Method and system for detecting hard material deformation defects in a partially occluded environment

Abstract

This invention proposes a method and system for detecting deformation defects in hardware materials under partially occluded environments, belonging to the field of material inspection technology. The method includes acquiring an image of the target hardware material and parsing the image to obtain the visible area of ​​the target hardware material; calculating the contribution of the visible area to the complete material contour based on the visible area and a pre-stored standard contour of the target hardware material; and comparing the contribution with a preset selection threshold. This invention acquires the image to be inspected, parses the visible area of ​​the target hardware material, calculates the contribution of the visible area to the complete contour, and intelligently selects and switches between a direct detection strategy and a reconstruction-assisted detection strategy based on a preset selection threshold. This enables the inference of deformation in occluded areas and the evaluation of the overall structural state of the material, thereby significantly improving the completeness and reliability of defect detection under complex occlusion environments.

Description

Technical Field

[0001] This invention relates to the field of material inspection technology, and in particular to a method and system for detecting deformation defects in hardware materials under partially obscured environments. Background Technology

[0002] Critical infrastructure in sectors such as power, communications, and energy utilizes a large number of hardware components, including insulators, fittings, and structural parts. Over long-term operation, these components are susceptible to defects such as surface corrosion, cracks, overall deformation (e.g., bending, expansion, collapse), or installation displacement due to mechanical stress, environmental corrosion, and accidental impacts. These defects directly affect equipment safety and system stability. Therefore, regular and efficient defect inspection of hardware components is a crucial step in ensuring the reliable operation of infrastructure.

[0003] Currently, computer vision-based automated inspection methods, due to their advantages of being non-contact, highly efficient, and low-cost, can be applied to the aforementioned scenarios. Typical existing technical solutions usually rely on cameras to capture images of materials and directly identify visible surface defects, such as scratches and stains, through image processing or deep learning models. However, in actual inspection environments, target materials are often obscured by other equipment, vegetation, or parts of their own structure, resulting in incomplete image information. This means that the above methods can only analyze the visible parts of the image and cannot know any state of the obscured areas, thus completely missing deformation defects occurring in the obscured areas. Furthermore, because the contour information of the obscured parts of the materials cannot be obtained, existing methods cannot determine whether the materials have undergone structural deformation as a whole, such as overall bending, asymmetrical expansion or collapse, or installation position misalignment. These overall deformations are often precursors to more serious safety hazards. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for detecting deformation defects in hardware materials under partially occluded environments. By acquiring the image to be detected and parsing the visible area of ​​the target hardware material, the contribution of the visible area to the complete contour is calculated. Based on a preset selection threshold, the system intelligently selects and switches between a direct detection strategy and a reconstruction-assisted detection strategy. This effectively overcomes the inherent limitations of single-vision detection methods in partially occluded environments. When the visibility of the material is high, a precise direct detection path is used; when occlusion is severe and visible information is insufficient, the system automatically switches to a reconstruction-assisted detection path. Through contour reconstruction and information fusion, the system can infer the deformation of the occluded area and evaluate the overall structural state of the material, thereby significantly improving the completeness and reliability of defect detection under complex occlusion environments.

[0005] The technical solution of this invention is implemented as follows: In a first aspect, the present invention provides a method for detecting deformation defects in hardware materials under partially obscured environments, comprising the following steps: Acquire an image containing the target hardware material to be inspected, and parse the image to be inspected to obtain the visible area of ​​the target hardware material; Based on the visible area and the standard outline of the pre-stored target hardware material, calculate the contribution of the visible area to the complete material outline. A preset selection threshold is set, the contribution is compared with the selection threshold, and one of the direct detection strategy and reconstruction-assisted detection strategy is selected. If the contribution is higher than the selection threshold, a direct detection strategy is adopted, which performs defect analysis based on the visible area and outputs the first defect detection result. If the contribution is not higher than the selection threshold, a reconstruction-assisted detection strategy is adopted, and a second defect detection result is output. The reconstruction-assisted detection strategy includes: Local geometric features are extracted from the visible area and combined with the pre-stored three-dimensional geometric priors of the target hardware material category to reconstruct a complete material outline model of the target hardware material. The complete material outline model is compared with the standard outline to obtain the outline deformation defect. The outline deformation defect is then fused with the visible surface defect to obtain the second defect detection result.

