Image processing-based liquid crystal glass substrate polaroid quality detection method and device

Abstract

This invention discloses a method and apparatus for quality inspection of polarizers on liquid crystal glass substrates based on image processing, belonging to the field of image inspection technology. The method includes: firstly, structural identification and region modeling of the liquid crystal glass substrate under test are performed, and a detection benchmark is established; then, multi-condition images of the same detection area are acquired under different detection states, and a reference image set and a detection image set are constructed; subsequently, spatial registration, background decoupling, defect enhancement, and layer discrimination are performed on the images; finally, quality evaluation results and process feedback are output based on defect type, layer position, and location information. This scheme can distinguish and detect surface defects, bonding interface defects, conductive layer-related anomalies, and liquid crystal functional anomalies, improving detection accuracy, evaluation consistency, and defect traceability in the manufacturing process, while also adapting to the differentiated detection needs of the central optical region, edge bonding region, conductive connection region, and partition boundary region.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, specifically to a method and apparatus for detecting the quality of polarizers on liquid crystal glass substrates based on image processing. Background Technology

[0002] Liquid crystal glass substrate polarizers are widely used in smart dimming glass, privacy partitions, display windows, automotive side windows, and architectural lighting components. These products are generally not formed by combining a glass substrate, conductive layer, liquid crystal functional layer, encapsulation layer, and attached optical layer. They can exhibit different light transmission responses in a power-off scattering state, a power-on transparent state, or a zone-driven state. The manufacturing process involves conductive layer deposition, etching, edge lead generation, liquid crystal layer lamination, polarizer or optical film attachment, encapsulation and curing, and power-on testing. Deviations at any step will result in bubbles, foreign matter, scratches, attachment misalignment, edge lifting, localized delamination, abnormal conductive layer boundaries, uneven brightness in zones, or abnormal transmittance in the finished product.

[0003] Conventional inspection methods typically involve visual inspection under single illumination conditions and sampling of products to measure macroscopic optical parameters such as transmittance and haze. While these methods can detect some local defects, they are not reliable for identifying interface defects located within multi-layered structures, functional anomalies related to the etching boundaries of conductive layers, or local anomalies that only appear under power-on / off switching or zone control. Furthermore, glass surface reflection, interlayer refractive index differences, edge encapsulation shadows, and zone boundary transition areas can cause the same defect to exhibit significantly different image features under different observation conditions, increasing the difficulty of inspection and judgment and making it easy for discrepancies to arise between sampling inspection conclusions and online judgment results.

[0004] In the above application scenarios, the existing technology has the following technical problems: For multilayer composite, variable transmittance, and zone-driven liquid crystal glass substrate polarizers, there is a lack of a standardized detection basis for acquiring and judging defect characterization around the same area under different operating states and imaging conditions. This makes it easy to confuse surface defects, interface defects, conductive layer-related anomalies, and functional optical anomalies, making it difficult to determine the true location, cause, and quality risk of defects. In cases with large product size, many zones, complex edge electrodes, or high requirements for transmittance consistency, if detection cannot be carried out for a long time, there will be quality risks such as the wrong release of defective products, the inability to trace process reasons, local devitrification in end-use, abnormal zone display, and inconsistent privacy effects. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method and apparatus for quality inspection of polarizers on liquid crystal glass substrates based on image processing. The method performs spatial registration, background decoupling, defect enhancement, and layer discrimination on the image, and finally outputs quality evaluation results and process feedback based on defect type, layer position, and location information. This solution can differentiate and detect surface defects, bonding interface defects, conductive layer-related anomalies, and liquid crystal functional anomalies, improving detection accuracy and defect traceability in the manufacturing process. It also adapts to the differentiated detection needs of the central optical region, edge bonding region, conductive connection region, and partition boundary region, thus solving the technical problems described in the background art.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A method for inspecting the quality of polarizers on liquid crystal glass substrates based on image processing includes the following steps: Step 1: Imaging the entire liquid crystal glass substrate under test, extracting at least one feature from the outer contour edge, encapsulation edge, conductive lead-out area, partition control boundary, and polarizer attachment boundary, establishing a substrate coordinate system, dividing the central optical detection area, edge attachment detection area, conductive connection detection area, and partition boundary detection area, and establishing structural reference information and state reference under at least two detection states; Step 2: Acquiring at least two types of images from the same detection area under different detection states, including transmission images, reflection images, bright field images, dark field images, and images with different polarization directions, constructing a reference image set and a detection image set; Step 3: Spatial registration and background decoupling of the detection image set, and extracting candidate defect regions based on the response differences under different detection states and different optical conditions, and identifying the defect type and layer; Step 4: Outputting quality evaluation results based on defect type, layer, and location information.

[0007] Furthermore, in step one, a substrate coordinate system is established, including determining the relative positions of the outer contour edge, package edge, conductive lead-out area, partition control boundary and polarizer attachment boundary based on the overall imaging results, and dividing the central optical detection area, edge attachment detection area, conductive connection detection area and partition boundary detection area according to the relative positions.

[0008] Furthermore, in step one, structural reference information and state reference are established, including establishing the positional correspondence between the glass substrate layer, the corresponding area of ​​the conductive layer, the corresponding area of ​​the liquid crystal functional layer and the corresponding area of ​​the polarizer attachment layer and the detection area, and recording the state reference of each detection area in the power-off state, the power-on state and the independent driving state of each zone.

[0009] Furthermore, in step two, image acquisition includes acquiring transmission images, images with different polarization directions, and reflection images sequentially according to a preset time sequence while maintaining the same spatial position in the same detection area. Dark field images are also acquired in the detection state corresponding to the partition boundary detection area to form a detection image corresponding to the detection area.

[0010] Furthermore, in step two, a reference image set is constructed, which includes classifying and storing normal sample images according to detection area, detection state, and image category, and writing normal sample images that are consistent with the current detection area, current detection state, and current image category into the reference image set as reference condition images.

[0011] Furthermore, in step three, spatial registration is performed, including global alignment and local boundary alignment of each image in the detection image set based on the corresponding reference features in the outer contour edge, conductive lead-out area, partition control boundary and polarizer attachment boundary, so that each image is uniformly mapped to the same detection area in the substrate coordinate system.

[0012] Furthermore, in step three, background decoupling is performed, including performing low-frequency brightness field correction, specular reflection suppression, strip smoothing along the polarizer attachment boundary, and bilateral background fitting along the partition control boundary on the registered image according to the category of the detection area, and using the decoupled abnormal foreground as input for extracting defect candidate areas.

[0013] Furthermore, in step three, the defect type and layer are determined by combining the response differences of the defect candidate region in the transmission image, reflection image and images under different detection states, as well as the spatial relationship between the defect candidate region and the polarizer attachment boundary, partition control boundary and conductive lead-out area, to determine the defect type and layer.

[0014] Furthermore, in step four, the quality evaluation results are output, including regional evaluation and whole-piece evaluation of the liquid crystal glass substrate under test by combining defect type, layer, location information, detection area category and response consistency under detection state, and outputting one of the evaluation results: good product, limited release product, reworkable product and defective product.

[0015] Furthermore, step four also includes generating a test report bound to the liquid crystal glass substrate under test. The test report includes the test area identifier, defect type, defect coordinates, layer determination result, response differences under test status and corresponding image evidence, and outputs the feedback results of the corresponding bonding process, conductive layer formation process, partition etching process, packaging process or electric drive debugging process according to the test area type, defect type and layer.

[0016] A method for inspecting the quality of polarizers on liquid crystal glass substrates based on image processing includes: a support platform, an overall imaging mechanism, a state application mechanism, a multi-condition image acquisition mechanism, a memory, and an image processor. The support platform supports the liquid crystal glass substrate under test. The overall imaging mechanism is used to image the entire liquid crystal glass substrate under test. The state application mechanism is used to apply at least two detection states to the liquid crystal glass substrate under test. The multi-condition image acquisition mechanism is used to acquire at least two types of images from the following categories for the same detection area under different detection states: transmission image, reflection image, bright field image, dark field image, and images with different polarization directions. The memory is used to store a reference image set, a detection image set, and program instructions. The image processor is connected to the overall imaging mechanism, the state application mechanism, the multi-condition image acquisition mechanism, and the memory, and is configured to execute program instructions to: extract at least one feature from the outer contour edge, encapsulation edge, conductive lead-out area, partition control boundary, and polarizer attachment boundary; establish a substrate coordinate system; divide the central optical detection area, edge attachment detection area, conductive connection detection area, and partition boundary detection area; and establish structural reference information and state reference. Construct a reference image set and a detection image set; perform spatial registration and background decoupling on the detection image set, and extract candidate defect regions based on response differences under different detection states and optical conditions to determine defect types and layers; output quality evaluation results based on defect type, layer, and location information.

[0017] (III) Beneficial Effects This invention provides a method and apparatus for quality inspection of polarizers on liquid crystal glass substrates based on image processing, which has the following beneficial effects: By performing structural identification, region modeling, and establishing substrate coordinate system, structural reference, and state reference for the liquid crystal glass substrate under test, the central optical detection area, edge attachment detection area, conductive connection detection area, and partition boundary detection area are described under the same semantic location, thereby avoiding the defect identification of multi-layer composite structures as single-layer sheets.

