A Method for Detecting Surface Defects in Alumina Plates Based on Image Recognition

Abstract

This invention discloses an image recognition-based method for detecting surface defects in alumina plates, relating to the field of image processing technology. It addresses the problem of misjudging defects under reflective interference. The invention constructs a detection process that includes multi-illumination imaging, co-domain registration, reflective suppression, texture decomposition, directional response enhancement, and false defect removal. This process effectively distinguishes between specular reflections, processing textures, and real defects, improving the visibility of anomalies such as scratches, pits, and blemishes, as well as the accuracy of boundary extraction. It outputs defect types, boundary information, and plate distribution results, facilitating quality traceability and process analysis, thereby improving the accuracy, stability, and engineering application effectiveness of alumina plate surface defect detection.

Description

Technical Field

[0001] This invention relates to the field of thin image processing technology, and more specifically, to a method for detecting surface defects in alumina plates based on image recognition. Background Technology

[0002] In the field of electronic packaging substrate manufacturing, image recognition-based surface defect detection technology for alumina plates has been widely used. This technology is usually set up in an online inspection station after the alumina plate has been sintered, ground and polished and before it enters the metallization or thin film circuit fabrication process. The plate surface is continuously acquired by an industrial camera, lighting device and image processing program, and grayscale analysis, edge extraction, texture recognition and abnormal area determination are performed on the acquired images to identify surface defects such as scratches, pits, blemishes and particle adhesion.

[0003] However, after grinding and polishing, the surface of anodized aluminum sheets is relatively flat and has obvious local reflections, which can easily form bright reflective areas under certain lighting conditions. At the same time, the surface of the sheet is often accompanied by processing textures, local gray-scale gradual changes and fine structural undulations, which make real defects and background textures mixed together in the image. Since the reflective response, processing texture and weak defects have similar appearances in a single image, existing methods are prone to misjudging reflective residues or texture continuations as defects, and can also easily submerge real anomalies such as shallow scratches and small pits in the background. This leads to inaccurate extraction of defect boundaries, unstable type identification and distortion of the sheet distribution results, which in turn affects the reliability of subsequent sorting, quality traceability and process analysis of anodized aluminum sheets. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the following solution is proposed to solve the problem of easy misjudgment of defects under reflective interference in the above-mentioned background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for detecting surface defects in alumina plates based on image recognition includes the following steps: Surface images of the alumina plate under different lighting conditions are acquired, and distortion correction, viewing angle correction and grayscale normalization are performed to generate a group of images in the same domain. Reflective regions are identified based on the brightness order, saturation distribution, and neighborhood continuity of corresponding pixels in a co-domain image group, and reflective suppression images are generated by combining boundary texture extension. Background texture decomposition is performed on the reflection suppression image to construct an abnormal response layer, and directional response maps are generated according to the differences in row, column and diagonal responses. Based on the directional response map, enhancement, segmentation, connectivity extraction, and boundary refinement are performed on linear anomalies, point anomalies, and patch anomalies, respectively, to generate a set of candidate defect regions; Based on the directional continuity, grayscale transition consistency, contour closure relationship, and neighborhood coordination relationship of the candidate defect region, false defects are eliminated, and the defect type identifier, defect boundary information, and board surface defect distribution results are output.

[0006] Furthermore, surface images of the alumina plate under inspection were acquired under different lighting conditions, including: Determine the target acquisition area of ​​the surface to be inspected, and set up lighting conditions with different incident directions or different lighting modes around the target acquisition area; Surface images of the target acquisition area are acquired under various lighting conditions to obtain a set of original surface images corresponding to the same target acquisition area; Each original surface image is associated with the target acquisition area identifier and written into the image entry to form an input image set for generating a co-domain image group.

[0007] Furthermore, distortion correction, viewpoint correction, and grayscale normalization are performed to generate a group of images in the same domain, including: A primary surface image is selected as the reference image, and its edge contour features and surface texture localization features are extracted. Based on edge contour features and surface texture localization features, geometric distortion correction, position offset correction and viewpoint correction are sequentially performed on the remaining original surface images to generate a corrected image that corresponds to the reference image space. Gray-level normalization processing is performed on the reference image and each calibration image to obtain a group of images in the same domain with consistent pixel positions and gray-level scales.

[0008] Furthermore, reflective regions are identified based on the brightness order relationship, saturation distribution relationship, and neighborhood continuity relationship of corresponding pixels in the same image group, including: Extract the brightness values ​​of each pixel in the same image group under different lighting conditions, construct a pixel brightness sequence according to the order of brightness change, and determine the brightness order relationship based on the pixel brightness sequence; The spatial clustering and boundary expansion states of saturated pixels in a group of images in the same domain are statistically analyzed to generate saturated distribution results. Based on the brightness order relationship and saturation distribution results, the reflective candidate region is determined by combining the grayscale continuity state and boundary connectivity state in the neighborhood of saturated pixels. Then, the reflective candidate region is connected and merged and the boundary is corrected to generate the reflective region.

[0009] Furthermore, a reflection-suppressed image is generated by combining boundary texture extension, including: Extract continuous texture fragments outside the boundary of the reflective area, and determine the texture extension direction based on the consistency of texture direction and the continuity of grayscale transition; Map continuous texture fragments along the texture extension direction to the interior of the reflective area to generate a compensation texture for the reflective area. The reflective area is subjected to brightness correction and boundary smoothing based on the compensated texture, and then stitched with the non-reflective image area outside the reflective area to generate a reflection-suppressed image.

[0010] Furthermore, background texture decomposition is performed on the reflection suppression image to construct an abnormal response layer, and directional response maps are generated according to the differences in row, column, and diagonal responses, including: Local texture statistics and region structure analysis are performed on the reflection-suppressed image to extract the periodic distribution results of the surface background texture and the gray-level change baseline; Based on the periodic distribution results and the grayscale change baseline, the reflection suppression image is decomposed into a background texture layer and an anomaly response layer. The background texture layer is used to characterize the stable texture structure of the alumina plate surface, and the anomaly response layer is used to characterize local anomaly changes that deviate from the background texture structure. Extract the grayscale variation amplitude, edge extension state, and local response density state of the abnormal response layer in the row, column, and diagonal directions respectively, and generate the corresponding response results in each direction; Based on the response results in each direction, the dominant response direction of the abnormal response layer is determined, and a directional response map is generated.

[0011] Furthermore, based on the directional response map, enhancement, segmentation, connectivity extraction, and boundary refinement are performed on linear anomalies, point anomalies, and patch anomalies, respectively, to generate a candidate defect region set, including: Based on the directional extension characteristics, local clustering characteristics, and boundary expansion characteristics of each abnormal region in the directional response map, the abnormal regions are divided into linear abnormal regions, point-like abnormal regions, and patchy abnormal regions. For linear anomalous regions, continuous enhancement and break-and-repair processing along the dominant response direction are performed; for point-like anomalous regions, local abrupt enhancement and isolated response highlighting processing are performed; and for patchy anomalous regions, regional closure enhancement and boundary expansion suppression processing are performed to generate anomalous enhancement results. The abnormal enhancement results are sequentially subjected to threshold segmentation, connected component extraction and boundary refinement to obtain candidate defect regions corresponding to each type of abnormality; The candidate defect regions corresponding to each type of anomaly are merged to generate a candidate defect region set.

