Industrial vision-based defect detection system for shoe production lines

Abstract

This invention relates to the field of industrial vision inspection and footwear production automation technology, specifically to a defect detection system for a footwear production line based on industrial vision. The system includes a control terminal and two-dimensional vision sensors and three-dimensional structured light sensors connected to it. The control terminal incorporates a multi-source data acquisition module, a flexible adaptive image registration module, a physical property conversion module, a visual stress coupling calculation module, and a dynamic tolerance decision feedback module. These modules are used to acquire two-dimensional visual image data of the footwear, three-dimensional structural point cloud data, and a baseline stress topology heatmap; extract visually abnormal area data; calculate the local strength attenuation matrix and failure risk index; and perform quality classification decisions and process closed-loop feedback to determine whether visual differences are sufficient to translate into actual failures. This invention overcomes the shortcomings of traditional isolated data analysis and improves the accuracy of abnormal area extraction.

Description

Technical Field

[0001] This invention relates to the field of industrial vision inspection and footwear production automation technology, specifically a defect detection system for footwear production lines based on industrial vision. Background Technology

[0002] In the field of defect detection in footwear production lines, industrial vision technology has gradually become the main means of quality control in final assembly processes such as shoe bonding and pressing, because it can achieve online, non-contact and continuous detection. Currently, defect detection in shoe production lines generally supports the identification of apparent anomalies based on two-dimensional images or three-dimensional contour data. If a more in-depth quality assessment of shoe defects is required, it is usually necessary to combine manual re-inspection, experience-based judgment, or offline durability testing to obtain a further evaluation of the consequences of the defects. However, when using the above methods to detect defects on a shoe production line, it can usually only determine whether there are appearance differences. This leads to fragmented information on abnormalities such as seams, glue lines, textures, and height variations, making it impossible to combine them with the flexible deformation characteristics of footwear for unified analysis. Furthermore, it cannot correlate abnormal areas with material physical properties and the distribution of dynamic stress during wear, resulting in difficulty in distinguishing between harmless appearance deviations and actual failure risks. At the same time, this approach is prone to false positives and false negatives when dealing with complex conditions such as reflection, occlusion, posture fluctuations, and missing point clouds. It is also difficult to provide feedback and adjustments to upstream gluing, pressing, and positioning processes based on the spatial concentration trend of high-risk defects, resulting in a disconnect between quality judgment and process optimization. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a defect detection system for shoe production lines based on industrial vision. Specifically, the technical solution of this invention includes: The control terminal and the two-dimensional vision sensor and three-dimensional structured light sensor that are connected to the control terminal in communication. The control terminal is equipped with a multi-source data acquisition module, a flexible adaptive image registration module, a physical property conversion module, a visual stress coupling calculation module, and a dynamic tolerance decision feedback module. The multi-source data acquisition module is used to acquire two-dimensional visual image data and three-dimensional structural point cloud data of the target object through a two-dimensional vision sensor and a three-dimensional structured light sensor, and to obtain a pre-generated benchmark stress topology heat map from a preset database. The flexible adaptive image registration module is used to map two-dimensional visual image data and three-dimensional structural point cloud data to a preset standard topology unfolded map, and compare the difference with the standard features of the standard topology unfolded map to extract visual abnormal region data. The physical property conversion module is used to calculate the local strength attenuation matrix based on visual anomaly area data and preset material physical parameters including adhesive viscosity parameters; The visual stress coupling calculation module is used to perform spatial coordinate alignment and convolution operation between the local intensity attenuation matrix and the benchmark stress topology heatmap to obtain the failure risk index. The dynamic tolerance decision feedback module is used to perform quality classification decision operations and process closed-loop feedback operations based on the failure risk index.

[0004] Preferably, the process of the multi-source data acquisition module acquiring the benchmark stress topology heat map includes: acquiring the computer-aided design model corresponding to the target object; performing dynamic stress simulation on the computer-aided design model based on a preset finite element analysis algorithm to generate the benchmark stress topology heat map; wherein, the benchmark stress topology heat map includes the dynamic stress weight coefficients of each region of the target object.

[0005] Preferably, the process of the flexible adaptive image registration module extracting visually abnormal region data includes: performing topological mapping between two-dimensional visual image data and three-dimensional structural point cloud data and a standard topological unfolded map based on a non-rigid image registration algorithm to obtain mapping feature data; extracting feature variation regions from the mapping feature data and marking the feature variation regions as visually abnormal region data; wherein, the visually abnormal region data includes abnormal area data and abnormal depth data.

[0006] Preferably, the process of the physical property conversion module calculating the local intensity attenuation matrix includes: extracting the adhesive viscosity parameter from the material physical parameters; multiplying the abnormal area data and abnormal depth data to obtain abnormal volume data, and converting the adhesive viscosity parameter into a dimensionless equivalent viscosity parameter according to a pre-established mapping relationship; quotienting the abnormal volume data and the equivalent viscosity parameter and performing normalization processing to obtain the intensity reduction percentage of each pixel or voxel point, wherein the normalization processing adopts the max-min normalization algorithm, using a pre-calibrated reference volume deviation limit value as the denominator, mapping the quotient value to the range of 0 to 100 to obtain a dimensionless value as the intensity reduction percentage, so as to generate a local intensity attenuation matrix with the same spatial dimension as the visual abnormal area data; wherein the local intensity attenuation matrix contains the intensity reduction percentage of each pixel or voxel point.

[0007] Preferably, the process of aligning the local intensity attenuation matrix with the reference stress topology heatmap in the visual stress coupling calculation module includes: extracting the geometric center coordinates and contour boundary coordinates of the local intensity attenuation matrix on the standard topology unfolded diagram as the first spatial coordinate data; extracting the coordinates of reference points on the reference stress topology heatmap that match the geometric center coordinates and contour boundary coordinates as the second spatial coordinate data; and performing coordinate system registration between the first spatial coordinate data and the second spatial coordinate data based on an affine transformation algorithm to obtain coordinate alignment data.

[0008] Preferably, the process of obtaining the failure risk index by the visual stress coupling calculation module includes: extracting the dynamic stress weight coefficient corresponding to the local intensity attenuation matrix in the benchmark stress topology heatmap based on coordinate alignment data; performing convolution operation on the local intensity attenuation matrix and the dynamic stress weight coefficient to obtain the risk matrix; summing the values ​​in the risk matrix to output the failure risk index.

[0009] Preferably, the process of the dynamic tolerance decision feedback module performing quality classification decision operation includes: preset a first risk threshold and a second risk threshold, wherein the first risk threshold is less than the second risk threshold; determining the relationship between the failure risk index and the first and second risk thresholds; if the failure risk index is less than or equal to the first risk threshold, generating a qualified release instruction; if the failure risk index is greater than the first risk threshold and less than or equal to the second risk threshold, generating a downgrade processing instruction; if the failure risk index is greater than the second risk threshold, generating a rework and scrap instruction.

