A method and system for detecting damage of inverted-plug plastic parts during demolding based on multi-view imaging

Abstract

The present application relates to the technical field of visual image processing, and discloses a kind of based on multi-view imaging's inverted buckle plastic part demolding damage detection method and system.The present application is based on three-dimensional design model and demolding direction division surface unit, establishes inverted buckle surface coordinate system and qualified surface reference;The multi-view image to be measured is registered with three-dimensional design model, and surface unit image set is formed;According to geometric visible state and local imaging state, determine main view angle and effective auxiliary view angle;Correct local image and compare with qualified surface reference, map candidate abnormal area to inverted buckle surface coordinate system;Surface abnormal information is associated and fused across view angle, and multi-view abnormal representation is formed;Damage identification and consistency check are carried out in combination with the spatial relationship of abnormal representation and inverted buckle boundary, demolding direction, and demolding damage detection result is formed.

Description

Technical Field

[0001] This invention belongs to the field of visual image processing technology, specifically relating to a method and system for detecting demolding damage of undercut plastic parts based on multi-view imaging. Background Technology

[0002] After the undercut plastic part is demolded, multi-view images are typically obtained using imaging devices positioned at different locations. The position of the plastic part in each image is determined by fixing fixtures, the plastic part's outline, structural features, or positioning marks. After brightness correction, noise reduction, edge enhancement, or geometric correction, the images from each viewpoint are compared with images of qualified plastic parts or input into a damage classification model to identify surface anomalies such as scratches, tears, and defects. The detection results from multiple imaging views are usually merged using independent judgment, fixed priority, result voting, or approximate matching of image planar positions.

[0003] The surface orientation of the undercut plastic parts changes significantly, and there are local occlusions and geometric obstructions in the demolding direction. The visibility, projection shape, and abnormal response of the same physical damage vary from different imaging perspectives. Selecting the imaging perspective based on the overall image quality, fixed camera orientation, or preset detection area cannot reflect the projected area, surface orientation, occlusion status, sharpness, and exposure status of specific surface locations, and may easily cause local damage to be compressed, obscured, or distorted in the corresponding image.

[0004] Candidate anomaly regions in different imaging perspectives are represented by their respective image pixel coordinates. Affected by perspective projection, surface curvature, pose deviation, and local occlusion, it is difficult to accurately correspond to the same entity anomaly, which may result in duplicate output, incorrect merging, or conflicting results. Normal textures, local reflections, contamination, and processing marks may also create significant image differences. It is difficult to distinguish between demolding damage and non-demolding anomalies based solely on appearance differences or general image features. Furthermore, when imaging information is insufficient or results from different perspectives conflict, there is a lack of basis for controlling the output of deterministic detection results. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and system for detecting demolding damage of undercut plastic parts based on multi-view imaging.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for detecting demolding damage of undercut plastic parts based on multi-view imaging, comprising the following steps:

[0007] S1: Based on the three-dimensional design model of the undercut plastic part and the demolding direction, divide the surface units, establish the undercut surface coordinate system, and map the qualified multi-view images to the undercut surface coordinate system to form a qualified surface reference;

[0008] S2: Based on the multi-view imaging calibration parameters, the multi-view image of the under-duck plastic part to be tested is registered with the three-dimensional design model, and each surface unit is projected onto the multi-view image to be tested to form a set of surface unit images.

[0009] S3: Based on the geometric visibility state and local imaging state of the surface unit under each imaging viewpoint, determine the main viewpoint and effective auxiliary viewpoint of each surface unit to form an effective viewpoint relationship;

[0010] S4: Perform imaging correction on the local image corresponding to the effective viewpoint relationship and compare it with the qualified surface reference to extract candidate anomaly regions. Map the candidate anomaly regions to the inverted surface coordinate system to form surface anomaly information.

[0011] S5: Based on the correspondence between entity locations and the continuity of abnormal morphology, cross-view correlation is performed on surface abnormal information, and the correlation results are fused using effective view relationships to form a multi-view abnormal representation.

[0012] S6: Based on the multi-view anomaly representation and the spatial relationship between the multi-view anomaly representation and the undercut boundary and demolding direction, the demolding damage identification model is used to identify damage, and the identification results are verified according to the cross-view consistency to form the demolding damage detection results.

[0013] Preferably, the local draft state and undercut boundary are determined based on the angular relationship between the local surface normal and the demolding direction, as well as the geometric blocking relationship along the demolding direction. Surface units are divided along the undercut boundary, model edges, curvature change lines, and surface topological boundaries. The position of the outer surface of the plastic part is represented by the surface unit number and the local coordinates within the unit. The transformation relationship between the local coordinates within the unit and the coordinates of the three-dimensional design model, as well as the positional correspondence between adjacent surface units, are established. The effective image information in the qualified multi-view images is mapped to the undercut surface coordinate system. Based on the appearance statistics of multiple qualified samples at the same surface position, a qualified surface benchmark is formed, including normal appearance characterization, normal fluctuation range, effective reference viewpoint, and reference imaging quality.

[0014] Preferably, the geometric visibility state of each imaging viewpoint is determined based on the projected area, the observation angle, and the occlusion ratio; the local imaging state is determined based on the local sharpness, exposure state, and effective texture preservation; imaging viewpoints that do not meet the geometric validity conditions or local imaging validity conditions are excluded, and the status evaluation results of the remaining imaging viewpoints are combined into a viewpoint quality value; the imaging viewpoint with the largest viewpoint quality value is determined as the main viewpoint, and the imaging viewpoints that meet the auxiliary viewpoint conditions and can provide different observation directions or supplement the surface coverage are determined as effective auxiliary viewpoints.

[0015] Preferably, local images corresponding to the main viewpoint and effective auxiliary viewpoint are extracted from the surface unit image set, and imaging correction parameters are determined using stable surface position, adjacent surface units, or imaging calibration area; the corrected local images are transformed to the inverted surface coordinate system and compared with the same entity position in the qualified surface reference for color or grayscale, local gradient, edge structure, and texture to form a surface difference response; candidate anomalous regions are extracted based on the normal fluctuation range and the adjacency relationship of the entity surface to form surface anomalous information including entity position, entity coverage, difference response, anomalous morphology, main extension direction, imaging viewpoint, and viewpoint quality value.

[0016] Preferably, the positional association is determined based on the overlapping relationship of the entity coverage of candidate anomaly regions or the connection relationship of adjacent surface units in different imaging perspectives, and the anomaly morphology continuity relationship is determined based on the main extension direction, boundary orientation, differential response distribution and local texture changes; candidate anomaly regions that simultaneously satisfy positional association and anomaly morphology continuity relationship are divided into the same multi-view anomaly group, and candidate anomaly regions that cannot establish association with candidate anomaly regions in other imaging perspectives are retained as single-view anomaly groups.

[0017] Preferably, the fusion weights of each imaging viewpoint are determined based on the viewpoint quality value and the surface mapping reliability determined by the local projection scale, pose registration error, and surface coordinate mapping error. The fusion weights are used to fuse the surface difference responses at the same entity surface location to form evidence of fusion anomalies, fusion anomaly regions, entity coverage of fusion anomaly regions, main extension direction, and fusion local features. The cross-viewpoint consistency is determined based on the position, shape, and anomaly response relationship between the candidate anomaly regions and the fusion anomaly regions in each imaging viewpoint. Among these, imaging viewpoints that cannot effectively observe the corresponding entity surface location are not considered as negative evidence.

[0018] Preferably, the surface distance between the fusion anomaly region and the undercut boundary, as well as the position of the fusion anomaly region on the normal draft side, the undercut side, or across the undercut boundary, are determined based on the adjacency relationship of the surface units; the demolding direction is projected onto the tangent plane of the surface where the fusion anomaly region is located to form a local demolding direction, and the directional relationship between the main extension direction and the local demolding direction is determined; the surface distance, position, and directional relationship are combined with evidence of fusion anomaly, local fusion features, effective viewpoint information, and cross-viewpoint consistency to form a candidate characterization of demolding damage.

[0019] Preferably, the surface evidence map, including fused anomaly evidence, the boundary of the fused anomaly region and the inverted boundary, and the numerical features including entity coverage, position status, directional relationship, effective viewpoint information, cross-viewpoint consistency degree and feature validity identifier are input into the demolding damage identification model; the demolding damage identification model encodes the surface evidence map and numerical features respectively, fuses the surface spatial features and numerical relationship features, and outputs the non-demolding anomaly and the category probability of each demolding damage category.