[0012] Furthermore, false defects are eliminated based on the directional continuity, grayscale transition consistency, contour closure relationship, and neighborhood coordination relationship of the candidate defect region, including: Extract the main direction extension state, gray level change state on both sides of the boundary, region contour closure state, and neighborhood texture coordination state of each candidate defect region. Based on the main direction extension state, determine whether the candidate defect area maintains a continuous extension relationship with the surface processing texture; based on the gray change state on both sides of the boundary, determine whether the candidate defect area belongs to the reflective residual area; based on the region contour closure state and the neighborhood texture coordination state, determine whether the candidate defect area belongs to the random noise area. Candidate defect regions that satisfy the relationships of continuous extension of processing texture, residual reflection, or random noise are marked as pseudo-defect regions. After removing pseudo-defect regions, a set of target defect regions is generated.

[0013] Furthermore, the output includes defect type identifiers, defect boundary information, and board surface defect distribution results, including: Extract the center position, boundary contour, extension direction, length parameter, area parameter, and gray-scale variation characteristics of each target defect region in the target defect region concentration; Defect description entries are constructed based on center location, boundary contour, extension direction, length parameter, area parameter, and regional grayscale variation characteristics. The defect description entries are matched with the preset defect discrimination rule set to determine the defect type identifier corresponding to each target defect area; The defect distribution results of the board surface are generated based on the positional distribution relationship of each target defect area on the surface of the alumina board to be inspected.

[0014] Furthermore, the distribution results of defects on the board surface are generated, including: The center position of each target defect area is mapped to the plate coordinate position of the alumina plate to be inspected, forming a set of defect positions; A defect distribution map of the board surface is established based on the defect location set and defect type identifier; Write the defect type identifier, boundary contour, plate surface coordinate position, and plate surface defect distribution map into the inspection result entries; Output the test results entries and create defect test records associated with the corresponding alumina plate identifier.

[0015] The technical effects and advantages of the image recognition-based method for detecting surface defects in alumina plates according to this invention are as follows: This invention constructs a closed-loop detection process for the surface of anodized alumina plates, including multi-illumination imaging, co-domain registration, reflection recognition and suppression, background texture decomposition, directional response enhancement, and false defect removal. This process achieves stable identification and accurate localization of surface defects such as scratches, pits, and blemishes. By utilizing the brightness sequence and saturation distribution relationships under different illumination conditions, it distinguishes between specular reflections and real anomalies, significantly reducing false and missed detections caused by residual reflections, processing texture continuation, and random noise. Furthermore, based on differentiated processing strategies for linear, dotted, and patchy anomalies, it improves the visibility of weak defects and the accuracy of boundary extraction. On this basis, it further outputs defect type identification, boundary information, and plate distribution results, facilitating quality traceability and process analysis, thereby improving the accuracy, stability, and engineering applicability of anodized alumina plate surface inspection. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of an image recognition-based method for detecting surface defects in alumina plates according to the present invention. Detailed Implementation

[0017] 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.

[0018] In order to achieve the above objectives, Figure 1 A schematic diagram of the structure of an image recognition-based method for detecting surface defects in alumina plates is provided, which specifically includes the following steps; Surface images of the alumina plate under different lighting conditions are acquired, and distortion correction, viewing angle correction and grayscale normalization are performed to generate a group of images in the same domain. Reflective regions are identified based on the brightness order, saturation distribution, and neighborhood continuity of corresponding pixels in a co-domain image group, and reflective suppression images are generated by combining boundary texture extension. Background texture decomposition is performed on the reflection suppression image to construct an abnormal response layer, and directional response maps are generated according to the differences in row, column and diagonal responses. Based on the directional response map, enhancement, segmentation, connectivity extraction, and boundary refinement are performed on linear anomalies, point anomalies, and patch anomalies, respectively, to generate a set of candidate defect regions; Based on the directional continuity, grayscale transition consistency, contour closure relationship, and neighborhood coordination relationship of the candidate defect region, false defects are eliminated, and the defect type identifier, defect boundary information, and board surface defect distribution results are output.

[0019] This embodiment limits the application scenario to the online surface inspection station of polished alumina ceramic substrates for electronic packaging before entering the thin film circuit fabrication. Alumina material has long been used as a substrate and packaging material in the electronics field. The surface of the polished ceramic substrate is relatively flat, and thin film processes can be carried out on the polished surface. Therefore, defect screening of its surface before entering the subsequent process has a clear industrial application basis. Meanwhile, for flat and highly reflective test surfaces, machine vision inspection usually requires combining different lighting methods such as vertical coaxial illumination and low-angle illumination to obtain responses to uniform backgrounds and minor surface defects, respectively.

[0020] Surface images of the alumina plate under inspection are acquired under different lighting conditions. Distortion correction, viewing angle correction, and grayscale normalization are performed to generate a group of images in the same domain. The specific implementation steps include: In this embodiment, the object to be tested is an alumina plate that has been sintered, ground and polished. After the alumina plate is sent into the testing station by the conveying mechanism, the polished surface is used as the surface to be tested. A fixed support platform, an image acquisition camera and an illumination assembly are set on the testing station. The image acquisition camera is located above the surface to be tested, and the illumination assembly is arranged around the image acquisition camera. The surface to be inspected is first divided into several target acquisition areas according to the coordinates of the board surface. Each target acquisition area is assigned a unique area identifier, which includes at least the alumina board number, the row and column position of the area, and the acquisition order information.

[0021] For each target acquisition area, different lighting conditions are set around the target acquisition area. The lighting conditions include at least one set of forward lighting conditions and two sets of oblique lighting conditions. The forward lighting conditions are formed by coaxial illumination incident along the normal direction of the surface to be inspected, which is used to obtain the overall outline of the target acquisition area and a relatively uniform background texture image. The first set of oblique lighting conditions is formed by strip lighting incident at a low angle to the surface to be inspected, used to enhance fine scratches, drag marks and local undulation boundaries extending along the first direction; The second set of oblique illumination conditions is formed by low-angle illumination with a different incident direction than the first set of oblique illumination, and is used to enhance the anomalous response extending in another direction; The above lighting conditions are set because a flat, highly reflective surface is more likely to obtain a uniform reflective background under vertical lighting, while low-angle lighting is more likely to show edge damage, surface undulations and minor scratches as differences in brightness, thus providing a basis for identifying and separating reflective interference from real defects.

[0022] After completing the lighting conditions settings, imaging is triggered sequentially for the same target acquisition area according to the preset acquisition sequence, and the original surface images of the target acquisition area under each lighting condition are obtained to form a group of original surface images. Each time an original surface image is acquired, an image entry is generated. The image entry contains the alumina plate number, target acquisition area identifier, lighting condition identifier, acquisition time sequence identifier, and corresponding image data. In this embodiment, the image data is stored in the form of a grayscale image matrix. Each pixel position in the grayscale image matrix corresponds to the imaging response of a local surface position within the target acquisition area. After imaging is completed under all illumination conditions for the same target acquisition area, the corresponding multiple image entries are merged into the input image set of the target acquisition area.