[0006] In some embodiments, acquiring an image to be detected containing the target hardware material and parsing the image to be detected to obtain the visible area of ​​the target hardware material specifically includes: Semantic segmentation is performed on the image to be detected to segment the foreground region of the material and the occlusion region, and the depth information of the image to be detected is obtained. By integrating the results of semantic segmentation with depth information, the part of the unobstructed area in the foreground region of the material is determined to be occluded in the depth direction and is taken as the initially visible area. The boundary of the initially visible area is optimized to output the visible area of ​​the target hardware material.

[0007] In some embodiments, calculating the contribution of the visible region to the complete material outline based on the visible region and the pre-stored standard outline of the target hardware material specifically includes: The standard outline of the pre-stored target hardware material is projected in two dimensions according to the posture of the target hardware material in the image to be detected, generating a reference standard outline aligned with the image space. Extract the visible outline of the target hardware material from the visible area; Calculate the geometric overlap measure between the visible portion of the contour and the reference standard contour, wherein the geometric overlap measure includes at least one of the contour intersection union ratio and the region area coverage ratio. Identify multiple predefined key geometric features in the reference standard contour and count the proportion of key geometric features covered by the visible part of the contour; Based on the geometric overlap metric and the proportion of key geometric features covered by the visible outline, a weighted calculation is performed to obtain the contribution of the visible area to the complete material outline.

[0008] In some embodiments, the direct detection strategy, which performs defect analysis based on the visible area and outputs a first defect detection result, specifically includes: Image analysis is performed on the visible area to extract multi-dimensional visual features that characterize the surface state of the material. The multi-dimensional visual features include at least two of the following: surface texture statistical features, edge gradient distribution features, and color space anomaly features. The extracted multi-dimensional visual features are input into a pre-trained defect recognition model, which outputs a judgment result on the category, location, and severity level of defects existing in the visible area. The judgment result is optimized, which includes filtering the judgment result based on the confidence level output by the defect recognition model and performing morphological operations on the detected defect area to optimize its boundary. The integrated and optimized judgment results generate a first defect detection result that includes the defect type, location, size, and severity.

[0009] In some embodiments, the step of extracting local geometric features from the visible area and combining these local geometric features with pre-stored three-dimensional geometric priors of the target hardware material's category to reconstruct a complete material outline model of the target hardware material specifically includes: The visible region is processed by a deep convolutional neural network to extract deep feature maps, and the deep feature maps are encoded into a set of feature vectors that represent local geometric structure and spatial context relationships. Based on the category information of the target hardware material, the corresponding basic 3D model is retrieved from the pre-stored 3D geometric prior library. The feature vector is matched with the projection features of the basic 3D model under different viewpoints to estimate the corresponding part and pose of the visible area in the complete 3D model. Using feature vectors and matched part and pose information as conditions, a pre-trained generative model is driven to generate an implicit representation of the complete two-dimensional contour of the target hardware material in pose. The implicit representation of the complete two-dimensional contour is decoded into an explicit set of contour points or polygons, and the decoded contour is geometrically normalized to ensure that it conforms to the overall topological constraints of the category's three-dimensional geometric prior, thus outputting a complete material contour model.

[0010] In some embodiments, comparing the complete material outline model with a standard outline to obtain outline deformation defects, and fusing the outline deformation defects with visible surface defects to obtain the second defect detection result, specifically includes: The complete material outline model is spatially aligned with the standard outline, and the geometric difference field between the two is calculated. Based on the geometric difference field, the location, deformation direction and deformation amount of the outline deformation area are identified and quantified. Within the visible area, perform anomaly detection of surface texture and structure to identify and locate surface defects; Establish the spatial relationship between contour deformation regions and surface defects. Based on predefined fusion rules, perform information fusion on contour deformation defects and surface defects to obtain defect information. The fusion rules are used to handle the defect priority of spatially overlapping regions and the merging logic of different types of defects. Based on the fused defect information, a structured second defect detection result is generated, which includes at least a description of the overall contour deformation, a list of local surface defects, and inferred defect hints for occluded areas.

