Vision inspection enhancement techniques based on physical geometry models

Abstract

The application discloses a visual detection enhancement method based on a physical geometric model. A defect carrier containing a surface support domain, a geometric disturbance field, a material disturbance field and a pedigree identifier is established, and defects are injected according to imaging conditions of a view angle, illumination, background and scale sampling; a normal model and a defect model are physically rendered in pairs under the same imaging condition, and an image and an identifier, depth and normal buffer are output; a depth visible area of surface projection, a defect identifier buffer area and a photometric response area of a pair of images are fused to generate multi-level labels of pixels, instances, categories, severity, scenes and pedigrees; according to visible coverage, boundary consistency and photometric observability indexes, samples are gated, unqualified samples are resampled according to failure causes, and qualified samples are used to update conditional binning weights. The application can reduce inconsistency between synthetic defect images and labels, and provide supervised training samples with multiple shooting conditions and traceability.

Description

Technical Field

[0001] This invention relates to the fields of industrial visual inspection training data generation and computer graphics rendering technology, specifically to a visual inspection enhancement method based on a physical geometry model. Background Technology

[0002] Visual defect detection in industrial products typically relies on training images labeled with defect categories, locations, or regions. In actual production, defect events are characterized by low frequency, long tails, and batch correlation. Some defects only produce recognizable image responses under specific viewing directions and incident lighting conditions. Therefore, real defect images are difficult to cover different product postures, lighting conditions, background conditions, and imaging scales.

[0003] One existing method involves continuously acquiring images from the production line and manually labeling defect categories, bounding boxes, or pixel regions. This method requires waiting for defects to occur naturally, and pixel boundaries are also affected by the labeler's experience, image contrast, and occlusion. Furthermore, when product structure, camera position, or lighting is adjusted, the existing sample distribution may not match the new inspection conditions.

[0004] Another existing approach is to paste textures and draw defect areas onto a normal 2D image, or to synthesize defect images using image generation models. This method can increase the number of defect appearances, but defects often lack a physical basis consistent with product surfaces, component relationships, occlusion states, and lighting responses, and the input mask may be offset from the actual visible defect areas in the final image.

[0005] Another existing approach uses 3D models and procedural rendering to generate training images, expanding scene variations by changing camera, light source, material, or background parameters, and having the renderer output segmentation, depth, or normal ground truth values. If the 3D defect extent or defect markers in the renderer are directly used as the final pixel labels, product occlusion, occlusion by other objects, or surface glancing observation may make the marked defects invisible in the image.

[0006] For metals, coatings, or highly reflective surfaces, whether a defect forms a detectable response is also related to the direction of the light source, the reflection state, and the camera orientation. Even if the defective surface is geometrically visible, there may not be enough color or brightness difference under the current illumination; in this case, treating the complete defect area as a positive sample will introduce supervisory noise into the training set, where the image is labeled as a defect but lacks corresponding evidence.

[0007] Therefore, a technical solution is needed that unifies defect physical characterization, imaging condition variations, normal and defective paired images, rendering buffers, multi-level labels, and sample coverage feedback within the same parameter spectrum, and determines the defect regions that can be used for training based on actual visibility and actual photometric response. Summary of the Invention

[0008] This invention provides a visual inspection enhancement method based on a physical geometric model. This technology leverages the programmable rendering capabilities of the physical geometric model to incorporate the product model, defect carrier, imaging conditions, paired images, rendering buffer, and label records into the same system. Under the same imaging conditions, a defect-free physical geometric model and a defect model are rendered separately. The effective defect area is determined based on geometric visibility, defect identification, and photometric response. Subsequent sampling is then adjusted through sample quality gating and conditional coverage feedback.

[0009] Through the above processing, the present invention can generate visual samples covering multiple angles, multiple lighting conditions, multiple backgrounds and multiple sizes, and make the physical representation of defects, image response and multi-level labels correspond to each other, thereby solving the problems of uncertain production shooting conditions, scarcity of real defect samples in terms of variety and quantity, and lack of a complete mechanism from defect generation to fine annotation. It provides rich and quality-controllable supervised training samples for visual inspection models, and reduces missed detections and false detections under complex production conditions.