[0023] When generating a set of images from the same domain, a primary surface image is first selected from the input image set as a reference image, specifically including: The reference image is preferably selected as an original surface image with less overall reflective area, complete boundary of target acquisition area and relatively stable texture transition. After determining the reference image, edge contour features and surface texture localization features are extracted from the reference image. Edge contour features are used to characterize the outer perimeter boundary, boundary turning position and local edge direction of the target acquisition area. Surface texture localization features are used to characterize relatively stable texture change points, local gray-level inflection points and microstructure positions within the target acquisition area. By simultaneously extracting edge contour features and surface texture localization features, it is possible to ensure both overall position matching and local position matching within the area, thereby establishing a stable spatial correspondence between each original surface image and the reference image. For the remaining original surface images, geometric distortion correction is first performed. In this embodiment, geometric distortion correction uses the lens distortion correction parameters obtained by pre-calibration at the detection station to correct the edge curvature and proportional stretching in the original surface images, so that each original surface image returns to a unified reference state in terms of geometric imaging relationship. After geometric distortion correction is completed, position offset correction and viewpoint correction are performed. Position offset correction determines the translation offset based on the correspondence between the edge contour features of the reference image and the image to be corrected, so that the image to be corrected is aligned with the reference image in the overall position of the region. Viewpoint correction corrects projection differences caused by minute changes in the board's pose or camera installation deviations based on the local mapping relationship between surface texture positioning features, aligning the image to be corrected with the reference image in the texture position within the region. After the above processing, a corrected image corresponding to the reference image space is obtained.

[0024] In the gray-level normalization process, a background texture area without obvious reflection or anomaly is first selected in the reference image as the gray-level reference area. Then, the reference pixel set corresponding to the gray-level reference area is located in each calibration image. Subsequently, according to the gray-level distribution relationship of the reference pixel set, the gray-level scale is adjusted for each calibration image so that the gray-level fluctuation range of each calibration image in the background area is consistent with that of the reference image. For regions with high grayscale but not yet saturated, the grayscale distribution is restored by interval compression. For regions with low grayscale, an interval stretching method is used to improve their correspondence with the reference image; For texture edges that have been identified as stable background regions, their relative grayscale levels are preserved. After grayscale normalization, the corrected images obtained under different lighting conditions are comparable in grayscale, while preserving the original contrast relationship of the board surface texture and local anomalies.

[0025] After the above geometric distortion correction, position offset correction, viewpoint correction and grayscale normalization processing, a group of images in the same domain is generated. A co-occurrence image group refers to an image group in which any pixel position corresponds to the same physical surface position within the target acquisition area, but the brightness values ​​are different under different lighting conditions. In other words, the pixel comparison in a co-occurrence image group is based on the actual co-located surface positions, rather than on adjacent or misaligned pixels.

[0026] Here is a specific example: When inspecting a polished alumina board used for thin-film circuit fabrication, a target acquisition area is first selected in the middle of the board surface. Three images are obtained by using vertical coaxial illumination, left low-angle illumination, and right low-angle illumination respectively. The image with the most uniform background is selected as the reference image. Geometric distortion correction, position offset correction, and viewing angle correction are performed on the other two images in sequence. Then, the stable background area in the reference image is used as the grayscale reference area, and grayscale normalization is performed on the two corrected images. After processing, the three images correspond to each other in pixel position and are comparable in grayscale scale, thus forming a co-domain image group. This co-domain image group is then input into the processing flow to identify the brightness response differences of the same surface position under different lighting conditions, and further complete the identification of reflective areas and the display of real defects.

[0027] Reflective regions are identified based on the brightness order, saturation distribution, and neighborhood continuity of corresponding pixels in a co-domain image group. A reflection-suppressed image is then generated by combining this with boundary texture extension. Specific implementation steps include: In this embodiment, the same target acquisition area is first obtained by the aforementioned steps. The images in the same target acquisition area have already corresponded in spatial position. Therefore, the same pixel position can be regarded as the same physical surface point on the surface of the corresponding alumina plate to be inspected. Using pixel positions as units, extract the brightness value of each pixel in the corresponding image under various lighting conditions, and construct a pixel brightness sequence according to the order of brightness value change from low to high. For example, for any pixel in the target acquisition area, if its brightness is low under front lighting, significantly increases under left oblique incident lighting, and continues to be bright under right oblique incident lighting, then the corresponding brightness change sequence can be formed accordingly. The order of brightness changes is written into the position record of the pixel as the basis for subsequent determination of the brightness order relationship. The brightness order relationship is to compare the brightness response order of the same physical location under different illumination effects, and is used to reflect the reflection sensitivity of the location to light from different incident directions.

[0028] After establishing the brightness order relationship, the saturated pixels in the same domain image group are statistically analyzed. In the specific processing, the pixels that reach the preset saturation threshold are first detected in each same domain image. Then, these pixels are mapped back to the same plate coordinate system, and their spatial aggregation state and boundary expansion state are statistically analyzed. The saturation threshold is determined according to the upper limit of the gray output of the image acquisition camera. It is preferred to take the gray value close to the saturation output range of the sensor as the upper limit of the judgment. When the background brightness in the same target acquisition area is generally high, the high gray tail of the background gray distribution in the target acquisition area is used as a dynamic reference to adjust the saturation threshold in the same direction to avoid misjudging normal bright textures as saturated pixels. If several saturated pixels appear consecutively in adjacent positions and expand outward along the same boundary under multiple lighting conditions, then this area is recorded as a high reflectivity response cluster area. Furthermore, a preset neighborhood is searched outward from the outer edge of the high reflectivity response cluster area. The grayscale transition within the neighborhood is counted to determine whether the saturated pixels maintain connectivity and whether the boundary shape exhibits a sheet-like unfolding trend. Based on this, a saturation distribution result is generated. Through this step, the bright areas caused by specular reflection or local strong reflection are separated from the ordinary bright texture.

[0029] It should be noted that the value of the preset neighborhood is determined based on the expected minimum defect width, background texture repetition spacing, and image spatial resolution within the target acquisition area, so that the preset neighborhood can cover the continuous background texture outside the reflective boundary without crossing adjacent independent abnormal areas.

[0030] When determining candidate reflective regions, three types of information are used simultaneously: brightness order relationship, saturation distribution results, and neighborhood continuity relationship. Specifically, these include: For a candidate region, if most of its pixels show obvious high brightness order consistency in the brightness sequence of multiple images in the same domain, and the region is also located in the high reflectivity response cluster area corresponding to the saturation distribution result, and the grayscale continuity relationship inside and outside the boundary of the region shows a sudden interruption, and the boundary connectivity relationship shows a closed or semi-closed sheet-like expansion, then the region is determined to be a reflective candidate region. Adjacent reflective candidate regions are connected and merged to avoid the same reflective region being split into multiple isolated small blocks. For reflective candidate regions with many jagged edges or local breaks, boundary correction is performed based on the grayscale continuity of the neighborhood to make the boundary more closely match the outer edge of the actual reflective region, thus obtaining the reflective region.

[0031] A concrete example of the actual effect of the above process is as follows: When a shallow scratch and a localized specular reflective area exist within the target acquisition area, the scratch typically exhibits a thin, elongated variation in brightness extending in a single direction under a certain oblique illumination condition. The specular reflective area, on the other hand, often displays localized high brightness, patchy saturation, and outward expansion of its boundaries under multiple illumination conditions. By comparing the brightness sequences of the same location under different illuminations, and combining this with the spatial aggregation results of saturated pixels and the connectivity of neighboring boundaries, the shallow scratch can be retained in the anomaly candidate information, while the patchy specular reflective area can be identified as a reflective region. This processing prevents subsequent steps from misclassifying the patchy reflective area as a planar defect.