[0020] Preferably, the completeness of perspective evidence is determined based on the number of effective imaging perspectives, perspective quality values, and differences in observation directions. Specifically, effective imaging perspectives with the same observation direction are grouped into the same observation direction group, and the highest perspective quality value within the group represents the effective evidence for that observation direction. The reliability of the result is determined based on model recognition confidence, cross-perspective consistency, and the completeness of perspective evidence. When the model candidate category with the highest category probability belongs to the demolding damage category, the reliability of the result meets the output conditions, and there is no insufficient imaging information or cross-perspective conflict, a demolding damage detection result is formed. When perspective evidence is insufficient, there is insufficient imaging information, or cross-perspective conflict occurs, the corresponding surface location and abnormal range are determined as areas to be reviewed.

[0021] Another technical solution provided by the present invention: a multi-view imaging-based demolding damage detection system for undercut plastic parts, comprising:

[0022] The surface datum construction module is used to divide the surface into units, establish an inverted surface coordinate system, and form a qualified surface datum.

[0023] The imaging registration module is used to perform pose registration between the multi-view images to be tested and the 3D design model, and to form a set of surface unit images;

[0024] The perspective evaluation module is used to determine the main perspective and effective auxiliary perspective of each surface unit, and to form an effective perspective relationship;

[0025] The anomaly extraction module is used to correct the corresponding local image and compare it with a qualified surface reference to form surface anomaly information;

[0026] The anomaly fusion module is used to correlate and fuse surface anomaly information across perspectives to form a multi-perspective anomaly representation.

[0027] The damage identification module is used to identify and verify damage based on multi-view anomaly representations and the spatial relationship between multi-view anomaly representations and the undercut boundary and demolding direction, forming demolding damage detection results or areas to be verified.

[0028] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0029] This invention establishes an undercut surface coordinate system and a qualified surface benchmark based on the three-dimensional design model and demolding direction of the undercut plastic part. Through pose registration, local images from different imaging perspectives are mapped to specific surface units. An effective imaging perspective is selected based on the local visibility state, and anomaly extraction, cross-view correlation, anomaly fusion, and demolding damage identification are completed at a unified entity surface location. This processing transforms scattered anomaly evidence from different image coordinates into multi-view anomaly representations with definite entity positions and spatial relationships, reducing the impact of surface orientation changes, local occlusion, and viewpoint differences on the detection results. Based on the evidence state, demolding damage detection results or areas awaiting verification are formed.

[0030] This invention evaluates the projected area, viewing direction, occlusion status, local sharpness, and exposure status of each surface unit, eliminating imaging perspectives that cannot effectively reflect the corresponding surface state, and determining the primary viewing perspective and effective auxiliary viewing perspectives. Thus, the applicable range of the imaging perspective is transformed from a fixed camera orientation to a locally effective range corresponding to the specific entity surface location. This ensures that the local image, which plays a primary detection role, maintains the necessary projection scale, visible range, and surface information, while the effective auxiliary viewing perspective provides different viewing directions or supplementary coverage, reducing abnormal omissions and distortion responses on the inverted surface caused by projection compression, geometric occlusion, defocusing, or exposure anomalies.

[0031] This invention utilizes an inverted surface coordinate system to convert candidate anomaly regions in each effective imaging viewpoint into a unified entity location representation. Cross-viewpoint correlations are established based on entity location correspondences and anomaly morphological continuity. The fusion contribution of different imaging viewpoints is adjusted according to viewpoint quality and surface mapping reliability. This processing ensures that the projection position and morphological changes of the same entity damage in different images correspond to surface topological relationships, merging anomaly evidence with continuous position and consistent morphology into a unified fused anomaly region. Simultaneously, it preserves single-viewpoint anomaly states that cannot be reliably correlated, thereby reducing duplicate output of the same damage and erroneous merging of adjacent unrelated anomalies. Furthermore, it ensures that the resulting entity coverage and fused anomaly evidence match the actual observation capabilities of each imaging viewpoint.

[0032] This invention determines the surface distance and positional state between the fusion anomaly region and the undercut boundary, and converts the demolding direction into a local demolding direction within the corresponding surface tangent plane, establishing a spatial correspondence between the main extension direction of the anomaly and the demolding action direction. Furthermore, the corresponding spatial relationship, along with fusion anomaly evidence, fusion local features, effective viewpoint information, and cross-viewpoint consistency, is used for demolding damage identification. Thus, damage judgment is simultaneously constrained by the anomaly appearance, the undercut structure position, the anomaly extension state, and the multi-viewpoint display state, enabling the differentiation between damage caused by demolding action and normal texture, reflection, contamination, and other non-demolding anomalies. The identification results are verified based on model recognition confidence, cross-viewpoint consistency, and viewpoint evidence completeness. It can also prevent the output of deterministic damage results when imaging information is insufficient or when anomaly information from different viewpoints conflicts, and transfer the corresponding area for verification. Attached Figure Description

[0033] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0034] Figure 1 This is a flowchart of a method for detecting demolding damage in undercut plastic parts based on multi-view imaging;

[0035] Figure 2 This is an architecture diagram of a multi-view imaging-based demolding damage detection system for undercut plastic parts. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Example 1:

[0038] like Figure 1 As shown, a method for detecting demolding damage of undercut plastic parts based on multi-view imaging includes the following steps:

[0039] S1: Based on the three-dimensional design model of the undercut plastic part and the demolding direction, divide the surface units, establish the undercut surface coordinate system, and map the qualified multi-view images to the undercut surface coordinate system to form a qualified surface reference;

[0040] S2: Based on the multi-view imaging calibration parameters, the multi-view image of the under-duck plastic part to be tested is registered with the three-dimensional design model, and each surface unit is projected onto the multi-view image to be tested to form a set of surface unit images.

[0041] S3: Based on the geometric visibility state and local imaging state of the surface unit under each imaging viewpoint, determine the main viewpoint and effective auxiliary viewpoint of each surface unit to form an effective viewpoint relationship;

[0042] S4: Perform imaging correction on the local image corresponding to the effective viewpoint relationship and compare it with the qualified surface reference to extract candidate anomaly regions. Map the candidate anomaly regions to the inverted surface coordinate system to form surface anomaly information.

[0043] S5: Based on the correspondence between entity locations and the continuity of abnormal morphology, cross-view correlation is performed on surface abnormal information, and the correlation results are fused using effective view relationships to form a multi-view abnormal representation.

[0044] S6: Based on the multi-view anomaly characterization and its spatial relationship with the undercut boundary and demolding direction, the demolding damage identification model is used to identify damage, and the identification results are verified according to the cross-view consistency to form demolding damage detection results.

[0045] In step S1, image pixels from different imaging perspectives cannot directly represent the same physical location on the surface of the undercut plastic part, and the normal surface appearance will also change with the surface orientation and imaging perspective. Therefore, based on the three-dimensional design model of the undercut plastic part and the demolding direction, surface units are divided, an undercut surface coordinate system independent of the imaging perspective is established, and the normal appearance state and allowable fluctuation range are formed in the undercut surface coordinate system.

[0046] Specifically, taking the 3D design model of the undercut plastic part, the demolding direction, the multi-view imaging calibration parameters, and qualified multi-view images as input, the 3D design model is converted into a model form that can represent the surface position, surface normal, and topological relationship. Surface units are divided according to the surface geometric changes and the relationship between the surface and the demolding direction. Local coordinates are established for each surface unit and the transformation relationship between them and the coordinates of the 3D design model to form the undercut surface coordinate system. The effective image information in the qualified multi-view images is mapped to the undercut surface coordinate system. After imaging normalization and statistical analysis of multiple qualified samples, a qualified surface reference is formed.

[0047] In practice, the 3D design model originates from the product design data of the undercut plastic part, and can at least determine the solid position, surface normals, geometric boundaries, and topological connections of the outer surface of the plastic part. When the 3D design model is a solid model, the outer surface of the plastic part is converted into a triangular mesh, parametric surface, or other model form that can perform surface sampling and projection calculations, while maintaining the geometric dimensions, surface orientation, and topological connections of the outer surface of the plastic part. The demolding direction is determined based on the mold opening direction, the core withdrawal direction, or the actual movement direction of the plastic part after detaching from the molding structure, and is based on the same spatial coordinate reference as the 3D design model.