[0011] In some embodiments, the step of fusing information from contour deformation defects and surface defects based on predefined fusion rules to obtain defect information specifically includes: Based on the category of the target hardware material and the type of the contour deformation defect, the corresponding rule group is called from the predefined defect fusion rule library. The rule group contains at least defect conflict resolution rules and defect merging conditions. For regions where contour deformation areas and surface defects overlap or are adjacent in space, analysis is performed based on defect conflict resolution rules. These rules include prioritizing the use of contour deformation defects when macroscopic contour deformation and microscopic surface defects coexist at the same location, and labeling surface defects as accompanying or derived phenomena of contour deformation. For non-conflicting defects, information is aggregated according to defect merging conditions. These conditions include a weighted calculation based on deformation and the severity level of surface defects to determine whether multiple discrete defects are caused by the same mechanical cause and to decide whether to merge them into a comprehensive defect description. Each defect information generated after fusion is labeled with a confidence level. The confidence level is determined based on the confidence level of the original evidence in the visible area on which the defect is based, as well as the integrity of the logical links it undergoes during the reconstruction and fusion process.

[0012] Secondly, the present invention provides a system for detecting deformation defects in hardware materials under partially obscured environments, for implementing the above method, comprising: The image acquisition and parsing module is used to acquire and parse the image to be detected in order to obtain the visible area of ​​the target hardware material; The visibility assessment module is used to calculate the contribution of the visible area to the complete material outline based on the visible area and the pre-stored standard outline. The strategy decision module is used to compare the contribution with a preset selection threshold and select a direct detection strategy or a reconstruction-assisted detection strategy accordingly. The direct detection module is used to perform defect analysis based on the visible area and output a first defect detection result when the direct detection strategy is selected; A reconstruction-assisted detection module, used to perform detection when the reconstruction-assisted detection strategy is selected, includes: The contour reconstruction unit is used to reconstruct and generate a complete material contour model of the target hardware material based on the local geometric features of the visible area and the pre-stored category three-dimensional geometric priors. The fusion detection unit is used to compare the complete material outline model with the standard outline to obtain the outline deformation defect, and fuse it with the surface defect of the visible area to output a second defect detection result.

[0013] Thirdly, the present invention provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the above-described method.

[0014] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0015] The method and system for detecting deformation defects in hardware materials under partially obscured environments of the present invention have the following advantages over the prior art: 1. By acquiring the image to be detected and parsing the visible area of ​​the target hardware material, the contribution of the visible area to the complete contour is calculated. Based on a preset selection threshold, the system intelligently selects and switches between direct detection strategy and reconstruction-assisted detection strategy. This effectively overcomes the inherent limitations of single vision detection methods in partially occluded environments. When the visibility of the material is high, a precise direct detection path is adopted. When the occlusion is severe and the visible information is insufficient, the system automatically switches to the reconstruction-assisted detection path. Through contour reconstruction and information fusion, the system can infer the deformation of the occluded area and evaluate the overall structural state of the material, thereby significantly improving the integrity and reliability of defect detection in complex occluded environments. 2. By employing the fusion of semantic segmentation results and depth information to determine the visible region, and comprehensively calculating the contribution based on 2D projection alignment, geometric overlap measurement, and key feature visibility assessment, a precise and robust perception and quantification foundation is established. This allows for accurate differentiation between visual overlap and real 3D occlusion, thus obtaining the true visible region. Furthermore, the contribution calculation not only reflects the proportion of visible area but also assesses the visibility of key structural features, making subsequent path decisions based not only on quantity but also on quality. This significantly reduces the risk of missed detections due to insufficient structural information in the visible parts but still misusing direct detection strategies, providing accurate prerequisites for the reliable operation of the entire system. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the method and system for detecting deformation defects in hardware materials under partially obscured environments according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] like Figure 1 As shown, the core of the present invention for detecting deformation defects in hardware materials under partially occluded environments lies in constructing a dual-path intelligent detection framework to overcome the inherent limitations of traditional visual detection due to incomplete information under occluded environments. This method intelligently determines the most suitable detection path by quantitatively evaluating the sufficiency of visible information, thereby ensuring efficiency while maximizing the integrity and reliability of the detection results. The main steps of the method are described in detail below.