[0032] After obtaining the reflective areas, the reflection-suppressed image is generated. First, continuous texture segments are extracted outward from the boundary of the reflective areas. Here, continuous texture segments refer to background image areas located outside the boundary of the reflective areas, where grayscale changes remain continuous, texture direction is relatively stable, and no obvious abnormal abrupt changes occur. During extraction, the search can be performed in segments along the boundary of the reflective area. Pixel blocks with relatively gentle texture changes and consistent with the original background of the board can be located outside each boundary segment and recorded as candidate continuous texture segments. Subsequently, the consistency of texture direction and the continuity of grayscale transition of these continuous texture segments can be analyzed. If a certain continuous texture segment extends more stably along the row direction outside the boundary, the row direction can be determined as the texture extension direction of the segment. If a continuous texture segment is more stable in the column or diagonal direction, the corresponding direction is determined as the texture extension direction. In this way, the compensation direction that best matches the surrounding background structure is selected for reflective areas of different shapes.

[0033] After determining the texture extension direction, continuous texture segments are mapped along this direction to the interior of the reflective area to generate a compensation texture. During mapping, the texture outside the boundary is not simply copied as a whole into the interior of the reflective area. Instead, it is extended inward segment by segment according to the boundary adjacency relationship, so that the compensation texture first connects with the original background texture near the boundary and then gradually covers the highlight response inside the reflective area. For a narrow and elongated reflective area, the texture is extended along the normal direction of its short side first in order to restore the continuity of the texture on both sides of the area as soon as possible. For blocky reflective areas, priority should be given to partitioning and extending along the main direction of the background texture to avoid generating obvious artificial splicing marks. After the compensation texture is generated, a brightness callback is performed on the reflective area to reduce the excessively high brightness response in the reflective area to a grayscale range that is consistent with the surrounding background. Simultaneously, boundary smoothing is performed on the edges of the reflective area to create a continuous transition between the compensation texture inside the reflective area and the non-reflective image area on the outside, ultimately resulting in a reflection-suppressed image.

[0034] It should be noted that in this embodiment, the goal of generating the reflection suppression image is not to eliminate all bright areas, but to suppress the unrealistic abnormal response caused by specular reflection while preserving the continuity of the original background texture of the target acquisition area.

[0035] For example, when there is a local specular reflective area in the target acquisition area, first extract the smoothed texture fragments that are not affected by the reflection along the outer edge of the reflective area, and then gradually extend it into the reflective area along the main direction of the smoothed texture to obtain the compensation texture. Then, the brightness in the reflective area is adjusted from the saturated state to a gray level similar to the surrounding smoothed background, and the boundary is smoothed. After processing, the area no longer appears as a patchy high-brightness abnormality, but is restored to a normal texture area consistent with the surrounding background, thus providing input for subsequent background texture decomposition and real defect enhancement.

[0036] Background texture decomposition is performed on the reflection suppression image to construct an abnormal response layer, and directional response maps are generated according to the differences in row, column, and diagonal responses. The specific implementation steps include: In this embodiment, after obtaining the reflection-suppressed image, instead of directly segmenting the entire image uniformly, the image content within the target acquisition area is first divided into a stable background portion and an abnormal portion deviating from the background. Specifically: Using local windows in the reflection suppression image as the basic analysis unit, the size of the local window is preferably set to be able to cover more than one background texture repetition cycle and smaller than the side length of the target acquisition area; an overlapping scanning method is used between adjacent local windows so that the texture continuity at the window boundary can be preserved and used for subsequent background texture area determination; Gray-level fluctuations, texture repetition spacing, local edge direction, and gray-level transition density are extracted window by window along the row and column directions to form the texture statistics results corresponding to each local window; If adjacent local windows maintain continuity in grayscale fluctuation period, texture repetition spacing and edge direction, then these local windows are determined to constitute a stable background texture area. If a local window deviates significantly from its neighboring windows in at least one of the above aspects, then the local window is marked as an abnormal candidate region; The above process involves first establishing a background texture reference for the target acquisition area itself, rather than directly applying an external uniform template, thereby adapting to the subtle background differences that may still exist after different alumina plates have been ground and polished.

[0037] When extracting the periodic distribution results of the surface background texture, the gray-level sequence within each local window is first expanded directionally to form gray-level change sequences along the row, column, and two diagonal directions. Then, the repetition intervals in the gray-level change sequences in each direction are statistically analyzed, specifically including: The repeated fluctuation intervals that remain continuous in adjacent windows are recorded as the periodic distribution results of the region. At the same time, stable gray centers and normal fluctuation ranges are extracted from the gray areas within each local window that are not affected by local anomalies, forming gray change baselines. The gray change baseline is not a single gray value, but rather the background gray change range that is allowed to appear in the target acquisition area under normal surface conditions. For a plate surface with relatively uniform polishing, the baseline of this grayscale change is usually relatively flat; For panels that retain slight processing textures, the grayscale variation baseline allows for continuous gradual changes in local directions. By establishing the periodic distribution results and the grayscale variation baseline simultaneously, the repeatability of the normal surface texture and the allowable range of normal grayscale fluctuations are fixed together.

[0038] After completing the above analysis, the reflection suppression image is decomposed into a background texture layer and an anomaly response layer. During the decomposition, the periodic distribution result is used as a constraint. In each local window, the image components that are consistent with the background repetition interval, background edge direction and gray-level change baseline are retained and summarized to form the background texture layer. Then, the local variation components that do not meet the consistency constraint are extracted from the original image and summarized to form the abnormal response layer; The background texture layer is used to characterize the stable texture structure of the alumina plate surface within the current target acquisition area, such as the smooth texture formed after polishing, the weak periodic background left over from the grinding process, and the slight light and dark gradient consistent with the overall plate surface. The anomaly response layer is used to preserve local variations that deviate from the background texture structure, such as narrow strip grayscale abrupt changes caused by fine scratches, local dense dark spots or bright spots caused by dot-like pits, and surface texture mismatch caused by blemish anomalies.

[0039] Through the above decomposition process, the detection step only processes the abnormal response components that deviate from the background for subsequent analysis, thereby reducing the interference of background texture on defect identification.

[0040] When generating the directional response map, the grayscale variation amplitude, edge extension state and local response density state are extracted from the abnormal response layer along the row direction, column direction and diagonal direction respectively. The grayscale variation amplitude is used to characterize the intensity of the brightness fluctuation of the abnormal response in a certain direction. Edge extension status is used to characterize whether an abnormal edge continues to expand along a certain direction; Local response density is used to characterize whether anomalous response pixels form a continuous cluster in a certain direction.

[0041] In practical implementation, a local abnormal region can be selected in the abnormal response layer, and row-direction response matrices, column-direction response matrices, and diagonal-direction response matrices can be established for it. Then, the continuous response length, gray-level abrupt change depth, and response point clustering degree of the abnormal region in each direction can be counted. If the continuous response length of an abnormal region in the row direction is significantly greater than that in other directions, and the gray-level change amplitude remains continuous in the row direction, then the abnormal region is determined to have row-direction dominant characteristics. If a certain anomalous region has a more stable response extension in the column direction or diagonal direction, it is determined to be column-dominant or diagonal-dominant, respectively. After writing the dominant direction results of each anomalous region into the direction marker, a directional response map can be formed.