[0018] By acquiring multi-condition images of the detection area under different detection states, a reference image set and a detection image set are established, allowing the transmission response, reflection response, polarization response, and state switching response to be compared within the same object, thereby improving the recognition rate of appearance defects, attachment defects, and functional abnormalities. By performing spatial registration and background decoupling on the detection image set, images obtained under different optical conditions are mapped to the substrate coordinate system, while simultaneously separating the inherent texture of the glass substrate, ambient reflection, packaging edge shadows, and partition control boundary transitions.

[0019] By combining the detection status and response differences under different optical conditions, the defect types and layers in the candidate areas are distinguished, making it possible to differentiate between bubbles, foreign objects, scratches, attachment misalignment, edge lifting, local delamination, conductive layer boundary anomalies, and inconsistent zone responses on the polarizer surface, attachment interface, corresponding areas of the conductive layer, and corresponding areas of the liquid crystal functional layer. Combining defect type, layer, and location information with the detection area category and state response consistency for quality evaluation transforms the quality evaluation results from simple pass / fail judgments into hierarchical conclusions tailored to the actual application scenarios of liquid crystal glass substrates. This strengthens the correlation between detection results and product release decisions, further linking detection results to the attachment process, conductive layer formation process, zone etching process, encapsulation process, and electric drive debugging process, forming an overall solution for the collaborative construction of the detection and manufacturing links. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of the liquid crystal glass substrate quality inspection equipment of the present invention; Figure 2 This is a schematic diagram illustrating the establishment of the substrate coordinate system and the division of the detection area in this invention; Figure 3 This is a schematic diagram of the timing of the detection state application and multi-condition image acquisition in this invention; Figure 4 This is a schematic diagram illustrating the construction of the reference image set and the encapsulation of the detection image set in this invention; Figure 5 This is a schematic diagram illustrating the multi-image registration, background decoupling, and abnormal foreground extraction of the present invention; Figure 6 This is a schematic diagram illustrating the defect identification, quality evaluation, and process closed-loop feedback of the present invention. Figure 7 This is a schematic diagram of the overall process of the liquid crystal glass substrate quality inspection method of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-7 This invention provides a method for quality inspection of polarizers on liquid crystal glass substrates based on image processing, including: Step 1: Before collecting defect information, the geometric position, composite layer position, and working state position of the liquid crystal glass substrate to be tested are set as repeatedly callable detection benchmarks. Subsequent steps rely on the same detection object and will not cause semantic drift.

[0023] After the LCD glass substrate enters the inspection station, even after it is fed to the designated area by the loading machine, it will still shift in the plane, and the edges will lift up due to the difference in the thickness of the encapsulant. Furthermore, the conductive lead-out area, the partition control boundary, and the polarizer attachment boundary are not collinear. If the image pixel position is used directly as the judgment basis, the same physical point will not be stably aligned in different shooting batches.

[0024] The LCD glass substrate quality inspection equipment first feeds the LCD glass substrate to be tested onto a support platform. The platform uses a combination of an upper support surface and lateral limiting surfaces to restrict its posture. The upper support surface suppresses localized sagging, while the lateral limiting surfaces constrain large-scale swaying. Based on this, a top imaging lens acquires an overall image vertically, and a lateral supplementary lighting unit projects oblique incident light from the long side, creating a distinguishable transition between light and dark areas at the encapsulation edge, conductive lead-out area, and polarizer attachment boundary within the overall image. Subsequently, instead of using a single contour as the identification result, the LCD glass substrate quality inspection equipment simultaneously extracts at least four of six structural features: outer contour edge, corner features, encapsulation edge, conductive lead-out area, partition control boundary, and polarizer attachment boundary. These features are then used together to construct the substrate coordinate system and structural reference information.

[0025] Relying solely on a single type of feature is susceptible to glare, occlusion, or edge damage. When multiple types of features are used in combination, even if any local feature becomes blurred, the remaining features can still maintain the continuity of the coordinate definition.

[0026] For example, the overall image first undergoes edge energy suppression reflection processing, and then searches for the outer contour edge according to edge direction consistency. After completing the outer contour edge search, the encapsulation edge and polarizer attachment boundary are traced inward, using corner features as turning points. Here, the brightest line in the image is not directly used as the physical boundary; instead, the line is required to simultaneously satisfy continuous length, corner consistency, and relative orientation with the conductive lead-out area, thereby excluding reflective stripes. Subsequently, the pixel positions in the overall image are mapped to the substrate coordinate system, and the mapping relationship is represented by a planar homography transformation as follows:

[0027] Among them, the horizontal coordinate of the substrate : The lateral position of the liquid crystal glass substrate under test in the substrate coordinate system, with values ​​ranging from non-negative values ​​within the short side length range of the liquid crystal glass substrate; the longitudinal coordinate of the substrate. The vertical position of the liquid crystal glass substrate under test in the substrate coordinate system is taken as a non-negative value within the range of the long side length of the liquid crystal glass substrate; the horizontal coordinate of the pixel is... : The horizontal pixel position in the overall image, with a value range corresponding to the pixel sequence of the imaging width; Pixel ordinate : Vertical pixel position in the overall image, with values ​​ranging from the pixel sequence corresponding to the imaging height; coordinate mapping matrix The planar mapping relationship is determined by the outer contour edge, corner features and conductive lead-out area. Each element is obtained by combining the calibration pattern and the current recognition result.

[0028] After adopting this mapping, any subsequent image that identifies the same physical features can be represented in the same substrate coordinate system. The substrate coordinate system is bound to the actual boundary of the liquid crystal glass substrate under test, and subsequent images can be repeatedly mapped to the same physical framework; the joint locking of the outer contour edge, corner features and conductive lead-out areas reduces the interference of local reflections and local missing edges on the coordinate definition; in step two, there is no need to re-estimate the entire orientation, only the substrate coordinate system needs to be called to cut out the corresponding detection area.

[0029] After the substrate coordinate system is locked, the planar position alone is not enough to support subsequent defect identification, because the same planar area may correspond to the glass substrate layer, the boundary of the conductive layer, the change area of ​​the liquid crystal functional layer, or the polarizer attachment interface.

[0030] In the substrate coordinate system, the quality inspection equipment for liquid crystal glass substrates uses the outer contour edge as the external constraint and the encapsulation edge, conductive lead-out area, and partition control boundary as the internal reference line to delineate the central optical inspection area, edge attachment inspection area, encapsulation transition inspection area, conductive connection inspection area, and partition boundary inspection area segment by segment. Subsequently, the positional correspondence between each inspection area and the corresponding areas of the glass substrate layer, conductive layer, liquid crystal functional layer, protective layer, and polarizer attachment layer is established.

[0031] The device also records the boundary origin for each detection zone, that is, whether the zone is formed by the inward offset of the outer contour edge, by the expansion of both sides of the partition control boundary, or by the envelope surrounding the conductive lead-out area.

[0032] Based on this source of boundary, the region attribution quantity is defined as:

[0033] Among them, the amount of regional affiliation : No. The structural attribution score for each detected region is a non-negative real number, which integrates different boundary sources into a single region determination criterion; edge distance item : No. The normalized distance from each detection region to the nearest outer contour edge or package edge, with a value ranging from 0 to 1. A value closer to 0 indicates that the region is closer to the edge; control boundary terms. : No. The degree of proximity between each detection area and the partition control boundary, with a value ranging from 0 to 1; Attach boundary items : No. The degree of overlap between each detection area and the polarizer attachment boundary, with values ​​ranging from 0 to 1; distance weight. Control weights and attach weight These represent the proportions of the three types of boundary information in the defined area, with values ​​being real numbers greater than 0 and less than 1, and the sum of the three is 1.

[0034] After adopting this area affiliation, the device can clearly distinguish between areas that are near the edge but far from the partition line and areas that are near the partition line but not near the edge.

[0035] For example, when the conductive lead-out area is located on a short side of the liquid crystal glass substrate, the area adjacent to the conductive lead-out area is defined as the conductive connection detection area; when the partition control boundaries are distributed in a longitudinal strip, partition boundary detection areas are simultaneously generated on both sides of each partition control boundary; when the polarizer attachment boundary is closed along the perimeter, the strip-shaped area from the polarizer attachment boundary to the outer contour edge is defined as the edge attachment detection area. The entire liquid crystal glass substrate under test is thus divided into several detection areas with clearly defined physical meanings.

[0036] When used, firstly, the structural reference information associates the planar region with the position of the composite layer, so that even if the gray levels of the same image are similar, its observation significance can be determined based on the region it is located in. Secondly, the region assignment quantity unifies the three types of information, namely edge, partition and attachment, into a single decision quantity, so that the region division has a reusable engineering expression.

[0037] However, static geometric position and layer position alone cannot cover the functional differences of liquid crystal glass substrates. This is because the liquid crystal functional layer exhibits a scattering response when the power is off and a transmission response when the power is on. Near the boundary of the partition control, changes in the electric field distribution will also show a transition between light and dark. The same interface defect may be masked by the scattered background when the power is off, but will show a sudden change in light transmission when the power is on. If the detection state is not bound to each detection area in step one, the multi-condition image obtained in step two can only indicate what was seen, but cannot indicate under what operating conditions this phenomenon was seen.

[0038] After generating the substrate coordinate system and structural reference information, the liquid crystal glass substrate quality inspection equipment applies a predetermined state sequence to the conductive lead-out areas sequentially by a driving power supply. This predetermined state sequence includes at least a power-off state and a power-on state, and further includes independent driving states for each zone in the case of a multi-zone design. To ensure consistency between state semantics and image time, the equipment does not immediately capture images after each state switch. Instead, it waits for the liquid crystal functional layer response to stabilize within a predetermined range before triggering the top imaging lens and the state recording unit to work synchronously. The state recording unit then adds a current inspection state identifier to each overall image. Subsequently, the equipment extracts the basic optical response within each inspection area to form a state reference.

