlayer boundary trajectory line. The length direction of the texture coordinate zone is laid out along the extension direction of the interlayer boundary trajectory line, and the width direction of the texture coordinate zone is laid out along the expansion direction perpendicular to the interlayer boundary trajectory line. According to the horizontal pixel position and vertical acquisition position in the detection coordinate, the surface texture images on both sides of the interlayer boundary trajectory line are sequentially arranged into the texture coordinate zone, and then... The distance between the surface texture image and the interlayer boundary trajectory line is used to arrange the texture images in order of proximity to the center line and then to the outer texture position. For the interlayer boundary trajectory line position with boundary discontinuity markers, the extension direction before and after the discontinuity is used as the basis for continuation to complete the arrangement order of the surface texture images at the position with boundary discontinuity markers, while maintaining the original grayscale of the surface texture images. After the interlayer boundary trajectory line and the surface texture images on both sides are arranged, the arrangement relationship between the interlayer boundary trajectory line, the surface texture images on both sides and the detection coordinates in the layer texture coordinate zone is sorted out to form a layer texture unfolding diagram.

[0032] S1.5. Perform full-resolution texture flattening on the surface texture image in the layer texture unfolding image, reducing minor gray-level fluctuations in the surface texture image, and preserving the width variations and brightness transitions on both sides of the interlayer boundary trajectory lines to generate a layer texture flattening image. Specifically, The original grayscale of the surface texture image is read along the length of the texture coordinate band from the unfolded layer map. Without changing the pixel size of the unfolded layer map, full-resolution texture flattening is performed on the fine grayscale fluctuations near the center line and on the outer texture positions. Full-resolution texture flattening uses existing image smoothing methods to weaken short-distance grayscale jumps between adjacent texture points at the original pixel positions, making the fine textures in the surface texture image into a continuous grayscale distribution. At the same time, it preserves the contour differences formed by width changes and the grayscale direction differences formed by light and dark transitions on both sides of the interlayer boundary trajectory line. If the grayscale fluctuations in the surface texture image are only distributed between local texture points, the local texture points are flattened into the adjacent continuous grayscale distribution. If the grayscale changes continue to extend along both sides of the interlayer boundary trajectory line, the position and direction of the grayscale changes in the texture coordinate band are preserved. After the full-resolution texture flattening is completed, the surface texture image in the unfolded layer map is organized into a flattened layer map image. The width changes and light and dark transitions on both sides of the interlayer boundary trajectory line are still preserved at the corresponding positions of the original detection coordinates, forming a flattened layer map.

[0033] In this process, the smoothing direction for full-resolution texture flattening is the length direction of the layer texture coordinate band. Texture points located at the same width and continuously arranged along the length direction of the layer texture coordinate band are used as the smoothing objects. A smoothing window is formed by the current texture point, the texture points adjacent to the current texture point in front of it, and the texture points adjacent to the current texture point behind it. The original gray level of each texture point within the smoothing window is read and the gray level average is calculated. When the original gray level of the current texture point suddenly increases or decreases relative to the gray level average, and the sudden increase or decrease is only distributed within the range of the current texture point, the texture points adjacent to the current texture point in front of it, and the texture points adjacent to the current texture point behind it, the original gray level of the current texture point is flattened to the gray level average. When the smoothing window crosses the interlayer boundary trajectory line, the texture points on both sides of the interlayer boundary trajectory line are processed separately and are not included in the same smoothing window.

[0034] S1.6. Map the layer texture flattening map to the detection coordinates of the composite substrate image, and overlay it with the interlayer boundary trajectory map to form a layer texture reference map. Specifically, Along the detection coordinates preserved for each pixel in the lamination flattening image, the same detection coordinate position in the composite substrate image is searched point by point, and the lamination flattening image in the lamination flattening image is re-implanted into the detection coordinate grid of the composite substrate image, so that the horizontal pixel position and vertical acquisition position of the lamination flattening image in the composite substrate image are restored to the corresponding position before flattening processing; for positions in the composite substrate image with coordinate anomaly markers, the coordinate anomaly markers are retained, and the continuous positions of the lamination flattening image before and after the coordinate anomaly markers are recorded continuously to avoid jumps and confusion in the detection coordinates of the lamination flattening image.

[0035] After completing the position playback, the interlayer boundary trajectory lines and the surface texture images on both sides of the interlayer boundary trajectory map are overlaid onto the already played-out layer texture flattening image according to the same detection coordinates. The interlayer boundary trajectory lines maintain the boundary position indication function, the surface texture images on both sides maintain the layer texture distribution indication function, and the layer texture flattening image maintains the flattened grayscale distribution, so that the interlayer boundary position, layer texture distribution position, and flattened grayscale shape are preserved simultaneously under the same detection coordinates. After completing the detection coordinate alignment and overlay, a layer texture reference image is obtained.

[0036] S1.7. Align the composite substrate image and the layer texture reference image segment by segment in the detection coordinates to form a reference image pair; perform a dual-path feature encoding method to perform a split-path transformation on the reference image pair, converting the composite substrate image in the reference image pair into a measured feature field, and converting the layer texture reference image in the reference image pair into a reference feature field, generating a dual-path feature field. Specifically, The detection coordinates preserved in the composite substrate image are compared segment by segment with the detection coordinates restored in the layer texture reference image. According to the image segment number, horizontal pixel position and vertical acquisition position, the original image grayscale in the composite substrate image and the flattened grayscale in the layer texture reference image are arranged to the same detection coordinate position. When the original image grayscale and the flattened grayscale exist at the same detection coordinate position, the corresponding segment of the composite substrate image and the corresponding segment of the layer texture reference image are matched to form a reference image pair. When there are coordinate anomaly markers, missing original image grayscale, or missing flattened grayscale at the same detection coordinate position, the corresponding detection coordinate position is marked as an alignment anomaly marker, and the segment-by-segment arrangement of the reference image pair is continued.

[0037] After segment-by-segment alignment is completed, the composite substrate image and the layer texture reference image in the reference image pair are respectively sent to the dual-path feature encoding method for split conversion. The dual-path feature encoding method extracts the original image gray level, gray level change direction and texture continuity position in the corresponding path of the composite substrate image and converts it into a measured feature field. In the corresponding path of the layer texture reference image, it extracts the flattened gray level, interlayer boundary trajectory line position and layer texture continuity position and converts it into a reference feature field. The measured feature field and the reference feature field are arranged along the same detection coordinates and together form the dual-path feature field.

[0038] Among them, the dual-path feature encoding method is a conventional and feasible feature processing method in the field of image recognition. The main operation is to extract features, encode features, and arrange them in the same coordinates for the composite substrate image and the layer texture reference image, respectively, to form a measured feature field and a reference feature field.