[0010] The visual detection enhancement technology based on physical geometry models includes the following six steps:

[0011] Step S1: Obtain the defect-free physical geometric model, reference material parameters, and imaging constraints of the industrial product, and establish a surface coordinate system on the product surface.

[0012] Step S2: Construct a defect carrier within the surface coordinate system, which includes a surface support domain, a geometric perturbation field, a material perturbation field, a defect category, a severity parameter, and a spectrum identifier.

[0013] Step S3: The imaging conditions are binned and sampled according to the observation direction, incident direction, background type, imaging scale and defect response difficulty. A defect model is generated based on the defect carrier, and the transformation chain from the surface support domain to the rendering coordinate system is saved.

[0014] Step S4: Under the same imaging conditions, perform physical rendering on the defect-free physical geometry model and the defect model respectively to obtain a normal reference image, a defect image, and a marker buffer, depth buffer, and surface normal buffer registered with the defect image.

[0015] Step S5: Determine the geometrically visible mask based on the transformation chain and depth buffer, determine the photometric response region based on the pixel difference between the normal reference image and the defect image, and fuse the geometrically visible mask, the buffered defect region, and the photometric response region into an effective defect region. Generate multi-level labels from the effective defect region.

[0016] Step S6: Perform quality gating on the samples based on visible coverage, boundary consistency, and photometric observability indicators; perform directional resampling on samples that fail the gating according to the failure indicators; encapsulate the samples that pass the gating and update the subsequent sampling weights of the corresponding imaging condition bins.

[0017] Corresponding to the method described above, the present invention also provides a visual inspection enhancement system based on a physical geometric model. The system includes a model and constraint management module, a defect carrier construction module, an imaging condition sampling module, a defect injection and pairwise physical rendering module, a visibility and photometric response fusion annotation module, a sample quality gating module, and a coverage feedback and sample encapsulation module. The method can also be executed by the processor and graphics rendering processing unit of an electronic device and stored in a computer-readable storage medium in the form of a program.

[0018] The features of this invention are:

[0019] (1) This invention utilizes a programmable rendering tool based on a physical geometry model to construct a visual sample generation mechanism with multiple angles, lighting conditions, backgrounds, and sizes. Specifically, the observation direction, incident direction, background type, imaging scale, and defect response difficulty are divided into imaging condition bins. Within the parameter range defined by each bin, the camera, light source, background, and output resolution are jointly sampled, and the subsequent sampling weights are adjusted according to the number of qualified samples. This can supplement insufficient shooting conditions under a limited rendering budget and alleviate the objective limitations on visual inspection caused by the uncertainty of viewing angle, lighting, background, and product imaging size in the production environment.

[0020] (2) This invention clarifies the basic representation morphology of product visual defects and the corresponding rendering strategy. The support domain of the product surface is used to limit the scope of the defect, and geometric perturbation field and material perturbation field are used to represent shape changes and changes in surface optical properties, respectively. The defect category, severity parameters and spectrum identifier are written into the defect carrier together with the perturbation, so that scratches, dents, protrusions, stains, discoloration, coating abnormalities and component relationship abnormalities can be generated in a controlled manner according to the corresponding physical representation, and form an image response that is coordinated with the product surface, occlusion relationship and lighting state under different imaging conditions, thereby alleviating the objective constraint of the scarcity of real defect sample types and quantities.

[0021] (3) This invention follows the forward generation process of product visual defects, establishing a traceable association between defect carrier construction, imaging condition sampling, defect injection, paired physical rendering under the same conditions, determination of effective defect areas, sample quality gating, and coverage feedback. By fusing the geometrically visible mask obtained from surface projection and depth judgment, the buffered defect area corresponding to the defect identifier buffer, and the photometric response area between the normal reference image and the defect image, pixel-level, instance-level, category-level, severity-level, imaging condition-level, and spectral-level labels are generated; samples that fail the quality gating are resampled according to the reason for failure. Thus, a complete mechanism from defect generation to multi-level fine annotation is constructed, providing a supervised reverse training environment for visual inspection models where image evidence and labels are consistent. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of the visual detection enhancement method based on a physical geometry model according to the present invention.