[0048] Surface sampling points are set on the outer surface of the 3D design model to obtain the entity position, local surface normal, and curvature of each sampling point. The local draft state is determined based on the angle between the local surface normal and the demolding direction, and ray intersection or envelope interference detection is performed from the corresponding surface position along the demolding direction. When the ray along the demolding direction intersects with other model surfaces of the plastic part, or when there is geometric obstruction as the plastic part surface moves along the demolding direction, the corresponding surface position is determined as the undercut surface position. The position where the local draft state or geometric obstruction state transitions forms the undercut boundary.

[0049] Surface unit boundaries are set along the undercut boundary, model edges, curvature change lines, and surface topology boundaries to maintain continuous surface orientation and defined demolding relationships within the same surface unit. Smaller surface units are used at locations with rapid surface orientation changes, large curvature, or near the undercut boundary, while larger surface units are used at locations that are approximately planar and have continuous surface states. The minimum effective projected area required for appearance comparison is determined based on the number of pixels corresponding to the minimum damage size of the target at the object-side resolution of the imaging device, ensuring that a surface unit has a visible projected area not less than the minimum effective projected area at at least one reference imaging viewpoint. Each surface unit records its unit number, solid surface extent, representative point position, surface normal, curvature information, adjacent units, local demolding state, and surface distance from the undercut boundary.

[0050] The inverted surface coordinate system uses a combination of surface element numbers and local coordinates within the element to represent the position of the outer surface of the plastic part. For each surface element, the representative point of the surface element is taken as the origin of the local coordinate system. The first and second local coordinate axes are determined based on the tangents of two independent surfaces, and the orientation of the local coordinate system is determined by the surface normal. When the 3D design model uses a triangular mesh, the surface position is represented by the triangular face number and the coordinates of the centroid within the triangular face; when using a parametric surface, the surface position is represented by the surface patch number and the surface parametric coordinates.

[0051] Based on the local coordinate origin, local coordinate axes, and model surface equations, a transformation relationship is established between the local coordinates within an element and the coordinates of the 3D design model. This allows any position in the inverted surface coordinate system to be converted into the solid's 3D coordinates, surface normal, and associated surface element in the 3D design model. For adjacent surface elements with a common boundary, the positional correspondence between the local coordinates on both sides of the common boundary is recorded, enabling the continuous tracking of solid positions spanning adjacent surface elements according to surface topology.

[0052] Select plastic parts with the same specifications as the undercut plastic part to be tested, and confirmed to be free of demolding damage through visual inspection, magnified observation, or dimensional inspection, as qualified samples. Qualified samples are used to obtain qualified multi-view images in camera, lighting, and imaging stations with the same parameters as the undercut plastic part to be tested or with corresponding calibration. Based on the pose of the qualified plastic part in the imaging station and the camera's internal, external, and lens distortion parameters for each imaging angle, the surface units are projected onto the qualified multi-view images. Combining projection depth, surface orientation, and ray occlusion relationships, projected pixels facing away from the imaging device, obstructed by other structures, or located outside the effective imaging range are excluded, and the retained image information is reverse-mapped to the undercut surface coordinate system.

[0053] The mapping results are normalized for brightness, color, and contrast. The normalization parameters are determined based on the imaging background, the imaging calibration area, or the location of a stable, defect-free surface in a qualified plastic part. These parameters are used to eliminate overall exposure and color response differences between different qualified samples and different imaging perspectives, while preserving local edges, textures, and color difference structures. Image information that remains overexposed, underexposed, out of focus, or continuously occluded after normalization is not included in the normal appearance statistics for the corresponding surface locations.

[0054] For the same surface sampling location in the inverted surface coordinate system, normalized image information from multiple qualified samples is collected, and brightness or color, gradient, texture, and edge response are statistically analyzed. Normal appearance characteristics are determined using the mean, median, or robust statistics, while normal fluctuation ranges are determined using standard deviation, quantile intervals, or median absolute deviation. When multiple effective reference viewpoints exist, the normal appearance statistical results for each imaging viewpoint are retained, and the projection scale, viewing direction, occlusion status, sharpness, and exposure status of the corresponding reference viewpoint are recorded. The normal appearance characteristics, normal fluctuation ranges, effective reference viewpoints, and reference imaging quality for each surface location collectively form a qualified surface benchmark.

[0055] In step S2, the actual position and orientation of the undercut plastic part under test in the imaging station will be affected by clamping deviation, and the surface units in the 3D design model cannot be directly mapped to the specific pixel positions in the multi-view images under test. Therefore, pose registration is performed between the multi-view images under test and the 3D design model under a unified imaging coordinate reference, and the projection correspondence between each surface unit and different images under test is established based on the registration results.

[0056] In this step, multi-view images of the under-duck plastic part under test are obtained under the same clamping state based on the camera intrinsic parameters, camera extrinsic parameters, and lens distortion parameters of each imaging angle. Positioning information that can correspond to the 3D design model is extracted from the multi-view images. The pose transformation relationship is solved and verified based on the correspondence between the positions of multiple 2D images and the positions of the 3D model. Using the pose transformation relationship and multi-view imaging calibration parameters, each surface unit is projected onto the multi-view images under test to determine the projection contour, pixel correspondence, projection depth, and occlusion state, forming a set of surface unit images.

[0057] Multi-view imaging calibration parameters include the camera's intrinsic parameters for each imaging viewpoint, the camera's extrinsic parameters relative to the imaging station, and lens distortion parameters. Multiple imaging views are calibrated using the same station coordinate system, enabling pixel rays in different images to be transformed to a unified three-dimensional coordinate reference. The calibration reprojection error should be less than the allowable positioning error corresponding to the minimum damage size of the target in the image. When the positional relationship of the imaging device, lens, or imaging station changes, the calibration of the corresponding imaging viewpoint must be re-performed.

[0058] The under-exposed plastic part under test maintains the same clamping position and orientation throughout all image acquisition. When using synchronous acquisition, the trigger times and exposure periods for each imaging viewpoint are synchronized; when using sequential acquisition, the positional relationship between the plastic part and the imaging device remains unchanged until all images under test are acquired. Distortion correction is performed on the original images according to lens distortion parameters, and each corrected image is associated with an imaging viewpoint number, image resolution, and exposure parameters to form a multi-view image under test.

[0059] Extract the outer contour, structural edges, holes, corner points, or positioning marks of the plastic part from the multi-view images to be tested. Based on feature type, relative position, and geometric relationships, match the image positioning features with corresponding features in the 3D design model to form multiple 2D-to-3D correspondences. Positioning features should be distributed across different surface locations and collectively constrain the translational and rotational states of the undercut plastic part under test. When positioning features are concentrated in a small area or are in a collinear state where pose cannot be effectively constrained, supplementary constraints are provided using positioning features, the outer contour of the plastic part, or structural edges from other imaging perspectives.

[0060] The rotation matrix and translation vector of the 3D design model relative to the imaging station are calculated based on the 2D-to-3D correspondence. The distance between the projected position of the model's positioning features and the actual position of the image's positioning features is used as the reprojection error. The pose parameters are jointly corrected using the reprojection errors from multiple imaging perspectives, and erroneous matching points where the reprojection error significantly deviates from other correspondences are eliminated. When the plastic part under test has sufficient surface structure, the surface point set reconstructed from multi-view images can also be used for point cloud registration with the 3D design model, and the registration result is verified using the reprojection error of the image contour or structural features. When the registration error meets the conditions determined by the imaging resolution, the minimum damage size of the target, and the allowable pixel positioning deviation, a pose transformation relationship is formed; when the registration error does not meet the corresponding conditions, the current plastic part is recorded as a registration anomaly.

[0061] For 3D surface points within a surface unit, the pose transformation relationship is used to convert the 3D surface points to the corresponding camera coordinate system, and then the coordinates are projected into image pixels according to the camera's intrinsic parameters. The 3D surface points are represented as homogeneous coordinates. Representing image pixels as homogeneous coordinates When, the projection relationship is ,in, The camera intrinsic parameter matrix, and These are the rotation matrix and translation vector from the model coordinate system to the camera coordinate system, respectively. After homogeneous coordinate normalization, the image pixel positions and projection depth are obtained.