[0042] Once the directional response map is generated, it can provide a direct basis for distinguishing anomaly types; For example, when there is a fine scratch extending horizontally along the board surface within the target acquisition area, this area in the abnormal response layer usually shows a continuous gray-scale abrupt change in the row direction, only a narrow width boundary in the column direction, and no significant continuous extension in the diagonal direction. In this case, the directional response map will mark this area as the row-dominant response area. If there is a local point-like depression in the target acquisition area, the response range of this area in the row direction, column direction and diagonal direction is short. In the directional response map, it will appear as an abnormal point-like area with no obvious single dominant direction but a high local response density. If there is a localized contamination patch within the target collection area, the area will expand in a planar manner in multiple directions. However, the outer edge of its boundary may have a more obvious diffusion trend in a certain diagonal upward direction. In this case, the directional response map can retain its multi-directional response characteristics and record its relatively dominant boundary expansion direction. After processing, differentiated enhancement and segmentation can be performed for different dominant directions and different response patterns, instead of using the same processing method for all abnormal areas.

[0043] For example, in the central region of a polished alumina plate used for subsequent thin-film metallization, two types of anomalous responses are still preserved in the reflection suppression image: One type is thin, elongated dark lines extending from the upper left to the lower right, and the other type is localized bright spots located near the edges of the board. After performing local texture statistics on the image, it was found that most background areas have smooth and repetitive gray-level fluctuations in both the row and column directions. Based on this, a background texture layer can be established. However, the thin, elongated dark lines and bright spots deviate significantly from the periodic distribution results and gray-level change baseline, and are therefore retained in the abnormal response layer. After further calculation of the response in each direction, the slender dark bars showed the strongest continuous response length and edge extension in the diagonal direction, and were therefore marked as the diagonal dominant anomaly in the directional response map. The response lengths of the clustered bright spots are similar in all directions, but the local response density is the strongest, so they are marked as locally clustered anomalous regions in the directional response map. Thus, subsequent steps can use continuous enhancement along the dominant direction for the former and local abrupt enhancement and region closure for the latter.

[0044] Based on the directional response map, enhancement, segmentation, connectivity extraction, and boundary refinement are performed on linear anomalies, point anomalies, and patch anomalies, respectively, to generate a candidate defect region set. The specific implementation steps include: In the aforementioned steps, background texture decomposition has been performed on the reflection suppression image to obtain an anomaly response layer for characterizing local anomaly changes, and a directional response map reflecting the dominant directional features of each anomaly region has been further formed. In this embodiment, subsequent processing no longer treats all abnormal regions as the same type of target and processes them uniformly. Instead, it utilizes the differences in directional extension, local aggregation, and boundary expansion of each abnormal region in the directional response map to divide the abnormal regions into linear abnormal regions, point abnormal regions, and patch abnormal regions. Then, it performs enhancement, segmentation, connectivity extraction, and boundary refinement processing adapted to their morphological features, and finally generates a set of candidate defect regions. This approach ensures that different morphological defects such as fine scratches, point pits, and local blemishes can be stably displayed under their respective suitable enhancement paths, avoiding the distortion caused by uniform processing, such as the breakage of slender defects, the submersion of point defects, or the outward expansion of patch defect boundaries.

[0045] In practice, the directional extension features, local clustering features and boundary expansion features are extracted for each abnormal region in the directional response map. The directional extension features are used to characterize the degree of continuous expansion of the abnormal region in the dominant direction, which can be characterized by the trend of the change of response length and width of the abnormal region along the dominant direction and the discontinuity distance between adjacent response segments. In this embodiment, an abnormal feature entry is established for each abnormal region. The abnormal feature entry records at least the dominant direction, the continuous response length of the dominant direction, the local response density, the boundary range, and the edge extension trend. Subsequent abnormal type classification is performed based on the abnormal feature entry. Local clustering features are used to characterize whether anomalous responses form a concentrated distribution within a small area. They can be characterized by local response pixel density, local gray-level abrupt change intensity, and the number of response centers within the anomalous region. Boundary expansion features are used to characterize whether the boundary of an abnormal region continues to expand into the surrounding surface region. They can be characterized by changes in boundary curvature, the extent of boundary encirclement, and the trend of edge extension.

[0046] After completing the above feature extraction, type classification is performed according to the feature combination relationship of each abnormal region: When the directional extension feature of an anomalous region is significantly stronger than the local clustering feature and the boundary expansion feature, and the anomalous region expands continuously in the dominant direction and its width remains relatively unchanged in the direction perpendicular to the dominant direction, it is identified as a linear anomalous region. When a certain abnormal region has obvious local clustering characteristics, and its extension length in all directions is relatively short, and the local response center is concentrated in a limited range, it is identified as a point-like abnormal region. When the boundary expansion feature of an anomalous region is significantly stronger than the other features, and the anomalous region forms a continuous coverage within a local area and has a relatively complete enclosing boundary, it is identified as a patch anomalous region. After the above processing, each anomalous region in the directional response map is classified into a category consistent with its shape.

[0047] For linear anomaly regions, in this embodiment, continuous enhancement and break-repair processing are performed along the dominant response direction. The goal of continuous enhancement is to improve the response consistency of linear anomalies in the dominant direction, so that the slender defects that were originally weak due to insufficient grayscale fluctuations can be continuously displayed.

[0048] Specifically, taking the dominant direction of the linear anomaly region as the extension axis, the continuity, direction consistency and width consistency of gray-level abrupt changes of adjacent response positions are compared point by point along this direction. Serial enhancement is performed on response segments that meet the continuity condition to make their gray-level differences in the dominant direction more concentrated. The break-and-connection processing is used to solve the problem of breakpoints in fine scratches in local areas due to polishing texture interference, background fluctuations, or local reflection residues. During processing, if two adjacent linear response segments are consistent in the dominant direction, and the interval between them is continuous with the segments at both ends in terms of grayscale change trend and edge direction, then the interval is used as the connection area, and directional connection is performed on it, so that the originally separated linear response segments are reconnected into a continuous whole. This allows shallow scratches, fine cracks, or drag marks extending in the same direction to form a more complete linear response structure before segmentation.

[0049] For point-like anomalous regions, this embodiment performs local mutation enhancement and isolated response highlighting processing, specifically including: The purpose of local mutation enhancement is to amplify the local gray-level abrupt changes in abnormal regions such as dot-like pits, attached particles, and localized collapse points relative to the surrounding background, thus significantly separating them from a smooth background. Specifically, a local neighborhood is established around the dot-like abnormal region, and the gray-level difference between the neighborhood center and the surrounding background, the edge contraction trend, and the local response concentration are compared. When the gray-level abrupt change depth in the central region exceeds the normal fluctuation range of the surrounding background, the abnormal response in the central region is amplified, so that the local point-like abrupt change remains compact in space and prominent in gray level. Isolated response highlighting processing is used to suppress the scattered weak responses around the point-like anomalous region that are not related to it, and only the anomalous part that is tightly clustered in the center is retained. During processing, the degree of clustering of response points in the local neighborhood is judged, and the scattered points on the periphery that are not continuously related to the central anomalous region and whose gray-level changes are unstable are weakened, so that the real point-like anomalous region appears in the image as a local anomalous point with clear boundaries and a well-defined center.

[0050] For patchy anomaly areas, this embodiment performs region closure enhancement and boundary expansion suppression. Region closure enhancement is used to improve the overall integrity of patchy anomalies, allowing surface contamination marks, localized planar damage, or flaky corrosion to form a complete planar representation.