[0023] Figure 2 This is a flowchart illustrating the defect carrier construction, imaging condition sampling, and defect injection process of the present invention.

[0024] Figure 3 This is a flowchart of the process for pairwise physical rendering and the fusion annotation of visibility and luminance response in the same scene according to the present invention.

[0025] Figure 4 This is a flowchart of the sample quality gating, targeted resampling, and conditional coverage feedback of the present invention.

[0026] Figure 5 This is a structural diagram of the visual inspection enhancement system of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings. The embodiments described are used to illustrate the technical principles of the present invention and are not intended to limit the scope of protection. Without departing from the common concept of defect carrier, paired rendering of the same scene, three-source fusion annotation, and quality feedback, adaptive adjustments can be made to the data structure, renderer interface, and parameter range.

[0028] The physical geometry model in this specification refers to a digital model that can describe the surface shape, component hierarchy, spatial transformations, and material properties of a product, and can be read by the graphics rendering processing unit. Physical rendering refers to the rendering process that calculates images based on camera models, light source models, material reflection characteristics, and geometric occlusion relationships.

[0029] like Figure 1As shown, the method executes steps S1 to S6 sequentially. The following describes each step in the order of model and imaging constraint initialization, defect carrier construction, imaging condition sampling and defect injection, pairwise physical rendering under the same conditions, visibility and photometric response fusion annotation, sample quality gating and coverage feedback.

[0030] Step S1: Model and Imaging Constraint Initialization

[0031] Step S1 includes reading the mesh model, parametric surface model, or assembly model of the industrial product, unifying the length unit and coordinate direction of the model, recording the product's enclosed area, and establishing local coordinates or texture coordinates for the product surface; at the same time, configuring the reference material parameters and imaging constraints to define the model basis and parameter boundaries for subsequent defect construction and rendering.

[0032] Step S1 further includes configuring reference material parameters. These reference material parameters may include base color, reflectivity, roughness, metallic properties, transparency, normal texture, and surface texture. The defect carrier only generates perturbations relative to the reference material within its surface support domain.

[0033] The imaging constraints can be obtained from the design parameters of the cameras and lighting devices at the production site, or they can be configured according to the production conditions to be covered. The imaging constraints should at least specify the camera focal length or field of view, the distance range from the camera to the product, the observation direction range, the position and energy range of the light source, the background set, and the output image size.

[0034] Step S2: Defect Carrier Construction

[0035] like Figure 2 As shown, step S2 includes determining a surface support domain on the product surface, binding a geometric perturbation field and a material perturbation field to the surface support domain, and writing the defect category, severity parameter and spectrum identifier, thereby forming a defect carrier that can simultaneously carry the spatial location, physical characterization and label attributes of the defect.

[0036] A surface support domain can consist of a continuous set of facets on the product surface, or it can consist of a texture coordinate region, a parametric surface region, or a neighborhood defined by surface geodesic distance. The surface support domain also records the product component to which it belongs, ensuring that surface defects do not cross non-adjacent components without constraint.

[0037] The geometric perturbation field defines the direction and magnitude of displacement at each location within the surface support domain. For scratches, dents, or cracks, the displacement direction can be inward along the surface normal; for protrusions or burrs, the displacement direction can be outward along the surface normal; fine surface undulations can also be characterized by changing the local normal.

[0038] The material perturbation field defines the change in material parameters within a surface support domain relative to a reference material. By altering the base color, reflectivity, roughness, transparency, or texture state, stains, discoloration, rust, coating peeling, and ablation marks can be characterized. Geometric and material perturbation fields can be used individually or in combination on the same surface support domain.

[0039] The severity parameter controls at least one of the following: defect length, width, depth, area, color difference, roughness difference, or part offset. The defect category and severity parameter are written to the defect carrier before rendering and become category-level and parameter-level labels directly during sample encapsulation.