[0062] Projected contours are formed according to the boundaries of surface units, and the projected area is calculated. Occlusion status is determined using depth buffering, surface depth comparison, or intersection of pixel rays with model surfaces. When the same pixel ray intersects with multiple model surfaces, the model surface closest to the camera is identified as the visible surface. If the projection depth of the current surface unit is greater than the depth of the nearest visible surface and exceeds the tolerance determined by the model discrete error and pose registration error, the corresponding position is identified as an occluded pixel. During projection, the surface unit number and local coordinates within the unit corresponding to the effective image pixels are recorded simultaneously, establishing a bidirectional correspondence between image pixels and the coordinates of the inverted surface. The surface unit number, imaging viewpoint number, projection contour, projected area, effective pixel range, pixel correspondence, projection depth, and occlusion ratio together form the surface unit image set.

[0063] In step S3, the surface orientation, curvature, and occlusion state of the inverted structure change with the entity's position, and the same imaging viewpoint cannot maintain the same observation capability for all surface units. Therefore, the geometrically visible state and local imaging state are evaluated separately for each surface unit, imaging viewpoints that cannot truly reflect the corresponding surface state are eliminated, and effective viewpoints that play the main detection role and supplementary detection role are determined.

[0064] Specifically, based on the surface unit image set and qualified surface reference, the geometric visibility state of each imaging viewpoint is first determined according to the projected area, observation angle and occlusion ratio. Then, the local imaging state is determined according to the local sharpness, exposure state and effective texture preservation. The various state parameters that meet the necessary validity conditions are normalized and combined into a viewpoint quality value. The main viewpoint and effective auxiliary viewpoint are determined in the effective imaging viewpoint to form an effective viewpoint relationship.

[0065] For the The surface unit and the first Each imaging viewpoint obtains the projection contour, effective pixel range, surface orientation, and occlusion pixel range from the surface unit image set. The projection area is determined by the number of effective pixels within the projection contour. Based on the image resolution, the object-space resolution of the imaging device, and the minimum damage size of the target, the minimum effective projection area required for local appearance comparison is determined. When the effective projection area is less than the minimum effective projection area, the corresponding local image cannot stably represent the minimum damage of the target, and the corresponding imaging viewpoint is determined to be a geometrically invalid viewpoint. A normalized projection area with a value between 0 and 1 is formed based on the ratio between the effective projection area and the reference projection area. .

[0066] Observe the included angle For the first The surface normal and pointing of the first surface unit The angle between the observation directions of each camera. When the surface unit is curved, the representative observation angle is determined based on the surface normal and effective projected area at each sampling position within the unit. This satisfies the frontal observation condition. Normalized to 0 to 1 to form the observation direction evaluation value When the surface unit faces away from the camera or the viewing angle causes the minimum damage to the target to be compressed to an unstable resolution in the projection direction, the corresponding imaging angle is determined to be a geometrically invalid angle.

[0067] The occlusion ratio is determined by the ratio of the number of projected pixels occluded by other model surfaces to the total number of projected pixels of the surface unit. The no-occlusion ratio... The difference between 1 and the occlusion ratio. In addition to the number of occluded pixels, it also checks whether unoccluded pixels can form a continuous local region covering the minimum damage size of the target. When unoccluded pixels are divided into multiple small regions that cannot individually support appearance comparison, the corresponding imaging viewpoint is also determined to be a geometrically invalid viewpoint. Projected area, surface orientation, and continuous unoccluded regions are mandatory geometric conditions that must be met; deficiencies in geometric visibility are not compensated for by other state parameters.

[0068] Local images of surface units are extracted from the imaging viewpoint determined by geometric validity, excluding occluded pixels, pixels outside the projected contour, and invalid boundary pixels. Local sharpness is determined based on gradient energy, Laplacian response variance, or edge spread width, and compared with the sharpness of a reference at the same entity location in a qualified surface benchmark. For normal, weakly textured surfaces, sharpness is evaluated using adjacent regions of the projected contour, the imaging calibration region, or adjacent surface units with stable structures under the same imaging viewpoint. Evaluation results that meet the sharpness criteria are normalized to 0 to 1 to form the local sharpness. .

[0069] Exposure status is determined based on the proportion of overly dark pixels, overly bright pixels, and local grayscale dynamic range in the local image. The boundaries for judging overly dark and overly bright pixels are determined based on the effective response range of the imaging device and the normal pixel distribution of qualified samples. The exposure evaluation result is then normalized to a range of 0 to 1 to form the exposure evaluation value. The degree of effective texture preservation is determined by comparing the local image with the normal edge and texture structure in a qualified surface reference. This verifies whether the local image retains the surface information required to perform the corresponding comparison. When overexposure, underexposure, or defocus makes the normal surface structure unrecognizable, the corresponding imaging viewpoint is determined to be an invalid local imaging viewpoint. When the local image only has overall brightness, color, or contrast changes that can be recovered through imaging correction, the corresponding imaging viewpoint is preserved, but the corresponding state evaluation value is reduced.

[0070] For imaging viewpoints judged by geometric validity and local imaging validity, the normalized projected area is... Evaluation value of observation direction Unobstructed ratio Local sharpness and exposure rating Combined into visual quality value ,Right now Evaluation coefficient , , , and All values ​​are non-negative and the sum of their coefficients is 1. The evaluation coefficients are calibrated based on the impact of each state parameter on the stability of anomaly detection. The viewing angle quality value is used to compare viewing angles that can already effectively image, but is not used to convert viewing angles that do not meet the necessary validity conditions into effective viewing angles.

[0071] For the same surface unit, the effective imaging viewpoint with the highest spectral quality value is determined as the primary viewpoint. Besides the primary viewpoint, imaging views whose spectral quality values ​​meet the auxiliary viewpoint criteria and can provide different observation directions or supplement areas not fully covered by the primary viewpoint are determined as effective auxiliary viewpoints. Repeated viewpoints with observation directions approximately the same as the primary viewpoint and which do not increase the effective coverage are not prioritized due to increased quantity. Effective viewpoint relationships include the correspondence between surface units and primary and effective auxiliary viewpoints, spectral quality values, and various state parameters. Surface units without effective imaging viewpoints are recorded as areas with insufficient imaging information; when there is only one effective imaging viewpoint, the primary viewpoint and single-viewpoint evidence state are retained.

[0072] In step S4, local images in the main view and the effective auxiliary view may still exhibit differences in brightness, color, and contrast. Direct comparison with a qualified surface reference can cause normal imaging fluctuations to produce abnormal responses. Therefore, imaging deviations in the effective view are corrected while preserving the local changes of the damage, and local changes exceeding the range of normal surface fluctuations are converted into surface anomaly information with definite entity locations.

[0073] To complete the above processing, local images corresponding to the main viewpoint and effective auxiliary viewpoint are extracted from the surface unit image set according to the effective viewpoint relationship. Imaging correction parameters are determined using stable surface position, adjacent surface units, or imaging calibration area. The corrected local images are transformed to the inverted surface coordinate system and compared with the appearance of the same entity position with the qualified surface reference to form the surface difference response. Candidate abnormal regions are extracted according to the normal fluctuation range and the adjacency relationship of the entity surface to form surface abnormality information corresponding to each effective viewpoint.

[0074] Local images are extracted based on the projected contours of surface units, excluding occluded pixels, pixels outside the projected contours, pixels exceeding the effective response range of the image, and pixels with unstable interpolation at the projected boundaries. From the retained local images, surface locations with minimal fluctuations in normal surface reference and without continuous high differential responses, reflection saturation, or occlusion changes in the image under test are selected as stable surface locations. When calculating correction parameters, pixels that significantly deviate from the normal appearance distribution are excluded, and the robust statistical relationships of the remaining pixels are used to obtain the correction parameters, avoiding the participation of candidate damage areas in the correction parameter calculation.

[0075] For the The surface unit at the first Local images in each imaging viewpoint, according to Image correction is performed, among which, These are the pixel values ​​of the original local image. These are the corrected pixel values. This is the gain coefficient. The gain and bias are determined for each color channel of the color image. The gain and bias are obtained through the median, quantiles, or robust regression relationships of stable surface locations, and are limited to the parameter range formed by normal imaging variations across multiple qualified samples. Imaging correction employs a consistent or gradually varying correction relationship within the surface unit area, without local compensation for individual high-difference pixels.

