A method and system for defect detection in footwear production processes

Abstract

The application discloses a kind of defect detection method and system for footwear production process, it is related to defect optical detection technical field, including collecting footwear product multi-angle polarization detection image, and generating footwear type partition benchmark atlas.Based on the construction material reflection residual graph of footwear type partition benchmark atlas, generate defect candidate graph.According to the process correlation discrimination of defect candidate graph, output defect detection result.The application combines multi-angle polarized optical measurement, footwear type partition benchmark atlas, material reflection residual calculation and process correlation discrimination, forms the online defect detection scheme suitable for footwear production process, improves defect recognition stability, process tracing capability and production line closed loop control capability.

Description

Technical Field

[0001] This invention relates to the field of defect optical inspection technology, specifically a defect detection method and system for footwear products during production. Background Technology

[0002] Footwear manufacturing typically involves continuous processes such as cutting, sewing, bonding, pressing, spraying, molding, and packaging. Defect types include cracks, delamination, stains, color differences, indentations, excess glue, and abnormal stitching. With the development of machine vision, polarization imaging, and 3D contour detection technologies, defect detection on production lines is gradually shifting from manual visual inspection to optical inspection and algorithmic judgment. Optical inspection methods can characterize material states using reflection, scattering, color, and height variations, making them suitable for online inspection of upper materials, bonding boundaries, and sole sidewalls under non-destructive conditions.

[0003] Existing defect detection methods for footwear products largely rely on ordinary RGB images or single contour data, typically employing object detection networks, template matching, or fixed thresholds for defect identification. However, due to the natural textures of shoe upper materials, fabric weaving cycles, leather reflectivity, and rubber sidewall patterns, variations in brightness and darkness in ordinary images do not necessarily represent defects. Normal stitching, normal glue lines, and the edges of sole patterns can produce local responses similar to cracks, glue detachment, or glue overflow, making it difficult for algorithms to reliably distinguish between normal structures and true defects. Even when combining multi-angle imaging, polarization imaging, 3D contouring, and neural network classification, without a unified benchmark map corresponding to shoe type, material batch, color coding, and process flow coding, the detection process can easily misjudge material differences as defects, or detect defects spatially but fail to correlate them to specific production processes, making it difficult to support sorting, rework, and process adjustments on the production line. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that existing methods for detecting defects in footwear products have the following drawbacks: it is difficult to distinguish between material texture and actual defects; normal stitching, normal glue lines, and the edges of sole patterns are easily misjudged; the detection results are difficult to correlate with specific production processes; and how to achieve online defect detection and process traceability without damaging footwear products.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a defect detection method for footwear products during the production process, comprising acquiring multi-angle polarization detection images of footwear products and generating a footwear type partitioning reference map.

[0007] Material reflection residual maps are constructed based on shoe-type zoning benchmark maps to generate defect candidate maps.

[0008] Based on the defect candidate image, perform process association discrimination and output the defect detection results.

[0009] As a preferred embodiment of the defect detection method for footwear products during production as described in this invention, the acquisition of multi-angle polarization detection images of the footwear products includes: after the footwear products enter the closed detection station with the conveyor mechanism, the defect detection system reads the shoe type code, material batch code, color code, and process flow code from the footwear product production order; controls a ring-shaped multi-angle light source to project linearly polarized detection light in the low grazing direction, medium grazing direction, and normal direction respectively; and controls a cross-polarization camera and a coaxial camera to acquire orthogonal polarization images, parallel polarization images, and contour height maps of the upper, sole sidewall, toe cap, and heel joint within the same conveyor stop cycle. Each image is spatiotemporally aligned according to camera calibration parameters, conveyor mechanism encoder position, and shoe last positioning marks; and the aligned images are subjected to dark field response subtraction, whiteboard response normalization, and lens distortion correction to obtain multi-angle polarization detection images of the footwear products.

[0010] As a preferred embodiment of the defect detection method for footwear products in the production process described in this invention, the generation of the shoe type partitioning reference map includes: inputting a multi-angle polarization detection image of the footwear product into a shoe body partitioning network corresponding to the shoe type code; the shoe body partitioning network includes a shared convolutional encoder, a shoe upper material branch, a shoe sole edge branch, and a seam positioning branch; and outputting pixel-level partitioning results for the main upper area, splicing boundary area, gluing transition area, shoe sole sidewall area, toe abrasion-resistant area, and heel pressure area. Reference reflection templates of the same shoe type, material, and color are read from a qualified sample library according to the material batch code. Non-rigid registration of the reference reflection templates is performed using shoe last positioning marks and contour height maps to generate a shoe type partitioning reference map containing partition boundaries, material reflection mean, material reflection dispersion, normal seam width, and normal glue line position.