[0039] For example, the driving power supply establishes an electrical connection with the conductive layer through the conductive lead-out area, maintaining a power-off state in the first time segment, applying a power-on state in the second time segment, and applying a partitioned independent driving state in the third time segment according to the area divided by the partition control boundary. The top imaging lens does not capture images freely, but is uniformly activated by a trigger signal issued by the state recording unit; only when the trigger signal confirms that the current detection state has entered the corresponding time segment and the overall brightness change of the liquid crystal functional layer tends to be stable, does the top imaging lens output the overall image. The reason for this setting is that if image acquisition occurs during the state transition process, the same detection area contains both the residue of the previous state and the initial response of the next state, and subsequent analysis cannot distinguish whether it is caused by state instability or defects. For example, when switching from a power-off state to a power-on state, the state recording unit first records the time of occurrence of the switching command, and then monitors whether the global brightness change in the overall image enters the gradual change range; when the gradual change range is established, the state recording unit sends an image acquisition trigger to the top imaging lens. Therefore, the same liquid crystal glass substrate corresponds to different overall images in the power-off state, power-on state, and independent driving state of each zone, and each image has a unique detection status identifier.

[0040] In use, the state application and image acquisition establish a one-to-one correspondence through a unified trigger, and the subsequent comparison yields state differences rather than timing errors; secondly, the partition-independent driving state is incorporated into step one, so that the observed objects near the partition control boundary have clear state semantics before image acquisition; thirdly, step two does not require additional explanation of the image acquisition time.

[0041] After the detection state is applied, it is necessary to further precipitate the basic response of each detection area under each detection state into a unified parameter; otherwise, although step two can continue to collect multi-condition images, it lacks an initial reference.

[0042] Therefore, the LCD glass substrate quality inspection equipment extracts three basic quantities within each inspection area: the average transmittance response, the edge transition response, and the interface continuity response, and writes them into a state reference table according to the inspection status. To enable the state reference to have a callable single-value expression, the area state fingerprint is defined as:

[0043] Among them, regional state fingerprint : No. The detection area in the first The comprehensive response identifier under each detection state takes a non-negative real number value; through the response : No. The detection area in the first The region transmittance performance under the first detection state, with values ​​ranging from 0 to 1 after normalization; from the first... The detection area in the first The average gray level of the region within the preset sampling window is normalized in the transmission image under each detection state; edge transition response. : No. The detection area in the first The steepness of the boundary brightness change under each detection state, with values ​​ranging from non-negative real numbers; obtained by integrating the grayscale gradient in the normal direction of the encapsulation edge, partition control boundary, or polarizer attachment boundary; continuous interface response. : No. The detection area in the first The degree of local texture continuity under each detection state is a non-negative real number; it is obtained from the directional consistency of local texture continuity in cross-polarized or transmission images. (Through weights) Transition weights and continuous weights These represent the proportions of the three basic quantities in the regional state fingerprint, with values ​​ranging from 0 to 1. The sum of the three is set to 1.

[0044] After using the regional state fingerprint, in step two, when acquiring transmission images, reflection images, bright field images, or dark field images, the starting response of the corresponding region under the corresponding detection state can be directly called.

[0045] As a supplement: the overall imaging is preferably completed by an area array camera located above the liquid crystal glass substrate under test, and the optical axis of the area array camera is perpendicular to the support platform; the preset illumination conditions preferably include at least one of coaxial surface illumination and lateral strip illumination, wherein the coaxial surface illumination is used to extract the outer contour edge and corner features, and the lateral strip illumination is used to enhance the light and dark transition of the packaging edge, conductive lead-out area and polarizer attachment boundary; after the liquid crystal glass substrate under test enters the detection station, the feeding conveyor stops the conveying, the support platform clamps or vacuum adsorbs the liquid crystal glass substrate under test, and then the overall imaging is triggered.

[0046] The structural reference information is not obtained through direct tomographic measurement, but rather through a joint mapping of the product design layout, the location of the conductive lead-out area, the location of the partition control boundary, the location of the package edge, and the location of the polarizer attachment boundary. The planar projection boundaries of the conductive layer and the liquid crystal functional layer are preferably pre-stored according to the conductive pattern design file or process layout, and then registered with the substrate coordinate system during testing. The corresponding areas of the protective layer or composite layer, and the polarizer or optical attachment layer, are determined based on the relative positions between the attachment boundary and the outer contour edge. The detection state is applied to the conductive layer by the driving power supply through the conductive lead-out area. After each state switch, the liquid crystal glass substrate quality inspection equipment determines whether the current state has entered a stable acquisition window based on the overall brightness change and boundary transition change in the preview image. Only after entering a stable acquisition window is the state registered as the state reference for the corresponding detection area.

[0047] For example, for the central optical detection area, the regional state fingerprint in the power-off state mainly includes the transmission response and the interface continuity response; for the partition boundary detection area, the regional state fingerprint in the power-on state and the partition independent driving state mainly includes the edge transition response; for the conductive connection detection area, the interface continuity response and the edge transition response jointly determine the state reference of the area. After step one is completed, the device obtains a reference table with area number, structural hierarchy mark and state fingerprint. Any subsequent image pointing to a certain detection area and a certain detection state can retrieve the initial structural semantics and initial state semantics of that area.

[0048] In use, firstly, the state reference is no longer limited to text labels, but is compressed into a regional state fingerprint that can be directly referenced in steps two and three; secondly, the basic performance of the liquid crystal glass substrate under different detection states is described through response, edge transition response and interface continuity response; thirdly, the state reference table integrates the substrate coordinate system, structural reference information and detection state into the same data object, so that step one and step two are seamlessly connected.

[0049] Step 2: Based on the unified geometric position, structural hierarchy and working status in Step 1, generate a multi-condition image and a reference image for the same detection area that can be compared laterally, tracked vertically, and directly sent to Step 3 for processing.

[0050] Among these issues, defects related to the polarizer on the liquid crystal glass substrate are not fully exposed under any single imaging condition. Scratches on the polarizer surface typically appear as bright lines in reflected images, while bubbles at the attachment interface typically appear as dark rings with dark edges in transmitted images. Anomalies near the partition control boundary will show inconsistent brightness in adjacent areas under partition-independent driving conditions. If images are taken under only single illumination and single driving conditions, anomalies from different sources are likely to fall into the same grayscale variation, making it difficult to distinguish between surface anomalies, interface anomalies, and functional anomalies even with subsequent registration and differentiation.

[0051] In this process, the liquid crystal glass substrate quality inspection equipment, after calling the inspection area division results from step one, does not capture a series of mixed images of the entire liquid crystal glass substrate under test at once. Instead, it first generates an image acquisition condition sequence for each inspection area. This image acquisition condition sequence consists of five elements: inspection state, illumination method, incident direction, polarization direction, and exposure level. Each element corresponds to the structural reference information of the inspection area. The condition sequence located in the central optical inspection area prioritizes transmission images and images with different polarization directions; the condition sequence located in the edge attachment inspection area prioritizes oblique incident reflection images and dark field images; and the condition sequences located in the conductive connection inspection area and the partition boundary inspection area prioritize transmission images and low-angle reflection images under the independent driving state of the partition. Thus, image acquisition is not simply about increasing the number of images, but about ensuring that each image corresponds to a specific observation task.

[0052] The LCD glass substrate quality inspection equipment calls the first [unclear] in the substrate coordinate system. The boundary of each detection area and read the same area in the first Regional state fingerprint under each detection state Subsequently, the lower transmitted light source, the upper annular reflected light source, the lateral strip oblique light source, the rotatable polarizer holder, and the exposure controller are controlled to operate in a predetermined sequence.

[0053] Taking a dimming glass with longitudinal partition control boundaries as an example, the device first illuminates the lower transmission light source while the power is off, and turns off the upper annular reflection light source and the lateral strip oblique light source to acquire the first transmission image of the central optical detection area; then, keeping the same detection area stationary, the rotatable polarizer holder is rotated to the cross polarization position to acquire the second polarized transmission image of the same area; then, the lower transmission light source is turned off, the upper annular reflection light source is illuminated, and the third reflection image of the detection area is acquired; finally, the lateral strip oblique light source is illuminated for the edge-attached detection area to acquire the edge scattering image against a dark background.