[0076] When a local image lacks sufficiently stable surface locations, the correction parameters are determined using adjacent surface units with consistent surface material and imaging response relationships from the same imaging viewpoint. If adjacent surface units also lack stable surface locations, the correction parameters are determined using the imaging calibration region. After correction, the brightness distribution, edge response, and texture structure of the normal structure location are checked to ensure they have recovered to the normal fluctuation range of a qualified surface reference. If the normal surface structure still cannot be identified after correction, deterministic candidate anomalous regions are not extracted using the corresponding local image, and the incomplete imaging state is retained.

[0077] By utilizing the correspondence between image pixels and the coordinates of the inverted surface, the corrected local image is transformed to the coordinate system of the inverted surface. For cases where a surface sampling location corresponds to multiple image pixels or where image pixels fall between adjacent surface sampling locations, interpolation is performed based on the projection scale, the centroid coordinates within the triangular face, or the local coordinates of the parametric surface. This is applied to the surface location. Obtain normal appearance characteristics, normal fluctuations, and effective reference angles from qualified surface references; when the same reference angle exists, use the statistical results of the same imaging angle; when the same reference angle does not exist, select an effective reference angle that matches the observation direction and projection scale.

[0078] The comparison between the corrected local image and a qualified surface reference includes differences in color or grayscale, local gradient, edge structure, and texture. Surface location. Standardized appearance differences at the location are ,in, The corrected image value of the surface to be measured. This is the standard value for normal appearance. This is a normal fluctuation. This is the stability constant determined based on the lower limit of normal fluctuations. Local gradient differences. Texture differences are determined based on gradient magnitude, gradient direction, and edge position deviation. Determined based on the local texture description results within the same surface neighborhood.

[0079] Standardize appearance differences Local gradient differences and texture differences Surface difference response is formed according to nonnegative combination coefficients. Each difference term is first transformed to a comparable response scale according to the normal difference distribution of qualified samples. The combination coefficient is determined based on the distinguishing ability of each difference term in qualified samples and confirmed damaged samples. Normal weak texture surfaces reduce the combined contribution of texture differences, while surfaces with stable textures or structural edges increase the combined contribution of local gradient differences or texture differences.

[0080] Anomaly thresholds are determined based on the surface difference response distribution formed by multiple qualified samples at the same surface location. The anomaly threshold is determined using the upper quantile of the difference response of the qualified samples or a combination of the median and median absolute deviation of the difference response. When there are significant differences in the degree of normal fluctuation at different surface locations, anomaly thresholds are determined separately for surface units or surface regions with the same normal appearance.

[0081] Surface sampling locations exceeding the anomaly threshold are used to extract connected regions based on local adjacency relationships within the unit and the correspondence between common boundaries of adjacent surface units. The minimum resolvable anomaly area is determined based on the object-space resolution of the imaging device, the minimum damage size of the target, and the distribution of isolated noise in qualified samples. Isolated responses with areas smaller than the minimum anomaly area and lacking boundaries, texture, or continuity are excluded as imaging noise; multiple small-area responses located in adjacent surface units and maintaining continuity near a common boundary are retained based on the combined entity coverage area and continuity relationship.

[0082] For the retained candidate anomaly regions, record the surface location, entity coverage, differential response intensity, boundary morphology, local texture, associated imaging viewpoint, and viewpoint quality value. The extension direction of the candidate anomaly region is determined through analysis of the principal direction of the surface coordinates; when the candidate anomaly region is linear or has a branching structure, the principal extension direction is determined based on the main extension segment of the surface skeleton. The principal viewpoint and each effective auxiliary viewpoint separately generate surface anomaly information; candidate anomaly regions from different imaging viewpoints are not directly merged in this step.

[0083] In step S5, the same entity damage may form candidate anomalous regions with different locations, boundary morphologies, and anomalous intensities in different imaging viewpoints. Surface anomalous information from different viewpoints cannot be directly accumulated or substituted for each other. Therefore, the entity correspondence of anomalous information from different viewpoints is determined using an inverted surface coordinate system, and the fusion contribution is adjusted based on viewpoint quality and surface mapping reliability.

[0084] Based on the aforementioned technical state, candidate anomalous regions from different imaging perspectives are uniformly represented in an inverted surface coordinate system. Cross-view correlations are established based on the overlapping of entity positions, the connection of adjacent surface units, and the continuity of anomalous morphology to form a multi-view anomalous group. The fusion weight of each imaging perspective is determined for the surface position covered by the multi-view anomalous group, and the surface difference response, entity coverage range, and local features are fused. The position, morphology, and anomalous response relationship between the anomalous regions of each perspective and the fused anomalous regions are compared to determine the degree of cross-view consistency and form a multi-view anomalous characterization.

[0085] In the specific implementation process, for two candidate anomaly regions originating from different imaging perspectives, the surface area ratio between the intersection and union of the entity coverage areas is calculated to determine the degree of surface coordinate overlap. When the two candidate anomaly regions are located in adjacent surface units and there is insufficient overlap, the distance between the endpoints of the anomaly boundaries along the entity surface is calculated based on the correspondence of the common boundaries. If both candidate anomaly regions extend to the vicinity of the common boundary, and the surface distance between the boundary endpoints does not exceed the connection distance determined by the pose registration error, surface mapping error, and the minimum damage size of the target, then the two candidate anomaly regions are determined to have a basis for positional association.

[0086] Based on location correlation, the main extension direction of candidate anomaly regions is transformed to the same surface tangent plane at the local coordinates of the same unit or the common boundary, and the main extension direction, boundary orientation, differential response distribution, and local texture changes are compared. When the main extension direction and boundary orientation remain continuous and there is no substantial contradiction in the local morphology, the candidate anomaly region is determined to satisfy the anomaly morphology continuity relationship. When the same entity damage forms a local gap due to occlusion, cross-gap correlation is allowed when the extension directions of the anomaly boundaries on both sides of the gap point to the same gap region, the boundaries can be continuously connected on the entity surface, and the gap width is within the allowable interval jointly determined by the occlusion range and the surface mapping error.

[0087] Using candidate anomaly regions as nodes, and employing entity location correspondence and anomaly morphological continuity as connection conditions, an association graph is established, and the association graph is then grouped into connected groups. In addition to satisfying the association condition with at least one candidate anomaly region within the group, a newly added candidate anomaly region should also not substantially conflict with the overall entity coverage and main extension trend of the group. Each connected group that passes the intra-group consistency check forms a multi-view anomaly group. If a candidate anomaly region is formed only in one effective imaging view or cannot establish a reliable association with candidate anomaly regions from other views, the corresponding candidate anomaly region is retained as a single-view anomaly group.

[0088] Surface locations covered by multi-view anomaly groups To determine the effective observation location of the surface And a collection of effective perspectives for integration To obtain the first from an effective perspective relationship Viewpoint quality value of each imaging viewpoint The reliability of surface mapping is determined based on the local projection scale, pose registration error, and surface coordinate mapping error. When the local projection scale can stably represent the target with minimal damage, the pose registration error is small, and the reverse mapping error is small, the surface mapping reliability is high; when the projection scale decreases or the corresponding error increases, the surface mapping reliability decreases.

[0089] No. An imaging perspective at the surface position The normalized fusion weight at the location is Obtain the first from surface anomaly information An imaging perspective at the surface position Surface difference response ,according to Computational fusion of anomaly evidence Imaging perspectives with high quality and reliable surface mapping contribute significantly to the fusion of anomaly evidence, and effectively supplementary perspectives provide additional evidence based on the actual observation conditions.

[0090] For the ability to effectively observe surface position However, no imaging viewpoint for candidate anomaly regions was formed, preserving the original surface difference response for consistency analysis; surface locations could not be observed due to occlusion, insufficient projected area, or local imaging failure. From the imaging perspective, the lack of anomalous response is not used as negative evidence. The fusion anomalous region is determined based on the continuous distribution of fusion anomalous evidence in the inverted surface coordinate system. The fusion anomalous evidence is then used as statistical weights to determine the entity coverage, fusion boundary, main fusion extension direction, anomalous response intensity, and local fusion features. During the fusion process, the original candidate anomalous regions and original surface difference responses from each imaging perspective are simultaneously preserved.

[0091] The surface coordinates of candidate anomalous regions and fused anomalous regions in each imaging viewpoint are compared. Surface position consistency is assessed. Determined based on differences in surface intersection ratio, center-surface distance, and entity coverage; degree of consistency of anomalous morphology. Determined based on the bidirectional surface distance at the boundary, the difference in length-to-width ratio, and the difference in the main extension direction; consistency of abnormal response intensity. The consistency level was determined based on the response increment of the anomalous region relative to the surrounding normal surface, the location of high response values, and the trend of response variation along the main extension direction. Each consistency level was normalized to 0 to 1, and the influence of low-quality views on the consistency results was reduced using view quality values.