[0011] As a preferred embodiment of the defect detection method for footwear products in the production process described in this invention, the construction of the material reflection residual map based on the shoe shape partition reference map includes: using the shoe shape partition reference map as a constraint, mapping each pixel in the multi-angle polarization detection image of the footwear product to the corresponding shoe body partition, and extracting the diffuse reflection texture response of the orthogonal polarization image, the specular reflection response of the parallel polarization image, and the height gradient response of the contour height map. For the main body area of ​​the upper and the sidewall area of ​​the sole, the diffuse reflection texture response is compared with the mean of material reflection and normalized using the material reflection dispersion. For the splicing boundary area and the glued transition area, the specular reflection response is mutually constrained with the normal glue line position. For the abrasion-resistant area of ​​the toe and the pressure area of ​​the heel, the height gradient response is mutually constrained with the reference contour curvature. When the reflection difference of the same pixel under different projection directions is inconsistent with the normal reflection difference of the corresponding partition, the inconsistency is written into the residual channel of the corresponding pixel. When the inconsistency originates from the normal seam width, the normal glue line position, or the material texture period, the inconsistency is written into the structure suppression channel to obtain the material reflection residual map.

[0012] As a preferred embodiment of the defect detection method for footwear products in the production process described in this invention, the generation of the defect candidate map includes: jointly judging the residual channels and structural suppression channels in the material reflection residual map; establishing a residual sorting queue within each shoe body partition; marking pixels in the residual channels that exceed the upper boundary of the normal residual distribution in the same partition as initial abnormal pixels; and then using the structural suppression channel to delete initial abnormal pixels that are consistent with the normal seam width, normal glue line position, sole pattern edge, and material texture period. Connected component aggregation, thin crack path tracing, glue line overflow contour closure, and stain diffusion boundary fitting are performed on the retained initial abnormal pixels to obtain candidate regions for cracks, glue delamination, stains, color difference, and indentation. The region position, region area, region main direction, cross-partition relationship, polarization response difference, and height gradient difference of each candidate region are written into a candidate region attribute table, and the spatial boundary in the candidate region attribute table is used to cover the pixels to form a defect candidate map.

[0013] As a preferred embodiment of the defect detection method for footwear product manufacturing described in this invention, the step of performing process association discrimination based on the defect candidate image includes: matching each candidate region in the defect candidate image with the process flow code; reading the cutting process region, sewing process region, bonding process region, pressing process region, and spraying process region corresponding to the process flow code to form a process influence region table. When a candidate region is located in the splicing boundary area and the main direction of the region is consistent with the stitch direction of the sewing process region, the sewing process discrimination rule is invoked; when a candidate region is located in the gluing transition area and the polarization response difference is continuous along the glue line extension direction, the bonding process discrimination rule is invoked; when a candidate region is located in the abrasion-resistant area of ​​the toe or the pressure area of ​​the heel and the height gradient difference is concentrated along the pressing boundary, the pressing process discrimination rule is invoked. For each candidate region, positional consistency, morphological consistency, polarization consistency, and height consistency are calculated, and the candidate region is marked as a process-related candidate region or a non-process-related candidate region based on the consistency results.

[0014] As a preferred embodiment of the defect detection method for footwear product manufacturing described in this invention, the output defect detection result includes: performing defect category merging on process-related and non-process-related candidate areas respectively; and generating defect category, defect level, shoe body partition location, process source, detection image index, and re-inspection path based on the candidate area attribute table. When there are consecutive cross-partition candidate areas for delamination, cracks, or indentations in the same footwear product, the cross-partition candidate areas are merged into the same defect detection result. When there are separate candidate areas with discontinuous attribute tables in the same shoe body partition, the separate candidate areas are kept as different defect detection results. The defect detection system writes the defect detection results into the production traceability database and outputs structured inspection records containing footwear product identification, defect category, defect level, and re-inspection path to the conveying mechanism, sorting mechanism, and manual re-inspection terminal.

[0015] As a preferred embodiment of the defect detection system for footwear product manufacturing process described in this invention, it includes a partition mapping module, a residual defect screening module, and a process discrimination module.

[0016] The partition mapping module is used to acquire multi-angle polarization detection images of footwear products and generate a footwear partition reference map.

[0017] The residual screening module is used to construct a material reflection residual map based on the shoe-type partition benchmark map and generate a defect candidate map.

[0018] The process discrimination module is used to perform process association discrimination based on the defect candidate image and output the defect detection results.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps for a defect detection method in the production process of footwear products.

[0020] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements steps for a defect detection method in the production process of footwear products.

[0021] The beneficial effects of this invention are:

[0022] By acquiring multi-angle polarization detection images and generating shoe type partition reference maps, the shoe type code, material batch code, color code, and process flow code of footwear products are synchronously bound with the detection images. This enables partitioned optical detection of the upper, sole sidewall, toe edging, and heel joint. It then provides a reference for subsequent material reflection residual calculation with the same shoe type, material, and color, achieving the beneficial effect of reducing the interference of material texture differences and normal structural differences on defect detection.

[0023] By constructing a material reflection residual map and generating a defect candidate map, the orthogonal polarization response, parallel polarization response, and contour height response are constrained according to the shoe body partitions. This achieves the separation between the defect response and normal stitching, normal glue lines, shoe sole pattern edges, and material texture periods. Then, abnormal pixels are aggregated into candidate regions with category, location, orientation, and cross-partition relationships. This achieves the beneficial effect of improving the interpretability of candidate regions for cracks, glue delamination, stains, color difference, and indentation under optical inspection conditions.