[0040] The lineage identifier is unique within a single sample generation task. The lineage identifier is stored in normal model copies, defect model copies, surface support domains, imaging condition parameter sets, rendered images, rendering buffers, and label records, thereby enabling tracing back from any training sample to the corresponding defect carrier and rendering task.

[0041] For structural relationship defects, the surface support domain can be extended to the component support area. Component missingness is achieved by disabling the rendering visibility of the target component; positional offset or abnormal assembly gaps are achieved by changing the transformation of the target component relative to the product coordinate system. The component identifier and transformation amount serve as instance-level and parameter-level labels.

[0042] Step S3: Imaging condition sampling and defect injection

[0043] Step S3 includes establishing conditional bins according to the observation direction, incident direction, background type, imaging scale, and defect response difficulty; sampling camera, light source, background, and output parameters within the corresponding parameter range; modifying the defect-free physical geometry model based on the defect carrier to obtain the defect model; and recording the transformation chain from the surface support domain to the rendering coordinate system.

[0044] Each conditional bin records the number of samples that have passed the sample quality gating. During initialization, each conditional bin can be set with the same sampling weight; during generation, the sampling weights are updated based on the number of samples, giving bins with fewer samples a higher sampling probability. The defect response difficulty bin can be written back based on photometric observables after sample rendering.

[0045] After selecting a conditional bin, random or stratified sampling is performed from the continuous parameter range corresponding to that bin to obtain camera intrinsic parameters, camera extrinsic parameters, light source position, light source direction, light source size, light source energy, background objects, and output resolution. The sampling results are then checked for imaging constraints to form an imaging condition parameter set.

[0046] Furthermore, step S3 can also set parameter-related constraints. For example, camera distance and focal length together ensure that the product projection area does not exceed the image boundary; light source energy and exposure parameters together ensure that normal surfaces are not overexposed; and background objects and products do not geometrically intersect. Parameter groups that do not meet the relevant constraints are resampled before rendering.

[0047] Step S3 involves replicating the defect-free physical geometric model to form a normal model and a model to be modified. The surface support domain of the defect carrier is located in the model to be modified. Geometric and material perturbation fields are applied within the surface support domain to obtain the defect model. The normal model does not receive defect perturbations.

[0048] During defect injection, the transformations at each level from the local coordinate system of the surface support domain to the product coordinate system, scene coordinate system, and camera coordinate system are recorded. These transformations form a transformation chain in the order of execution. For surface positions changed by geometric perturbations, the correspondence between vertices or surface sampling points before and after the perturbation is also recorded.

[0049] Step S4: Paired Physical Rendering under the Same Conditions

[0050] like Figure 3 As shown, step S4 includes placing the normal model and the defect model under the same product pose, camera, light source, background and output settings to perform physical rendering, obtain the normal reference image and the defect image, and simultaneously output the object identifier buffer, the defect identifier buffer, the depth buffer and the surface normal buffer.

[0051] Paired rendering preserves camera intrinsic and extrinsic parameters, product pose, light source parameters, background parameters, output resolution, random sampling seed, and image post-processing parameters. If the renderer uses random ray sampling, the normal model and the defect model reuse the random sequence to reduce random noise differences unrelated to defects.

[0052] The defect model also performs identifier buffer rendering. The object identifier buffer writes a unique object value for each visible product or part; the defect identifier buffer writes a defect value for visible fragments belonging to the defect carrier; the depth buffer records the depth from the nearest visible surface to the camera; and the surface normal buffer records the camera coordinate system normal of the nearest visible surface.

[0053] Color images and all render buffers use the same resolution and pixel origin. If the renderer uses multiple render passes, each pass reuses the same camera and scene state, and scene parameters are locked until all passes are complete.

[0054] Step S5: Visibility and photometric response fusion annotation

[0055] Step S5 includes projecting the surface support domain along the transform chain onto the pixel plane and determining the geometrically visible mask by combining it with the depth buffer, determining the photometric response region based on the pixel difference of the paired images, and then fusing the geometrically visible mask, the buffered defect region corresponding to the defect identifier buffer, and the photometric response region into an effective defect region, thereby generating a multi-level label.