[0039] S1.8. Within the dual-path feature field, the reference feature field and the measured feature field are aligned to the same feature coordinates. The deviation positions of the layer patterns in the measured feature field are repositioned to a reference position. The unrepositioned deviation positions of the layer patterns are then connected to form a deviation feature map. Specifically, Within the dual-path feature field, the detection coordinates are used as a common arrangement basis to categorize the reference feature field and the measured feature field into the same feature coordinates. The flattened grayscale, interlayer boundary trajectory line position, and texture continuity position in the reference feature field are used as the categorization benchmark to compare the original image grayscale, grayscale change direction, and texture continuity position in the measured feature field. The texture deviation position in the measured feature field is compared with the texture continuity position near the same feature coordinate, and the original image grayscale, grayscale change direction, flattened grayscale of the nearby texture continuity position, and extension direction of the nearby texture continuity position are read. The neighboring pixel range is the range formed by extending two pixels to each side of the nearby texture continuity position. When the texture deviation position is within the neighboring pixel range... Within the specified range, if the grayscale difference between the original image grayscale of the deviated layer texture position and the flattened grayscale of the adjacent continuous layer texture position does not exceed twice the average grayscale difference of adjacent layer texture positions in the reference feature field, and the direction of the grayscale change at the deviated layer texture position is continuously distributed along the extension direction of the adjacent continuous layer texture position for no less than five pixels, the deviated layer texture position is included in the repositioning benchmark. If any of the following occurs: the deviated layer texture position does not fall within the range of neighboring pixels, the grayscale difference exceeds twice the average grayscale difference of adjacent layer texture positions in the reference feature field, and the direction of the grayscale change does not continuously distribute along the extension direction of the adjacent continuous layer texture position for no less than five pixels, the deviated layer texture position is retained as an unrepositioned deviated layer texture position. Connect adjacent unrepositioned deviated layer texture positions along the detection coordinates to form a deviation feature map.

[0040] The setting of at least five pixels is to exclude single-point noise and short-distance grayscale jitter, and only include the deviation of the layer texture with continuous extension shape into the positioning reference.

[0041] S1.9. Based on the misaligned layer texture positions within the deviation feature map, define the corresponding image range in the composite substrate image, and merge adjacent corresponding image ranges into a deviation image region. Specifically, The misaligned layers within the feature map are repositioned into the composite substrate image according to the detection coordinates. First, the image position with the same horizontal pixel position and vertical acquisition position as the misaligned layer deviation position is located in the composite substrate image. Then, the image position is expanded outward along the grayscale changes of the original image around the image position to obtain the corresponding image range covering the layer deviation position. When the coordinate boundaries of adjacent corresponding image ranges are connected and the grayscale changes of the original image are continuous, the adjacent corresponding image ranges are merged into the same deviation image region. When there is a coordinate interval between adjacent corresponding image ranges and the grayscale changes of the original image are discontinuous, the adjacent corresponding image ranges are treated as independent deviation image regions, thus completing the delineation of the deviation image region in the composite substrate image.

[0042] S1.10. Apply layer ridges to the deviated image regions. Referring to the clipping constraints of the interlayer boundary trajectory lines in the image, deviated image regions that cross the interlayer boundary trajectory lines are designated as interlayer clipping windows, and deviated image regions that do not cross the interlayer boundary trajectory lines are designated as surface clipping windows, generating a candidate clipping window sequence. Specifically... The detection coordinate range of the deviated image region is compared with the position of the interlayer boundary trajectory line in the layer texture reference. When the interlayer boundary trajectory line passes through the deviated image region and divides the deviated image region into image parts located on both sides of the interlayer boundary trajectory line, the deviated image region is classified as an interlayer clipping window. When the interlayer boundary trajectory line does not pass through the deviated image region and the entire deviated image region is located on one side of the interlayer boundary trajectory line, the deviated image region is classified as a surface clipping window. When the interlayer boundary trajectory line is close to the edge of the deviated image region but does not separate the deviated image region, it is classified according to the extension direction of the gray-level change of the original image within the deviated image region. The original image gray-level change extending along both sides of the interlayer boundary trajectory line is classified as an interlayer clipping window, and the original image gray-level change extending only on one side of the interlayer boundary trajectory line is classified as a surface clipping window. The interlayer clipping windows and surface clipping windows are arranged in the order of detection coordinates to generate a candidate clipping window sequence.

[0043] S1.11. The composite substrate image is cropped using the interlayer cropping window and surface cropping window in the candidate cropping window sequence. The original image grayscale and detection coordinates within the cropped area are preserved and identified as candidate image segments for defects. Specifically, Along the detection coordinate range in the candidate clipping window sequence, image regions corresponding one-to-one with interlayer clipping windows and surface clipping windows are extracted from the composite substrate image. Each extracted image region retains the original image grayscale and detection coordinates in the composite substrate image. The image region extracted by the interlayer clipping window retains the image portions on both sides of the interlayer boundary trajectory line, while the image region extracted by the surface clipping window retains the image portion on one side of the interlayer boundary trajectory line. The image regions are arranged according to the order of the candidate clipping window sequence and determined as candidate defect image segments.

[0044] S2. Input the candidate defect image fragments into the image recognition model, separate the residual layer image and the gray-scale abrupt edge image, remove the residual layer image, close the gray-scale abrupt edge image into the defect boundary image, extract the transmission abnormal image from the composite substrate image under the same detection coordinates, and recombine the defect boundary image and the transmission abnormal image to generate a candidate defect parameter group.

[0045] S2.1. Based on the detection coordinates carried by the candidate defect image fragments, extract the corresponding layer texture reference fragments from the layer texture reference image. Arrange the candidate defect image fragments and the corresponding layer texture reference fragments into a candidate reference fragment set. Specifically, Using the detection coordinates carried by the candidate defect image fragment as an index, a reference area with the same image segment number, horizontal pixel position, and vertical acquisition position is found in the layer texture reference image. The layer texture reference image area with the same coordinate range as the candidate defect image fragment is extracted to form a co-position layer texture reference fragment. The co-position layer texture reference fragment retains the flattened grayscale, interlayer boundary trajectory line position, and layer texture continuity position in the layer texture reference image, and maintains the co-position arrangement with the original image grayscale and detection coordinates in the candidate defect image fragment. When the detection coordinates are complete and consistent, the candidate defect image fragment and the co-position layer texture reference fragment are arranged in pairs according to the same detection coordinates. When there are missing detection coordinates or alignment anomaly marks, the missing positions and alignment anomaly marks are retained and the pair arrangement is continued to form a candidate reference fragment set.