[0010] Preferably, the process of the dynamic tolerance decision feedback module executing the process closed-loop feedback operation includes: extracting the target spatial coordinate position corresponding to the visual abnormal area data that generates rework and scrap instructions; counting the cumulative number of times rework and scrap instructions are generated at the target spatial coordinate position within a preset time window or preset detection batch; setting a preset number threshold; judging the relationship between the cumulative number and the number threshold; if the cumulative number is greater than or equal to the number threshold, generating a process adjustment warning signal and outputting the process adjustment warning signal to external production control equipment for adjusting the production parameters of the upstream process of the target object; if the cumulative number is less than the number threshold, maintaining the current production state and continuing to execute data acquisition operations.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention acquires two-dimensional and three-dimensional data through a multi-source data acquisition module and maps them to a standard topological unfolded map using a flexible adaptive image registration module. This setup effectively integrates surface texture and spatial height information, eliminates interference from the normal elastic deformation of flexible footwear products, overcomes the defects of traditional isolated data analysis, and improves the accuracy of abnormal area extraction. 2. This invention combines the extracted abnormal area and depth data with material physical parameters such as adhesive viscosity through a physical property conversion module to calculate and generate a local strength attenuation matrix. This setting transforms simple geometric appearance anomalies into physical quantitative data reflecting a decrease in local bonding or support capacity, realizing an engineering transition from appearance differences to structural weakening. 3. This invention utilizes a visual stress coupling calculation module to spatially align and convolve the local strength attenuation matrix with the benchmark stress topology heatmap to obtain the failure risk index. This setting correlates defect assessment with the dynamic stress consequences during actual wear, enabling accurate differentiation between harmless apparent deviations and real structural failure risks, effectively reducing the misjudgment rate within the preset tolerance range. 4. This invention uses a dynamic tolerance decision feedback module to perform quality classification decisions based on the failure risk index, and outputs a process adjustment warning signal upstream when high-risk anomalies occur continuously at specific spatial coordinate positions. This setting solves the technical problem of the disconnect between quality judgment and process optimization, and realizes closed-loop control from single-piece defect screening to automated correction of production line parameters. Attached Figure Description

[0012] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a schematic diagram of the modules of the shoe production line defect detection system based on industrial vision provided in the embodiments of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0014] The defect detection system for shoe production lines based on industrial vision includes a control terminal and two-dimensional vision sensors and three-dimensional structured light sensors that communicate with the control terminal. The control terminal is equipped with a multi-source data acquisition module, a flexible adaptive image registration module, a physical property conversion module, a visual stress coupling calculation module, and a dynamic tolerance decision feedback module. The multi-source data acquisition module is used to acquire two-dimensional visual image data and three-dimensional structural point cloud data of the target object through a two-dimensional vision sensor and a three-dimensional structured light sensor, and to obtain a pre-generated benchmark stress topology heat map from a preset database. The flexible adaptive image registration module is used to map two-dimensional visual image data and three-dimensional structural point cloud data to a preset standard topology unfolded map, and compare the differences with the standard features of the standard topology unfolded map to extract visual abnormal area data. The physical property conversion module is used to calculate the local strength attenuation matrix based on visual anomaly area data and preset material physical parameters including adhesive viscosity parameters; The visual stress coupling calculation module is used to perform spatial coordinate alignment and convolution operation between the local intensity attenuation matrix and the benchmark stress topology heatmap to obtain the failure risk index. The dynamic tolerance decision feedback module is used to perform quality classification decision operations and process closed-loop feedback operations based on the failure risk index; This embodiment provides a defect detection mechanism for shoe production lines based on industrial vision, such as... Figure 1 As shown; specifically, this embodiment takes a sports shoe upper and outsole bonding assembly line as the main scenario. This production line continuously produces the same shoe model. The shoe upper material includes a fabric layer, a hot-melt reinforcement layer, and an adhesive layer. After the shoe body is glued, bonded, and pressed, it enters the online inspection station. The control terminal is arranged in the cabinet on the side of the production line, and two-dimensional vision sensors are set on the top and oblique side of the shoe body to collect two-dimensional appearance information such as seams, glue lines, surface texture, and color transition. The three-dimensional structured light sensor is set within the scanning field of the toe bending area, the side fitting area, and the heel covering area to collect the undulation height of the shoe upper, the thickness of the glue line accumulation, the collapse depth, and the contour deviation. After receiving the above data, the control terminal sequentially completes multi-source acquisition, flexible registration, physical property conversion, visual stress coupling calculation, and dynamic tolerance decision-making. Therefore, it does not simply judge whether there is a visual difference, but further determines whether the difference is sufficient to be transformed into a real structural failure during the wearing stress process. In practice, the multi-source data acquisition module collects multiple state information of the same shoe within the same detection cycle; the two-dimensional image reflects surface anomalies, such as missing glue edges, skipped stitches, material wrinkles, and contamination points; the three-dimensional point cloud reflects spatial anomalies, such as local pits, bulges, glue layer build-up, and bonding gaps; for flexible footwear products, relying solely on two-dimensional images may easily misjudge natural textures and differences in leather stretching as defects, while relying solely on three-dimensional point clouds may ignore thin-layer anomalies such as color discontinuities and broken stitches. Therefore, the simultaneous input of both is to characterize surface texture and boundary change information as well as spatial contour and height change information, respectively. The control terminal pre-stores a standard topology unfolded diagram of the shoe model. This unfolded diagram maps areas such as the toe, side, heel, and collar to a standard shape under a unified reference coordinate system, ensuring that different shoe bodies can still fall within the same analysis framework even when deformation errors exist within tolerance ranges. The flexible adaptive image registration module maps the actual acquired two-dimensional images and three-dimensional point clouds to this standard topology unfolded diagram to determine whether there are any abnormal areas deviating from the standard state in each region. The physical property conversion module further combines material physical parameters to convert visual anomalies into local strength attenuation matrices, characterizing the degree of decrease in bonding, support, or adhesion capabilities at that location compared to the ideal state. The visual stress coupling calculation module then overlays the local strength attenuation matrix with the dynamic stress distribution of the shoe during wear for analysis: if the anomaly is located in a low-stress area, the actual failure risk is lower than the first failure risk threshold; if the anomaly is located in the toe area of ​​repeated bending, the side stretching area, or the heel locking area, the probability of subsequent glue separation, thread breakage, blisters, or collapse is greater than the preset probability threshold; the dynamic tolerance decision feedback module outputs release, downgrade, or rework / scrap instructions based on the failure risk index, and feeds back the spatial location information of high-risk areas to upstream equipment for subsequent process correction; To facilitate understanding, a simplified example can be used to illustrate the data flow. Suppose that the standard topology unfolded diagram is divided into three main regions: the toe area, the side area, and the heel area. The two-dimensional image extracts a discontinuous band of glue line in the toe area, and the three-dimensional point cloud obtains a depression with a depth deviation within a preset range in the same band area. After flexible registration, this anomaly is jointly marked as the first anomaly region. If a color depth difference is detected in the two-dimensional image of the side panel area, but no height change is observed in the three-dimensional point cloud, the area is marked as the second abnormal area and assigned the first preset structural influence weight during subsequent conversion; the control terminal performs differential processing on the first and second abnormal areas, and performs classification calculations in combination with their material properties and stress locations, thereby decoupling the appearance difference from the functional failure risk. In addition, if the 2D vision sensor causes local image overexposure due to surface reflectivity exceeding the preset reflectivity threshold, the control terminal will call adjacent frame images of the same shoe body or oblique view images for compensation; if the surface reflectivity is less than or equal to the preset reflectivity threshold, the current 2D vision image data will be maintained and enter the subsequent registration process; if the 3D structured light sensor causes partial point cloud loss due to shoelace occlusion, wet glue reflection, or high-speed transport, risk analysis will only be performed on the valid areas outside the missing areas, and the missing areas will be marked as areas to be re-inspected, rather than being directly rejected. If the timestamp deviation between the two-dimensional and three-dimensional data within the same detection cycle exceeds the preset synchronization tolerance, the fusion judgment of the current shoe body will be paused, only the single-source warning result will be retained and re-collection will be triggered to avoid splicing data from different times into the same object; if the standard topology unfolding diagram or material parameters of the shoe model are not successfully obtained, the system will switch to the basic appearance detection mode and output a label for appearance-only judgment to prevent outputting a life risk conclusion with excessive uncertainty under the condition of no physical prior; During the middle shift of the sports shoe assembly line, a right shoe that has just been bonded and pressed enters the inspection station. The upper two-dimensional vision sensor detects an interruption in the width of the glue line boundary on the outer side of the toe, which is less than the first preset width. At the same time, the oblique two-dimensional vision sensor identifies irregular reflections near the seam. The three-dimensional structured light sensor further detects a shallow depression at this location. The control terminal maps the three types of information to the toe bending area of ​​the standard topology unfolded diagram. After physical property conversion, it is determined that the local bonding ability has decreased. The analysis is then combined with the high dynamic stress weight of the toe area of ​​the shoe and finally outputs a high failure risk index. If the other shoe only has glue line overflow with a width less than the second preset width in the low stress area on the outer side of the heel and the depth deviation is zero, the system determines that its impact on durability and comfort is limited and does not trigger the scrapping action. The purpose of this step / mechanism is to upgrade the detection of footwear defects from static appearance comparison to functional evaluation that is related to the actual stress consequences, so as to achieve a quality judgment of flexible footwear products that is more in line with engineering practice and reduce the misjudgment rate of harmless appearance deviations.