[0054] The above process keeps the liquid crystal glass substrate under test stationary, changing only the optical path and detection state, thereby ensuring that multiple images point to the same physical area. The conditional image obtained in each acquisition is denoted as:

[0055] Among them, conditional images : No. The detection area in the first The detection status and the first Under individual image acquisition conditions, the lateral coordinate of the substrate Longitudinal coordinates of the substrate The image response formed at the location, whose value range is a discrete value within the camera gray quantization interval, serves as the original input for step three; Detection area boundary : No. Each detection region is a closed region in the substrate coordinate system, and its value is a two-dimensional region composed of a set of boundary points; region state fingerprint. Step 1 generates the first The detection area in the first The comprehensive response indicator under each detection state has a non-negative real number value; exposure level. : No. The exposure control value under each image acquisition condition is within the range of the set of discrete exposure levels supported by the device. angle of incidence : No. The angle between the light ray and the normal direction of the liquid crystal glass substrate under each sampling condition is within the range of 0 degrees and less than 90 degrees; polarization direction. : No. The transmission direction of the polarizer under each sampling condition, with a value range of 0 degrees to 180 degrees; Imaging Operator : No. The optical path imaging relationship corresponding to each image acquisition condition is specifically determined by the on / off state of the light source, the incident direction of the light source, the polarization direction, and the exposure level, mapping the same detection area into a conditional image with specific observation semantics. Imaging Operator Corresponding to the A fixed image acquisition link under certain acquisition conditions does not represent an abstract mathematical black box, but rather the imaging process jointly formed by the on / off state of the light source, the incident direction of the light source, the transmission direction of the polarizer, the exposure level, and the camera acquisition settings; among them, when the first... When the image acquisition condition is transmission imaging The exposure is determined by the following factors: the lower transmitted light source is turned on, the upper reflected light source is turned off, the polarizer's transmission direction is fixed, and the camera's exposure setting is fixed. When the image acquisition condition is reflectance imaging The exposure setting is determined by the activation of the upper annular reflective light source, the deactivation of the lower transmissive light source, and the fixed camera exposure level; when the... When the image acquisition condition is polarization imaging It also includes the action of rotating the rotatable polarizer holder to a preset angle. Therefore, The input comes from the device execution state, and the output is the corresponding condition image.

[0056] When used, the conditional image is not an abstract picture, but a collection result jointly defined by the detection area, detection state and optical path conditions, so it has a clear source when compared later; secondly, transmission, reflection, dark field and polarization changes are applied sequentially in the same area, so that defects in different layers are revealed respectively.

[0057] The quality inspection equipment for LCD glass substrates does not employ a free-shooting method; instead, image acquisition is uniformly triggered by a status recording unit. Specifically, after the driving power supply applies a power-off state, a power-on state, or an independent driving state to the conductive lead-out area, the status recording unit first reads the area's status fingerprint. The corresponding initial state is then used to extract the overall brightness and darkness changes, boundary transition changes, and local continuity changes of the detection area from the preview image; only when these three types of changes simultaneously enter the gradual change range will the state recording unit send a valid trigger to the exposure controller and the camera.

[0058] When the liquid crystal functional layer switches from the power-off state to the power-on state, it undergoes a transition from scattering to transmission. If an image is captured during this transition, unstable stripes will be superimposed on the image. Furthermore, the edge-attached detection area is easily affected by short-term reflection drift under oblique incident light. Without stable window control, adjacent condition images of the same detection area will be mixed with state transition traces, causing step three to misjudge the state afterimage as a defect.

[0059] As shown in the example below, when the partition boundary detection area adopts partition-independent driving state image acquisition, the left partition is powered on and off. After the brightness and darkness distribution on both sides of the boundary is no longer expanded, a transmission image and a low-angle reflection image are acquired respectively; then the same partition combination is used. The state recording unit adds a detection area number, detection state number, image acquisition condition number, and trigger time to each conditional image, forming a traceable time sequence chain.

[0060] In use, the image acquisition trigger is uniformly issued by the status recording unit, so that the same detection area can share the same status judgment caliber under different image acquisition conditions; the stable window control reduces the pollution of the condition image by status switching ghosting and short-term reflection drift; the detection status number and the image acquisition condition number are written into the image record at the same time, so that the image can be retrieved in an accurate time sequence when performing registration and difference in step three.

[0061] Furthermore, even if multiple conditional images are obtained, without a corresponding reference object in the same region, state, and condition, step three can only perform intra-image comparisons and cannot determine whether a certain grayscale change originates from the structure itself or from abnormal behavior deviating from the normal range. The composite structure of the liquid crystal glass substrate determines that different detection areas have naturally different background textures and boundary transitions. There are already light-dark transitions near the boundary of the partition control, and there are already edge shadows in the encapsulation transition detection area. If the entire product uses a single standard image as a reference, normal boundary transitions will also be treated as deviations.

[0062] Therefore, the reference object is refined to the first... The first detection area, in the first The detection status, at the... Each layer under a given image acquisition condition forms a reference image set that corresponds one-to-one with the conditional image.

[0063] In this process, the LCD glass substrate quality inspection equipment, after completing multi-condition image acquisition for the same inspection area, incorporates conditional images with the same number from normal samples or normal areas into a reference construction process. This process does not use a single image as a reference; instead, it merges images of multiple batches of normal samples, multiple acquisitions of the same normal sample, or adjacent areas confirmed as normal within the same LCD glass substrate under test, according to the inspection area number, inspection status number, and image acquisition condition number. The merged images are first aligned in the substrate coordinate system, then edge-obstructed segments and unstable image segments are removed, finally generating a reference conditional image. Simultaneously, all conditional images obtained for the current LCD glass substrate under test are encapsulated into an inspection image set, along with region boundaries, structural reference information, inspection status number, and image acquisition condition number. Thus, not only are the original images obtained, but also semantic metadata explaining why the images were captured at this time and place.

[0064] Furthermore, the LCD glass substrate quality inspection equipment does not arbitrarily select one of the normal images after merging. Instead, it first groups the images according to the image acquisition conditions, then layers them according to the inspection area numbers, and performs sorting, truncation, and fusion within the same group and layer.

[0065] Specifically, the device sorts the response values ​​of the same pixel location in multiple normal images from low to high, removes the brightest and darkest edge samples, and then fuses the intermediate samples according to the region center weights to generate a reference condition image. In this system, the center weight is set to a higher value in the central optical detection area, while in the edge-attached detection area, it gradually changes along the direction from the outer contour edge to the polarizer attachment boundary. This ensures that the reference condition image maintains both the normal central uniform area characteristics and the edge band transition. For example, when constructing a reference image for the oblique incident reflection image of the edge-attached detection area, the device first aligns the bright edge positions of multiple batches of normal products in this area, then removes extreme samples affected by a single spot shift, retains intermediate samples for fusion, and finally obtains a reference condition image that contains normal edge bright bands but does not contain random high-brightness spots. For the partition boundary detection area, the reference image sets are established separately in the power-off state, power-on state, and partition-independent driving state, and are not mixed to avoid folding state differences into the background.

[0066] When used, firstly, the reference image set is refined to three dimensions: detection area, detection state, and image acquisition conditions, so that normal structural features are accurately preserved; secondly, sorting and truncating fusion suppresses accidental reflections and local dust interference in single image acquisition.

[0067] For the same detection area, it is preferable to acquire images in the order of transmission image - polarized transmission image - reflection image - dark field image. When performing independent driving state image acquisition, it is preferable to complete all transmission images first, and then complete all reflection images, in order to reduce the impact of frequent optical path switching on registration stability. The construction of the reference image set preferably includes: mapping conditional images acquired by normal samples of the same type under the same detection state and the same image acquisition conditions to the same substrate coordinate system; cropping regional images according to the detection area; removing regional images with occlusion, high brightness saturation, or unstable state; and performing sorting, truncation, fusion or median fusion on the retained images to form reference conditional images corresponding to the detection area, detection state, and image acquisition conditions.

[0068] The LCD glass substrate quality inspection equipment merges and encapsulates all condition images, reference condition images, and corresponding metadata for the same inspection area, and generates a condition deviation for each pair of condition images with the same name.

[0069] The conditional deviation is not simply the average pixel difference, but rather assigns different weights to key areas within the detection region during calculation: for the central optical detection area, the device increases the weight of the central uniform region; for the edge-attached detection area, the device increases the weight near the polarizer attachment boundary; for the partition boundary detection area, the device increases the weight of the narrow band regions on both sides of the boundary. Its expression is:

[0070] Among them, the conditional deviation amount : No. The detection area in the first The detection status and the first Under each image acquisition condition, the weighted deviation relative to the reference image is represented by a non-negative real number. Weighting function : No. The spatial weights of each location within a detection region are continuous functions with values ​​greater than 0, increasing the contribution of key components in deviation calculation; when the detection region is an edge-attached detection region... A higher value is taken near the polarizer attachment boundary; when the detection area is the zone boundary detection area. Take the higher value on both sides of the partition control boundary;

[0071] Among them, the weight gain coefficient Indicates the first The weighting value of key parts in each detection area should be a real number greater than or equal to 0, which is used to increase the contribution of key parts to the deviation of conditions.

[0072] Region mask function Indicates the first Important parts of the detection area are masked with 0 or 1, which are used to mark the central observation zone, the zone adjacent to the polarizer attachment boundary, the zone adjacent to the control boundary, or the zone adjacent to the conductive lead-out area, etc.

[0073] When the When the detection area is the central optical detection area, the region mask function Take one value within the central observation zone and zero values ​​at other locations; when the... When the detection region is an edge-attached detection region, the region mask function Take one within the zone adjacent to the polarizer attachment boundary; when the first... When the detection region is a partition boundary detection region, the region mask function Take one within the narrow band area on both sides of the partition control boundary.

[0074] Conditional Image The original condition image of the current liquid crystal glass substrate under test has the same meaning as described above and provides the response to be tested in the deviation calculation; the reference condition image... A reference image constructed based on normal samples or normal regions, whose value range is the same as that of the conditional image, provides a normal response baseline in deviation calculation; Normalized difference function The normalized difference between two pixel responses is specifically taken as: The numerator represents the absolute difference, while the denominator is used to suppress the amplification and imbalance between bright and low-brightness areas under the same absolute difference, unifying the deviations under different brightness backgrounds to a comparable scale; the substrate's lateral coordinates Longitudinal coordinates of the substrate The meaning is consistent with step one, which is to mark the spatial position of the deviation in the integral operation; For example, when a certain edge-attached detection area is slightly warped, the obliquely incident reflection image and the reference condition image show continuous band-like deviations near the attachment boundary, with a large amount of conditional deviation. When there are small bubbles in the central optical detection area, the transmitted image and the reference condition image show closed-loop deviations at the bubble edge under power-off conditions, and the cross-polarized image shows additional deviations at the same position. The equipment is also encapsulated in the detection image set. Step three no longer needs to start searching for anomalies from scratch; registration refinement and defect enhancement can be performed from the position with the largest conditional deviation.