[0046] S2.2. Using an image recognition model, isotopic texture resolution is performed on the candidate reference segment set to separate the residual texture images extending along the isotopic texture reference segments and the gray-level abrupt edge images deviating from the extension direction of the isotopic texture reference segments in the defect candidate image segments. Specifically, Each candidate defect image fragment and its corresponding layer texture reference fragment in the candidate reference fragment set are input into the image recognition model at the same detection coordinates. The image recognition model performs parallel comparison between the original image grayscale of the candidate defect image fragment and the flattened grayscale of the corresponding layer texture reference fragment at the corresponding detection coordinate position. At the same time, it combines the position of the interlayer boundary trajectory line and the continuous position of the layer texture to determine the texture extension relationship. Among them, the image part with grayscale changes continuously distributed along the extension direction of the corresponding layer texture reference fragment and consistent with the continuous position of the layer texture is classified as the residual layer texture image. The image part with grayscale changes deviating from the extension direction of the corresponding layer texture reference fragment and forming edge transitions and grayscale abrupt changes at local positions is classified as the grayscale abrupt change edge image. The original markings are retained at the detection coordinate missing position and the alignment anomaly mark position. Only the corresponding position with complete and consistent detection coordinates is subjected to corresponding texture resolution. Finally, the residual layer texture image and the grayscale abrupt change edge image are separated within the candidate defect image fragment.

[0047] The training process of the image recognition model is as follows: Historical defect candidate image fragments and historical isotopic layer texture reference fragments were used as training samples. The training samples came from historical composite substrate images, historical layer texture references, manual review records, and defect detection history records. The training sample annotations included the layer texture residual image annotation area, the gray-scale abrupt change edge image annotation area, the layer texture extension direction annotation, and the edge turning position annotation. The layer texture residual image annotation area came from the delineation of normal layer texture residual positions in the manual review records. The gray-scale abrupt change edge image annotation area came from the delineation of defect edge positions in the manual review records. The layer texture extension direction annotation came from the interlayer boundary trajectory line position and layer texture continuity position in the historical isotopic layer texture reference fragment. The edge turning position annotation came from the review results of gray-scale abrupt change positions in the defect detection history records. This enabled the image recognition model to distinguish the layer texture residual image that extends with the isotopic layer texture reference fragment and the gray-scale abrupt change edge image that deviates from the extension direction of the isotopic layer texture reference fragment after training.

[0048] The image recognition model adopts the existing convolutional neural network semantic segmentation structure. The input layer receives defect candidate image segments and co-located layer texture reference segments arranged according to the same detection coordinates. The convolutional layer extracts the original image grayscale, the flattened grayscale, the continuous position of the layer texture, and the direction of grayscale abrupt change. The decoding layer restores the pixel positions in the detection coordinates. The segmentation output layer outputs a probability map of residual layer texture images and a probability map of grayscale abrupt change edge images. During training, the probability map of residual layer texture images output by the segmentation output layer is compared with the labeled areas of residual layer texture images, and the probability map of grayscale abrupt change edge images output by the segmentation output layer is compared with the labeled areas of grayscale abrupt change edge images. The cross-entropy loss function is used to form the pixel classification error, and the Dice loss function is used to constrain the overlap deviation of the labeled areas of residual layer texture images and grayscale abrupt change edge images. Then, the pixel classification error and overlap deviation are propagated back to the decoding layer and the convolutional layer through backpropagation, and the weights of the convolutional kernel and the decoding layer are updated according to the direction of decrease of training error.

[0049] After the training error continuously decreases and the probability maps of layered residual images and gray-scale abrupt edge images are consistent with the annotations of the training samples, the hierarchical connection relationship and weight parameters of the image recognition model are saved. After training, the candidate reference segment set is input into the image recognition model. The segmentation output layer of the image recognition model first outputs the probability maps of layered residual images and gray-scale abrupt edge images. The layered residual probability in the layered residual image probability map and the gray-scale abrupt edge probability in the gray-scale abrupt edge image probability map are read pixel by pixel according to the same detection coordinates. When the layered residual probability of the current pixel is greater than or equal to 0.5, the layered residual probability is higher than the gray-scale abrupt edge probability, and the absolute value of the difference between the two probabilities is greater than or equal to 0.1, the current pixel is assigned to the initial region of layered residual images. When the gray-scale abrupt edge probability of the current pixel is greater than or equal to 0.5, the gray-scale abrupt edge probability is higher than the layered residual probability, and the absolute value of the difference between the two probabilities is greater than or equal to 0.1, the current pixel is assigned to the initial region of gray-scale abrupt edge images.

[0050] When both the probability of residual texture and the probability of gray-scale abrupt edge of the current pixel are less than 0.5, the current pixel is assigned to the background region. When at least one of the probability of residual texture and the probability of gray-scale abrupt edge of the current pixel is greater than or equal to 0.5, and the absolute value of the difference between the two probabilities is less than 0.1, the direction of gray-scale abrupt edge of the current pixel is read, and the direction of gray-scale abrupt edge of the current pixel is compared with the extension direction of the corresponding layer texture reference segment. When the direction of gray-scale abrupt edge is distributed along the extension direction of the corresponding layer texture reference segment, the current pixel is assigned to the initial region of residual texture. When the direction of gray-scale abrupt edge deviates from the extension direction of the corresponding layer texture reference segment, the current pixel is assigned to the initial region of gray-scale abrupt edge. The initial region of residual texture and the initial region of gray-scale abrupt edge are connected and processed respectively. Isolated regions with an area of ​​less than nine pixels and not connected to adjacent regions of the same type are assigned to the background region. The initial region of residual texture after connection processing is determined as the residual texture image, and the initial region of gray-scale abrupt edge after connection processing is determined as the gray-scale abrupt edge image, which are used for subsequent residual coverage area expansion, counter-evidence occlusion area occlusion, and defect boundary image closure.

[0051] The threshold of 0.5 is used as the classification threshold for the probability of residual layer texture and the probability of gray-scale abrupt change edge because 0.5 indicates that the probability of the corresponding class reaches the median of binary classification, which can prevent low-confidence pixels from directly entering the initial area of ​​residual layer texture or the initial area of ​​gray-scale abrupt change edge. The threshold of 0.1 is used as the difference threshold between the probability of residual layer texture and the probability of gray-scale abrupt change edge because when the absolute value of the difference between the two probabilities is less than 0.1, the probabilities of the two classes are close, indicating that the current pixel is in a state of blurred class boundary, and it is necessary to combine the gray-scale change direction with the extension direction of the corresponding layer texture reference segment for secondary classification. Nine pixels are used as the threshold for culling isolated areas because nine pixels can form a minimum connected neighborhood of three rows and three columns, which is used to eliminate single-point noise and short-line noise, and retain the residual layer texture image and the gray-scale abrupt change edge image with local area support.

[0052] S2.3. Expand the pixel occupancy of the residual layer image in the defect candidate image segment into a residual coverage area. Use the residual coverage area as a counter-evidence masking area. Mask the counter-evidence masking area in the defect candidate image segment while retaining the gray-scale abrupt change edge image to obtain the edge candidate segment. Specifically, The residual image of the layer texture obtained by the image recognition model is returned to the corresponding pixel position in the defect candidate image segment. Starting from the pixel position covered by the residual image of the layer texture, it is extended outward along the adjacent continuous pixels to the position before the edge of the residual image of the layer texture meets the edge image of the gray-scale change, forming a residual coverage area.