[0020] Step S100: Obtain an image to be detected containing the target hardware material, and parse the image to be detected to obtain the visible area of ​​the target hardware material.

[0021] In some embodiments, the system first acquires an image of the target hardware material using fixed or mobile visual acquisition devices deployed on-site, such as high-definition cameras or industrial cameras. Due to the complex on-site environment, the target material may be obscured by adjacent equipment, supports, vegetation, or other parts of itself. Therefore, the key task is to accurately separate the unobscured visible portion of the target material from the entire image. To achieve this, the system analyzes the image to be detected. In one possible implementation, this analysis process combines semantic segmentation technology with spatial relationship analysis. First, a pre-trained semantic segmentation model is used to identify pixel regions in the image belonging to the target hardware material category and the occlusion category. Then, to further distinguish between visual overlap and actual 3D spatial occlusion, depth information can be used for analysis to determine which parts of the target material are truly located in front of the occlusion and not completely covered in 3D spatial depth. Finally, by combining the 2D segmentation results and 3D spatial relationships, the visible area of ​​the target hardware material is accurately extracted, that is, the surface portion of the material that can be directly observed in the actual 3D scene, providing accurate input for all subsequent analyses.

[0022] Step S200: Based on the visible area and the standard outline of the pre-stored target hardware material, calculate the contribution of the visible area to the complete material outline.

[0023] After obtaining the precise visible area, it is necessary to objectively and quantitatively evaluate the sufficiency of the current visible information for judging the overall state of the materials. The standard contour is a two-dimensional contour template or three-dimensional model projection pre-established for each type of hard material, representing its standard size and shape in its intact state. The process of calculating the contribution first involves spatial alignment, where the standard contour is affinely transformed or projected from the perspective of the target material in the current image to place it in the same coordinate system as the visible area in the image. Next, the system calculates the geometric relationship between the aligned visible area and the standard contour. The contribution is a comprehensive quantitative indicator that not only calculates the proportion of the visible area's pixel area to the total area of ​​the standard contour, but also incorporates the evaluation of the visibility of key geometric features. For example, the system will identify key points in the standard contour that are sensitive to deformation, such as corners, edge intersections, or feature lines, and count the proportion of these key features covered in the visible area. Finally, by weighted fusion of the area overlap rate and the key feature visibility rate, a comprehensive contribution value representing the information value of the currently visible part is calculated. The higher the value, the more reliable the judgment of the material's state is based solely on the visible part.

[0024] Step S300: Preset selection threshold, compare contribution with selection threshold, and select one of direct detection strategy and reconstruction-assisted detection strategy.

[0025] In this step, the system needs to preset one or a set of selection thresholds. These thresholds can be set or adaptively adjusted according to the importance of different material types and the requirements of detection accuracy. The contribution calculated in step S200 is compared with the selection threshold. If the contribution is higher than the selection threshold, it indicates that the currently visible area provides sufficient information, and the system selects the direct detection strategy. Conversely, if the contribution is not higher than the selection threshold, it indicates severe occlusion, and the visible information is insufficient to independently support a reliable overall state assessment. In this case, the system will select the reconstruction-assisted detection strategy, which compensates for the lack of visual information through information reconstruction and fusion. This decision-making mechanism ensures that the system can adaptively switch between efficiency-first and integrity-first modes according to the actual situation.

[0026] Step S400: If the contribution is higher than the selection threshold, a direct detection strategy is adopted to perform defect analysis based on the visible area and output the first defect detection result.

[0027] When the system decides to enter this path, it will perform in-depth analysis of the acquired visible area. Defect analysis based on the visible area can employ a deep learning-based defect detection model. This model, trained on a large amount of labeled data, can directly analyze the input visible area image, identifying and locating surface defects such as cracks, rust, damage, and stains. After analysis, the system integrates all detected defect information to generate the first defect detection result. This result is output in a structured data format, including the defect type, its specific location in the image, its geometric dimensions, severity level, and detection confidence. This path fully utilizes visible information, has a fast processing speed, and is suitable for most routine inspection scenarios with minimal occlusion.