[0092] Cross-perspective consistency ,in, , and All coefficients are non-negative and their sum is 1. Each coefficient is determined based on the influence of corresponding consistency features in confirmed damaged samples and non-demolding anomaly samples on the stability of the corresponding entity anomaly. For single-view anomaly groups, the cross-view consistency degree is considered. It is recorded as 0, and a single-view status flag is set to indicate that no cross-view supporting evidence has been obtained. The fusion of anomalous evidence, entity coverage, fusion of local features, imaging views involved in the fusion, quality values ​​of each view, original anomalous information of each view, cross-view consistency, and single-view status flag together form a multi-view anomalous characterization.

[0093] In step S6, multi-view anomaly characterization can represent the abnormal appearance and cross-view manifestation state of the entity surface. However, surface anomalies may also be formed by normal texture, contamination, reflection, or other non-demolding factors. Therefore, the spatial relationship between the abnormal appearance and the undercut boundary and demolding direction is used together for damage identification, and the output of the deterministic detection results is controlled according to the completeness and consistency of the cross-view evidence.

[0094] Specifically, for each multi-view anomaly representation, the positional relationship between the fused anomaly region and the inverted boundary, as well as the directional relationship between the main extension direction of the anomaly and the local demolding direction, are determined. The corresponding spatial relationships are combined with fused anomaly evidence, fused local features, effective viewpoint information, and cross-viewpoint consistency to form a demolding damage candidate representation. The trained demolding damage recognition model is used to output the non-demolding anomalies and the category probabilities of each demolding damage category. The credibility of the results is determined based on the model recognition confidence, cross-viewpoint consistency, and viewpoint evidence completeness to form the demolding damage detection result or the area to be reviewed.

[0095] Based on the adjacency relationships of surface units recorded in the undercut surface coordinate system, the shortest surface path length from each surface position within the fusion anomaly region to the undercut boundary is calculated, determining the surface distance between the fusion anomaly region and the nearest undercut boundary. When the fusion anomaly region intersects with the undercut boundary, the surface distance is set to 0; when the fusion anomaly region crosses the undercut boundary, the anomaly coverage area and cross-boundary ratio on both sides of the undercut boundary are recorded. Based on the local draft state of the surface units on both sides of the undercut boundary, it is further determined whether the fusion anomaly region is located on the normal draft side, the undercut side, or crosses the undercut boundary.

[0096] The main extension direction of the fused anomaly region is obtained from the multi-view anomaly representation. When the fused anomaly region spans multiple surface units, each local extension direction is transformed along the surface topology to the tangent plane of the representative location of the anomaly region. When the anomaly region is approximately circular, has a small area, or does not have a stable principal axis, the main extension direction is recorded as an indeterminate state, and a direction feature validity indicator is set.

[0097] The demolding direction is denoted as The unit surface normal of the location representing the anomaly region is denoted as... The projection of the demolding direction onto the tangent plane of the corresponding surface is: When the projection direction can be stably determined, for Normalization is performed to form a local demolding direction; when the demolding direction is close to the surface normal and the projection of the tangent plane cannot form a stable direction, the local demolding direction is recorded as an indeterminate state.

[0098] Let the main extension direction of the anomaly be denoted as... When both the abnormal main extension direction and the local demolding direction are effective, the included angle of the directions is: The absolute value is used to eliminate the influence of the selection of the positive and negative directions of the main axis of the linear anomaly. When the main extension direction of the anomaly or the local demolding direction is uncertain, the direction angle is set to a missing state while retaining the validity identifier of the direction feature. The surface distance between the anomaly area and the undercut boundary, the positional state of both sides of the undercut boundary, and the direction angle together represent the spatial relationship between the anomaly, the undercut structure, and the demolding action.

[0099] The intensity distribution, anomalous area, boundary morphology, local texture changes, main extension direction of the anomalous area, positional relationship of the inverted boundary, local demolding direction relationship, number of effective viewpoints, quality values ​​of each viewpoint, cross-viewpoint consistency, and feature validity indicators of the fused anomaly evidence are combined to form a candidate representation of demolding damage. For model inputs that require preservation of spatial structure, a local surface range covering the anomalous area and adjacent normal surfaces is extracted in the inverted surface coordinate system, centered on the representative location of the fused anomalous region. The fused anomaly evidence, anomalous boundaries, and inverted boundaries are then resampled at a uniform surface scale to form a fixed-size surface evidence map. The surface evidence map and numerical features are used together as inputs to the demolding damage identification model.

[0100] The demolding damage identification model comprises a surface evidence map feature encoding part, a numerical feature encoding part, and a classification output part. The surface evidence map feature encoding part employs a convolutional feature encoding structure to extract and fuse the local spatial distribution and morphological features of anomalous evidence, anomalous boundaries, and inverted boundaries. The numerical feature encoding part standardizes continuous numerical features, numerically encodes the positional state and feature validity identifiers of inverted boundaries, and forms numerical relationship features through fully connected feature transformation. The classification output part fuses surface spatial features and numerical relationship features to output the class probabilities of non-demolding anomalies and each demolding damage category.

[0101] The demolding damage identification model is trained and validated before being put into inspection. Training samples are derived from qualified plastic parts, non-demolding abnormal plastic parts, and demolding damaged plastic parts confirmed by manual visual inspection, magnified observation, or dimensional measurement. Each sample plastic part is used as training input through the same surface mapping, benchmark comparison, perspective evaluation, anomaly extraction, and cross-perspective fusion process as the actual inspection. Training labels include at least non-demolding abnormalities and demolding damage; when it is necessary to distinguish damage types, the demolding damage labels are further divided into tearing, laceration, undercut edge defects, and other damage categories that need to be identified.

[0102] Training samples are divided into training, validation, and test sets based on the physical plastic parts. Images from different viewpoints and multi-viewpoint anomaly representations of the same physical plastic part are not assigned across datasets. The model uses class cross-entropy loss for classification training, and class weights are set according to the number of samples in each class. The validation set is used to determine the classification output conditions, numerical feature standardization parameters, model parameters, and result confidence calculation parameters. The test set is used to test the recognition stability under different damage categories and different effective viewpoint states. After training, the model parameters are fixed to handle candidate representations of demolding damage formed by the undercut plastic parts to be tested.

[0103] During the detection process, candidate representations of demolding damage are input into the demolding damage recognition model to obtain the probabilities of each category. The category with the highest probability is designated as the model's candidate category. When a model's candidate category belongs to the demolding damage category, the corresponding category is determined as the candidate damage category, and the probability of that category is recorded as the model's recognition confidence level. When the probability of non-demolding anomaly category is the highest, no demolding damage output is generated, and the corresponding anomaly location and non-demolding anomaly identification results are retained.

[0104] The completeness of viewpoint evidence is determined by the number of imaging viewpoints that can effectively observe fusion anomalous regions, the quality values ​​of each viewpoint, and the differences in observation direction. Effective imaging viewpoints are grouped according to the observation direction. Multiple effective viewpoints with approximately the same observation direction are grouped into the same observation direction group, and the highest viewpoint quality value within the group represents the effective evidence for that direction. This avoids increasing the completeness of viewpoint evidence by repeatedly setting viewpoints with the same observation direction. The completeness of viewpoint evidence increases when the number of effective evidence from different observation directions increases and the representative viewpoint quality value also increases; conversely, the completeness of viewpoint evidence decreases when there is only a single effective viewpoint, multiple viewpoints are concentrated in the same observation direction, or the viewpoint quality is low. The number of effective evidence and the representative viewpoint quality value are normalized to 0 to 1 according to the reference viewpoint state that can form a stable identification in the validation set.

[0105] Model recognition confidence Cross-perspective consistency and the completeness of perspective evidence Combining results into credibility ,Right now ,in, , and All coefficients are non-negative and the sum of the coefficients is 1. The coefficients of each combination and the output conditions are determined by the validation set.

[0106] When a candidate model category belongs to the demolding damage category, the result confidence meets the output conditions, the fused anomalous region does not involve areas with insufficient imaging information, and there is no substantial conflict between anomalous information from different viewpoints, the candidate damage category is determined as the damage category. The location and extent of the damaged surface are determined using the entity coverage area in the multi-view anomalous characterization, resulting in a demolding damage detection result that includes the damage category, surface location, damage extent, effective viewpoints involved in the identification, and result confidence.