[0056]

[0057] in, Defect carrier The geometric visibility mask, Indicates pixel position, Represents the surface support domain. Indicates the imaging condition parameters A defined projection transformation, Indicates the projected depth of the defect surface. Indicates a deep buffer. Indicates depth tolerance. This indicates a conditional indicator function. This expression is used to exclude occluded surface support regions.

[0058] Step S5-1 also calculates the photometric difference between the pixel-registered normal reference image and the defect image. Since the two images reuse the same imaging conditions and random sampling state, the pixel difference near the defect support domain mainly corresponds to the geometric or material state changes caused by the defect carrier.

[0059]

[0060] in, Defect carrier In pixels Difference in light intensity at that location The color vector representing the defect image. This represents the color vector of the normal reference image. This represents the L2 norm of a color vector. Color vectors can be calculated in the renderer's output color space or a calibrated luminance / chrominance space.

[0061] Step S5-2 extracts the buffered defect regions belonging to the current spectrum identifier from the defect identifier buffer, and identifies pixels with photometric differences reaching a preset photometric response threshold as photometric response regions. Pixel-level fusion is performed on the geometrically visible mask, buffered defect regions, and photometric response regions to obtain the effective defect regions.

[0062]

[0063] in, Indicates the effective defect area. This indicates the buffered defect area obtained by the defect identifier buffer. This represents the photometric response threshold corresponding to the defect category or severity. This formula ensures that the final pixel label simultaneously incorporates surface projection, renderer identification, and photometric response data under the current imaging conditions.

[0064] After obtaining the effective defect region, the system generates a pixel-level defect mask and an instance-level defect identifier according to the connectivity relationship, and derives the defect boundary, circumscribed region, visible area, and center position from the effective defect region. The defect category, severity parameter, component identifier, imaging condition parameter group, and genealogy identifier are written into the tag record along with the region.

[0065] Furthermore, for defects with weak photometric response but still needing to be retained as difficult samples, low and high thresholds can be set. Pixels above the high threshold form defined defect regions, while pixels between the two thresholds that simultaneously satisfy geometric visibility and buffer identification constraints form difficult defect regions, and their difficulty level is written in the label.

[0066] Step S6: Sample quality gating, targeted resampling, and coverage feedback

[0067] Step S6 includes calculating the visible coverage of the effective defect area, the consistency between the geometric visible mask and the buffer defect area boundary, and the photometric observable index; samples that fail to pass the gate are adjusted according to the failure index to adjust the defect position or imaging parameters, and samples that pass the gate enter the encapsulation process and update the conditional bin count and subsequent sampling weight.

[0068]

[0069] in, Defect carrier Visible coverage The visible pixel mask representing the product object. This represents the summation over all pixel locations. Visible coverage is used to exclude samples where defects are completely invisible or occupy too small a proportion in the product projection.

[0070] Step S6-1 also calculates the boundary consistency between the surface projection result and the renderer defect identifier buffer.

[0071]

[0072] in, Indicates boundary consistency. and These represent the geometrically visible mask and the buffer defect region, respectively. The intersection and union are calculated on the pixel plane. A boundary consistency below a threshold indicates inconsistencies between projection, depth determination, or the identifier buffer.

[0073] Step S6-1 also calculates the average photometric response within the geometrically visible area to determine whether the injected defect provides sufficient image evidence under the current camera and lighting conditions.

[0074]

[0075] in, Indicates the observable photometric index, This indicates a positive number to prevent the denominator from being zero. The lower the photometric observable index, the weaker the difference in the image formed by the current lighting and viewing direction, even though the defect is geometrically visible.

[0076] Visible coverage, boundary consistency, and photometric observability are compared with their corresponding thresholds. The thresholds can be configured according to the defect type, severity, and training task; for fine scratches, a smaller visible coverage is allowed but a higher boundary consistency is required, while for specular surface defects, the photometric observability threshold can be set according to the upper limit of imaging noise.