[0075] When used, the detection image set not only contains the images themselves, but also reference condition images, region boundaries, and state condition metadata, which can be directly called in step three; the condition deviation amount expresses in advance where the abnormality occurs first compared with the normal, thus providing an entry point for background decoupling and candidate region extraction; spatial weights enable different detection regions to generate deviation results according to their physical interest parts, avoiding the masking of local abnormalities after averaging the whole region.

[0076] Step 3: Spatial unification, background stripping, and defect visualization are performed on the detection image set formed in Step 2, so that abnormal responses in the same detection area are separated from the composite optical background and further fall into a clear defect type and layer category.

[0077] Although step two ensures that the LCD glass substrate under test remains as still as possible during image acquisition, slight parallax, edge magnification differences, and brightness field shifts can still occur due to different optical path switching, polarizer rotation, exposure level changes, and state switching. For the edge attachment detection area and the partition boundary detection area, even a small amount of pixel-level shift can misinterpret normal boundaries as abnormal stripes; for the central optical detection area, non-uniform illumination and specular reflection can cut the originally continuous background into multiple bright and dark islands.

[0078] Therefore, we first unify the spatial positions of the images under each condition, and then perform decoupling processing with boundary constraints on the composite optical background so that subsequent defect enhancement is based on the same spatial semantics.

[0079] Among them, the liquid crystal glass substrate quality inspection equipment reads the first image from the inspection image set one by one. The detection area in the first The detection status and the first Conditional images under specific image acquisition conditions Reference condition image Deviation from conditions And call the detection area boundary given in step one. With regional state fingerprint .

[0080] First, using the outer contour edges, corner features, conductive lead-out areas, package edges, partition control boundaries, or polarizer attachment boundaries identified in Step 1 as registration anchor points, different condition images of the same detection area are transformed to the same substrate coordinate system. Then, background constraints are generated on the registered images based on the reference condition images and structural reference information. Finally, low-frequency brightness fields, edge overexposure bands, and large-area specular reflection bands are peeled off layer by layer. This background decoupling is not a blind removal of the background, but rather requires the background model to remain smooth and continuous in the central optical detection area, form a unidirectional transition along the polarizer attachment boundary in the edge attachment detection area, and retain normal brightness and darkness transitions in the partition boundary detection area, thereby avoiding flattening the normal structure itself.

[0081] The liquid crystal glass substrate quality inspection equipment does not directly perform a uniform translation of the entire image, but instead first moves the image along the boundary of the inspection area. Search for the most stable registration anchor point in the region. If the first... If the first detection area is the central optical detection area, then the outer contour edge projection, the fine line of the partition control boundary, and the far-end contour of the conductive lead-out area are preferentially selected as anchor points; if the second detection area is the central optical detection area, then the third detection area is the central optical detection area. If the detection area is an edge attachment detection area, then the polarizer attachment boundary, encapsulation edge and corner features are selected as anchor points.

[0082] Preferably, the image processing host generates anchor point templates on the reference condition image, then performs a layered search on the condition image. First, phase correlation is used to complete coarse alignment of the entire region, then bicubic interpolation local affine fitting is used to complete boundary refinement alignment. Finally, one-dimensional line feature constraints are added to the partition control boundaries to eliminate boundary offset caused by oblique incident light. The registered image is as follows:

[0083] Among them, the registered images : No. The detection area in the first The detection status and the first Under the given image acquisition conditions, the value range of the registered image response is consistent with that of the conditional image; detection region boundary : No. Each detection region is a closed region in the substrate coordinate system, and its value is a two-dimensional boundary point set, restricting the registration process to only be performed within the current detection region; region state fingerprint. Step 1 establishes the first The detection area in the first The comprehensive response identifier under each detection state takes a non-negative real number value, which constrains the interpretation of the bright-dark reversal region when searching the anchor point. Registration operator Regarding the first The first detection area, the first The detection status and the first The hierarchical registration relationship constructed by each mapping condition. The specific implementation consists of phase correlation, local affine fitting, and boundary line constraints. The conditional image is mapped to a unified substrate coordinate system; the registration operator... The preferred implementation is a two-stage approach: global coarse registration and local fine registration. Global coarse registration calculates the translation and rotation amounts based on the corresponding points of the outer contour edge, corner features, conductive lead-out areas, or partition control boundaries. Local fine registration refines and aligns the boundaries by using local affine or spline transformations based on the grayscale transition bands near the polarizer attachment boundary, package edge, or partition control boundary. If the detection area is in the edge attachment detection area, the polarizer attachment boundary is preferred as the primary constraint. If the detection area is in the partition boundary detection area, the partition control boundary is preferred as the primary constraint.

[0084] For example, the partition boundary detection area appears as a wide gray band in the transmission image when power is off, and as a narrow bright line in the reflection image when power is on. The image processing host first uses the intersection of the conductive lead-out area and the partition control boundary for coarse positioning, then aligns along the alternating bright and dark bands on both sides of the partition control boundary, and finally aligns the same partition control boundary simultaneously on both images. If the edge attachment detection area is locally reflective, then the encapsulation edge and the polarizer attachment boundary are used for double boundary constraints to avoid the highlight area for alignment.

[0085] In use, registration anchor points are selected according to the type of detection area to ensure that both the central optical detection area and the edge attached detection area can obtain a stable alignment reference. Layered registration eliminates the translation error of the whole area, the local affine error and the boundary line offset in turn, so that the multi-condition images coincide on the same physical area. The regional state fingerprint written in step two is used to constrain the interpretation of brightness and darkness to avoid misregistration caused by state flipping.

[0086] Furthermore, the LCD glass substrate quality inspection equipment does not employ a full-image uniform filtering method because such filtering would simultaneously weaken both the actual warping signal in the edge-attached detection area and the normal transition signal in the partition boundary detection area. Therefore, the image processing host determines the filtering method based on the boundaries of the detection area. Reference condition image and deviation from conditions Construct a regionalized context.

[0087] Preferably, in the central optical detection area, the background uses a path of curved surface fitting and guided filtering to remove large-scale transmittance fluctuations and eliminate low-frequency illumination fields; in the edge attachment detection area, the background uses strip smoothing along the polarizer attachment boundary direction to retain normal edge bright bands; in the partition boundary detection area, double-sided segmented fitting is used to fit the normal brightness on both sides of the partition-controlled boundary, and the transition groove can be located at the boundary.

[0088] Among them, background image : No. The detection area in the first The detection status and the first Under the given image acquisition conditions, the background response obtained after structural constraint decoupling has a value range consistent with the registered image, representing the normal large-scale optical performance that the region should have under the current conditions; reference condition image. The corresponding reference image constructed in step two provides normal boundary shapes and normal brightness distributions for background fitting; conditional deviation. The weighted deviation from the reference condition image generated in step two is a non-negative real number, which indicates local anomalies that should be retained or avoided during background fitting. Background operator The regionalized background relationship is constructed by combining the registered image, reference condition image, conditional deviation, and detection region boundary. Its specific implementation consists of one or a combination of surface fitting, guided filtering, strip smoothing, or two-sided piecewise fitting, separating the normal structure background from the abnormal foreground. The background operator... The preferred approach is to select different background modeling paths based on the type of detection area: for the central optical detection area, a low-order surface fitting combined with guided filtering is used to generate the background image; for the edge-attached detection area, a one-dimensional strip smoothing along the tangential direction of the polarizer attachment boundary is used to generate the background image; for the partition boundary detection area, a two-sided segmented fitting with the partition control boundary as the dividing line is used to generate the background image; for the encapsulation transition detection area, an edge-preserving smoothing along the encapsulation edge direction is used to generate the background image. After background modeling, locally high-deviation areas are removed from the fitted samples, and the fitting is repeated once to avoid defect areas being involved in the background.

[0089] Subsequently, the image processing host subtracts the background image from the registered image to form the foreground response. The low-energy residues with small absolute values ​​and continuous extension are re-incorporated into the background, while residues exhibiting closed-loop boundaries, thin line protrusions, local dark rings, or offset distributions along partition control boundaries are retained as anomalous foregrounds.

[0090] For example, when there is a local lifting in the edge-attached detection area, the background obtained by strip smoothing still maintains the normal edge bright band, while the lifting area presents an independent raised band in the foreground response; when there is a bubble in the central optical detection area, the curved surface fitted background retains the gradual change in transmittance, and the bubble edge is retained in the form of a closed dark ring.

[0091] When used, background decoupling is constrained by the structure of the detection area, and normal partition boundaries and normal edge bright bands will not be mistakenly removed as anomalies; the conditional deviation is introduced into the background fitting process, so that obviously deviated areas are not averaged back into the background.