[0107] When candidate anomaly regions in multiple high-quality, valid viewpoints cannot establish a correspondence between entity locations, the main extension direction or boundary orientation of the anomaly contradicts each other, high-quality viewpoints that can clearly observe the fused anomaly region consistently lack corresponding anomaly responses, or the cross-viewpoint consistency is lower than the conflict judgment criteria, it is determined that there is a substantial conflict between the anomaly information from different viewpoints. When a multi-viewpoint anomaly group has only a single valid viewpoint, the viewpoint evidence completeness does not meet the output requirements, there is a substantial conflict between the cross-viewpoint anomaly information, or the anomaly region involves an area with insufficient imaging information, no deterministic damage output is formed. The corresponding surface location and anomaly range are determined as areas to be reviewed, and the single-viewpoint state, low-quality viewpoint, location with insufficient imaging information, or cross-viewpoint conflict type that triggers the review are associated with it.

[0108] Example 2:

[0109] like Figure 2 As shown, a multi-view imaging-based demolding damage detection system for undercut plastic parts, used to implement the aforementioned method, includes:

[0110] The surface datum construction module is used to divide surface units based on the 3D design model and demolding direction of the undercut plastic part, establish the undercut surface coordinate system, and map qualified multi-view images to the undercut surface coordinate system to form a qualified surface datum.

[0111] The imaging registration module is used to obtain multi-view images of the test subject based on multi-view imaging calibration parameters, perform pose registration between the undercut plastic part to be tested and the three-dimensional design model, and project each surface unit onto the multi-view images of the test subject to form a set of surface unit images.

[0112] The viewing angle evaluation module is used to determine the main viewing angle and effective auxiliary viewing angle based on the geometric visibility state and local imaging state of each surface unit under different imaging angles, and to form an effective viewing angle relationship.

[0113] The anomaly extraction module is used to perform imaging correction on the local image corresponding to the effective viewpoint relationship and compare it with the qualified surface reference to extract candidate anomaly regions. The candidate anomaly regions are then mapped to the inverted surface coordinate system to form surface anomaly information.

[0114] The anomaly fusion module is used to perform cross-view correlation of surface anomaly information based on the correspondence between entity locations and the continuity of anomaly morphology, and to fuse the correlation results using effective view relationships to form a multi-view anomaly representation.

[0115] The damage identification module is used to identify damage based on multi-view anomaly representation and the spatial relationship between multi-view anomaly representation and undercut boundary and demolding direction, using demolding damage identification model, and to verify the identification results according to the cross-view consistency, forming demolding damage detection results or areas to be reviewed.

[0116] Each module is connected via data to form a connection structure with the surface reference construction module as the reference connection end and the imaging registration module, the viewing angle evaluation module, the anomaly extraction module, the anomaly fusion module, and the damage identification module as the sequential processing chain. The surface reference construction module is also connected to the viewing angle evaluation module, the anomaly extraction module, and the damage identification module, respectively. The imaging registration module is also connected to the anomaly extraction module, and the viewing angle evaluation module is also connected to the anomaly fusion module.

[0117] For the surface reference construction module, the surface reference construction module receives the three-dimensional design model of the undercut plastic part, the demolding direction, the multi-view imaging calibration parameters and qualified multi-view images. Based on the angle relationship between the local surface normal and the demolding direction and the geometric blocking relationship along the demolding direction, the local demolding state and undercut boundary are determined, and the surface unit is divided along the undercut boundary, model edge, curvature change line and surface topology boundary line.

[0118] The surface reference construction module represents the position of the outer surface of the plastic part using surface element numbers and local coordinates within the elements. It establishes the transformation relationship between local coordinates within the elements and coordinates of the 3D design model, as well as the coordinate correspondence between adjacent surface elements, forming an inverted surface coordinate system. The surface reference construction module maps the effective image information from qualified multi-view images to the inverted surface coordinate system, performs imaging normalization on the mapping results, and forms a qualified surface reference based on the appearance statistics of multiple qualified samples at the same surface position, including normal appearance characteristics, normal fluctuation range, effective reference viewing angle, and reference imaging quality.

[0119] The imaging registration module includes a multi-view imaging unit and a registration projection unit. The multi-view imaging unit includes multiple imaging devices arranged around the undercut plastic part to be tested and an imaging station for fixing the undercut plastic part to be tested. The multiple imaging devices are calibrated using the same station coordinate system, and the undercut plastic part to be tested is acquired under the condition that it is kept in the same clamping state.

[0120] The registration and projection unit extracts the outer contour, structural edges, holes, corners, or positioning marks of the plastic part from the multi-view images under test, establishes the correspondence between the image positioning features and the positioning features of the 3D design model, and solves and verifies the pose transformation relationship of the undercut plastic part under test. The registration and projection unit uses the pose transformation relationship and multi-view imaging calibration parameters to project each surface unit onto the multi-view images under test, determining the projection contour, projection area, projection depth, occlusion state, and the correspondence between image pixels and the coordinates of the undercut surface, forming a set of surface unit images; when the pose registration error does not meet the corresponding conditions, registration anomaly information is generated.

[0121] Based on a set of surface unit images and a qualified surface benchmark, the viewing angle evaluation module determines the geometric visibility state of each surface unit under different imaging angles according to the projected area, viewing angle, and occlusion ratio. It then determines the local imaging state based on local sharpness, exposure status, and effective texture preservation. The module first eliminates imaging angles that do not meet the necessary validity conditions, then combines the evaluation results into a viewing angle quality value. The effective imaging angle with the highest viewing angle quality value is determined as the primary viewing angle, and imaging angles that meet the auxiliary viewing angle conditions and can provide different viewing directions or supplement surface coverage are determined as effective auxiliary viewing angles, thus forming an effective viewing angle relationship. Surface units without effective imaging angles are recorded as areas with insufficient imaging information.

[0122] Based on the effective viewing angle relationship, the anomaly extraction module extracts local images corresponding to the main viewing angle and effective auxiliary viewing angle from the surface unit image set. It determines imaging correction parameters using stable surface positions, adjacent surface units, or imaging calibration regions, and corrects the brightness, color, and contrast of the local images. The anomaly extraction module transforms the corrected local images to an inverted surface coordinate system and compares them with the same entity positions in a qualified surface reference for color or grayscale, local gradient, edge structure, and texture to form a surface difference response. The anomaly extraction module extracts candidate anomaly regions based on the normal fluctuation range and the adjacency relationship of entity surfaces, and associates the surface position, entity coverage, difference response intensity, boundary morphology, local texture, main extension direction, imaging viewing angle, and viewing angle quality value of the candidate anomaly regions to form surface anomaly information.

[0123] For the anomaly fusion module, the module compares the overlap of entity positions, connection of adjacent boundaries, main extension direction and local morphology of candidate anomaly regions from different imaging perspectives in the inverted surface coordinate system. Based on the correspondence of entity positions and the continuity of anomaly morphology, it establishes cross-view correlation to form a multi-view anomaly group. Candidate anomaly regions that cannot establish reliable correlation with other candidate anomaly regions from different perspectives are retained as single-view anomaly groups.

[0124] The anomaly fusion module determines the fusion weights for different imaging perspectives based on the viewpoint quality value and surface mapping reliability. It fuses surface difference responses and local features at the same entity surface location to form fused anomaly evidence, entity coverage, and fused local features. The anomaly fusion module compares the position, morphology, and anomaly response relationships between candidate anomaly regions from each perspective and the fused anomaly region to determine the degree of cross-viewpoint consistency. It then combines the fused anomaly evidence, entity coverage, fused local features, effective viewpoint information, original anomaly information from each perspective, and cross-viewpoint consistency into a multi-view anomaly representation.

[0125] For the damage identification module, based on the entity coverage of the multi-view anomaly representation in the undercut surface coordinate system, it determines the surface distance and positional state between the fused anomaly region and the undercut boundary. It then projects the demolding direction onto the tangent plane of the surface containing the fused anomaly region to determine the directional relationship between the main anomaly extension direction and the local demolding direction. The damage identification module combines the corresponding spatial relationship with fused anomaly evidence, fused local features, effective viewpoint information, and cross-viewpoint consistency to form a candidate demolding damage representation.