[0077] When the visible coverage does not reach the threshold, the defect location, camera pose, or imaging scale is changed first; when the boundary consistency does not reach the threshold, the current sample is stopped from being encapsulated and the transformation chain, depth buffer, or defect labeling buffer is checked; when the photometric observability does not reach the threshold, the camera pose, light source direction, or light source energy is changed while maintaining the defect category and severity parameters.

[0078] Samples that pass quality gating enter the packaging process. Each supervised training sample contains at least a normal reference image, a defect image, a pixel mask of the effective defect region, an instance-level defect identifier, a defect category, a severity parameter, product and component identifiers, camera parameters, light source parameters, background parameters, imaging scale, quality indicators, and phylogenetic identifiers.

[0079] Sample encapsulation can use a combination of image files and structured label files, or it can be written to a dataset container that supports random access. Regardless of the encapsulation method, image files and label records are associated with sample identifiers and can be traced back to the defect carrier and rendering task through lineage identifiers.

[0080] When a sample passes through the quality gate, its observation direction, incident direction, background type, imaging scale, and photometric response difficulty are mapped to the corresponding condition bin, and the sample count of that condition bin is increased. The normalized sampling weights of the condition bins can be updated according to the following formula.

[0081]

[0082] in, Indicates conditional binning Sampling weights, This indicates the number of samples within the bin that have passed the quality gate. Represents the smoothing constant. This represents the summation over all conditional bins. Conditional bins with fewer samples have a higher probability of subsequent sampling.

[0083] like Figure 4 As shown, the quality gating result directly changes the parameter state of the next batch of samples. Samples that pass the gating update the condition coverage count; samples that fail the gating record failure indicators, and the failure reasons are mapped to the set of parameters to be adjusted, thereby avoiding indiscriminate resampling of all parameters.

[0084] like Figure 5 As shown, the model and constraint management module is used to maintain the normal model, reference material parameters and imaging constraints; the defect carrier construction module is used to establish the surface support domain, disturbance field, defect attributes and spectrum identifier; and the imaging condition sampling module is used to generate the imaging condition parameter set.

[0085] The Defect Injection and Paired Physical Rendering module is used to output the normal model, defect model, transformation chain, normal reference image, defect image, and rendering buffer; the Visibility and Photometric Response Fusion Annotation module is used to generate the geometric visibility mask, photometric response region, effective defect region, and multi-level labels.

[0086] The sample quality gating module calculates visible coverage, boundary consistency, and photometric observability metrics; the coverage feedback and sample encapsulation module triggers targeted resampling based on the cause of failure, or encapsulates qualified samples and updates conditional binning weights. Each module is configured according to... Figure 1 The steps shown are executed sequentially and in tandem.

[0087] In one embodiment of scratching a metal casing surface, a continuous patch is selected from the casing surface as a surface support region. A scratch region is formed by a centerline extending tangentially along the surface and a width varying with the centerline. The scratch depth is formed by an inward normal displacement, and the roughness and color parameters of the scratch region are varied. The system records the scratch length, width, and depth as severity parameters.

[0088] For the metal casing, the imaging condition sampling module samples from the allowed azimuth, pitch, camera distance, light source direction, and background set. The defect injection and paired physical rendering modules generate a normal reference image, a scratch image, and a rendering buffer under the same parameter set. If the scratch is located at the edge of the casing and is occluded, the visibility coverage gate prevents the current sample from entering the training set.

[0089] When a scratch is geometrically visible but the light source orientation results in insufficient color difference between it and a normal surface, photometric observables trigger resampling of the light source orientation or camera pose. Resampling preserves the scratch category and severity, allowing the same physical defect to form an image response that can be used for supervised training under new imaging conditions.

[0090] In one example of a coated stain, the defect carrier alters only the color, reflectivity, and roughness within the surface support domain without changing its geometric location. The transform chain is still used to determine the support domain projection, the depth buffer is used to cull occlusions, and the pairwise photometric difference is used to confirm the actual response of the material changes under the current illumination.