[0051] Specifically, the local weak responses within the anomalous patch region are first filled, restoring the internal fractured areas caused by background texture interference to continuous surface regions. Then, based on the adjacent edge relationships of the anomalous patch boundary, the unclosed areas of the boundary are filled, forming a complete closed contour for the entire anomalous region. Boundary expansion suppression processing is used to prevent the anomalous patch region from expanding outward without constraint during enhancement, mistakenly absorbing adjacent normal background regions as anomalous regions. During processing, the background texture layer outside the boundary of the patch anomaly region is used as a reference to compare the gray-level transition relationship and texture continuity relationship inside and outside the boundary. Suppression is performed on the weak edge response that is only brought out during the enhancement process, so that the patch anomaly boundary stays near the real gray-level abrupt change position. Therefore, by combining the processing of region closure enhancement and boundary expansion suppression, the integrity of the patch anomaly can be preserved while avoiding its false boundary expansion.

[0052] After completing the above enhancement processing for different types of anomalies, anomaly enhancement results are generated. At the same time, threshold segmentation, connected region extraction, and boundary refinement are sequentially performed on the anomaly enhancement results to form candidate defect regions corresponding to each type of anomaly. During threshold segmentation, a uniform threshold is not used for all anomaly enhancement results. Instead, the segmentation intervals are determined based on the differences in the enhanced responses of linear anomalies, point anomalies, and patch anomalies. This ensures that the thin and extended parts of linear anomalies are not excessively truncated, the local high response cores of point anomalies are not swallowed by the background, and the complete surface regions of patch anomalies are not broken up. Specifically, the segmentation intervals of linear anomalies prioritize the retention of continuous weak responses in the dominant direction, the segmentation intervals of point anomalies prioritize the retention of local high mutation centers, and the segmentation intervals of patch anomalies prioritize the retention of continuous response regions within the boundaries of closed surface regions. Each segmentation interval can be determined based on the enhanced response statistics of the corresponding anomaly type in the current target acquisition area. After segmentation is completed, connected component extraction is performed on the segmentation results corresponding to each anomaly type. Spatially continuous pixel sets whose gray-level responses belong to the same anomaly are extracted into separate regions. In linear anomaly processing, connected component extraction focuses on ensuring the merging of adjacent segments in the dominant direction. In handling point-like anomalies, the focus is on extracting connected regions to suppress peripheral discrete points; In patch anomaly processing, the focus of connected region extraction is to preserve the overall closed relationship of the surface region.

[0053] After obtaining each connected region, the boundary refinement process is performed. The purpose of boundary refinement is to correct the coarse boundary formed during the enhancement and segmentation process into a fine boundary that is more consistent with the real anomaly edge, so that the subsequent extraction of boundary contour, length parameter and area parameter has higher stability. For linear anomaly regions, boundary refinement mainly compresses the lateral redundant edge, so that the slender anomaly remains continuous along the dominant direction and maintains a narrow width along the vertical direction. For point-like anomaly regions, boundary refinement mainly involves removing the weak outer edges, so that the point-like anomalies retain a concentrated and compact outer contour; For patchy anomaly regions, boundary refinement mainly involves adjusting the edge direction along the actual grayscale transition position of the region boundary to make the outline of the patch anomaly consistent with its actual coverage area. After the boundary refinement is completed, candidate defect regions corresponding to linear anomalies, point anomalies, and patch anomalies can be obtained respectively. Finally, the candidate defect regions corresponding to the above types of anomalies are uniformly mapped back to the plate coordinate system of the same target acquisition area and merged to generate a set of candidate defect regions.

[0054] A specific example can be used to illustrate this: For an alumina plate that has undergone polishing and is about to enter the metallization process, three types of anomalies are simultaneously present in the directional response map of a certain target acquisition area: The first type consists of thin, elongated dark lines extending from the upper left to the lower right; the second type consists of small, high-density dark spots located in the middle of the region; and the third type consists of an irregular, planar bright spot located near the edge. During processing, firstly, based on the directional extension characteristics, the slender dark lines are identified as linear anomalous regions, and continuous enhancement and break repair are performed along their dominant direction to reconnect the slender dark lines that were originally interrupted by background interference. Then, based on the local aggregation characteristics, the small dark spots are identified as point-like anomalous regions, and local abrupt enhancement and isolated response highlighting are performed to make the small dark spots stand out from the background texture. Finally, based on the boundary expansion characteristics, the planar bright spots are identified as patch anomalous regions, and region closure enhancement and boundary expansion suppression are performed to form complete but not excessively expanded anomalous boundaries. Subsequently, threshold segmentation, connected region extraction, and boundary refinement were performed on the three types of abnormal enhancement results respectively, and candidate defect regions corresponding to fine scratches, dot-like pits, and local blemishes were obtained. These were then merged to generate a set of candidate defect regions for the target acquisition area.

[0055] False defects are eliminated based on the directional continuity, grayscale transition consistency, contour closure relationship, and neighborhood coordination relationship of candidate defect regions. The defect type identifier, defect boundary information, and board surface defect distribution results are then output. Specific implementation steps include: In the aforementioned steps, differential enhancement, segmentation, connectivity extraction, and boundary refinement of linear anomalies, point anomalies, and patch anomalies have been completed based on the directional response map, and a candidate defect region set corresponding to the current target acquisition area has been obtained. At this time, the candidate defect region set contains both real surface defects and pseudo-defect regions formed by processing texture continuation, local reflection residue, or random noise disturbance. Therefore, in this embodiment, the processing first performs pseudo-defect removal on the candidate defect region set, and then performs type identification, boundary information extraction, and board surface distribution output on the remaining real defect regions.

[0056] In the pseudo-defect removal stage, the main direction extension state, gray scale change state on both sides of the boundary, region outline closure state and neighborhood texture coordination state are extracted for each candidate defect region in the candidate defect region set. The main direction extension state is used to characterize whether the candidate defect region continues to expand along a certain fixed direction and whether the expansion direction is consistent with the main extension direction of the background texture of the board surface. The grayscale change state on both sides of the boundary is used to characterize whether the grayscale transition on both sides of the boundary of the candidate defect region is stable and whether there is a sudden change feature of rapid collapse from highlight to normal background. The region contour closure status is used to characterize whether the outer contour of the candidate defect region is completely closed, and whether there are a large number of broken openings or discontinuous swings in the local boundary. The neighborhood texture coordination state is used to characterize whether a continuous transition can be formed between the background texture surrounding the candidate defect region and the texture inside the region.

[0057] It should be noted that, in order to ensure the stability of subsequent judgment results, the above state does not rely on the judgment of a single pixel, but extracts the candidate defect region and its boundary neighborhood as a whole analysis object.

[0058] For the extraction of the main direction extension state, this embodiment first obtains the extension direction and extension length of each candidate defect region, and then compares the extension direction with the main direction of the processing texture recorded in the background texture layer of the current target acquisition area. If the extension direction of a candidate defect region is consistent with the main direction of the processing texture, and the length change of the candidate defect region in the main direction is consistent with the length change trend of adjacent stripes in the background texture, while its width remains small and the boundary undulation is gentle, then it is determined that the candidate defect region has a continuous extension relationship with the surface processing texture. Although such regions may be enhanced into slender abnormalities in the previous step, they are essentially a natural continuation of the processing texture, rather than a real surface defect. Therefore, in this embodiment, they are included in the scope of pseudo-defect judgment.

[0059] For the extraction of grayscale change states on both sides of the boundary, this embodiment extracts grayscale transition segments from the inner and outer sides along the boundary normal direction of each candidate defect region, and compares the symmetry, continuity and stability of grayscale changes on both sides of the boundary. If a candidate defect region has localized high-brightness collapse inside its boundary, and the gray-scale distribution outside its boundary remains adjacent to the residual features of the high-brightness region before reflection suppression, and if the candidate defect region has exhibited saturated pixel aggregation or saturated boundary expansion in images corresponding to different lighting conditions, then the candidate defect region is determined to be a reflection residue region. In other words, although such regions retain a certain local response after reflection suppression, the gray-scale transition relationship on both sides of their boundary does not conform to the stable edge change law of real scratches, pits, or blemishes, but is closer to the unstable boundary response caused by specular reflection residue.