[0015] Furthermore, the process of the multi-source data acquisition module acquiring the benchmark stress topology heat map includes: acquiring the computer-aided design model corresponding to the target object; performing dynamic stress simulation on the computer-aided design model based on the preset finite element analysis algorithm to generate the benchmark stress topology heat map; wherein, the benchmark stress topology heat map contains the dynamic stress weight coefficients of each region of the target object.

[0016] This embodiment provides a mechanism for generating a baseline stress topology heat map. Specifically, on the same sports shoe assembly line, if only real-time visual data is relied upon, local anomalies can be extracted, but the repeated stress distribution experienced by that area during actual use cannot be obtained. For glue defects of the same width, the structural impact is significantly different depending on whether they are located near the toe bend line or near the outer heel decorative piece. Therefore, this embodiment pre-establishes a baseline stress topology heat map corresponding to the structural usage scenario for the shoe before the production line officially starts operating, and uses it as a physical prior for subsequent online judgment. In the specific implementation process, the control terminal pre-acquires the computer-aided design model of the target shoe, which includes at least the upper outline, fitting boundary, reinforcement plate distribution, outsole edge structure and the geometric relationship of each material layer; and performs dynamic stress simulation on the model, with the simulated working conditions corresponding to common wearing stages such as normal walking, sudden stop and turn, forefoot push-off and heel strike. The mechanism is that the actual failure of shoes is caused by the accumulation of repeated bending, shearing and stretching. The output of finite element analysis is not a single absolute stress value, but the relative stress sensitivity of each region under various dynamic working conditions. Based on this, the system generates a reference stress topology heat map, which serves as a risk amplification base map: high-weight regions in the heat map represent defects smaller than the preset size threshold that are easily amplified by repeated wearing loads, while low-weight regions represent the stress tolerance of materials and structures to local disturbances that meet the preset safety threshold. For ease of explanation, the following specific application example can be used: Assume that the shoe body is divided into a forefoot flexion zone, a sidewall fitting zone, and a heel stability zone; after dynamic stress simulation, the forefoot flexion zone exhibits continuous high flexion and high shear during the walking cycle, the sidewall fitting zone exhibits moderate stretching during turning and lateral forces, and the heel stability zone only bears local pressure for a short time at the moment of landing; then the heat map assigns a first dynamic stress weighting coefficient to the forefoot flexion zone, a second dynamic stress weighting coefficient to the sidewall fitting zone, and a third dynamic stress weighting coefficient to the non-critical decorative areas in the heel stability zone, wherein the first dynamic stress weighting coefficient is greater than the second dynamic stress weighting coefficient, and the second dynamic stress weighting coefficient is greater than the third dynamic stress weighting coefficient; During subsequent online testing, if the anomaly is mapped to the forefoot flexion area, the system assigns it a higher failure risk weight; if the anomaly is mapped to the outer edge of the heel stability area, a preset tolerance range is allowed. In the anomaly handling mechanism, if the computer-aided design model lacks local structural information, such as incomplete heel lining thickness information, the system calls a template model of the same shoe size and material version for geometric completion and adds a low-confidence marker to the completed area to adjust the aggressiveness of the automatic judgment; if the simulation conditions are inconsistent with the actual usage conditions, the system needs to regenerate the heat map under the corresponding conditions to avoid risk measurement deviation caused by using the old heat map; if different shoe sizes exist for the same shoe model, heat maps are established separately for each shoe size family to match the nonlinear geometric changes of the toe bend line position and the fitting boundary; if the heat map is not ready, the system continues to perform appearance detection, but stops outputting high and low risk conclusions based on functional failure. Before introducing a running training shoe to the production line, engineers retrieved the shoe's digital structural model and conducted stress simulations for three conditions: forefoot push-off, normal gait, and heel cushioning. The results showed that the outer toe flex strip and the inner forefoot contact boundary were high-sensitivity areas under multiple conditions, while the outer heel decorative edge was a low-sensitivity area. The control terminal generated a corresponding baseline stress topology heat map based on this data, which was used as a unified reference in subsequent testing. Anomalies of the same degree of glue deficiency would trigger a high-risk warning when mapped to the outer toe flex strip, but would be handled with tolerance based on low sensitivity weights when mapped to the outer heel decorative edge. The purpose of this step / mechanism is to introduce stable and reusable force priors for online visual inspection, so that subsequent risk assessment has a clear structural mechanics background and achieves the engineering effect of differentiated interpretation of defect locations.