[0024] By identifying process associations and outputting defect detection results, the defect candidate image is matched with the influence area table of cutting, sewing, bonding, pressing and spraying processes. This achieves a comprehensive judgment on the consistency of the candidate area in terms of position, shape, polarization and height. Then, a structured inspection record containing defect category, defect level, shoe body partition location, process source and re-inspection path is output, achieving the beneficial effect of linkage between inspection, sorting, re-inspection and production traceability. Attached Figure Description

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is an overall flowchart of a defect detection method for footwear products during production, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0027] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0028] Example 1, referring to Figure 1 As an embodiment of the present invention, a defect detection method for footwear products during the manufacturing process is provided, comprising:

[0029] S1: Acquire multi-angle polarization detection images of footwear products and generate a footwear type zoning reference map.

[0030] After footwear products enter the closed inspection station via the conveyor mechanism, the defect detection system reads the shoe type code, material batch code, color code, and process flow code from the footwear production order. It controls a ring-shaped multi-angle light source to project linearly polarized detection light in the low grazing direction, medium grazing direction, and normal direction, respectively. Simultaneously, it controls a cross-polarized camera and a coaxial camera to acquire orthogonal polarized images, parallel polarized images, and contour height maps of the upper, sole sidewalls, toe cap, and heel joint within the same conveyor stop cycle. Each image is spatiotemporally aligned according to camera calibration parameters, conveyor encoder position, and last positioning marks. The aligned images are then subjected to dark field response subtraction, whiteboard response normalization, and lens distortion correction to obtain multi-angle polarized inspection images of the footwear products.

[0031] Furthermore, the enclosed inspection station includes a light-shielding box, a conveying and docking mechanism, a shoe last positioning fixture, a ring-shaped multi-angle light source, a cross-polarized camera, a coaxial camera, and an industrial control processor. The inner wall of the light-shielding box is equipped with an matting layer. The conveying and docking mechanism provides the industrial control processor with the footwear product's arrival signal via an encoder. The shoe last positioning fixture constrains the toe and heel directions via shoe last positioning marks. The low grazing direction of the ring-shaped multi-angle light source is used to enhance the shadow of crack edges, the medium grazing direction is used to enhance the outline of glue overflow, and the normal direction is used to acquire color and stain responses. The cross-polarized camera is used to suppress strong specular reflections, and the coaxial camera is used to acquire the outline height map. This ensures that the acquisition object, acquisition angle, polarization state, and hardware triggering relationship all correspond to the defect detection scenario of footwear products.

[0032] Furthermore, the incident angle in the low grazing direction is set to 15 degrees relative to the tangent plane of the shoe body's measured surface, the incident angle in the medium grazing direction is set to 45 degrees relative to the tangent plane of the shoe body's measured surface, and the incident angle in the normal direction is set to 90 degrees relative to the tangent plane of the shoe body's measured surface. These incident angles are determined by a shoe upper material reflection calibration test. The calibration test uses qualified shoe products of the same shoe type and calibrated shoe products with standard scratches, standard glue overflow, and standard stains. The three incident angles are determined under the condition that the grayscale difference at the defect edge in the orthogonal polarization image is greater than the grayscale difference of the normal texture in the same zone. The industrial control processor triggers the light source and camera in the order of low grazing direction, medium grazing direction, and normal direction during each transport stop cycle.

[0033] Furthermore, the shoe type code in the footwear production order is used to determine the shoe's outline and the location of the last positioning mark; the material batch code is used to determine the reflectance reference range of the same batch of materials; the color code is used to determine the color compensation matrix after the whiteboard response is normalized; and the process flow code is used to determine the process influence area table required for subsequent process association judgment. These codes are written into RFID tags or QR code tags by the production traceability database before the footwear products enter the closed inspection station. The code reader transmits these codes to the industrial control processor, which binds them to the image index of each inspection image to prevent the same inspection image from deviating from the corresponding shoe type and material conditions in subsequent processing stages.

[0034] Furthermore, dark field response subtraction involves acquiring a dark field image by turning off the ring multi-angle light source, and then subtracting the camera noise response of the corresponding pixels in the dark field image from the detected image. White board response normalization involves acquiring a diffuse reflection standard white board image under the same exposure parameters, dividing the corresponding pixel in the detected image by the standard white board response and multiplying it by the standard white board reflectance coefficient. Lens distortion correction calculates the camera intrinsic parameters and distortion parameters using a checkerboard calibration board, and then resamples the detected image to the shoe last positioning mark coordinate system. After correction, each pixel has three types of indices: shoe surface coordinates, camera pixel coordinates, and encoder position of the conveyor mechanism.