[0053] The remaining covered area is marked as the counter-evidence masking area, and the corresponding original image grayscale within the counter-evidence masking area is masked in the defect candidate image segment. Only the grayscale abrupt change edge image outside the counter-evidence masking area and at the edge of the counter-evidence masking area is retained. When there are discontinuities between the remaining covered areas and the discontinuity position is not penetrated by the grayscale abrupt change edge image, they are retained as independent counter-evidence masking areas. When the remaining covered areas are connected or overlap, they are merged into the same counter-evidence masking area. After masking is completed, the remaining grayscale abrupt change edge images in the defect candidate image segment form edge candidate segments.

[0054] The counter-evidence masking area is the masking range obtained by marking the residual coverage area. It is used to exclude pixels identified as residual layer texture images from the defect candidate image fragments, so as to prevent residual layer texture images from participating in the subsequent defect boundary image closure.

[0055] S2.4. Within the edge candidate segments, the broken endpoints of the gray-scale abrupt change edge image are reconnected and arranged to form a closed contour by piecing together the broken endpoints belonging to the same deviated image region, thus creating a defect boundary image. Specifically, In the edge candidate segment, the unclosed edge end positions are found along the extension direction of the gray-level abrupt change edge image, and these edge end positions are organized as break endpoints. The interval direction, gray-level change continuation direction, and spatial proximity relationship between each break endpoint within the same offset image region are compared sequentially. The detection coordinates of two break endpoints are read to determine the endpoint spacing between them, and the end direction of the gray-level abrupt change edge image where the two break endpoints are located is read. When the endpoint spacing is no greater than 5% of the short side pixel length of the defect candidate image segment, the directional angle between the two end directions is no greater than 20°, and the two break endpoints... When the two broken endpoints are not blocked by the counter-evidence masking area, they are determined to have the same direction of continuation and the interval span is allowed. They are then joined along the connection position between the two broken endpoints to make the originally broken gray-scale abrupt edge image within the same deviation image area enclosed end to end. If any of the following occurs: the distance between the endpoints is greater than 5% of the short side pixel length of the defect candidate image segment, the directional angle between the two ends is greater than 20°, or the two broken endpoints are blocked by the counter-evidence masking area, the broken state is retained and it does not participate in the current joining. After the broken endpoints are joined, the already enclosed gray-scale abrupt edge image is organized into a defect boundary image.

[0056] Among them, five percent of the short side pixel length of the defect candidate image segment is used as the endpoint spacing criterion to make the patching range change synchronously with the size of the defect candidate image segment and avoid cross-regional erroneous patching; 20° is used as the direction angle criterion to allow slight bending of the gray-scale abrupt edge image while excluding fracture endpoints with obviously opposite directions.

[0057] S2.5. Based on the detection coordinates of the defect boundary image, obtain the co-position transmission image region from the composite substrate image. Remove the transmission background image that changes with the co-position layer texture reference segment in the co-position transmission image region to form a transmission anomaly image. Specifically... The image range in the composite substrate image is defined by the detection coordinates of the defect boundary image. An image range with the same image segment number, horizontal pixel position and vertical acquisition position as the defect boundary image is extracted from the composite substrate image as the co-position transmission image region. The original image grayscale in the co-position transmission image region is used as the source of transmission judgment. The brightness and darkness distribution of the co-position transmission image region is compared with the flattened grayscale, interlayer boundary trajectory line position and layer continuity position in the co-position layer texture reference segment. The brightness and darkness distribution that changes continuously along the co-position layer texture reference segment is identified as the transmission background image and removed. The brightness and darkness changes that do not change continuously with the co-position layer texture reference segment and are located within the range of the defect boundary image are retained to form the transmission anomaly image.

[0058] S2.6. Organize the defect boundary image into candidate defect boundaries, and realign the brightness and darkness distribution positions of the transmission anomaly image with the interlayer boundary images in the co-position tomographic reference segment to obtain candidate defect layers. Then, reassemble the candidate defect boundaries and candidate defect layers to generate a candidate defect parameter set. Specifically, Closed contours in the defect boundary image are registered one by one according to the detection coordinate sequence. The boundary position, enclosed range, and image segment number of each closed contour are kept, and each closed contour is determined as a candidate defect boundary. The position of brightness and darkness in the transmission anomaly image is checked. The vertical proximity and horizontal correspondence of the position of brightness and darkness within the same tomographic reference segment with the interlayer boundary image are calibrated one by one. When the position of brightness and darkness falls within the plane enclosed by the same interlayer boundary image, the corresponding plane position is recorded. When the position of brightness and darkness crosses the adjacent interlayer boundary image, it is registered as a candidate defect layer according to the plane where the brightness and darkness are concentrated. When the position of brightness and darkness cannot fall into a clear plane position, the layer anomaly mark is retained. The candidate defect boundaries and candidate defect layers are matched one by one according to the detection coordinate consistency. When the candidate defect boundaries and candidate defect layers can correspond to the same defect candidate image segment, they are merged and recorded to generate a candidate defect parameter group.

[0059] S3. Perform physical forward simulation on each candidate defect parameter in the candidate defect parameter group, and convert multiple candidate defect parameters into multiple simulated defect images. Competitively verify the multiple simulated defect images through defect candidate image fragments to generate a parameter verification sequence.

[0060] S3.1. Pair the candidate defect boundaries and candidate defect layers in the candidate defect parameter group one by one to form a candidate defect parameter sequence, specifically, The candidate defect boundaries and candidate defect layers in the candidate defect parameter group are paired according to the detection coordinates. The candidate defect boundary is taken from the boundary position, enclosed range and image segment number of the closed contour. The candidate defect layer is taken from the layer position after the brightness and darkness distribution position is returned to its original position and the image segment number. The candidate defect boundary and candidate defect layer with the same image segment number, overlapping detection coordinates and the candidate defect layer falling within the enclosed range of the candidate defect boundary are combined to form a candidate defect parameter. When there are multiple candidate defect layers that can be paired, the candidate defect layers with the brightness and darkness distribution position concentrated within the enclosed range of the candidate defect boundary are selected to participate in the pairing. When the candidate defect boundary lacks a candidate defect layer, the candidate defect layer has a layer anomaly mark, or the detection coordinates of the candidate defect boundary and the candidate defect layer cannot overlap, the boundary pairing anomaly mark is retained and the candidate defect layer is not entered into the front-end pairing position for the time being. After completing the pairing item by item, the candidate defect parameters are arranged according to the detection coordinate order of the candidate defect image segments to form a candidate defect parameter sequence.