[0017] Furthermore, the process of the flexible adaptive image registration module extracting visually abnormal region data includes: performing topological mapping between two-dimensional visual image data and three-dimensional structural point cloud data and a standard topological unfolded map based on a non-rigid image registration algorithm to obtain mapping feature data; extracting feature variation regions from the mapping feature data and marking the feature variation regions as visually abnormal region data; wherein, the visually abnormal region data includes abnormal area data and abnormal depth data.

[0018] This embodiment provides a flexible adaptive image registration and anomaly extraction mechanism. Specifically, during the aforementioned production line operation, the shoe body undergoes natural deformation after pressing and cooling. If a rigid template is used for direct comparison, normal elastic deformation may be misjudged as geometric defects. Therefore, this embodiment adopts a non-rigid registration method, which maps the actual shoe body to a standard topological unfolded diagram while maintaining the topological relationship, and then extracts the feature variation region from the deviation region after mapping. In the specific implementation process, the two-dimensional visual image provides information on texture boundaries, glue line outlines, seam distribution and color change, while the three-dimensional structural point cloud provides information on height, slope, concavity and convexity and edge thickness. During flexible registration, the system prioritizes extracting stable structural features as anchor points, such as the center line of the toe, the boundary of the eyelet, the central axis of the heel, and the inflection point of the fitting edge. This allows for continuous deformation of the flexible area between anchor points, so that even if there is a stretching offset of the same upper material, it can still be mapped to the corresponding coordinates of the standard unfolded diagram. After registration, the system compares the actual mapped features with the standard features in a unified coordinate system. If a region shows both two-dimensional boundary anomalies and three-dimensional height anomalies, it is identified as a substantial feature variation. If only a texture offset with an offset less than a preset offset threshold occurs without accompanying structural depth changes, it is determined to be a natural material texture or optical acquisition deviation. Anomaly area data characterizes the coverage of the anomaly on the unfolded map, and anomaly depth data characterizes the deformation or porosity of the anomaly in the direction perpendicular to the shoe surface. Furthermore, to unify the definition of abnormal depth data, in this embodiment, after the three-dimensional structural point cloud is mapped to the standard topology unfolded map, the local reference surface around the abnormal region is used as the zero-point reference for depth. The local reference surface is obtained by fitting the continuous normal point cloud outside the abnormal region, which represents the surface morphology of the position in the normal fitting state. The normal deviation of each mapped point in the abnormal region relative to the local reference surface is recorded as the abnormal depth. Therefore, the overall shoe posture deformation, such as the toe being raised, is eliminated after registration and datum reconstruction. Only local wrinkles, gaps, collapses, and other relative deviations are included in the abnormal depth data. Correspondingly, the abnormal area data is obtained by performing connected component statistics on the continuously out-of-tolerance mapping units in the standard topology unfolded graph, ensuring that the area and depth originate from the same abnormal region and coordinate frame. A simplified example can be used to illustrate this; assuming that a standard rubber band exists in the toe bend area of ​​the standard topology unfolded diagram; during actual acquisition, due to the rebound of the shoe upper, the standard rubber band shifts as a whole in the image and is accompanied by local gaps; rigid comparison will determine both the overall shift and local gaps as anomalies, while non-rigid registration will correct the overall shift to the standard position and only extract the specific area corresponding to the local gap; if the 3D point cloud shows that the specific area is sunken relative to the local reference surface, then the anomalous area has both area and depth attributes; if another area only has color differences without sunken areas, it is recorded as a visual variation, but is given a lower weight in subsequent structural analysis; In addition, if the abnormal delivery posture causes the anchor point positioning to fail, the system performs posture normalization processing and then enters flexible registration; if normalization fails, it is marked as an abnormal posture and needs to be re-examined; if the matching degree between the two-dimensional image and the three-dimensional point cloud is insufficient in a local area, such as a shoelace occluding the image but not the point cloud, the system extracts high-confidence anomalies in the intersection area of ​​the two sources and retains low-confidence labels in the single-source effective area. If the standard topology unfolded map version is inconsistent with the current batch, the system will prevent automatic registration and prompt for template update; if the number of effective point clouds fitted to the local reference surface is insufficient, the abnormal depth of the region will not be directly taken as the extreme value, but will be marked as a depth unconfirmed area, only the two-dimensional abnormal area information will be retained and re-sampling will be triggered to avoid misjudging the missing point cloud as a depth anomaly. During continuous testing, a slight wrinkle appeared on the inner side of the toe area of ​​a left shoe. The 2D image identified an abnormal texture direction in this area, and the 3D point cloud showed that it formed a shallow groove relative to the surrounding area. After the system completed flexible registration through the eyelet boundary and the center line of the toe, it confirmed that the wrinkle was a real feature variation formed by local material stacking. The abnormal area was marked on the standard topology unfolded map, and its coverage and normal deviation were recorded. At the same time, an area on the outer side of the heel with a darker printed color but a flat surface was only retained as a low-structure influence variation. The purpose of this step / mechanism is to eliminate the interference of normal deformation of flexible products, extract abnormal regions with real structural significance, and achieve high-confidence anomaly identification.

[0019] Furthermore, the process of the physical property conversion module calculating the local intensity attenuation matrix includes: extracting the adhesive viscosity parameter from the material physical parameters; multiplying the abnormal area data and abnormal depth data to obtain abnormal volume data; converting the adhesive viscosity parameter into an equivalent viscosity parameter with one dimension according to the pre-established mapping relationship; and quotienting the abnormal volume data and the equivalent viscosity parameter and performing normalization processing to obtain the intensity reduction percentage of each pixel or voxel. The normalization process employs a max-min normalization algorithm, using a pre-calibrated baseline volume deviation limit as the denominator to map the quotient value to the range of 0 to 100, thereby obtaining a dimensionless value representing the percentage decrease in intensity. This generates a local intensity attenuation matrix with the same spatial dimension as the visually abnormal region data. The local intensity attenuation matrix contains the percentage decrease in intensity for each pixel or voxel.