[0035] Multi-angle polarization detection images of footwear products are input into a shoe body partitioning network corresponding to the shoe type code. This network includes a shared convolutional encoder, an upper material branch, a sole edge branch, and a seam positioning branch. It outputs pixel-level partitioning results for the main upper area, splicing boundary area, gluing transition area, sole sidewall area, toe abrasion-resistant area, and heel pressure area. Reference reflection templates of the same shoe type, material, and color are retrieved from a qualified sample library according to the material batch code. Non-rigid registration of these templates is performed using last positioning marks and contour height maps to generate a shoe type partitioning reference map containing partition boundaries, material reflection mean, material reflection dispersion, normal seam width, and normal glue line position.

[0036] Furthermore, the shared convolutional encoder of the shoe body partitioning network includes a first convolutional layer, a first normalization layer, a first activation layer, a second convolutional layer, a second normalization layer, a second activation layer, and a multi-scale feature fusion layer connected in sequence. The upper material branch connects to the multi-scale feature fusion layer and outputs the partitioning probabilities of the upper body area, the toe abrasion-resistant area, and the heel pressure area. The sole edge branch connects to the multi-scale feature fusion layer and outputs the partitioning probabilities of the sole sidewall area and the glued transition area. The seam positioning branch connects to the multi-scale feature fusion layer and outputs the centerline probability and width probability of the splicing boundary area. The training samples are obtained from manually labeled qualified footwear product inspection images. The loss function consists of partitioning cross-entropy loss, boundary smoothing loss, and seam centerline regression loss. The output maximum probability partitioning result is used as the pixel-level partitioning result.

[0037] Furthermore, the qualified sample library establishes a three-level index based on shoe type code, material batch code, and color code. Each reference reflection template is obtained by statistical analysis of multi-angle polarization detection images of footwear products that have passed manual re-inspection within the same production batch. The mean material reflection is the median of the pixel response at the same angle in the same partition. The material reflection dispersion is the median absolute deviation of the pixel response at the same angle in the same partition multiplied by 1.4826. The normal seam width is obtained by statistical analysis of the width probability peak value output by the seam positioning branch. The normal glue line position is obtained by statistical analysis of the center line of the continuous bright band of the parallel polarization response in the glue transition zone. All of the above statistical values ​​are written into the shoe type partition reference map.

[0038] Furthermore, the non-rigid registration uses the shoe last positioning mark as the starting point for rigid registration, and the highest point of the toe, the heel endpoint, the upper edge of the sole sidewall, and the center line of the splicing boundary in the contour height map as constraint points. A thin-plate spline transformation is used to map the reference reflection template to the current footwear product coordinate system. The bending energy weight of the thin-plate spline transformation is set to 0.01, derived from the contour height deviation statistics of 50 pairs of qualified footwear products of the same shoe type. When the weight is higher than 0.01, the sole sidewall will become over-smoothed; when the weight is lower than 0.01, the splicing boundary will produce local jitter. Therefore, 0.01 is used as the registration weight for the shoe type partitioning reference map.

[0039] It should be noted that this step, through multi-angle polarization acquisition and construction of shoe type partitioning reference maps, enables the detection data to carry material, shoe type and process indexes, providing a unified reference for reflection difference measurement in optical inspection and reducing the interference of shoe material texture and structural edges on subsequent defect judgment.

[0040] S2: Construct a material reflection residual map based on the shoe-type partition benchmark map to generate a defect candidate map.

[0041] Using the shoe-type partitioning reference map as a constraint, each pixel in the multi-angle polarization detection image of the footwear product is mapped to the corresponding shoe partition. The diffuse reflection texture response of the orthogonal polarization image, the specular reflection response of the parallel polarization image, and the height gradient response of the contour height map are extracted respectively. For the main body area of ​​the upper and the sidewall area of ​​the sole, the diffuse reflection texture response is compared with the mean material reflection and normalized using the material reflection dispersion. For the splicing boundary area and the glued transition area, the specular reflection response is mutually constrained with the normal glue line position. For the abrasion-resistant area of ​​the toe and the pressure-bearing area of ​​the heel, the height gradient response is mutually constrained with the reference contour curvature. When the reflection difference of the same pixel under different projection directions is inconsistent with the normal reflection difference of the corresponding partition, the inconsistency is written into the residual channel of the corresponding pixel. When the inconsistency originates from the normal seam width, the normal glue line position, or the material texture periodicity, the inconsistency is written into the structure suppression channel, resulting in a material reflection residual map.

[0042] Furthermore, the material reflection residual map includes a residual channel, a structural suppression channel, a partition index channel, and an angle source channel. The residual channel records the abnormal reflection intensity caused by defects; the structural suppression channel records the interpretable structural response generated by normal seams, normal glue lines, sole pattern edges, and material texture periods; the partition index channel records the main upper area, splicing boundary area, glue transition area, sole sidewall area, toe abrasion-resistant area, or heel pressure area to which each pixel belongs; and the angle source channel records that the residual mainly originates from the low grazing direction, medium grazing direction, or normal direction. The four channels use the same pixel coordinates, enabling subsequent defect candidate maps to simultaneously read abnormal intensity and normal structural interpretations.