[0028] Step S500: If the contribution is not higher than the selection threshold, then the reconstruction-assisted detection strategy is adopted and the second defect detection result is output.

[0029] When the system determines that there is insufficient visible information, it initiates this more complex innovation path. The core idea of ​​this path is to infer the invisible from the visible, and to restore the overall understanding of the material by reconstructing the geometric shape of the obscured part, thereby detecting deformation defects that cannot be found by relying solely on the visible part.

[0030] The reconstruction-assisted detection strategy specifically includes two core sub-steps, namely step S510 and step S520.

[0031] Step S510: Extract local geometric features from the visible area, and combine the local geometric features with the pre-stored three-dimensional geometric priors of the target hardware material category to reconstruct a complete material outline model of the target hardware material.

[0032] First, the system extracts local geometric features from the obtained visible area that characterize its local shape and structure, such as edge orientation, curvature, and corner distribution. Simultaneously, based on the material category, the system retrieves the corresponding category's 3D geometric prior from the knowledge base. This prior can be a parametric 3D CAD model or a set of basis vectors representing common shape variations of that type of material. Then, the system constructs and runs a generative model, such as a conditional generative adversarial network or a diffusion model. This model uses the extracted local geometric features as conditional input and the category's 3D geometric prior as shape constraints to learn the mapping relationship from the local to the global. The model's goal is to generate a complete material outline model that is geometrically continuous with the currently visible portion and conforms to the general shape rules of that type of material.

[0033] Step S520: Compare the complete material outline model with the standard outline to obtain the outline deformation defect, and fuse the outline deformation defect with the visible area surface defect to obtain the second defect detection result.

[0034] After obtaining the reconstructed complete outline, the system has the foundation for overall deformation analysis. This step first involves a detailed comparison between the generated complete material outline model and the standard outline in the same coordinate system. By calculating the distance field, Hausdorff distance, or key dimension differences between the two, the system can quantitatively identify whether the material exhibits overall outline deformation defects such as bending, expansion, collapse, or asymmetry, and determine their location and degree of deformation. Simultaneously, the system performs surface defect analysis on the visible area, similar to step S400. Finally, based on spatial relationships and preset logical rules, the system integrates the outline deformation defects and visible area surface defects into a unified framework, fusing them to generate a second defect detection result. This result not only includes details of visible surface defects but, more importantly, adds judgment and description of the overall deformation in invisible areas. This provides a comprehensive report that reflects the overall safety status of the material far better than direct detection, achieving a technological leap from local observation to overall assessment.

[0035] In some embodiments, the step of acquiring an image to be detected containing the target hardware material and parsing the image to be detected to obtain the visible area of ​​the target hardware material specifically includes steps S101-S103, which can achieve high-precision extraction of the truly unobstructed observable part of the target material in a three-dimensional space from complex scene images, laying a precise data foundation for subsequent evaluation and decision-making.

[0036] Step S101: Perform semantic segmentation on the image to be detected, segment the foreground region of the material and the occlusion region, and obtain the depth information of the image to be detected.

[0037] This step performs two key preprocessing tasks in parallel. First, a pre-trained semantic segmentation model, such as one based on DeepLabV3+ or Mask R-CNN, is used to perform pixel-level classification of the input image to be detected. The semantic segmentation model can accurately identify and segment all pixel sets belonging to the target hardware material category in the image, called the material foreground region. At the same time, it identifies and segments pixel sets belonging to the category of possible occlusions, such as "other equipment," "vegetation," and "packaging," called the occlusion region. This process achieves preliminary separation of different object categories in the image. Second, the depth information of the image to be detected is acquired or calculated simultaneously. The source of the depth information can be a directly equipped RGB-D camera, such as a structured light, ToF, or binocular camera, which can provide the actual distance value corresponding to each pixel. If only a monocular RGB camera is used, the relative or absolute depth map of the scene can be estimated from a single image through an advanced monocular depth estimation algorithm.

[0038] Step S102: Combine the results of semantic segmentation with depth information to determine the part of the unobstructed area in the foreground region of the material that is occluded in the depth direction, and take it as the initially visible area.