[0092] Furthermore, although the abnormal foreground is already apparent, the foreground morphologies of defects from different sources still overlap. Surface scratches appear as thin, bright lines in the reflective image, while bubbles at the attachment interface appear as closed dark rings in the transmission image. Anomalies in the corresponding regions of the conductive layer often spread along the partition control boundaries or conductive lead-out areas. If threshold segmentation is performed based solely on a single foreground response, only a coarse conclusion about the presence of abnormal regions can be obtained, making it difficult to further pinpoint the polarizer surface, attachment interface, corresponding regions of the conductive layer, corresponding regions of the liquid crystal functional layer, or the encapsulation edge regions. Therefore, multi-condition joint enhancement is performed on top of the foreground response, and the enhanced candidate regions are then jointly correlated with the structural reference information from step one and the conditional semantics from step two, thereby forming a layer-interpretable defect conclusion.

[0093] The liquid crystal glass substrate quality inspection equipment calls a set of registered and decoupled foreground responses for the same inspection area, including at least a transmissive foreground, a reflective foreground, and a state-changing foreground. The image processing host first performs directional enhancement based on the foreground morphology: directional filtering enhancement is used for linear remnants, ring edge enhancement is used for closed-loop remnants, and differential enhancement along the boundary normal is used for boundary-adjacent remnants. Next, the image processing host extracts candidate regions from the enhanced image and integrates the response strength, extension direction, boundary proximity, state-changing amplitude, and deviation from the reference condition image of each candidate region in the foreground under various conditions into discriminative features.

[0094] As a supplement, the layer discrimination determines which type of defect it may be in, such as the polarizer surface, the polarizer-substrate bonding interface, the conductive layer corresponding area, the liquid crystal functional layer corresponding area, or the packaging edge area, based on the different reactions of the defect under the detection state and the imaging state, combined with the planar projection correspondence in the structural reference information.

[0095] Finally, based on these discrimination features, the defect type and layer category are given, and candidate areas related to liquid crystal functional layers, conductive layer partitions or attachment interfaces are additionally marked as functionally associated anomalies.

[0096] Furthermore, the image processing host does not directly set a threshold on a single foreground response, but rather semantically combines the foreground responses of the same detection area under multiple image acquisition conditions. Preferably, the transmissive foreground focuses on changes in light transmission and dark ring closure features, the reflective foreground focuses on surface fine lines and bright spot protrusions, the cross-polarized foreground focuses on interface stress stripes and interlayer refraction disturbances, and the foreground in the partitioned independent driving state focuses on the asymmetry of the responses on both sides of the boundary.

[0097] To ensure that these foregrounds are included in the same discrimination criterion, the image processing host constructs candidate saliency values:

[0098] Among them, candidate significance : No. The degree of anomaly significance of each detection region under various detection states and mapping conditions is represented by a non-negative real number, providing a unified spatial saliency map for candidate region extraction; foreground response. : No. The detection area in the first The detection status and the first Anomalies in the foreground under certain image acquisition conditions, carrying anomalous information after the background has been stripped away; Conditional deviation Step two generates a weighted deviation level, which adjusts the confidence level of different foreground responses; regional state fingerprints. Step 1 establishes the state baseline, constraining the interpretation direction of the foreground state change under different detection states; conditional weights. : No. The detection status and the first The contribution ratio of each sampling condition to the joint significance is a real number greater than 0. The influence of key conditions is increased according to the type of detection area. In the central optical detection area, the weights of transmission condition and cross-polarization condition are higher. In the edge-attached detection area, the weights of reflection condition and oblique incidence dark field condition are higher. Enhancement Operator The relationship between foreground response, conditional deviation, and region state fingerprint is fused into a single conditional saliency map. Specific implementations include one or a combination of directional filtering, ring edge enhancement, boundary normal difference, and magnitude normalization, transforming foreground responses from different physical sources to a superimposed scale; the enhancement operator... The specific enhancement method is selected based on the candidate anomaly morphology: for elongated anomalies, directional filtering enhancement is used; for closed-loop anomalies, ring edge response enhancement is used; for boundary-adjacent anomalies, differential enhancement along the boundary normal direction is used; and for state-related anomalies, difference enhancement between foreground responses of different detection states is used. After enhancement, the enhanced responses at the same location in multi-condition images are normalized and fused.

[0099] The image processing host performs connected component extraction on candidate saliency quantities and requires that the candidate regions simultaneously satisfy at least one of the following: minimum extension length, boundary closure, boundary proximity, or consistency of state changes.

[0100] For example, when there is an attachment interface bubble in the central optical detection area, the closed-loop dark band in the transmission foreground and the local perturbation in the cross-polarization foreground are superimposed on the candidate saliency, forming a highly saliency closed-loop region with clear boundaries; when there is a scratch in the edge attachment detection area, the thin bright line in the reflection foreground dominates after directional filtering, while the transmission foreground hardly responds, and finally the linear candidate region is extracted.

[0101] When used, multi-condition joint enhancement can prevent single image thresholding from biasing towards a single class defect; the conditional deviation amount and the region state fingerprint are jointly included in the saliency calculation, so that the information in step one and step two still play a constraining role in step three; the candidate region has dual semantics of morphology and state, which can provide an entry point for type discrimination and layer discrimination.

[0102] Furthermore, the liquid crystal glass substrate quality inspection equipment calculates the response sequence of each candidate region in the transmission foreground, reflection foreground, cross-polarization foreground, and partitioned independent driving state foreground, and calls upon the spatial relationship between the candidate region and the polarizer attachment boundary, partition control boundary, conductive lead-out area, and encapsulation edge.

[0103] Preferably, if the candidate region is elongated and bright in the reflective foreground, has a low response in the transmissive foreground, and has no fixed relationship with the polarizer attachment boundary, it is judged as a scratch on the polarizer surface; if the candidate region is a closed-loop dark band in the transmissive foreground, is accompanied by local disturbance in the cross-polarization foreground, and its position falls in the corresponding area of ​​the polarizer attachment layer, it is judged as an attachment interface bubble or local delamination; if the candidate region extends along the partition control boundary and shows bilateral asymmetric changes in the partition independent driving state foreground, it is judged as an abnormality in the corresponding area of ​​the conductive layer or the corresponding area of ​​the liquid crystal functional layer; if the candidate region is close to the packaging edge and simultaneously exhibits edge protrusion in both the reflective and dark field foregrounds, it is judged as an abnormal packaging edge or edge lifting.

[0104] To ensure a consistent representation of the discrimination results, the image processing host outputs three results for each candidate region: a defect label, a layer label, and a function-related label. The layer label is limited to the polarizer surface, the attachment interface, the corresponding area of ​​the conductive layer, the corresponding area of ​​the liquid crystal functional layer, or the encapsulation edge area. The function-related label is limited to general appearance anomalies or state-related anomalies. For example, if a narrow band appears in the partition boundary detection area, its brightness alternating with the independent driving switch of the partition, and this narrow band extends along the partition control boundary, the image processing host classifies it as an anomaly in the conductive layer corresponding area and marks it as a state-related anomaly. If a short, strip-shaped protrusion parallel to the encapsulation edge appears in the edge attachment detection area, and it is highly significant in the oblique incidence reflection image and remains essentially unchanged during power-on / off switching, the image processing host classifies it as an anomaly in the encapsulation edge area.

[0105] When used, the candidate area no longer stays in the abnormal patch, but falls into a clear defect label and layer label; spatial relationship, condition response and state change are used simultaneously for discrimination, making the boundaries of surface abnormality, interface abnormality and functional abnormality clearer; the functional association label marks the abnormality related to power on / off or independent drive of the zone separately, and establishes a direct basis for the comprehensive quality evaluation and process feedback in step four.

[0106] Step 4: Transform the defect objects obtained in Step 3 into executable quality grading, traceable inspection reports, and implementable process feedback results, so that the inspection chain extends from discovering anomalies to making handling decisions.

[0107] Step three has provided candidate areas, defect labels, layer labels, and functional association labels. However, these results are still intermediate information at the technical interpretation level and do not directly answer the production questions of whether the LCD glass substrate can be released, whether it should be reworked, or whether it should be intercepted. LCD glass substrates are multilayer composite optical components. The short, strip-shaped warping in the edge-attached detection area and the closed-loop bubbles in the central optical detection area have different impact paths on the observer. Although the narrow-band anomalies in the conductive connection detection area are small in area, once associated with changes in power on / off states, their risk is often higher than that of ordinary surface scratches of the same area.

[0108] Therefore, the defective objects are further organized into quality risk quantities, and risk merging and level determination are completed at the detection area level and the whole area level, respectively.

[0109] Among them, the LCD glass substrate quality inspection equipment reads the data from step three. The first detection area For each candidate region, its defect label, layer label, functional association label, candidate region boundary, and boundary with the detection region are retrieved. The relative position of the defects and the strength of the response to the original image evidence are considered. Subsequently, the equipment first calculates the regional quality risk within the detection area, and then assigns different evaluation emphases to the central optical detection area, edge attachment detection area, encapsulation transition detection area, conductive connection detection area, and partition boundary detection area according to the detection area category. The central optical detection area emphasizes light transmission anomalies, scattering anomalies, and visual continuity; the edge attachment detection area emphasizes warping, delamination, and attachment misalignment; and the conductive connection detection area and partition boundary detection area emphasize consistency of state switching and the impact of boundary proximity. The quality grade of the entire liquid crystal glass substrate is not derived from a single maximum defect, nor from the sum of the total area, but rather from the aggregation of the regional quality risk quantities of each detection area according to the area's responsibility.

[0110] LCD glass substrate quality inspection equipment for the first Defect labels are read one by one from each candidate region in the detection area. Layer labels Functional association tags The candidate region area, the main extension length of the candidate region, and the normalized distance from the candidate region to the critical boundary of the detection region are considered. Specifically, the critical boundary of the central optical detection area is the central observation zone and the zoning control boundary; the critical boundary of the edge-attached detection area is the polarizer attachment boundary and the package edge; and the critical boundary of the conductive connection detection area is the conductive lead-out region boundary. Based on these factors, the device constructs the regional quality risk quantity.