[0043] Furthermore, the diffuse reflection texture response of the orthogonal polarization image is obtained by removing the low-frequency illumination field, which is calculated using a guided filter with a radius three times the normal seam width of the main body area of ​​the shoe upper. The specular reflection response of the parallel polarization image is obtained by the pixel difference between the parallel polarization image and the orthogonal polarization image, and a one-dimensional continuity constraint is applied along the normal glue line position direction within the glue transition zone. The height gradient response of the contour height map is calculated by the height difference between the coordinates of adjacent shoe surfaces, and a directional constraint is applied along the curvature direction of the reference contour within the abrasion-resistant area of ​​the toe and the pressure-bearing area of ​​the heel, ensuring that cracks, delamination, stains, color differences, and indentations enter the material reflection residual map in different response combinations.

[0044] Furthermore, the first [item] in the material reflection residual diagram The material reflection residual intensity of each pixel is calculated according to the following formula:

[0045]

[0046] in, Indicates the first Material reflection residual intensity per pixel. This indicates the pixel number in the multi-angle polarization detection image of footwear products. Indicates the projection direction number. This indicates the total number of projection directions involved in the calculation. Indicates the first The pixel in the first Orthogonal polarization response under each projection direction. Indicating the shoe type zoning reference map with the first The pixel and the The average material reflection value corresponding to each projection direction. Indicates the first The pixel in the first The polarization difference coefficient is calculated from parallel polarization images and orthogonal polarization images under each projection direction. Indicates the first The normal texture periodicity stability coefficient for each pixel in the shoe body partition. The material reflection residual intensity is written into the residual channel, and the normal texture periodicity stability coefficient is used to reduce the influence of natural leather texture or fabric texture on the residual channel.

[0047] Furthermore, the writing conditions for the structure suppression channel are determined by the normal seam width, normal glue line position, and material texture period in the shoe type partitioning reference map. When the distance from the residual pixel to the center line of the normal seam is less than half the normal seam width and the angle between the residual principal direction and the seam center line does not exceed 12 degrees, the structure suppression channel writes the seam structure identifier. When the distance from the residual pixel to the center line of the normal glue line is less than 0.6 times the normal glue line width and the parallel polarization response is continuous along the glue line direction, the structure suppression channel writes the glue line structure identifier. When the residual pixels exhibit a periodic repetition distribution in the material texture period direction, the structure suppression channel writes the texture structure identifier. The 12 degrees and 0.6 times are derived from the statistical analysis of the direction and position deviations of the normal structure boundaries in qualified samples of the same shoe type.

[0048] The residual channels and structural suppression channels in the material reflection residual map are jointly evaluated. First, a residual sorting queue is established within each shoe body partition. Pixels in the residual channels that exceed the upper bound of the normal residual distribution in the same partition are marked as initial anomalous pixels. Then, the structural suppression channel is used to delete initial anomalous pixels that are consistent with the normal seam width, normal glue line position, sole pattern edge, and material texture period. Connected component aggregation, thin crack path tracing, glue line overflow contour closure, and stain diffusion boundary fitting are performed on the retained initial anomalous pixels to obtain candidate regions for cracks, glue delamination, stains, color difference, and indentation. The region location, region area, region principal direction, cross-partition relationship, polarization response difference, and height gradient difference of each candidate region are written into the candidate region attribute table, and the spatial boundary in the candidate region attribute table is used to cover the pixels to form a defect candidate map.

[0049] Furthermore, the upper bound of the normal residual distribution within the same zone is calculated from qualified footwear products of the same style, material, and color in the qualified sample library. Specifically, it is the median of the material reflection residual intensity within the same zone plus 2.8 times the median of the absolute deviation. The 2.8 times factor is derived from the percentile statistics of the residuals of normal texture, normal glue lines, and normal stitching in the qualified samples. When the material batch code of the current production batch changes, the defect detection system recalculates the upper bound of the normal residual distribution within the same zone to avoid misjudging different leather, fabric, or rubber materials using a fixed threshold.

[0050] Furthermore, connected component aggregation employs an eight-neighbor rule to connect initial anomalous pixels within the same shoe body partition. Cross-partition connections are only allowed between the splicing boundary area and the main upper area, the glued transition area and the sole sidewall area, and the toe abrasion zone and the main upper area. Slender crack path tracing starts from the low-grazing-direction residual enhancement region and searches for anomalous paths along the main direction of the region whose width is less than half the normal seam width and whose length is greater than 5 times the width. Glue overflow contour closure starts from the continuous region of parallel polarization response and closes the boundary on both sides of the normal glue line position. Stain diffusion boundary fitting starts from the region of abrupt change in color response in the normal direction and fits a boundary where brightness or chromaticity gradually decreases. Indentation candidate regions are generated starting from the region of concentrated height gradient difference.