[0020] This embodiment provides a mechanism for converting visual anomalies into physical strength attenuation. Specifically, after extracting the anomaly region, if only the geometric dimension description is output, its impact on bonding reliability cannot be directly quantified. For anomalies of the same volume, the actual bonding strength attenuation caused by adhesives in the first viscosity range and adhesives in the second viscosity range are not the same. Therefore, this embodiment introduces the adhesive viscosity parameter in the material physical parameters to convert visual anomalies into a local strength attenuation matrix. In the specific implementation process, the abnormal area data represents the in-plane expansion range of the affected area, and the abnormal depth data represents the degree of voids or collapse in the thickness direction; the combination of the two approximately represents the volume of damage to the continuity of the adhesive layer caused by the abnormality; the adhesive viscosity parameter represents the ability of the adhesive to fill and wet the gaps; for adhesives with lower viscosity, the shallow voids are easily redistributed and filled during the pressing stage; for adhesives with higher viscosity, similar voids are more likely to form local voids; the physical property conversion module converts each pixel or voxel point of the abnormal area in space into the corresponding percentage of strength reduction, generating a local strength attenuation matrix of the same dimension as the original abnormal area, forming a quantitative distribution of local material capability weakening; Furthermore, to standardize the calculation method, this embodiment discretizes the abnormal regions mapped to the standard topology unfolded map into multiple area units; each area unit corresponds to a pixel or a micro-region composed of adjacent pixels; for any area unit, its unit area is multiplied by the abnormal depth to obtain the local abnormal volume, and then divided by the adhesive viscosity parameter to obtain the percentage of strength reduction at that location; the abnormal area data is unfolded into a unit area sequence according to the discrete units during calculation to ensure that the local strength attenuation matrix and the visual abnormal region data maintain a point-by-point correspondence; To eliminate dimensional differences in adhesive parameters between different batches, a calibrated equivalent viscosity parameter is used in the manufacturing process. This parameter is pre-calibrated to a uniform engineering scale using process test data from similar shoe models. The calibration process involves: extracting the fill rate data of different adhesives under standard gaps from historical process tests, converting the fill rate into a dimensionless numerical coefficient through linear mapping, and using this as the equivalent viscosity parameter; ensuring that the output relative strength reduction level has engineering interpretability; if the percentage of strength reduction at a certain location exceeds the theoretical upper limit, the control terminal performs a cutoff according to the preset upper limit; if the calculation result is close to zero, it is treated as zero attenuation to ensure that the matrix value is stable within a reasonable range and to eliminate the interference of extreme sampling noise. The following specific application example can be used for understanding: Suppose an abnormal region consists of a first segment, a second segment, and a third segment. The area of ​​the first segment is less than a first area threshold and its depth is greater than a first depth threshold. The area of ​​the second segment is greater than a second area threshold and its depth is less than a second depth threshold. The area and depth of the third segment are both within the corresponding preset intermediate intervals. If the current adhesive spreadability parameter is greater than the preset spreadability threshold, the shallow wide region corresponding to the second segment has limited strength weakening, while the deep void corresponding to the first segment has a significant impact. In the converted local strength attenuation matrix, the first segment shows a high attenuation percentage, the second segment is lower, and the third segment is moderate. The system analyzes the structural weakening distribution based on this rather than simply geometric anomalies. In the anomaly handling mechanism, if material physical parameters are missing, the system calls the most recently confirmed parameters as temporary values, adds parameters to be confirmed and marks them for re-inspection; if the abnormal depth data increases abnormally but is confirmed as an optical reflection artifact by point cloud quality inspection, the intensity conversion of the area is not performed and resampling is triggered; if the abnormal area is less than the preset area detection lower limit and the depth is close to the sensor resolution limit, it is classified into the low confidence perturbation zone, and risk amplification is only triggered when high stress coincides; if the abnormal area spans different material sub-regions, the control terminal calls the corresponding equivalent viscosity parameters for each sub-region and then splices them to generate a local intensity attenuation matrix to avoid the single processing of cross-interface material properties; One shoe has a strip-shaped abnormal area on the outer rubber strip of the toe. The image shows that the coverage area is less than the preset area threshold, and the point cloud shows that the local depression is clear. The control terminal reads the viscosity parameters of the hot melt adhesive and determines that such depressions are not easy to fill under the current process. It converts the depression into a significant local strength reduction distribution. Another shoe has a visually abnormal area on the heel decorative edge that is larger than the preset maximum area value. However, the depth deviation is less than the preset depth threshold and the material has high cushioning. The converted local strength reduction distribution is in the preset low attenuation range. The purpose of this step / mechanism is to map two-dimensional and three-dimensional geometric anomalies into physical quantities that reflect the decay of local adhesive capacity, thereby achieving an engineering transition from geometric features to structural effects.

[0021] Furthermore, the process of aligning the local intensity attenuation matrix with the reference stress topology heatmap in the visual stress coupling calculation module includes: extracting the geometric center coordinates and contour boundary coordinates of the local intensity attenuation matrix on the standard topology unfolded map as the first spatial coordinate data; extracting the coordinates of the reference point on the reference stress topology heatmap that match the geometric center coordinates and contour boundary coordinates as the second spatial coordinate data; and performing coordinate system registration between the first spatial coordinate data and the second spatial coordinate data based on the affine transformation algorithm to obtain the coordinate alignment data.