[0061] S3.2, Project the candidate defect boundaries from the candidate defect parameter sequence onto the substrate layer corresponding to the candidate defect layer, and set the image area enclosed by the candidate defect boundaries as the defect occupancy area to form a layer perturbation patch. Specifically, Taking the same candidate defect parameter in the candidate defect parameter sequence as the processing object, the candidate defect boundary is placed on the substrate layer marked by the candidate defect layer, while keeping the boundary position and enclosed range of the candidate defect boundary in the detection coordinates unchanged; when the candidate defect boundary is closed and there is no layer anomaly mark for the candidate defect layer, the image range enclosed by the candidate defect boundary is set as the defect occupancy area, so that the defect occupancy area simultaneously carries the detection coordinates of the candidate defect boundary and the layer position of the candidate defect layer; when the candidate defect boundary has a boundary pairing anomaly mark or the candidate defect layer has a layer anomaly mark, the original arrangement position of the candidate defect parameter in the candidate defect parameter sequence is retained, and no defect occupancy area is generated; after completing the placement of the candidate defect boundary on the substrate layer and the setting of the defect occupancy area, a layer perturbation sheet is formed.

[0062] S3.3. Within the layer perturbation sheet, the edge region of the defect-occupying area is assigned to the reflection imaging path, and the inner region of the defect-occupying area is assigned to the transmission imaging path, forming an optical perturbation sheet. Specifically, Within the layer perturbation sheet, the edge region of the defect-occupying area is divided along the closed contour of the candidate defect boundary, and the edge region is assigned to the reflection imaging path, so that the reflection imaging path can carry the reflection grayscale changes of the candidate defect boundary in the composite substrate image; the image range within the defect-occupying area that is enclosed by the candidate defect boundary but does not adhere to the closed contour is divided into the inner region, and the inner region is assigned to the transmission imaging path, so that the transmission imaging path can carry the transmission brightness and darkness changes of the candidate defect layer in the composite substrate image; when the candidate defect boundary does not form a closed contour, the corresponding candidate defect parameter does not form a defect-occupying area, and the boundary pairing anomaly mark is retained. After the defect-occupying area completes the assignment of the edge region and the inner region, it forms an optical perturbation sheet.

[0063] Among them, the reflection imaging path is the imaging branch formed in the optical perturbation sheet after the edge region of the defect-occupying area enters the surface reflection direction, carrying the reflection grayscale change of the candidate defect boundary in the composite substrate image; the transmission imaging path is the imaging branch formed in the optical perturbation sheet after the inner region of the defect-occupying area enters the interlayer transmission direction, carrying the transmission brightness and darkness change of the candidate defect layer in the composite substrate image.

[0064] S3.4. Through physical forward simulation, the reflection imaging path and transmission imaging path in the optical perturbation sheet are derived respectively, completing the simulation of the reflection imaging response in the edge region and the transmission imaging response in the inner region, forming the reflection simulation image and the transmission simulation image. Specifically, Imaging simulations of the reflection and transmission imaging paths in the optical perturbation sheet were performed using physical forward simulation. Within the reflection imaging path, the grayscale changes of the edge region after entering the surface reflection direction were simulated along the detection coordinates of the candidate defect boundary. The original image grayscale, flattened grayscale, and candidate defect boundary direction on both sides of the edge region were organized into a reflection grayscale distribution to form a reflection simulation image. Within the transmission imaging path, the brightness and darkness changes of the internal region after entering the interlayer transmission direction were simulated according to the layer position of the candidate defect layer. The positions of the transmission anomaly image, interlayer boundary trajectory lines, and layer texture continuity within the internal region were organized into a transmission brightness and darkness distribution to form a transmission simulation image. When there are boundary pairing anomaly markers in the reflection imaging path, the reflection simulation image retains the anomaly position markers. When there are layer anomaly markers in the transmission imaging path, the transmission simulation image retains the anomaly position markers. This completes the simulation of the reflection imaging response of the edge region and the transmission imaging response of the internal region.

[0065] Among them, the edge region is the image range of the defect occupancy area that is adjacent to the closed contour of the candidate defect boundary, and the inner region is the image range of the defect occupancy area that is enclosed by the candidate defect boundary but is not adjacent to the closed contour.

[0066] It should be noted that the reflection imaging path adopts an edge reflection grayscale response model. The input parameters of the edge reflection grayscale response model include the detection coordinates of the candidate defect boundary, the direction of the candidate defect boundary, the original image grayscale on both sides of the edge region, and the flattened grayscale under the same detection coordinates in the layer texture reference. In the edge reflection grayscale response model, the flattened grayscale is first used as the defect-free reference grayscale. Then, the difference between the original image grayscale on both sides of the edge region and the defect-free reference grayscale is processed to obtain the reflection grayscale perturbation. Subsequently, the reflection grayscale perturbation is arranged into the detection coordinates of the edge region according to the direction of the candidate defect boundary to form the reflection grayscale distribution. Transmission imaging path The interlayer transmission brightness-darkness response model is adopted. The input parameters of the interlayer transmission brightness-darkness response model include the layer position of the candidate defect layer, the position of the transmission anomaly image in the internal region, the position of the interlayer boundary trajectory line, the position of the layer texture continuity, and the flattened grayscale of the co-position layer texture reference segment. In the interlayer transmission brightness-darkness response model, the flattened grayscale of the co-position layer texture reference segment is first used as the transmission background grayscale. Then, the brightness and darkness changes generated by the transmission anomaly image position in the candidate defect layer are organized into the transmission brightness-darkness perturbation amount. Subsequently, the transmission brightness-darkness perturbation amount is arranged in the internal region according to the layer position of the candidate defect layer and the detection coordinates of the internal region to form the transmission brightness-darkness distribution.

[0067] S3.5. Map the simulated reflection image and the simulated transmission image to the detection coordinates of the composite substrate image, and overlay them with the normal layer texture image in the layer texture reference to form a simulated defect image. Specifically, The candidate defect boundary detection coordinates carried by the reflection simulation image and the candidate defect layer detection coordinates carried by the transmission simulation image are respectively matched to the same detection coordinate position of the composite substrate image, so that the reflection grayscale distribution falls on the edge region and the transmission brightness and darkness distribution falls on the inner region. The normal layer texture image is taken from the layer texture image that is not occupied by the deviated image area under the corresponding detection coordinate in the layer texture reference image. The normal layer texture image is first retained as the bottom layer reference, and then the reflection simulation image and the transmission simulation image are superimposed on the corresponding edge region and inner region in the normal layer texture image. When the detection coordinates are consistent, the superimposed reflection grayscale distribution and transmission brightness and darkness distribution are retained. When there are abnormal position markers in the detection coordinates, the abnormal position markers are retained and do not participate in the current superposition. After the superposition is completed, a simulated defect image is formed.