[0051] Furthermore, the region position in the candidate region attribute table is represented by both shoe surface coordinates and image pixel coordinates. The region area uses the registered shoe surface area rather than the number of image pixels. The main direction of the region is determined by the major axis of the smallest circumscribed ellipse of the candidate region. The cross-zone relationship records whether the candidate region crosses the zone boundary. The polarization response difference records the difference between parallel polarization response and orthogonal polarization response. The height gradient difference records the difference between the height gradient response and the baseline contour curvature within the candidate region. The defect candidate image is stored using the candidate region attribute table as an index, and each pixel can be traced back to the corresponding candidate region, the corresponding shoe zone, and the corresponding projection direction.

[0052] It should be noted that this step combines multi-angle polarization response, shoe shape partitioning reference map, and structural suppression channel to solve the problem that ordinary image detection is difficult to distinguish between defects, textures, seams, and glue lines. Material reflection residual intensity provides measurable residuals in the sense of optical detection, and defect candidate map provides traceable spatial boundaries, enabling subsequent process correlation judgments to be performed simultaneously based on physical response and shoe structure.

[0053] S3: Perform process association judgment based on the defect candidate map and output the defect detection results.

[0054] Each candidate region in the defect candidate map is matched with the process flow code. The cutting process region, sewing process region, bonding process region, pressing process region, and spraying process region corresponding to the process flow code are read to form a process influence region table. When the candidate region is located in the splicing boundary area and the main direction of the region is consistent with the stitch direction of the sewing process region, the sewing process discrimination rule is called. When the candidate region is located in the gluing transition area and the polarization response difference is continuous along the extension direction of the glue line, the bonding process discrimination rule is called. When the candidate region is located in the abrasion-resistant area of ​​the toe or the pressure area of ​​the heel and the height gradient difference is concentrated along the pressing boundary, the pressing process discrimination rule is called. For each candidate region, positional consistency, morphological consistency, polarization consistency, and height consistency are calculated, and the candidate region is marked as a process-related candidate region or a non-process-related candidate region based on the consistency results.

[0055] Furthermore, the process influence area table is generated from the process route database. The cutting process area corresponds to the material edge of the main body area of ​​the upper and the boundary of the abrasion-resistant area of ​​the toe. The sewing process area corresponds to the center line of the stitch in the splicing boundary area and the normal seam width range on both sides. The bonding process area corresponds to the normal glue line position in the gluing transition area and the glue line width range on both sides. The pressing process area corresponds to the pressure boundary of the abrasion-resistant area of ​​the toe, the pressure area of ​​the heel, and the sidewall area of ​​the sole. The spraying process area corresponds to the color coverage area of ​​the main body area of ​​the upper and the sidewall area of ​​the sole. The defect detection system projects the regional location of the candidate areas onto the process influence area table, enabling the candidate areas to establish a one-to-one or one-to-many correspondence with the actual production processes.

[0056] Furthermore, the sewing process discrimination rules read the main direction, width, and cross-zone relationship of the candidate area. When the candidate area is discontinuously distributed along the stitch direction and its width is less than the normal stitch width, it is marked as a candidate for sewing tear. When the candidate area has a symmetrical high residual distribution along both sides of the stitch and the height gradient difference is not concentrated, it is marked as a candidate for stitch stain. The bonding process discrimination rules read the continuity of the parallel polarization response of the candidate area and the distance to the normal glue line position. When the candidate area is continuous along the glue line direction and crosses the bonding transition zone and the sidewall area of ​​the sole, it is marked as a candidate for glue delamination. When the candidate area is located outside the glue line and the polarization response difference is higher than the upper limit of the normal residual distribution in the same zone, it is marked as a candidate for glue overflow. The pressing process discrimination rules read the height gradient difference of the candidate area and the pressing boundary distance. When the height gradient difference is concentrated along the pressing boundary and the color response in the normal direction does not change abruptly, it is marked as a candidate for indentation.

[0057] Furthermore, the process association consistency value between the candidate region and the process influence region is calculated according to the following formula:

[0058]

[0059] in, Indicates the first The process association consistency value of each candidate region. This indicates the candidate region number in the defect candidate map. Indicates the sequence number of the process-related features. This represents the total number of process-related features. Indicates the first The candidate region in the first The detection attribute values ​​on the process association features include positional consistency, morphological consistency, polarization consistency, and height consistency. The table showing the influence area of ​​a process is related to the first... The candidate region corresponds to the first process. Each process is associated with a baseline value. The process association consistency value is between 0 and 1, with the closer it is to 1, the more the candidate area matches the defect pattern generated by the corresponding process.

[0060] Furthermore, the classification rule for process association consistency values ​​is as follows: when the process association consistency value is not less than 0.72 and the location of the candidate region falls within the corresponding process influence area table, the candidate region is marked as a process-related candidate region. When the process association consistency value is less than 0.72 or the location of the candidate region does not fall within the corresponding process influence area table, the candidate region is marked as a non-process-related candidate region. 0.72 is calculated from process defect samples confirmed by manual re-inspection in historical production batches. The calculation method is to obtain the boundary value that ensures a process-related defect recall rate of not less than 95% and a non-process-related false association rate of minimal. The defect detection system confirms the above boundary value based on no less than 30 pairs of samples when each shoe model code is first launched.