[0111] Among them, regional quality risk : No. The overall risk level of a detection area, a non-negative real number, which calculates the impact of multiple candidate areas within the same area to the same evaluation scale; number of candidates. : No. The number of candidate regions confirmed in step three within each detection area, a positive integer, limiting the cumulative range of regional risk quantities; Defect type weight : No. The first detection area The defect label weights corresponding to each candidate region are real numbers with values ​​greater than 0, reflecting the differences in the basic impact of bubbles, foreign objects, scratches, creases, attachment offsets, edge lifting, local delamination, abnormal encapsulation edges, abnormal conductive layer boundaries, inconsistent partition responses, and abnormal local light transmission in the evaluation. Area characterization : No. The first detection area Normalized area of ​​each candidate region, a non-negative real number; length representation. : No. The first detection area The principal extension length of each candidate region, a non-negative real number, represents the elongated shape of scratches, boundary strip anomalies, and anomalies extending along the partition control boundary; long area conversion factor. The conversion factor for length characteristics to area scale is a real number greater than 0. Slender anomalies and blocky anomalies are included in the same formula for comparison. Layer weight : No. The first detection area The layer label weights corresponding to each candidate region are real numbers greater than 0, reflecting the differences in structural risk among the polarizer surface, attachment interface, corresponding region of the conductive layer, corresponding region of the liquid crystal functional layer, and encapsulation edge region; functional weights : No. The first detection area The functional association label weights corresponding to each candidate region are real numbers greater than 0, reflecting the difference between general appearance anomalies and state-related anomalies in the evaluation. Boundary distance : No. The first detection area The normalized distance from each candidate region to the critical boundary of that region, with values ​​ranging from real numbers greater than or equal to 0, improves the significance of anomalies near the polarizer attachment boundary, partition control boundary, conductive lead-out region boundary, and encapsulation edge in the evaluation. For example, when a short, band-like uplift appears in the edge attachment detection area along the polarizer attachment boundary, the length characterization quantity... Distance from boundary It will simultaneously affect the detection area, exhibiting a small area but a high level of quality risk. When a closed-loop bubble appears in the central optical detection area, the area characterization quantity... With functional weight This will increase the risk profile of the region.

[0112] When in use, the regional quality risk level The area, length, layer, and functional information of the candidate region are uniformly converted to avoid evaluation based solely on area or quantity; boundary distance measurement. By directly incorporating the proximity of critical structures into risk representation, anomalies near attachment boundaries, zoning control boundaries, and conductive lead-out regions are fully reflected.

[0113] Furthermore, the LCD glass substrate quality inspection equipment does not determine the overall grade of the substrate based on a single area. Instead, it reweights the quality risk of each area according to the category of the inspection area, and introduces a state consistency penalty to separately increase the grade of anomalies that affect the power-on / off state and the independent driving state of each zone. This results in the overall substrate quality assessment metric:

[0114] Among them, the overall quality assessment quantity The overall assessment result of the current LCD glass substrate under test is a non-negative real number, used for subsequent processing level classification; total number of test areas. The number of detection areas participating in the overall evaluation, with a positive integer value, limits the cumulative range of the overall quality judgment quantity; Regional responsibility weight : No. The weight of each inspection area in the overall evaluation is a real number greater than 0, reflecting the different importance of the central optical inspection area, edge attachment inspection area, packaging transition inspection area, conductive connection inspection area, and partition boundary inspection area in the use of the finished product; regional quality risk level. The meaning remains consistent with the above, indicating that the entire evaluation is being input; State consistency quantity : No. The anomaly intensity related to the power-on / off state or the independent drive state of a zone within a detection area, with values ​​ranging from non-negative real numbers; state penalty weight. : No. State consistency in each detection region The penalty weight in the overall evaluation is a real number greater than 0. During the project implementation, the LCD glass substrate quality inspection equipment pre-stores a grading rule table, which is used to determine the overall quality of the entire substrate. This is mapped to one of four handling levels: good, limited release, reworkable, and defective. Preferably, when the anomaly is mainly concentrated in the edge attachment detection area and the layer label falls on the packaging edge area or attachment interface, the equipment will determine the overall quality. Corresponding to reworkable or limited-release items; when the anomaly falls within the central optical inspection area and is accompanied by a consistency measure. When raised, the equipment will measure the overall quality of the piece. Corresponding to defective products.

[0115] For example, if a closed-loop bubble is detected in the central optical detection area of ​​a liquid crystal glass substrate, and a narrow band abnormality related to the independent driving state of the partition is detected in the partition boundary detection area, the device will directly give a defective product conclusion on the display interface and output a sorting signal to the downstream interception mechanism; if a local edge warping is detected only in the encapsulation transition detection area of ​​another liquid crystal glass substrate, the device will give a reworkable product conclusion and guide the sample to the rework buffer station.

[0116] When using it, the overall quality assessment quantity Incorporating both regional risks and state consistency highlights state-related anomalies that directly impact usability; regional responsibility weights. This allows different detection areas to participate in the overall grading according to their actual responsibilities, avoiding the situation where minor edge anomalies and central optical anomalies are treated the same.

[0117] As a supplement: the quality grading rules are preferably pre-stored in the liquid crystal glass substrate quality inspection equipment and are jointly determined based on the inspection area category, defect label, layer label, functional association label, regional quality risk level and overall quality judgment level; among them, candidate areas located in the central optical inspection area and whose functional association label is a state-related abnormality are preferably directly upgraded to the overall risk level; candidate areas located in the edge attachment inspection area or packaging transition inspection area and whose layer label is an attachment interface or packaging edge area are preferably included in the rework judgment path.

[0118] Furthermore, while overall grading can support sorting, if the grading conclusion is not linked to specific candidate areas, specific image evidence, and specific process origins, the manufacturing end cannot know why it was intercepted, which process step should be investigated, or if there is an entry point for rework. The sources of anomalies in LCD glass substrates span the bonding process, conductive layer formation process, partition etching process, encapsulation process, and electric drive debugging process; a single pass or fail mark cannot support subsequent processing.

[0119] Therefore, the entire grade is expanded in reverse to the report field and process source field, forming a result object that can be read by quality personnel and called by the manufacturing system.

[0120] Among them, the quality inspection equipment for liquid crystal glass substrates obtains the overall quality judgment value. After determining the treatment level, first record the sample identifier, time identifier, substrate coordinate system version number, detection area division results, and regional quality risk of each detection area. Overall quality assessment Defect labels, layer labels, and functional association labels are centrally organized. Then, original image evidence, the foreground response image from step three, and the defect contour marker image are extracted from the inspection image set to generate an inspection report bound to the sample. Subsequently, based on the inspection area location, layer labels, functional association labels, and the prior association between defect type and process, the equipment calculates the attribution strength of each candidate process source, and uses the process with the highest attribution strength as the primary feedback object, while retaining secondary feedback objects, and writes this information into the manufacturing database and process dashboard.

[0121] The LCD glass substrate quality inspection equipment generates a unique report record for each LCD glass substrate under test, and encapsulates the fields in the following order: sample information - area information - defect information - evidence information - handling information. Sample information includes sample code, batch number, inspection time, and inspection station number; area information includes the boundaries of five types of inspection areas and the area quality risk level. Consistency quantity of state Defect information includes defect labels, layer labels, functional association labels, and substrate lateral and longitudinal coordinates for each candidate region; evidence information includes the original condition image number, foreground response image number, and marked image number with the defect outline superimposed; handling information includes overall wafer quality assessment. The report specifies the treatment level and subsequent destination. Preferably, after the inspection report is generated by the image processing host, one copy is written to the manufacturing database in the form of a structured field file, and the other copy is displayed on the workstation screen in the form of a graphic page for operators to view.

[0122] For example, when the equipment determines that a sample is reworkable, the workstation screen will display the sample code, treatment level, and a thumbnail of the defect location. Downstream rework stations can directly retrieve the same inspection report. Anomalies are displayed in the attachment inspection area, the layer label is the attachment interface, and the main defect label is edge lifting.

[0123] When in use, the test report links all the results of steps one and four onto the same record object; image evidence and treatment level are bound on the same page, allowing on-site personnel to match conclusion, location, and image; structured field files and graphic pages are displayed in parallel, and test results can be read by both equipment and personnel.

[0124] Furthermore, the liquid crystal glass substrate quality inspection equipment pre-stores a process association table, which is established by the applicant during the sample development stage based on defect labels, layer labels, inspection area categories, and functional association labels. The process association table is preferably established from known defective samples, rework samples, and manual verification results from the trial production stage, and is continuously updated during production based on rework confirmation results. The process association table records at least the relationships between defect labels, layer labels, inspection area categories, and the bonding process, conductive layer formation process, zone etching process, encapsulation process, and electric drive debugging process.

[0125] For example, attachment misalignment, edge lifting, and localized delamination are highly correlated with the attachment process; conductive layer boundary anomalies are highly correlated with the conductive layer formation process and the partition etching process; package edge anomalies are highly correlated with the packaging process; and state-related anomalies falling within the conductive connection detection area or the partition boundary detection area are highly correlated with the electric drive debugging process. The process attribution quantity is calculated based on the process correlation table.