[0091] In one embodiment of assembly defects, the supporting domain of the defect carrier is the target component, and the defect injection process disables the rendering visibility of the target component. The normal reference image retains the target component, while the defect image does not contain the target component; the effective defect area can be determined by the visible projection of the target component in the normal reference image, the object identifier buffer, and the paired photometric difference.

[0092] Multi-level labels can be used to train classification, object detection, instance segmentation, semantic segmentation, severity regression, or pairwise differential learning models. The training task selection does not alter the processing relationships of defect vectors, paired rendering within the same scene, three-source fusion labels, and quality feedback.

[0093] Through the above implementation method, how defects are generated, whether they are visible in the current scene, whether sufficient light intensity response is formed, how the final label is determined, and how the next round of samples are sampled are all connected by traceable intermediate states, thereby forming a closed loop of supervised training sample generation for complex production shooting conditions.

Claims

1. A visual detection enhancement method based on a physical geometric model, characterized in that, include: S1. Obtain the defect-free physical geometry model, reference material parameters, and imaging constraints of the industrial product, and establish a surface coordinate system on the product surface; S2. Construct a defect carrier within the surface coordinate system. The defect carrier includes a surface support domain, a geometric perturbation field, a material perturbation field, a defect category, a severity parameter, and a spectrum identifier; S3. Sample the camera, light source, background, and imaging scale parameters according to the imaging conditions, generate a defect model based on the defect carrier, and save the transformation chain from the surface support domain to the rendering coordinate system; S4. Under the same imaging conditions, perform physical rendering on the defect-free physical geometry model and the defective model respectively to obtain a normal reference image, a defective image, a defect identifier buffer, and a depth buffer. S5. Determine the geometrically visible mask based on the transformation chain and depth buffer, determine the photometric response region based on the pixel difference between the normal reference image and the defect image, and fuse the geometrically visible mask, the buffered defect region corresponding to the defect identifier buffer, and the photometric response region into an effective defect region. Generate a multi-level label containing defect location, defect type, severity parameter, imaging condition parameter, and spectral identifier from the effective defect region. S6. Perform quality gating on the samples based on visible coverage, boundary consistency, and photometric observability indicators. Resample the samples that fail the gating according to the failure index, encapsulate the samples that pass the gating, and update the subsequent sampling weights of the corresponding imaging condition bins.

2. The method according to claim 1, characterized in that, The surface support domain is determined by a set of facets, texture coordinate regions, parametric surface regions, or component support regions on the product surface; the geometric perturbation field is used to change the position or normal of vertices within the surface support domain, and the material perturbation field is used to change the reflectivity, roughness, color, or texture state of the surface support domain; the spectrum identifier is used to associate the defect carrier, defect model, rendered image, and multi-level labels.

3. The method according to claim 1, characterized in that, In step S4, the physical rendering maintains the same camera parameters, product pose, light source parameters, background parameters, and output resolution, and makes the defect image, defect marker buffer, and depth buffer use the same pixel origin and resolution.

4. The method according to claim 1, characterized in that, Step S5 includes: projecting the surface support domain along the transform chain onto the pixel plane; comparing the projection depth with the pixel depth in the depth buffer to obtain the geometrically visible mask; calculating the color difference between the normal reference image and the defect image at the same pixel position to obtain the photometric response region; and finding the intersection of the geometrically visible mask, the buffered defect region, and the photometric response region to obtain the effective defect region.

5. The method according to claim 1, characterized in that, When the visible coverage does not reach the threshold, adjust the defect location, camera pose, or imaging scale; when the boundary consistency does not reach the threshold, regenerate the transformation chain, depth buffer, or defect identification buffer; when the photometric observability index does not reach the threshold, adjust the camera pose or light source parameters; when a sample passes through the gating, increase the subsequent sampling weight of the imaging condition bin with a small number of samples.

6. A visual inspection enhancement system based on a physical geometric model, characterized in that, It includes a model and constraint management module, a defect carrier construction module, an imaging condition sampling module, a defect injection and pairwise physical rendering module, a fusion annotation module, a sample quality gating module, and a coverage feedback and sample encapsulation module; each module is configured to perform the method described in any one of claims 1 to 5.

Images