[0060] For the extraction of the region contour closure state and the neighborhood texture coordination state, this embodiment sequentially tracks the boundary contour of the candidate defect region, records whether the boundary forms a closed loop, whether there are a large number of fine branches, and whether it repeatedly turns back within a short distance. At the same time, it compares the consistency between the background texture in the preset neighborhood outside the candidate defect region and the residual texture inside the candidate defect region. If the contour closure of a candidate defect region is poor, manifested as frequent interruptions in local boundaries or the inability of the outer contour to form a stable enclosing relationship, and the coordination of its neighboring textures is also poor, manifested as the absence of stable boundaries and clear structural differences between the textures inside the region and the background textures outside, and only exhibiting discrete, occasional, and random weak responses, then it is judged as a random noise region. Such regions are often caused by acquisition noise, local gray-scale perturbations, or isolated residual responses during the enhancement process, and do not possess the morphological integrity and structural independence that a real defect should have.

[0061] After completing the above analysis, any candidate defect region that satisfies the relationship of continuous extension of processing texture, reflection residue, or random noise is marked as a pseudo-defect region and removed from the candidate defect region set. After the removal, the remaining regions constitute the target defect region set. The target defect region set retains only the abnormal regions that still have independent boundaries, stable grayscale changes, and clear spatial locations after multi-condition screening, thus providing reliable input for subsequent type identification and board distribution output.

[0062] When the same candidate defect region simultaneously satisfies two or more pseudo-defect judgment relationships, the judgment is made first based on the continuous extension relationship of the processing texture. If the continuous extension relationship of the processing texture is not satisfied, the judgment is made based on the reflection residue relationship. If it is still not satisfied, the judgment is made based on the random noise relationship. Only when the candidate defect region does not satisfy any of the above pseudo-defect judgment relationships is it retained as the target defect region.

[0063] Taking a specific example, in a certain target acquisition area, there exists simultaneously a weak, elongated response extending along the grinding direction, a sheet-like residual bright area near the reflective edge, and a locally closed small dark spot. During processing, the former two are marked as pseudo-defect areas and removed because they are consistent with the continuous extension of the processing texture and with the reflective residue, respectively. However, the latter is retained in the target defect area because of its stable closed boundary, poor coordination with neighboring textures, and independent grayscale changes. Through this processing, the interference of processing texture and reflective residue on the final detection results can be significantly reduced.

[0064] After obtaining the target defect region set, the process proceeds to the defect type identification and boundary information output stage, which specifically includes: First, for each target defect region in the target defect region set, the center position, boundary contour, extension direction, length parameter, area parameter, and regional grayscale variation features are extracted. The center position is used to characterize the geometric center of the target defect region in the current target acquisition area. The boundary contour is used to characterize the actual outer edge shape of the target defect region. The extension direction is used to characterize the main development direction of the target defect region. The length parameter is used to characterize the spatial span of the target defect region along the main direction. The area parameter is used to characterize the coverage area of ​​the target defect region on the board surface. The regional grayscale variation features are used to characterize the light and dark variation patterns inside and near the boundary of the target defect region. After the above features are extracted, they are uniformly written into the defect description entries. The defect description entries can be understood as data records that provide a structured description of each target defect area. They include at least the alumina plate identifier, target acquisition area identifier, target defect area number, center position, boundary contour, extension direction, length parameter, area parameter, and area grayscale variation characteristics.

[0065] In the defect type identification stage, a defect discrimination rule set is pre-established, and each defect description entry is matched with the rule set. Each rule entry in the defect discrimination rule set contains at least the applicable defect type, direction condition, relationship condition between length parameter and area parameter, boundary contour condition, and region grayscale change characteristic condition. When a defect description entry simultaneously meets the above conditions of the same rule entry, the defect type identifier corresponding to that rule entry is output. In this embodiment, the defect discrimination rule set covers at least three common surface defects: scratches, pits, and blemishes. If a defect description item has a single extension direction, a length parameter that is significantly greater than the area parameter, a thin and elongated boundary outline, and a gray-scale change feature that shows a continuous abrupt change along a single direction, then the target defect area is identified as a scratch defect. If a defect description item has a small area parameter, a relatively compact boundary outline, and regional grayscale variation features concentrated in the local center, and does not show long-distance expansion in any direction, then it is judged as a point-like pit defect. If a defect description entry has a relatively large area parameter, a clear boundary contour, and a persistent grayscale variation feature within the surface region with numerous changes in boundary direction, then it is classified as a blemish-type defect. After matching is complete, each target defect area is assigned a corresponding defect type identifier.

[0066] In the stage of generating the defect distribution results on the board surface, the center position of each target defect area is first mapped to the board surface coordinate position of the alumina board to be inspected, forming a set of defect positions. That is, according to the aforementioned implementation process, each target acquisition area already has a unique area identifier, and the correspondence between the target acquisition area and the board surface coordinates has been established during image acquisition. Therefore, the center position of the target defect area in the local image can be directly converted into the board surface coordinate position on the entire alumina board to be inspected. During the conversion, the starting position of the target acquisition area in the plate coordinate system is used as the area reference point, and the pixel size correspondence obtained by the image acquisition camera calibration is used as the proportional conversion basis. The center position and boundary contour coordinates of the target defect area in the local image are converted into the plate coordinate position and plate contour coordinates. After mapping is completed, a defect distribution map of the board surface is established based on the defect location set and the corresponding defect type identifier. This defect distribution map is used to intuitively represent the spatial distribution of different types of defects on the entire alumina board, such as which defects are concentrated in the edge area of ​​the board surface, which defects are concentrated in the middle area of ​​the board surface, and whether there is a local clustering phenomenon among different types of defects.

[0067] During the test result organization stage, the defect type identifier, boundary contour, plate surface coordinate position, and plate surface defect distribution map are written into the test result entries. In addition to the above, the test result entries can also include the alumina plate identifier, test time identifier, and target acquisition area source information to form a complete test result record. In this embodiment, the test result entries include at least the alumina plate identifier, test time identifier, target acquisition area identifier, defect type identifier, boundary contour, plate surface coordinate position, and corresponding index information of the plate surface defect distribution map, so as to facilitate subsequent quality review, process traceability, and batch statistics retrieval. Subsequently, the test result entries are output, and a defect detection record associated with the corresponding alumina plate identifier is established. This defect detection record can be directly called by the online screening station to determine whether the alumina plate should enter the subsequent process, or it can be called by the quality analysis system to statistically analyze the distribution of defect types and plate surface locations of a specific batch of alumina plates.

[0068] For example, if an alumina plate has multiple scratch-like defects along its edge, while only a few dot-like pits appear in the center, the defect distribution map will show that the scratches are concentrated at the edge and the dot-like defects are scattered in the center. The test results not only record the specific type and boundary outline of each defect, but also associate these defects with specific process steps through the plate coordinates, thereby helping to determine whether they are related to handling contact, edge clamping, or local particulate contamination.

[0069] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0070] Those skilled in the art will recognize that the modules 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.

[0071] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0072] 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 scope of the technology 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.