[0061] The system performs defect category merging separately for process-related and non-process-related candidate areas, generating defect category, defect level, shoe body partition location, process source, inspection image index, and re-inspection path based on the candidate area attribute table. When there are consecutive cross-partition candidate areas for delamination, cracks, or indentations within the same footwear product, these cross-partition candidate areas are merged into a single defect detection result. When there are separate candidate areas within the same shoe body partition with discontinuous attribute tables, these separate candidate areas are maintained as separate defect detection results. The defect detection system writes the defect detection results into the production traceability database and outputs structured inspection records containing footwear product identification, defect category, defect level, and re-inspection path to the conveying mechanism, sorting mechanism, and manual re-inspection terminal.

[0062] Furthermore, defect category merging is performed according to the candidate region attribute table in the defect candidate map. Crack candidate regions are merged based on the continuity of the main direction of the region and the path of the slender crack; delamination candidate regions are merged based on the cross-zone continuity between the glue transition zone and the sole sidewall area; stain candidate regions are merged based on the overlap of the color response diffusion boundary in the normal direction; color difference candidate regions are merged based on the color compensation difference corresponding to the color code within the main area of ​​the upper; and indentation candidate regions are merged based on the continuity of the height gradient difference along the pressing boundary. Even if candidate regions of different defect categories are spatially adjacent, separate defect detection results are generated to avoid merging adjacent stains and delamination into a single defect.

[0063] Furthermore, defect levels are determined based on defect type, area, cross-zone relationship, and process origin. Crack and delamination candidate areas have higher priority than stain and color difference candidate areas. Continuous candidate areas across zones have higher priority than single-zone candidate areas. Indentation candidate areas originating from the pressing process area and exhibiting a high concentration of gradient differences are assigned separate indentation levels. The re-inspection path includes the footwear product's stop number on the conveyor mechanism, the inspection image index, the shoe surface coordinates, the defect candidate image boundary, and the corresponding process origin. The manual re-inspection terminal displays a magnified view, the corresponding projection direction image, and the candidate area attribute table according to the re-inspection path.

[0064] Furthermore, when defect detection results are written into the production traceability database, they are linked to footwear product identification, production orders, material batch codes, process flow codes, inspection time, and inspection equipment numbers. The conveying mechanism executes options for continued conveying, slow conveying, or stop for re-inspection based on the defect level. The sorting mechanism switches between qualified, rework, or isolated channels based on the defect level. The manual re-inspection terminal displays the candidate area location and process origin based on the re-inspection path. The re-inspection conclusion is written back to the production traceability database and used to update the qualified sample library and process defect sample library.

[0065] It should be noted that this step binds the candidate defect image with the actual production process area, process flow code, and re-inspection path, so that the detection results not only indicate whether a defect exists, but also indicate the corresponding location of the defect and its possible process origin. The process association consistency value overcomes the problem that simple image classification cannot trace the production cause, and improves the usability and closed-loop control capability of defect detection results in the footwear product production inspection system.

[0066] Example 2, an embodiment of the present invention, provides a defect detection system for footwear product manufacturing process, including a partition mapping module, a residual defect screening module, and a process discrimination module.

[0067] The zoning mapping module is used to acquire multi-angle polarization detection images of footwear products and generate a zoning reference map of the footwear type.

[0068] The residual screening module is used to construct material reflection residual maps based on shoe-type partition benchmark maps and generate defect candidate maps.

[0069] The process discrimination module is used to perform process association discrimination based on the defect candidate image and output the defect detection results.

Claims

1. A method for defect detection during the production of footwear products, characterized in that, include: Collect multi-angle polarization detection images of footwear products and generate a footwear type zoning reference map; Based on the shoe-type zoning baseline map, a material reflection residual map is constructed to generate a defect candidate map; Based on the defect candidate image, perform process association discrimination and output the defect detection results.

2. The defect detection method for footwear products as described in claim 1, characterized in that: The acquisition of multi-angle polarization detection images of footwear products includes... After the footwear products enter the closed inspection station with the conveyor, the defect detection system reads the shoe type code, material batch code, color code and process flow code from the footwear product production order. It controls the ring multi-angle light source to project linearly polarized detection light in the low grazing direction, medium grazing direction and normal direction respectively, and controls the cross-polarized camera and coaxial camera to acquire orthogonal polarized images, parallel polarized images and contour height maps of the upper, sole sidewall, toe edging and heel joint within the same conveyor stop cycle. Each image is spatiotemporally aligned according to camera calibration parameters, encoder position of conveyor mechanism, and shoe last positioning mark. Dark field response subtraction, whiteboard response normalization, and lens distortion correction are performed on the aligned images to obtain multi-angle polarization detection images of footwear products.