[0126] The damage identification module inputs candidate representations of demolding damage into the trained demolding damage identification model, obtains the class probabilities of non-demolding anomalies and each demolding damage category, and forms the result credibility based on the model identification confidence, cross-viewpoint consistency, and the viewpoint evidence completeness determined by the number of effective viewpoints, viewpoint quality, and differences in observation direction. When the model output belongs to the demolding damage category, the result credibility meets the output conditions, and there is no substantial conflict between the anomaly information from different viewpoints, the damage identification module forms a demolding damage detection result including the damage category, surface location, damage range, effective viewpoints involved in the identification, and result credibility. When viewpoint evidence is insufficient, cross-viewpoint anomaly information conflicts, or the anomaly area involves an area with insufficient imaging information, the damage identification module determines the corresponding surface location and anomaly range as the area to be reviewed.

[0127] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for detecting demolding damage of undercut plastic parts based on multi-view imaging, characterized in that, Includes the following steps: S1: Based on the three-dimensional design model of the undercut plastic part and the demolding direction, divide the surface units, establish the undercut surface coordinate system, and map the qualified multi-view image to the undercut surface coordinate system to form a qualified surface reference; S2: Based on the multi-view imaging calibration parameters, the multi-view image of the under-duck plastic part to be tested is registered with the three-dimensional design model, and each surface unit is projected onto the multi-view image to be tested to form a set of surface unit images. S3: Based on the geometric visibility state and local imaging state of the surface unit under each imaging viewpoint, determine the main viewpoint and effective auxiliary viewpoint of each surface unit to form an effective viewpoint relationship; S4: Perform imaging correction on the local image corresponding to the effective viewpoint relationship and compare it with the qualified surface reference to extract candidate abnormal regions. Map the candidate abnormal regions to the inverted surface coordinate system to form surface abnormal information. S5: Based on the correspondence between entity locations and the continuity of abnormal morphology, the surface abnormal information is correlated across perspectives, and the correlation results are fused using the effective perspective relationships to form a multi-perspective abnormal representation. S6: Based on the multi-view anomaly representation and the spatial relationship between the multi-view anomaly representation, the undercut boundary, and the demolding direction, the demolding damage identification model is used to identify damage, and the identification results are verified according to the cross-view consistency to form demolding damage detection results.

2. The method according to claim 1, characterized in that, The local draft state and the undercut boundary are determined based on the angular relationship between the local surface normal and the demolding direction, as well as the geometric blocking relationship along the demolding direction. The surface units are divided along the undercut boundary, model edges, curvature change lines, and surface topology boundaries. The position of the outer surface of the plastic part is represented by the surface unit number and the local coordinates within the unit. The transformation relationship between the local coordinates within the unit and the coordinates of the three-dimensional design model, as well as the positional correspondence between adjacent surface units, are established. The effective image information in the qualified multi-view images is mapped to the undercut surface coordinate system. Based on the appearance statistics of multiple qualified samples at the same surface position, the qualified surface reference is formed, including normal appearance characterization, normal fluctuation range, effective reference viewpoint, and reference imaging quality.

3. The method according to claim 1, characterized in that, The geometric visibility state of each imaging viewpoint is determined based on the projected area, the observation angle, and the occlusion ratio. The local imaging state is determined based on the local sharpness, exposure state, and effective texture preservation. Imaging viewpoints that do not meet the geometric validity conditions or local imaging validity conditions are excluded. The status evaluation results of the remaining imaging viewpoints are combined into a viewpoint quality value. The imaging viewpoint with the largest viewpoint quality value is determined as the main viewpoint. The imaging viewpoints that meet the auxiliary viewpoint conditions and can provide different observation directions or supplement the surface coverage are determined as the effective auxiliary viewpoints.

4. The method according to claim 1, characterized in that, Local images corresponding to the main viewpoint and the effective auxiliary viewpoint are extracted from the surface unit image set. Imaging correction parameters are determined using stable surface position, adjacent surface units, or imaging calibration region. The corrected local images are transformed to the inverted surface coordinate system and compared with the same entity position in the qualified surface reference for color or grayscale, local gradient, edge structure, and texture to form a surface difference response. Candidate abnormal regions are extracted based on the normal fluctuation range and the entity surface adjacency relationship to form the surface abnormality information including entity position, entity coverage, difference response, abnormal morphology, main extension direction, imaging viewpoint, and viewpoint quality value.

5. The method according to claim 4, characterized in that, The positional association is determined based on the overlap of the entity coverage of the candidate anomaly regions or the connection relationship of adjacent surface units in different imaging perspectives. The anomaly morphology continuity is determined based on the main extension direction, boundary orientation, differential response distribution and local texture changes. Candidate anomaly regions that simultaneously satisfy the positional association and the anomaly morphology continuity are divided into the same multi-view anomaly group, and candidate anomaly regions that cannot be associated with candidate anomaly regions in other imaging perspectives are retained as single-view anomaly groups.

6. The method according to claim 5, characterized in that, Based on the viewpoint quality value and the surface mapping reliability determined by the local projection scale, pose registration error and surface coordinate mapping error, the fusion weight of each imaging viewpoint is determined; the fusion weight is used to fuse the surface difference response at the same entity surface location to form fusion anomaly evidence, fusion anomaly region and the entity coverage, main extension direction and fusion local features of the fusion anomaly region; The degree of cross-view consistency is determined based on the position, shape, and abnormal response relationship between the candidate abnormal region in each imaging view and the fused abnormal region, wherein the imaging view that cannot effectively observe the corresponding entity surface position is not considered as negative evidence.

7. The method according to claim 6, characterized in that, The surface distance between the fusion anomaly region and the undercut boundary is determined based on the adjacency relationship of the surface units, as well as the position of the fusion anomaly region on the normal draft side, the undercut side, or across the undercut boundary. The demolding direction is projected onto the tangent plane of the surface where the fusion anomaly region is located to form a local demolding direction, and the directional relationship between the main extension direction and the local demolding direction is determined. The surface distance, the position state, and the directional relationship are combined with the fusion anomaly evidence, the fusion local features, the effective viewpoint information, and the cross-viewpoint consistency to form a candidate characterization of demolding damage.

8. The method according to claim 7, characterized in that, The surface evidence map, including the fusion anomaly evidence, the boundary of the fusion anomaly region, and the inverted boundary, as well as the numerical features including the entity coverage area, the positional state, the directional relationship, the effective viewpoint information, the cross-viewpoint consistency degree, and the feature validity identifier, are input into the demolding damage identification model. The demolding damage identification model encodes the surface evidence map and the numerical features respectively, fuses them to form surface spatial features and numerical relationship features, and outputs the non-demolding anomaly and the category probability of each demolding damage category.

9. The method according to claim 8, characterized in that, The completeness of perspective evidence is determined based on the number of effective imaging perspectives, perspective quality values, and differences in observation directions. Effective imaging perspectives with the same observation direction are grouped into the same observation direction group, and the highest perspective quality value within the group represents the effective evidence for that observation direction. The reliability of the result is determined based on model recognition confidence, the degree of cross-perspective consistency, and the completeness of perspective evidence. When the model candidate category with the highest category probability belongs to the demolding damage category, the reliability of the result meets the output conditions, and there is no insufficient imaging information or cross-perspective conflict, the demolding damage detection result is formed. When perspective evidence is insufficient, there is insufficient imaging information, or cross-perspective conflict occurs, the corresponding surface location and abnormal range are determined as areas to be reviewed.

10. A demolding damage detection system for undercut plastic parts based on multi-view imaging, characterized in that, include: A surface reference construction module is used to divide the surface unit, establish the inverted surface coordinate system, and form the qualified surface reference. An imaging registration module is used to perform pose registration between the multi-view image to be tested and the three-dimensional design model, and to form the surface unit image set. A perspective evaluation module is used to determine the main perspective and the effective auxiliary perspective for each of the surface units, and to form the effective perspective relationship; An anomaly extraction module is used to correct the corresponding local image and compare it with the qualified surface reference to form the surface anomaly information; An anomaly fusion module is used to perform cross-view correlation and fusion of the surface anomaly information to form the multi-view anomaly representation. The damage identification module is used to identify and verify damage based on the multi-view anomaly characterization and the spatial relationship between the multi-view anomaly characterization and the undercut boundary and the demolding direction, forming the demolding damage detection result or the area to be verified.

Images