[0073] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting surface defects in alumina plates based on image recognition, characterized in that: The specific steps include: Surface images of the alumina plate under different lighting conditions are acquired, and distortion correction, viewing angle correction and grayscale normalization are performed to generate a group of images in the same domain. Reflective regions are identified based on the brightness order, saturation distribution, and neighborhood continuity of corresponding pixels in a co-domain image group, and reflective suppression images are generated by combining boundary texture extension. Background texture decomposition is performed on the reflection suppression image to construct an abnormal response layer, and directional response maps are generated according to the differences in row, column and diagonal responses. Based on the directional response map, enhancement, segmentation, connectivity extraction, and boundary refinement are performed on linear anomalies, point anomalies, and patch anomalies, respectively, to generate a set of candidate defect regions; Based on the directional continuity, grayscale transition consistency, contour closure relationship, and neighborhood coordination relationship of the candidate defect region, false defects are eliminated, and the defect type identifier, defect boundary information, and board surface defect distribution results are output.

2. The method for detecting surface defects of alumina plates based on image recognition according to claim 1, characterized in that: Surface images of the alumina plate under inspection were acquired under different lighting conditions, including: Determine the target acquisition area of ​​the surface to be inspected, and set up lighting conditions with different incident directions or different lighting modes around the target acquisition area; Surface images of the target acquisition area are acquired under various lighting conditions to obtain a set of original surface images corresponding to the same target acquisition area; Each original surface image is associated with the target acquisition area identifier and written into the image entry to form an input image set for generating a co-domain image group.

3. The method for detecting surface defects of alumina plates based on image recognition according to claim 2, characterized in that: Perform distortion correction, viewpoint correction, and grayscale normalization to generate a group of images in the same domain, including: A primary surface image is selected as the reference image, and its edge contour features and surface texture localization features are extracted. Based on edge contour features and surface texture localization features, geometric distortion correction, position offset correction and viewpoint correction are sequentially performed on the remaining original surface images to generate a corrected image that corresponds to the reference image space. Gray-level normalization processing is performed on the reference image and each calibration image to obtain a group of images in the same domain with consistent pixel positions and gray-level scales.

4. The method for detecting surface defects of alumina plates based on image recognition according to claim 1, characterized in that: Reflective regions are identified based on the brightness order, saturation distribution, and neighborhood continuity of corresponding pixels in a co-domain image group, including: Extract the brightness values ​​of each pixel in the same image group under different lighting conditions, construct a pixel brightness sequence according to the order of brightness change, and determine the brightness order relationship based on the pixel brightness sequence; The spatial clustering and boundary expansion states of saturated pixels in a group of images in the same domain are statistically analyzed to generate saturated distribution results. Based on the brightness order relationship and saturation distribution results, the reflective candidate region is determined by combining the grayscale continuity state and boundary connectivity state in the neighborhood of saturated pixels. Then, the reflective candidate region is connected and merged and the boundary is corrected to generate the reflective region.

5. The method for detecting surface defects of alumina plates based on image recognition according to claim 4, characterized in that: The reflection-suppressed image is generated by combining boundary texture extension, including: Extract continuous texture fragments outside the boundary of the reflective area, and determine the texture extension direction based on the consistency of texture direction and the continuity of grayscale transition; Map continuous texture fragments along the texture extension direction to the interior of the reflective area to generate a compensation texture for the reflective area. The reflective area is subjected to brightness correction and boundary smoothing based on the compensated texture, and then stitched with the non-reflective image area outside the reflective area to generate a reflection-suppressed image.

6. The method for detecting surface defects of alumina plates based on image recognition according to claim 1, characterized in that: Background texture decomposition is performed on the reflection-suppressed image to construct an abnormal response layer, and directional response maps are generated according to the differences in row, column, and diagonal responses, including: Local texture statistics and region structure analysis are performed on the reflection-suppressed image to extract the periodic distribution results of the surface background texture and the gray-level change baseline; Based on the periodic distribution results and the grayscale change baseline, the reflection suppression image is decomposed into a background texture layer and an anomaly response layer. The background texture layer is used to characterize the stable texture structure of the alumina plate surface, and the anomaly response layer is used to characterize local anomaly changes that deviate from the background texture structure. Extract the grayscale variation amplitude, edge extension state, and local response density state of the abnormal response layer in the row, column, and diagonal directions respectively, and generate the corresponding response results in each direction; Based on the response results in each direction, the dominant response direction of the abnormal response layer is determined, and a directional response map is generated.

7. The method for detecting surface defects of alumina plates based on image recognition according to claim 6, characterized in that: Based on the directional response map, enhancement, segmentation, connectivity extraction, and boundary refinement are performed on linear anomalies, point anomalies, and patch anomalies, respectively, to generate a set of candidate defect regions, including: Based on the directional extension characteristics, local clustering characteristics, and boundary expansion characteristics of each abnormal region in the directional response map, the abnormal regions are divided into linear abnormal regions, point-like abnormal regions, and patchy abnormal regions. For linear anomalous regions, continuous enhancement and break-and-repair processing along the dominant response direction are performed; for point-like anomalous regions, local abrupt enhancement and isolated response highlighting processing are performed; and for patchy anomalous regions, regional closure enhancement and boundary expansion suppression processing are performed to generate anomalous enhancement results. The abnormal enhancement results are sequentially subjected to threshold segmentation, connected component extraction and boundary refinement to obtain candidate defect regions corresponding to each type of abnormality; The candidate defect regions corresponding to each type of anomaly are merged to generate a candidate defect region set.

8. The method for detecting surface defects of alumina plates based on image recognition according to claim 1, characterized in that: False defects are eliminated based on the directional continuity, grayscale transition consistency, contour closure relationship, and neighborhood coordination relationship of the candidate defect region, including: Extract the main direction extension state, gray level change state on both sides of the boundary, region contour closure state, and neighborhood texture coordination state of each candidate defect region. Based on the main direction extension state, determine whether the candidate defect area maintains a continuous extension relationship with the surface processing texture; based on the gray change state on both sides of the boundary, determine whether the candidate defect area belongs to the reflective residual area; based on the region contour closure state and the neighborhood texture coordination state, determine whether the candidate defect area belongs to the random noise area. Candidate defect regions that satisfy the relationships of continuous extension of processing texture, residual reflection, or random noise are marked as pseudo-defect regions. After removing pseudo-defect regions, a set of target defect regions is generated.

9. The method for detecting surface defects of alumina plates based on image recognition according to claim 8, characterized in that: Output defect type identifier, defect boundary information, and board surface defect distribution results, including: Extract the center position, boundary contour, extension direction, length parameter, area parameter, and gray-scale variation characteristics of each target defect region in the target defect region concentration; Defect description entries are constructed based on center location, boundary contour, extension direction, length parameter, area parameter, and regional grayscale variation characteristics; The defect description entries are matched with the preset defect discrimination rule set to determine the defect type identifier corresponding to each target defect area; The defect distribution results of the board surface are generated based on the positional distribution relationship of each target defect area on the surface of the alumina board to be inspected.

10. The method for detecting surface defects of alumina plates based on image recognition according to claim 9, characterized in that: Generate the defect distribution results on the board surface, including: The center position of each target defect area is mapped to the plate coordinate position of the alumina plate to be inspected, forming a set of defect positions; A defect distribution map of the board surface is established based on the defect location set and defect type identifier; Write the defect type identifier, boundary contour, plate surface coordinate position, and plate surface defect distribution map into the inspection result entries; Output the inspection results entries and create defect inspection records associated with the corresponding alumina plate identifier.

Images