[0126] Among them, the amount of process attribution : No. The attribution strength between each candidate process and all anomalies in the current sample, a non-negative real number, is used to determine the primary and secondary feedback processes; process number. : One of the following processes: bonding process, conductive layer formation process, partition etching process, packaging process, or electric drive debugging process, with a discrete number; Correlation coefficient : No. The candidate process and the first The first detection area Defect labels for each candidate region and layer labels The correlation strength between them, taking values ​​that are real numbers greater than or equal to 0; regional coupling coefficient. : No. The candidate process and the first The degree of spatial coupling between detection regions, taking values ​​that are real numbers greater than or equal to 0; Candidate contribution : No. The first detection area The contribution of each candidate region to the process attribution is represented by a non-negative real number, and a regional quality risk factor is comprehensively introduced. The impact of functional association tags and boundary proximity relationships on the determination of process origin;

[0127] Among them, candidate contribution Indicates the first The first detection area The contribution of each candidate region to the process source determination is represented by a non-negative real number. Its function is to uniformly calculate the regional quality risk, functional correlation, and boundary proximity into the process feedback loop. (Regional quality risk) Functional weight Distance to boundary The meaning is the same as described above.

[0128] Correlation coefficient The process association table is derived from the trial sample, the rework sample and the results of manual review. It records the correspondence between defect labels, layer labels and the attachment process, conductive layer formation process, partition etching process, packaging process and electric drive debugging process.

[0129] Regional coupling coefficient The spatial distribution patterns of anomalies that are prone to occur in different detection areas for each candidate process are predetermined.

[0130] When the process is assigned quantity After the calculation is completed, the LCD glass substrate quality inspection equipment selects the candidate process with the largest number of belongings as the main feedback process, and writes the process, along with the relevant inspection area, defect label and image evidence, into the manufacturing database, and sends the feedback record to the corresponding process board.

[0131] For example, if the anomalies of a sample are mainly concentrated in the edge attachment detection area and the encapsulation transition detection area, and the layer labels mainly fall on the attachment interface and the encapsulation edge area, then the equipment will treat the attachment process and the encapsulation sequence as feedback objects. If the anomalies are continuously distributed along the partition control boundary, and the function-related labels are status-related anomalies, then the equipment will treat the partition etching process and the electric drive debugging sequence as feedback objects. The equipment displays two results on the workstation screen: the primary feedback process and the secondary feedback process, which manufacturing personnel will then use for on-site verification.

[0132] When used, the process allocation quantity Defect labels, layer labels, and inspection area categories are mapped to specific processes to avoid manual judgment of the source of each piece based on experience; the main feedback process and the secondary feedback process output simultaneously, enabling complex anomalies to have a multi-process collaborative investigation entry point; inspection reports, sorting results, and process feedback share the same data object, forming a closed-loop link from inspection to manufacturing disposal.

[0133] Please see Figures 1-7 This invention provides a quality inspection device for polarizers on liquid crystal glass substrates based on image processing, comprising: a support platform, an overall imaging mechanism, a state application mechanism, a multi-condition image acquisition mechanism, a memory, and an image processor; the support platform supports the liquid crystal glass substrate under test, the overall imaging mechanism images the entire liquid crystal glass substrate under test, the state application mechanism applies at least two detection states to the liquid crystal glass substrate under test, the multi-condition image acquisition mechanism acquires at least two types of images from the following categories for the same detection area under different detection states: transmission image, reflection image, bright field image, dark field image, and images with different polarization directions; the memory stores a reference image set, a detection image set, and program instructions; the image processor is integrated with the overall imaging mechanism. The system includes a mechanism, a state application mechanism, a multi-condition image acquisition mechanism, and a memory connection, and is configured to execute program instructions to: extract at least one feature from the outer contour edge, package edge, conductive lead-out area, partition control boundary, and polarizer attachment boundary; establish a substrate coordinate system; divide the central optical detection area, edge attachment detection area, conductive connection detection area, and partition boundary detection area; and establish structural reference information and state reference; construct a reference image set and a detection image set; perform spatial registration and background decoupling on the detection image set; extract candidate defect regions based on response differences under different detection states and optical conditions; determine defect type and layer; and output quality evaluation results based on defect type, layer, and location information.

[0134] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0135] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0136] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0137] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0138] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for quality inspection of polarizers on liquid crystal glass substrates based on image processing, characterized in that: include, Step 1: Imaging the entire LCD glass substrate under test, extracting at least one feature from the outer contour edge, encapsulation edge, conductive lead-out area, partition control boundary, and polarizer attachment boundary, establishing a substrate coordinate system, dividing the central optical detection area, edge attachment detection area, conductive connection detection area, and partition boundary detection area, and establishing structural reference information and state reference under at least two detection states; Step 2: Acquiring at least two types of images from the same detection area under different detection states, including transmission images, reflection images, bright field images, dark field images, and images with different polarization directions, constructing a reference image set and a detection image set; Step 3: Spatial registration and background decoupling of the detection image set, and extracting candidate defect regions based on the response differences under different detection states and optical conditions, identifying defect types and layers; Step 4: Outputting quality evaluation results based on defect type, layer, and location information.

2. The method for quality inspection of polarizers on liquid crystal glass substrates according to claim 1, characterized in that: Step one establishes a substrate coordinate system, including determining the relative positions of the outer contour edge, package edge, conductive lead-out area, partition control boundary and polarizer attachment boundary based on the overall imaging results, and dividing the central optical detection area, edge attachment detection area, conductive connection detection area and partition boundary detection area according to the relative positions.

3. The method for quality inspection of polarizers on liquid crystal glass substrates according to claim 2, characterized in that: Step one establishes structural reference information and state reference, including establishing the positional correspondence between the glass substrate layer, the corresponding area of ​​the conductive layer, the corresponding area of ​​the liquid crystal functional layer and the corresponding area of ​​the polarizer attachment layer and the detection area, and recording the state reference of each detection area in the power-off state, the power-on state and the independent driving state of each zone.

4. The method for quality inspection of polarizers on liquid crystal glass substrates according to claim 3, characterized in that: Step two involves acquiring images, including acquiring transmission images, images with different polarization directions, and reflection images sequentially according to a preset time sequence while maintaining the same spatial position within the same detection area. Additionally, dark field images are acquired in the detection state corresponding to the partition boundary detection area to form a detection image corresponding to the detection area.

5. The method for detecting the quality of polarizers on liquid crystal glass substrates according to claim 4, characterized in that: Step two involves constructing a reference image set, which includes classifying and storing normal sample images according to detection area, detection state, and image category, and writing normal sample images that are consistent with the current detection area, current detection state, and current image category into the reference image set as reference condition images.

6. The method for detecting the quality of polarizers on liquid crystal glass substrates according to claim 5, characterized in that: In step three, spatial registration is performed, including global alignment and local boundary alignment of each image in the detection image set based on the corresponding reference features in the outer contour edge, conductive lead-out area, partition control boundary and polarizer attachment boundary, so that each image is uniformly mapped to the same detection area in the substrate coordinate system.

7. The method for detecting the quality of polarizers on liquid crystal glass substrates according to claim 6, characterized in that: Step 3 involves background decoupling, which includes performing low-frequency brightness field correction, specular reflection suppression, strip smoothing along the polarizer attachment boundary, and bilateral background fitting along the partition control boundary on the registered image according to the category of the detection area. The decoupled abnormal foreground is then used as input for extracting candidate defect regions.

8. The method for quality inspection of polarizers on liquid crystal glass substrates according to claim 7, characterized in that: Step three involves identifying the defect type and layer, including combining the response differences of the defect candidate region in the transmission image, reflection image, and images under different detection states, as well as the spatial relationship between the defect candidate region and the polarizer attachment boundary, partition control boundary, and conductive lead-out area, to determine the defect type and layer.

9. The method for quality inspection of polarizers on liquid crystal glass substrates according to claim 8, characterized in that: Step four outputs quality evaluation results, including regional and whole-piece evaluations of the liquid crystal glass substrate under test by combining defect type, layer, location information, detection area category and response consistency under detection conditions, and outputs one of the following evaluation results: good product, limited release product, reworkable product and defective product.

10. A device for detecting the quality of polarizers on liquid crystal glass substrates based on image processing, characterized in that: include: The system comprises a support platform, an overall imaging mechanism, a status application mechanism, a multi-condition image acquisition mechanism, a memory, and an image processor. The support platform is used to support the liquid crystal glass substrate under test. The overall imaging mechanism is used to image the liquid crystal glass substrate under test as a whole. The state application mechanism is used to apply at least two detection states to the liquid crystal glass substrate under test. The multi-condition image acquisition mechanism is used to acquire at least two types of images from the following categories for the same detection area under different detection states: transmission image, reflection image, bright field image, dark field image, and images with different polarization directions. The memory is used to store the reference image set, the detection image set, and the program instructions. The image processor is connected to the overall imaging mechanism, the state application mechanism, the multi-condition image acquisition mechanism, and the memory, and is configured to execute program instructions to: extract at least one feature from the outer contour edge, the encapsulation edge, the conductive lead-out area, the partition control boundary, and the polarizer attachment boundary; establish a substrate coordinate system; divide the central optical detection area, the edge attachment detection area, the conductive connection detection area, and the partition boundary detection area; and establish structural reference information and state reference. Construct a reference image set and a detection image set; perform spatial registration and background decoupling on the detection image set, and extract candidate defect regions based on response differences under different detection states and optical conditions to determine defect types and layers; Output quality evaluation results based on defect type, layer, and location information.

Images