[0022] This embodiment provides a spatial coordinate alignment mechanism between local intensity attenuation results and stress heatmaps. Specifically, after obtaining the local intensity attenuation matrix and the reference stress topology heatmap, if the two are not in the same coordinate system, it will lead to risk mapping misalignment. In order to eliminate the coordinate deviation caused by left and right foot mirroring, size changes and flexible deformation, this embodiment extracts the geometric center and contour boundary of the abnormal area and performs rigid and non-rigid coordinate system registration with the reference point coordinates in the heatmap. In the specific implementation process, the geometric center coordinates represent the overall landing point of the abnormal area, and the contour boundary coordinates represent the spatial expansion direction and coverage of the abnormality; the corresponding reference points in the heat map are selected from stable structural points such as the center of the toe, the metatarsophalangeal bend line, the starting point of the eyelet, and the intersection of the heel center axis; affine transformation is used to handle translation, scaling, rotation, and slight shearing relationships, accurately projecting the abnormal area into the reference space corresponding to the force map, ensuring that the system accurately identifies the spatial coupling relationship between the strength weakening area and the high stress zone; Furthermore, to satisfy the correspondence constraints of the affine transformation, the coordinates of the geometric center and the contour boundary serve as the location description of the abnormal region. The control terminal uses this location description as an index on the standard topology unfolded diagram to extract at least three non-collinear stable structural anchor points surrounding the abnormal region, and establishes a mapping between the coordinates of these anchor points in the local intensity attenuation matrix coordinate system and the coordinates of the corresponding anchor points in the reference stress topology heat map. The geometric center of the abnormal region is used to determine the local anchor point group, and the contour boundary is used to limit the registration range and orientation constraints. The coordinate pairs input to the affine transformation algorithm are the aforementioned set of anchor point coordinates, avoiding the registration under-registration problem caused by a single center point. Furthermore, the contour boundary coordinates are defined as the sequence of boundary points obtained by sampling along the abnormal contour at preset arc length intervals, or the set of feature corner points obtained by fitting. When the abnormal region has a regular shape, the major axis endpoints, minor axis endpoints, and geometric center are extracted to form a basic description. When the shape is irregular, the local registration relationship is determined by the contour boundary coordinates and the anchor points of the adjacent structures. The geometric center coordinates and contour boundary coordinates describe the location and shape of the abnormal region, and the corresponding reference point coordinates provide a stable reference on the heat map side, together serving high-precision coordinate system registration. The following specific application example can be used to illustrate this: Assume that the geometric center of the first anomalous region in the standard topological unfolded diagram is located at the front of the shoe toe, and its outline is a long and thin strip; the first high-stress zone in the thermal diagram is also a long and thin arc; if coordinate alignment is not performed, the first anomalous region and the first high-stress zone may be misaligned due to differences in initial scale; after alignment, the system determines that the main outline of the first anomalous region is located inside the first high-stress zone; conversely, the second anomalous region has a similar area but its center is closer to the outer side of the heel, and after coordinate alignment, it is only mapped to the low-stress zone, achieving strict spatial attribution determination; In the anomaly handling mechanism, if the outline of the abnormal region is severely broken, causing unstable extraction of the geometric center, the system connects and merges adjacent sub-regions and then extracts the main outline; if the heat map reference points are shifted due to version upgrade, the system reloads the reference point set of the corresponding version; if there is confusion between the left and right feet, the system completes the identification based on the direction of the shoe tongue or the shape of the sole, and then calls the mirrored heat map; if the abnormal outline exceeds the effective boundary of the shoe body after affine transformation, the current registration is determined to be invalid and a re-collection review is triggered. The local strength attenuation matrix of one right shoe shows a thin weakening band on the outer side of the toe. The system extracts the center position and boundary envelope of the weakening band on the standard topology unfolded diagram and calls the reference point of the stress heat map of the right shoe version. After coordinate alignment, it is confirmed that the weakening band is highly coincident with the high stress band of the forefoot bending. The decorative edge of the heel of the other shoe has an abnormal boundary with a length greater than the preset upper limit. After alignment, it is mapped to the low stress area of ​​the heat map, and the system outputs a different risk assessment for the two. The purpose of this step / mechanism is to ensure that the local strength attenuation information and the actual structural stress region strictly correspond in the same coordinate system, thereby ensuring the spatial consistency of subsequent risk coupling calculations.

[0023] Furthermore, the process of obtaining the failure risk index by the visual stress coupling calculation module includes: extracting the dynamic stress weight coefficient corresponding to the local intensity attenuation matrix in the benchmark stress topology heatmap based on coordinate alignment data; performing convolution operation on the local intensity attenuation matrix and the dynamic stress weight coefficient to obtain the risk matrix; summing the values ​​in the risk matrix to output the failure risk index.

[0024] This embodiment provides a mechanism for generating a failure risk index. Specifically, after spatial alignment, local strength attenuation and dynamic stress distribution need to be comprehensively quantified. This embodiment obtains the risk matrix and outputs the failure risk index through convolution operation, uniformly expressing the probability of functional failure of the target object under the current abnormal conditions. In the specific implementation process, the dynamic stress weight coefficient characterizes the sensitivity of the region to repeated mechanical action, and the local strength attenuation matrix characterizes the degree of decline in the load-bearing and bonding capacity of the region; convolution operation, as a neighborhood coupling method, comprehensively calculates the overlap between the abnormal region and the stress field and the influence of the diffusion edge, avoiding the same judgment for the edge anomaly of the high stress zone and the core area anomaly; after the risk matrix is ​​generated, it is numerically summarized and a single failure risk index is output to quantitatively characterize the comprehensive risk level of structural failure of the shoe body in actual use; Furthermore, to ensure calculation accuracy, this embodiment extracts the stress weight distribution matrix corresponding to the coverage area of ​​the local intensity attenuation matrix; the control terminal extracts a local weight sub-map with the same coordinates and scale as the anomaly region from the benchmark stress topology heatmap; the weights at each position in this local weight sub-map together constitute the set of dynamic stress weight coefficients participating in the convolution operation; the distribution of weights depends not only on the anomaly center position, but also on the spatial coupling relationship between the anomaly contour and the edge of the high stress zone and the neighboring hotspots. Furthermore, the local intensity attenuation matrix and dynamic stress weight coefficients undergo mesh unification processing before entering the convolution calculation: the preset sampling units on the standard topology unfolded graph are used as common coordinate units, the local intensity attenuation matrix is ​​projected and discretized, and the local weight subgraph is resampled at the same scale to ensure that the two are strictly consistent in row and column size, sampling interval and region boundary, and to ensure that the convolution operation is performed at the same local coordinate resolution. Furthermore, the unit values ​​in the risk matrix represent the risk contribution within the corresponding sampling unit; when performing summation calculation, effective risk units that have passed connectivity screening and confidence screening are prioritized for aggregation; for isolated high-value points or scattered anomalies with insufficient continuity, the system performs minimum connectivity area constraint and neighborhood smoothing processing before incorporating the effective values ​​into the failure risk index to ensure that the final index reflects stable structural risks and filters out instantaneous acquisition noise. The following specific application example can be used to illustrate this; assuming that the aligned anomalous weakening region consists of the first adjacent grid, the second adjacent grid, and the third adjacent grid, with the second adjacent grid located at the high-stress core and the first and third adjacent grids located in the transition zone; after coupling, the second adjacent grid contributes the most to the risk matrix; if the first and second sub-regions of another anomalous region have similar intensity decreases but are located at the low-stress edge, then the overall risk contribution after coupling is lower; when summing, the former dominates the failure risk index, while the latter only provides a weak indication, and the output result closely matches the actual stress consequences; In the anomaly handling mechanism, if there are multiple separate abnormal areas on the shoe body, the system generates sub-risk matrices for each area and then summarizes them as a whole to prevent large low-risk areas from obscuring small high-risk core areas; if the dynamic stress weight coefficient of a local reference area is missing, the system uses the smooth weight of adjacent continuous areas for compensation; if the high value of the risk matrix originates from boundary noise points, the system performs noise removal to prevent misjudgment; if multiple abnormal areas fall on different stress zones, the system retains the area labels of each sub-risk matrix so that the risk source coordinates are included when outputting the scrap command. One shoe has localized weakening of the adhesive on the toe bend band, while the other shoe has a similar degree of weakening on the heel trim. The system extracts the corresponding dynamic stress weights and performs convolution operations. The former overlaps with the high-frequency bending area, forming a continuous high-value risk band at the forefoot position of the toe. The latter forms scattered low-value points in the low-stress area. Ultimately, the former outputs a high failure risk index, and the latter outputs a low failure risk index, avoiding judgment based solely on visual area. The purpose of this step / mechanism is to unify the weakening of local structures and the amplification of dynamic forces into the same risk dimension, so as to achieve a quantitative assessment oriented towards the actual consequences of use.