[0068] S3.6 After the candidate defect parameters in the candidate defect parameter sequence have undergone physical forward simulation, multiple simulated defect images are formed, specifically, The candidate defect parameters that have completed physical forward simulation in the candidate defect parameter sequence are summarized in their original order. When the reflection simulation image and transmission simulation image corresponding to the candidate defect parameter have been superimposed into a simulated defect image, the simulated defect image is kept under the same candidate defect parameter item as the candidate defect boundary, candidate defect layer, defect occupancy area and detection coordinates. When the candidate defect parameter has a boundary pairing anomaly mark, layer anomaly mark or anomaly position mark, the corresponding simulated defect image retains the anomaly position and continues to be in the original order in the candidate defect parameter sequence. After summarizing the candidate defect parameters in the candidate defect parameter sequence item by item, multiple simulated defect images are formed that are consistent with the order of the candidate defect parameter sequence.

[0069] S3.7. Based on the detection coordinates of the candidate defect image segments, multiple simulated defect images are superimposed onto the same coordinate positions of the candidate defect image segments to form a co-position verification piece. Specifically, The detection coordinates of the candidate defect image fragments are set as verification coordinates, and the detection coordinates carried by multiple simulated defect images are calibrated respectively. When the grayscale distribution of reflection and the brightness and darkness distribution of transmission in the simulated defect image are at the same image segment number, horizontal pixel position and vertical acquisition position as the original image grayscale in the candidate defect image fragment, the simulated defect image is superimposed on the same coordinate position of the candidate defect image fragment. When there are abnormal position marks in the simulated defect image or the detection coordinates cannot be aligned with the candidate defect image fragment, the original arrangement position of the simulated defect image is retained and the verification coordinate abnormality is marked. After multiple simulated defect images are superimposed respectively, a co-position verification piece is formed for the generation of boundary verification record and transmission verification record.

[0070] S3.8 In the co-position verification piece, the simulated defect outer edge of the simulated defect image is co-positioned with the gray-scale abrupt change edge image in the defect candidate image segment. The bonding position, offset position, and break position are preserved and compiled into a boundary verification record. Specifically, In the same-position verification piece, the edge verification direction is determined along the outer edge of the simulated defect image, and the edge verification direction is compared segment by segment with the gray-scale abrupt edge image in the defect candidate image segment; the image position where the simulated defect outer edge and the gray-scale abrupt edge image are located at the same detection coordinate and the edge direction is continuous is recorded as the fitting position; the image position where the simulated defect outer edge and the gray-scale abrupt edge image have similar directions but the horizontal pixel position and the vertical acquisition position are offset is recorded as the offset position; the image position where there is no continuous gray-scale abrupt edge image at the location of the simulated defect outer edge, or the gray-scale abrupt edge image is interrupted on the corresponding edge direction, is recorded as the break position; the fitting position, offset position, and break position, together with the corresponding candidate defect parameters, simulated defect image, and detection coordinates, are organized to form the boundary verification record.

[0071] S3.9. Along the boundary verification record, arrange the grayscale change direction between the candidate defect image fragments and the simulated defect image, and record the position of the overlap between the simulated transmission image and the abnormal transmission image to form a transmission verification record. Specifically, The order of the alignment, misalignment, and breakage positions in the boundary verification record is used to organize the original image grayscale change direction in the defect candidate image segment and the simulated grayscale change direction in the simulated defect image. Positions where the original image grayscale change direction matches the simulated grayscale change direction are designated as priority verification positions for the overlap between the transmission simulation image and the transmission anomaly image. Positions where the original image grayscale change direction does not match the simulated grayscale change direction are designated as pending verification positions for the overlap between the transmission simulation image and the transmission anomaly image. Within the priority verification positions and pending verification positions, the detection coordinates, brightness and darkness distribution positions, and candidate defect layers of the transmission simulation image and the transmission anomaly image are further compared. Positions with consistent detection coordinates and brightness and darkness distribution positions falling within the same candidate defect layer are retained as overlap positions. Positions with misaligned detection coordinates, broken brightness and darkness distribution positions, or inconsistent candidate defect layers are retained as transmission misalignment positions. The grayscale change direction, overlap positions, and transmission misalignment positions are arranged along with the candidate defect parameters to form a transmission verification record.

[0072] It should be noted that, Figure 6As the standard deviation of image noise σ and the fluctuation of transmission brightness gradually increase, the overall transmission overlap rate of the complete method of this invention decreases, but it still maintains 80.8% under the highest perturbation conditions. This result shows that the transmission simulation image formed after the candidate defect layer enters the transmission imaging path can continue to maintain a good overlap relationship with the transmission anomaly image even when the transmission background fluctuation is enhanced. This proves that the transmission verification record can effectively characterize the positional consistency between the candidate defect layer and the actual transmission anomaly, so that the parameter verification sequence can not only rely on the defect boundary, but also combine the layer imaging characteristics to complete the parameter screening of candidate defects.

[0073] S3.10. The boundary verification records and transmission verification records jointly constrain the sequential positions of candidate defect parameters. Candidate defect parameters with continuous boundary alignment and complete overlap are placed at the front end, while candidate defect parameters with offset and broken positions are placed at the back end, generating a parameter verification sequence. Specifically, Based on the original arrangement of candidate defect parameters in the candidate defect parameter sequence, the aligning, misaligning, and disconnecting positions in the boundary verification record, along with the grayscale change direction, coincident position, and transmission misalignment position in the transmission verification record, are all incorporated into the same candidate defect parameter item. Candidate defect parameters whose aligning positions are continuous in the boundary verification record and whose coincident positions are complete in the transmission verification record are retained at the beginning of the parameter verification sequence as priority candidates for subsequent correction. Candidate defect parameters whose aligning positions are continuous in the boundary verification record but whose transmission verification records show transmission misalignment positions are also considered. After the candidate defect parameters at the front end, the candidate defect boundaries are retained and awaiting candidate defect layer correction; candidate defect parameters whose boundary verification records show a concentration of offset and break positions but whose transmission verification records show complete overlapping positions are arranged after the candidate defect parameters whose candidate defect layers need correction, retaining the candidate defect layers and awaiting candidate defect boundary correction; candidate defect parameters whose boundary verification records show a concentration of offset and break positions, and whose transmission verification records show transmission offset positions, are moved to the back end of the parameter verification sequence; after completing the order and position arrangement of candidate defect parameters, a parameter verification sequence is generated.

[0074] S3.11 After the parameter verification sequence is generated, the front-end candidate defect parameters in the parameter verification sequence are checked in a linked manner, and the checked front-end candidate defect parameters are passed to S4. Specifically, Using the front-end candidate defect parameters in the parameter verification sequence as the access object, the boundary verification record, transmission verification record, candidate defect boundary, and candidate defect layer under the same candidate defect parameter item are retrieved. First, the boundary usability state and boundary correction state of the candidate defect boundary are determined by the fitting position, offset position, and disconnection position in the boundary verification record. Then, the layer usability state and layer correction state of the candidate defect layer are determined by the overlapping position and transmission offset position in the transmission verification record. When the candidate defect boundary is in the boundary usability state and the candidate defect layer is in the layer usability state, the front-end candidate defect parameter is maintained at the front end of the parameter verification sequence. When the candidate defect boundary is in the boundary correction state and the candidate defect layer is in the layer usability state, the enclosing position of the candidate defect boundary in the detection coordinate is corrected by the overlapping position in the transmission verification record. When the candidate defect boundary is in the boundary usability state and the candidate defect layer is in the layer correction state, the repositioning position of the candidate defect layer in the co-position tomographic reference segment is corrected by the fitting position in the boundary verification record.