3. The defect detection method for footwear products as described in claim 2, characterized in that: The generated shoe type zoning reference map includes: The multi-angle polarization detection image of the footwear product is input into the footwear partitioning network corresponding to the footwear type code. The footwear partitioning network includes a shared convolutional encoder, a footwear material branch, a sole edge branch, and a seam positioning branch. It outputs pixel-level partitioning results for the main body area of ​​the footwear, the splicing boundary area, the gluing transition area, the sole sidewall area, the toe abrasion-resistant area, and the heel pressure area. According to the material batch code, the same shoe type, material and color reference reflection template is read from the qualified sample library. The reference reflection template is non-rigidly registered using the shoe last positioning mark and contour height map to generate a shoe type partition reference map that includes partition boundaries, material reflection mean, material reflection dispersion, normal seam width and normal glue line position.

4. The defect detection method for footwear products as described in claim 3, characterized in that: The construction of the material reflection residual map based on the shoe-type zoning benchmark map includes... Using the shoe type partition reference map as a constraint, each pixel in the multi-angle polarization detection image of the footwear product is mapped to the corresponding shoe body partition, and the diffuse reflection texture response of the orthogonal polarization image, the specular reflection response of the parallel polarization image, and the height gradient response of the contour height map are extracted respectively. For the main body area of ​​the upper and the sidewall area of ​​the sole, the diffuse reflection texture response is compared with the mean material reflection and normalized by the material reflection dispersion. For the splicing boundary area and the glued transition area, the specular reflection response is mutually constrained with the normal glue line position. For the abrasion-resistant area of ​​the toe and the pressure area of ​​the heel, the height gradient response is mutually constrained with the reference contour curvature. When the reflection difference of the same pixel under different projection directions is inconsistent with the normal reflection difference of the corresponding partition, the inconsistency amount is written into the residual channel of the corresponding pixel. When the inconsistency amount comes from the normal seam width, normal glue line position or material texture period, the inconsistency amount is written into the structure suppression channel to obtain the material reflection residual map.

5. The defect detection method for footwear products as described in claim 4, characterized in that: The generated defect candidate map includes, The residual channels and structural suppression channels in the material reflection residual map are jointly judged. A residual sorting queue is established in each shoe body partition. Pixels in the residual channels that exceed the upper limit of the normal residual distribution in the same partition are marked as initial abnormal pixels. Then, the structural suppression channel is used to delete the initial abnormal pixels that are consistent with the normal seam width, normal glue line position, shoe sole pattern edge and material texture period. Connectivity aggregation, thin crack path tracing, glue line overflow contour closure, and stain diffusion boundary fitting are performed on the retained initial abnormal pixels to obtain candidate regions for cracks, glue delamination, stains, color difference, and indentation. The region location, region area, region main direction, cross-regional relationship, polarization response difference, and height gradient difference of each candidate region are written into the candidate region attribute table, and the pixels are covered by the spatial boundary in the candidate region attribute table to form a defect candidate map.

6. The defect detection method for footwear products as described in claim 5, characterized in that: The process association determination based on the defect candidate image is performed. include, Match each candidate region in the defect candidate map with the process flow code, and read the cutting process region, sewing process region, bonding process region, pressing process region and spraying process region corresponding to the process flow code to form a process influence region table; When the candidate area is located in the splicing boundary area and the main direction of the area is consistent with the stitch direction of the sewing process area, the sewing process discrimination rule is invoked. When the candidate area is located in the gluing transition area and the polarization response difference is continuous along the glue line extension direction, the bonding process discrimination rule is invoked. When the candidate area is located in the abrasion-resistant area of ​​the toe or the pressure area of ​​the heel and the height gradient difference is concentrated along the pressing boundary, the pressing process discrimination rule is invoked. For each candidate region, calculate positional consistency, morphological consistency, polarization consistency, and height consistency. Based on the consistency results, mark the candidate region as a process-related candidate region or a non-process-related candidate region.

7. The defect detection method for footwear products as described in claim 6, characterized in that: The output defect detection results include, Defect category merging is performed on candidate regions related to the process and candidate regions not related to the process, and defect category, defect level, shoe body partition location, process source, inspection image index and re-inspection path are generated based on the candidate region attribute table; When there are continuous candidate areas for delamination, cracks, or indentations across different zones in the same footwear product, the candidate areas across zones will be merged into the same defect detection result. When there are separate candidate areas with discontinuous attribute tables in the same shoe body zone, the separate candidate areas will be kept as different defect detection results. The defect detection system writes the defect detection results into the production traceability database and outputs structured inspection records containing footwear product identification, defect category, defect level, and re-inspection path to the conveying mechanism, sorting mechanism, and manual re-inspection terminal.

8. A defect detection system for footwear products during manufacturing, employing the defect detection method for footwear products as described in any one of claims 1 to 7, characterized in that: Includes a partition mapping module, a residual screening module, and a process identification module; The partition mapping module is used to acquire multi-angle polarization detection images of footwear products and generate a footwear partition reference map; The residual screening module is used to construct a material reflection residual map based on the shoe type partition benchmark map and generate a defect candidate map. The process discrimination module is used to perform process association discrimination based on the defect candidate image and output the defect detection results.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the defect detection method for footwear products as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the defect detection method for footwear products as described in any one of claims 1 to 7.

Images