[0025] Furthermore, the process of the dynamic tolerance decision feedback module performing quality classification decision operations includes: presetting a first risk threshold and a second risk threshold, wherein the first risk threshold is less than the second risk threshold; determining the relationship between the failure risk index and the first and second risk thresholds; if the failure risk index is less than or equal to the first risk threshold, generating a qualified release instruction; if the failure risk index is greater than the first risk threshold and less than or equal to the second risk threshold, generating a downgrade processing instruction; if the failure risk index is greater than the second risk threshold, generating a rework and scrap instruction.

[0026] This embodiment provides a quality classification decision mechanism based on a failure risk index. Specifically, after outputting the failure risk index, in order to avoid redundant processing and waste of system computing resources caused by using a single rigid threshold, this embodiment sets a first risk threshold and a second risk threshold to accurately classify products into three categories: qualified release, downgraded processing, and rework and scrap, thereby realizing dynamic tolerance management of flexible products. In the specific implementation process, the first risk threshold corresponds to the upper limit of functional acceptance, that is, the impact of the anomaly on the structural life and appearance consistency is within the engineering tolerance range; the second risk threshold corresponds to the upper limit of unacceptability, that is, the probability of the anomaly turning into structural failure under the preset working condition test is greater than the preset failure probability threshold; the interval between the two corresponds to the degradation handling strategy, which requires the introduction of manual review, local repair or specific transfer path; the dual threshold setting enables the risk assessment results to be directly mapped to standardized production line actions; To illustrate with a simplified example: assume that the failure risk index of three shoes detected consecutively corresponds to low, medium, and high levels, respectively; for low level, the system generates a release instruction and records the batch file; for medium level, the system generates a downgrade instruction and directs it to the re-inspection or repair channel; for high level, the system generates a rework and scrap instruction and intercepts it from entering the packaging process. In addition, if the failure risk index is at the threshold boundary and the original data has a low confidence mark, the system will trigger the re-sampling confirmation process first; if the first risk threshold and the second risk threshold need to be adjusted according to the shoe style or material system, the control terminal will automatically load the corresponding threshold template according to the production task order; if the sensor status fluctuation causes the data to fall into the threshold boundary, the system will trigger the equipment self-check; if the production line strategy does not support the downgrade flow, the intermediate products will be uniformly imported into the rework station, but the medium risk label will be retained to distinguish the target objects of the first quality defect level. One shoe has slight glue overflow on the outer heel, with a failure risk index below the first risk threshold, so the system generates a pass / fail instruction; another shoe has an abnormality on the inner forefoot that overlaps with a medium stress area, with the index between the two thresholds, so the system generates a downgrade instruction to import glue repair; a third shoe has a continuously weakened and overlapping high stress area in the toe flexure band, with the index above the second risk threshold, so the system generates a rework and scrap instruction; the decision-making process is based on precise triage based on quantified functional risk levels; The purpose of this step / mechanism is to convert the results of the coupled calculations into executable classification control instructions, and to establish a standardized closed loop for quality risk and manufacturing disposal strategies.

[0027] Furthermore, the process of the dynamic tolerance decision feedback module executing the process closed-loop feedback operation includes: extracting the target spatial coordinate position corresponding to the visual abnormal area data that generates rework and scrap instructions; counting the cumulative number of times rework and scrap instructions are generated at the target spatial coordinate position within a preset time window or preset detection batch; setting a preset number threshold; judging the relationship between the cumulative number and the number threshold; if the cumulative number is greater than or equal to the number threshold, generating a process adjustment warning signal and outputting the process adjustment warning signal to external production control equipment for adjusting the production parameters of the upstream process of the target object; if the cumulative number is less than the number threshold, maintaining the current production state and continuing to execute data acquisition operations.

[0028] This embodiment provides a closed-loop feedback mechanism for upstream processes. Specifically, based on the single-piece classification decision, in order to solve the problem of recurring defects caused by systemic process instability, this embodiment counts the target spatial coordinates of the rework and scrap instructions. When the cumulative number of times reaches a preset threshold, a process adjustment warning signal with spatial positioning attributes is output to the upstream equipment. In the specific implementation process, the target spatial coordinates correspond to the stable process position markers on the standard topology unfolded diagram, such as the outer bending band of the toe and the inner forefoot bonding edge; the system counts the frequency of rework and scrap triggered at each position within a preset time window or inspection batch; if a certain position repeatedly triggers a high-risk warning, the system determines that the phenomenon is not an isolated material difference, but rather indicates that there is a continuous deviation in the upstream process at that position; after the system reaches the number of times threshold, it outputs a process adjustment warning signal containing spatial coordinates, which allows external production control equipment to adjust production parameters such as glue application path, glue spraying pressure, and pressing temperature accordingly; Furthermore, to unify the positioning criteria, the target spatial coordinate position is jointly expressed by the region identifier and coordinate interval under the standard topology unfolded graph coordinate system; the region identifier indicates the subdivided process sub-region to which the anomaly belongs, and the coordinate interval defines the center position and boundary range of the anomaly; after the system performs left and right foot recognition and shoe size matching, it calls the corresponding version of the standard topology unfolded graph coordinates to ensure that the cumulative count is always based on the unified process position semantics, so that the statistical object is directly associated with the physical position of the upstream tooling or mold; Furthermore, if a single shoe has multiple separate rework and scrap anomaly areas, the system counts them independently according to the target spatial coordinates of each anomaly area; if the anomaly centers of multiple shoes are close, they are grouped into clusters with the same target spatial coordinates according to a preset spatial clustering radius and then the cumulative count is calculated. This processing method ensures that the cumulative count accurately reflects the spatial concentration trend and avoids the same process problem being counted separately due to small fluctuations in coordinates. The following specific application example can be used to illustrate this: Suppose there are 10 shoes that need to be reworked and scrapped in a batch of inspections, of which 7 high-risk anomalies are concentrated in the first target area on the outer side of the toe; Since the cumulative number of times in the first target area exceeds the threshold, the system sends an early warning for the first target area to the gluing and pressing unit; If the second target area only has 1 scrapping incident, the system maintains the observation state and does not output an early warning to prevent random cases from causing frequent dynamic adjustments to the production line parameters; In the anomaly handling mechanism, if the cumulative number of occurrences is close to the threshold but during a transition period, such as material change or shift switching, the system extends the observation window to filter out transitional fluctuations; if the upstream equipment does not support automatic instruction reception, the control terminal outputs early warning information to the central control panel for manual confirmation; if repeated scrapping at spatial coordinates originates from non-manufacturing source anomalies such as downstream handling and extrusion, the system combines the process timestamp to remove it without triggering upstream process adjustments; if the cumulative number of occurrences has not reached the threshold, the system maintains the current state and continues to collect data. The system detected that multiple shoes repeatedly triggered rework and scrap commands at the flexing strap on the outer side of the toe within the past twenty minutes. After statistical analysis confirmed that the cumulative number of times at this location reached a preset threshold, the system issued a process adjustment warning to the upstream glue coating robot, prompting it to adjust the glue spraying trajectory and glue dispensing pressure at this location, and also prompted the pressing equipment to check the local pressure of the corresponding mold. If the high-risk anomaly at this location does not decrease in subsequent batches, the system will further issue an equipment maintenance warning. For occasional scrapping that does not form a concentrated trend, the system only records the event without interfering with the overall line parameters. The purpose of this step / mechanism is to transform the end-of-line quality assessment results into a basis for upstream process optimization, thereby achieving a closed-loop feedback from individual component screening to overall line quality stability control.