[0075] When the candidate defect boundary is in a state of boundary correction and the candidate defect layer is in a state of layer correction, the front-end candidate defect parameter is moved to the back end of the parameter verification sequence. After the correction, the candidate defect boundary and candidate defect layer re-enter the in-situ verification piece and form boundary verification records and transmission verification records again. When the bonding position in the boundary verification record is continuous and the overlapping position in the transmission verification record is complete, the corrected front-end candidate defect parameter is retained. If there are still misaligned or broken positions in the boundary verification record, or if there are still transmission misaligned positions in the transmission verification record, the corresponding candidate defect parameter is moved to the back end of the parameter verification sequence. Finally, the candidate defect parameter maintained at the front end in the parameter verification sequence is passed to S4 to correct the candidate defect layer and candidate defect boundary in the candidate defect parameter group.

[0076] S4. Correct the candidate defect layer and candidate defect boundary of the candidate defect parameter group according to the parameter verification sequence, determine the target defect parameter, and map the target defect parameter to the detection coordinates in the composite substrate image to generate the defect detection result.

[0077] S4.1. Take the candidate defect parameters maintained at the front end of the parameter verification sequence in S3.11 as the correction parameters, retrieve the candidate defect boundary, candidate defect layer, boundary verification record and transmission verification record in the correction parameters, and find the candidate defect boundary and candidate defect layer with consistent detection coordinates and consistent image segment numbers in the candidate defect parameter group.

[0078] The fitting positions, offset positions, and break positions in the boundary verification record are written into the correction process of the candidate defect boundary. The coincident positions and transmission offset positions in the transmission verification record are written into the correction process of the candidate defect layer. When the boundary verification record shows that the fitting position is continuous and the transmission verification record shows that the coincident position is complete, the candidate defect boundary and candidate defect layer in the correction parameters are retained as the correction result. When the boundary verification record shows that the offset position or break position still exists, the boundary position and enclosure range of the candidate defect boundary are adjusted according to the image position of the continuous distribution of the fitting position. When the transmission verification record shows that the transmission offset position still exists, the layer position of the candidate defect layer is adjusted according to the image position of the continuous distribution of the coincident position.

[0079] After the candidate defect boundaries and candidate defect layers are corrected, the corrected candidate defect boundaries and candidate defect layers are merged into the target defect parameter. If the candidate defect boundary still has boundary pairing anomaly markers or the candidate defect layer still has layer anomaly markers, the anomaly markers are retained and the candidate defect parameter is not included in the current target defect parameter. The boundary position, enclosing range, layer position, and image segment number in the target defect parameter are mapped to the corresponding detection coordinate positions in the composite substrate image, so that the target defect parameter retains the complete defect localization relationship in the composite substrate image. The target defect parameter and the corresponding detection coordinates are organized into the defect detection result.

[0080] It should be noted that, Figure 5 When performing repeated detections under the same perturbation level, the repeat detection consistency rate of the complete method of this invention is consistently higher than that of other comparative methods. In the local magnified results in the D4 to D5 interval, the repeat detection consistency rate of the complete method of this invention decreases from 87.2% to 83.1%, but is still significantly higher than the single gray-scale anomaly identification method, the reflection imaging path verification method, and the transmission imaging path verification method. This result shows that this invention does not rely on a single gray-scale anomaly to determine the defect result, but uses boundary verification records and transmission verification records to jointly constrain the sorting position of candidate defect parameters. This allows the target defect parameters to still have good repeatability even under conditions of increased noise, increased layer complexity, and increased transmission brightness fluctuations, thus demonstrating the stability of the defect detection results.

[0081] This embodiment also provides a computer device applicable to the image recognition-based composite substrate defect detection method, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the image recognition-based composite substrate defect detection method proposed in the above embodiment.

[0082] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0083] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the image recognition-based composite substrate defect detection method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0084] In summary, this invention achieves consistent screening of candidate defect parameters through physical forward simulation and competitive verification with simulated defect images. After the candidate defect boundary and candidate defect layer enter the reflection imaging path and transmission imaging path respectively, they form a reflection simulation image and a transmission simulation image, which are then verified against the candidate defect image fragments under the same detection coordinates. This ensures that the candidate defects obtained from image recognition no longer rely solely on a single grayscale anomaly. The boundary verification record reflects the fit between the simulated defect outer edge and the grayscale abrupt change edge image, while the transmission verification record reflects the overlap between the transmission simulation image and the transmission anomaly image. The parameter verification sequence further places candidate defect parameters with continuous boundary fit and complete overlap at the forefront, thereby improving the accuracy, stability, and interpretability of target defect parameter determination.

[0085] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting defects in composite substrates based on image recognition, characterized in that: include, Acquire composite substrate images and record the detection coordinates of the composite substrate images. Perform texture benchmarking on the surface texture and interlayer boundary images in the composite substrate images to form a layer texture reference. Cut the image areas that deviate from the layer texture reference into defect candidate image fragments. The candidate image fragments of defects are fed into the image recognition model to separate the residual image of layer texture and the gray-scale abrupt edge image. The residual image of layer texture is removed and the gray-scale abrupt edge image is closed into the defect boundary image. The transmission abnormal image is extracted from the composite substrate image under the same detection coordinate and the defect boundary image and the transmission abnormal image are recombined to generate the candidate defect parameter group. Physical forward simulation is performed on the candidate defect parameters in the candidate defect parameter group, and multiple candidate defect parameters are converted into multiple simulated defect images. The multiple simulated defect images are competitively verified by defect candidate image fragments to generate a parameter verification sequence. The candidate defect layer and candidate defect boundary of the candidate defect parameter group are corrected according to the parameter verification sequence, the target defect parameter is determined, and the target defect parameter is mapped to the detection coordinates in the composite substrate image to generate the defect detection result.

2. The image recognition-based composite substrate defect detection method as described in claim 1, characterized in that: The formation of the texture reference is specifically as follows: The composite substrate image is divided into multiple composite substrate image segments according to the detection coordinate order, and arranged sequentially to generate a continuous image band. The interlayer boundary images that span multiple composite substrate image segments are traced within a continuous image band. The discontinuous interlayer boundary images are connected into an interlayer boundary trajectory line, and the surface texture images on both sides of the interlayer boundary trajectory line are preserved to generate an interlayer boundary trajectory map. Establish a texture coordinate zone with the interlayer boundary trajectory line in the interlayer boundary trajectory map as the center line, and arrange the surface texture images on both sides of the interlayer boundary trajectory line along the extension direction of the interlayer boundary trajectory line to the texture coordinate zone to generate a texture unfolding map. Full-resolution texture flattening is performed on the surface texture image in the layer texture unfolding image to reduce the fine gray-level fluctuations in the surface texture image, while preserving the width variation and brightness transition on both sides of the interlayer boundary trajectory line, thus generating a layer texture flattening image. The lamination flattening map is mapped to the detection coordinates of the composite substrate image and superimposed with the interlayer boundary trajectory map to form a lamination reference map.