[0029] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A defect detection system for shoe production lines based on industrial vision, characterized in that, It includes a control terminal and a two-dimensional vision sensor and a three-dimensional structured light sensor that are communicatively connected to the control terminal. The control terminal is equipped with a multi-source data acquisition module, a flexible adaptive image registration module, a physical property conversion module, a visual stress coupling calculation module, and a dynamic tolerance decision feedback module. The multi-source data acquisition module is used to acquire two-dimensional visual image data and three-dimensional structural point cloud data of the target object through the two-dimensional vision sensor and the three-dimensional structured light sensor, and to obtain a pre-generated benchmark stress topology heat map from the preset database. The flexible adaptive image registration module is used to map the two-dimensional visual image data and the three-dimensional structural point cloud data to a preset standard topology unfolded map, and compare the difference with the standard features of the standard topology unfolded map to extract visual abnormal region data. The physical property conversion module is used to calculate the local strength attenuation matrix based on the visual anomaly area data and preset material physical parameters including adhesive viscosity parameters. The visual stress coupling calculation module is used to perform spatial coordinate alignment and convolution operation between the local intensity attenuation matrix and the benchmark stress topology heatmap to obtain the failure risk index. The dynamic tolerance decision feedback module is used to perform quality classification decision operations and process closed-loop feedback operations based on the failure risk index.

2. The defect detection system for shoe production lines based on industrial vision according to claim 1, characterized in that, The process by which the multi-source data acquisition module acquires the reference stress topology heat map includes: Obtain the computer-aided design model corresponding to the target object; The computer-aided design model is subjected to dynamic stress simulation based on a preset finite element analysis algorithm to generate the reference stress topological thermogram. The reference stress topology heatmap includes dynamic stress weighting coefficients for each region of the target object.

3. The defect detection system for shoe production lines based on industrial vision according to claim 1, characterized in that, The process by which the flexible adaptive image registration module extracts data from the visually abnormal region includes: Based on a non-rigid image registration algorithm, the two-dimensional visual image data and the three-dimensional structural point cloud data are topologically mapped to the standard topological unfolded diagram to obtain mapping feature data. Extract the feature variation regions from the mapped feature data and mark the feature variation regions as the visual abnormality region data; The visual anomaly region data includes anomaly area data and anomaly depth data.

4. The defect detection system for shoe production lines based on industrial vision according to claim 3, characterized in that, The process by which the physical property conversion module calculates the local intensity attenuation matrix includes: Extract the adhesive viscosity parameter from the physical parameters of the material; The abnormal area data is multiplied by the abnormal depth data to obtain abnormal volume data. The adhesive viscosity parameter is converted into an equivalent viscosity parameter with one dimension according to the pre-established mapping relationship. The abnormal volume data is divided by the equivalent viscosity parameter and normalized to obtain the intensity reduction percentage of each pixel or voxel. The normalization process employs a maximum-minimum normalization algorithm, using a pre-calibrated baseline volume deviation limit as the denominator to map the quotient value to the range of 0 to 100, thereby obtaining a dimensionless value representing the percentage decrease in intensity, and generating the local intensity attenuation matrix with the same spatial dimension as the visual anomaly region data. The local intensity attenuation matrix includes the percentage decrease in intensity for each pixel or voxel.

5. The defect detection system for shoe production lines based on industrial vision according to claim 4, characterized in that, The process by which the visual stress coupling calculation module aligns the local intensity attenuation matrix with the reference stress topology heatmap in spatial coordinates includes: The geometric center coordinates and contour boundary coordinates of the local intensity attenuation matrix on the standard topology unfolded diagram are extracted as the first spatial coordinate data. The coordinates of reference points on the reference stress topology heatmap that match the coordinates of the geometric center and the contour boundary are extracted as the second spatial coordinate data; The first spatial coordinate data and the second spatial coordinate data are registered using an affine transformation algorithm to obtain coordinate aligned data.

6. The defect detection system for shoe production lines based on industrial vision according to claim 5, characterized in that, The process by which the visual stress coupling calculation module obtains the failure risk index includes: Based on the coordinate alignment data, extract the dynamic stress weighting coefficients corresponding to the local intensity attenuation matrix in the benchmark stress topology heatmap; The risk matrix is ​​obtained by convolving the local intensity attenuation matrix with the dynamic stress weighting coefficient, and the values ​​in the risk matrix are summed to output the failure risk index.

7. The defect detection system for shoe production lines based on industrial vision according to claim 1, characterized in that, The process by which the dynamic tolerance decision feedback module performs the quality classification decision operation includes: A first risk threshold and a second risk threshold are preset, wherein the first risk threshold is less than the second risk threshold; Determine the relationship between the failure risk index and the first risk threshold and the second risk threshold; If the failure risk index is less than or equal to the first risk threshold, a qualified release instruction is generated; If the failure risk index is greater than the first risk threshold and less than or equal to the second risk threshold, a degradation processing instruction is generated. If the failure risk index is greater than the second risk threshold, a rework and scrap instruction is generated.

8. The defect detection system for shoe production lines based on industrial vision according to claim 7, characterized in that, The process by which the dynamic tolerance decision feedback module performs process closed-loop feedback includes: Extract the target spatial coordinates of the visually abnormal region data that generated the rework and scrap instruction; The cumulative number of times the rework and scrap instructions are generated at the target spatial coordinate position within a preset time window or preset detection batch; Preset threshold number of attempts; Determine the relationship between the cumulative number of times and the number threshold. If the cumulative number of times is greater than or equal to the number of times threshold, a process adjustment warning signal is generated and the process adjustment warning signal is output to an external production control device for adjusting the production parameters of the upstream process of the target object; If the cumulative number of times is less than the number of times threshold, the current production status is maintained and the data collection operation continues.

Images