3. The image recognition-based composite substrate defect detection method as described in claim 2, characterized in that: Specifically, cropping the image region that deviates from the layer texture reference into defect candidate image segments involves... The composite substrate image and the layer texture reference image are aligned segment by segment under the detection coordinates to form a reference image pair; The reference image pair is split and converted by a dual-path feature encoding method. The composite substrate image in the reference image pair is converted into a measured feature field, and the layer texture reference in the reference image pair is converted into a reference feature field, thus generating a dual-path feature field. Within the dual-path feature field, the reference feature field and the measured feature field are assigned to the same feature coordinates. The deviation positions of the layer texture in the measured feature field are repositioned to the reference position, and the unrepositioned deviation positions of the layer texture are connected to form a deviation feature map. Based on the non-positioned layer texture deviation position within the deviation feature map, the co-position image range is delineated in the composite substrate image, and adjacent co-position image ranges are merged into the deviation image region. Apply layer texture to the deviated image region to refer to the clipping constraint of the interlayer boundary trajectory line in the image. Deviated image regions that cross the interlayer boundary trajectory line are divided into interlayer clipping windows, and deviated image regions that do not cross the interlayer boundary trajectory line are divided into surface clipping windows, generating a candidate clipping window sequence. The composite substrate image is cropped using interlayer cropping windows and surface cropping windows in the candidate cropping window sequence. The original image grayscale and detection coordinates within the cropped area are preserved and identified as candidate image fragments for defects.

4. The image recognition-based composite substrate defect detection method as described in claim 3, characterized in that: The generation of candidate defect parameter sets specifically refers to... Based on the detection coordinates carried by the defect candidate image fragments, the corresponding layer texture reference fragments are extracted from the layer texture reference image, and the defect candidate image fragments and the corresponding layer texture reference fragments are arranged into a candidate reference fragment set. Image recognition models are used to perform isotopic texture resolution on the candidate reference fragment set, separating the residual layer texture image that extends with the isotopic layer texture reference fragment and the gray-scale abrupt edge image that deviates from the extension direction of the isotopic layer texture reference fragment in the defect candidate image fragment. The pixel occupancy of the residual layer image in the defect candidate image segment is expanded into a residual coverage area. The residual coverage area is used as a counter-evidence masking area. The counter-evidence masking area is masked in the defect candidate image segment while retaining the gray-scale change edge image to obtain the edge candidate segment. Within the edge candidate segments, the broken endpoints of the gray-scale abrupt edge image are connected and arranged to form a closed contour by connecting the broken endpoints belonging to the same deviated image region, thus forming a defect boundary image. Based on the detection coordinates of the defect boundary image, the in-situ transmission image region is obtained from the composite substrate image. The transmission background image that changes with the in-situ layer texture reference segment in the in-situ transmission image region is removed to form a transmission abnormality image. The defect boundary image is organized into candidate defect boundaries. The position of the brightness and darkness distribution of the transmission anomaly image is aligned with the interlayer boundary image in the co-position layer reference segment to obtain the candidate defect layer. The candidate defect boundary and candidate defect layer are then recombined to generate a candidate defect parameter group.

5. The image recognition-based composite substrate defect detection method as described in claim 4, characterized in that: The process of converting multiple candidate defect parameters into multiple simulated defect images specifically involves: The candidate defect boundaries and candidate defect layers in the candidate defect parameter group are paired one by one to form a candidate defect parameter sequence; The candidate defect boundary in the candidate defect parameter sequence is projected onto the substrate layer corresponding to the candidate defect layer, and the image range enclosed by the candidate defect boundary is set as the defect occupancy area to form a layer perturbation sheet. Within the layer perturbation sheet, the edge region of the defect-occupying area is assigned to the reflection imaging path, and the inner region of the defect-occupying area is assigned to the transmission imaging path, thus forming an optical perturbation sheet. By physically forward simulation, the reflection imaging path and the transmission imaging path in the optical perturbation sheet are respectively derived, and the reflection imaging response simulation of the edge region and the transmission imaging response simulation of the inner region are completed, thus forming the reflection simulation image and the transmission simulation image. The simulated reflection image and the simulated transmission image are mapped to the detection coordinates of the composite substrate image and superimposed with the normal layer texture image in the layer texture reference to form a simulated defect image; After the candidate defect parameters in the candidate defect parameter sequence are physically simulated, multiple simulated defect images are formed.

6. The image recognition-based composite substrate defect detection method as described in claim 5, characterized in that: The reflection imaging path is an imaging branch formed in the optical perturbation sheet after the edge region of the defect-occupying area enters the surface reflection direction, carrying the reflection grayscale change of the candidate defect boundary in the composite substrate image. The transmission imaging path is an imaging branch formed in the optical perturbation sheet after the internal region of the defect-occupying area enters the interlayer transmission direction, carrying the transmission brightness and darkness changes of the candidate defect layer in the composite substrate image.

7. The image recognition-based composite substrate defect detection method as described in claim 6, characterized in that: The generated parameter verification sequence is specifically as follows: Based on the detection coordinates of the defect candidate image fragments, multiple simulated defect images are superimposed onto the same coordinate positions of the defect candidate image fragments to form a co-position verification piece; In the co-position verification film, the simulated defect outer edge of the simulated defect image is co-positioned with the gray-scale abrupt change edge image in the defect candidate image segment, and the bonding position, offset position and disconnection position are preserved and organized into a boundary verification record. The grayscale change direction between the candidate defect image fragments and the simulated defect image is arranged along the boundary verification record, and the position of the overlap between the transmission simulated image and the transmission abnormal image is marked to form a transmission verification record. The candidate defect parameters are constrained by the boundary verification records and the transmission verification records. Candidate defect parameters with continuous boundary fit and complete overlapping positions are placed at the front end, and candidate defect parameters with offset and broken positions are placed at the back end, thus generating a parameter verification sequence.

8. The image recognition-based composite substrate defect detection method as described in claim 7, characterized in that: The determination of the target defect parameter is achieved by correcting the candidate defect level and candidate defect boundary in the candidate defect parameter group by modifying the front-end candidate defect parameter in the parameter verification sequence, and then synthesizing the corrected candidate defect level and candidate defect boundary into the target defect parameter.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the image recognition-based composite substrate defect detection method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the image recognition-based composite substrate defect detection method according to any one of claims 1 to 8.

Images