[0039] This step distinguishes between overlap on a 2D image and occlusion in real 3D space. The system performs pixel-level registration and fusion analysis on the semantic segmentation mask obtained in step S101, namely the material foreground region and the occluding object region, with the depth map. For each pixel in the material foreground region, the system queries its corresponding depth value in the depth map, i.e., the distance from the point to the camera, and compares it with the depth value of the occluding object region located at the same image coordinate position. According to the 3D spatial relationship, if a point belongs to the material foreground and its depth value is less than the depth value of any overlapping occluding object region, then the point is considered to be unoccluded and visible from the current viewpoint. By traversing the entire material foreground region and applying this 3D occlusion judgment logic, the system can filter out all pixels that win in the depth direction. The set of these points is determined as the preliminary visible region. This process effectively excludes those material parts that are covered by occluding objects in the 2D image but are actually located behind the occluding objects, ensuring the 3D authenticity of the visible region determination.

[0040] Step S103: Perform boundary optimization processing on the initially visible area and output the visible area of ​​the target hardware material.

[0041] The preliminary visible region obtained after step S102 may have jagged, hollow, or discontinuous boundaries due to noise in the depth map, inaccurate segmentation edges, or minor judgment errors. To improve the stability and accuracy of subsequent feature extraction and analysis, this step performs boundary optimization processing on the preliminary visible region. Optimization processing typically includes morphological operations, such as performing a closing operation to fill small holes caused by noise within the region, followed by an opening operation to smooth the boundaries and remove isolated noise points. In addition, edge-based connection algorithms or conditional dilation / erosion operations can be applied to ensure that the boundaries of the visible region are coherent and smooth, and can better fit the actual physical edges of the target material. After optimization, the final output is a clearly defined visible region of the target hard material, which reflects the surface part of the target material that can be directly observed by the camera in the actual 3D scene.

[0042] In some embodiments, the step of calculating the contribution of the visible area to the complete material outline based on the visible area and the pre-stored standard outline of the target hardware material specifically includes steps S201-S205. This step is used to construct a multi-dimensional evaluation system to quantify the information value of the current visible part in terms of geometry for judging the overall state of the material.

[0043] Step S201: Project the pre-stored standard contour of the target hardware material in two dimensions according to the pose of the target hardware material in the image to be detected, and generate a reference standard contour aligned with the image space.

[0044] The pre-stored standard contour is typically a standard view of an ideal two-dimensional contour template or a three-dimensional CAD model of the material in its intact and undeformed state. To effectively compare it with the actual target in the current image, the standard contour must first be aligned with the target's pose in the image. The system first estimates the approximate spatial pose of the target material in the image to be detected, such as through feature point matching or coarse-grained pose estimation algorithms. Then, based on this estimated pose, the system performs corresponding two-dimensional projection transformations on the three-dimensional standard model, such as perspective projection or affine transformation, or performs corresponding geometric deformations on the two-dimensional template, thereby generating a reference standard contour that is perfectly aligned with the current image coordinate system. The reference contour represents the complete shape that an ideal, unobstructed target should present under the current shooting perspective, providing an accurate benchmark for subsequent geometric comparisons.

[0045] Step S202: Extract the visible outline of the target hardware material from the visible area.

[0046] The system tracks and analyzes the boundaries of the visible area, extracting continuous curves or polygons that represent the visible shape of the target material, i.e., the visible outline. This process involves edge detection and outline tracking algorithms, and the results are smoothed and simplified to obtain the main outline that can clearly express the geometric features. The extracted outline is the direct object for subsequent geometric measurement.

[0047] Step S203: Calculate the geometric overlap metric between the visible portion of the contour and the reference standard contour, wherein the geometric overlap metric includes at least one of the contour intersection union ratio and the region area coverage ratio.

[0048] Geometric overlap measures are used to evaluate the degree of spatial agreement or coverage between a visible profile and a reference standard profile. Specifically, they can include the union ratio of profile intersections and the area coverage ratio of a region.

[0049] For the union ratio of the intersection points of the contours, the regions enclosed by the visible part contour and the reference standard contour can be calculated separately. Then, the ratio of the intersection area to the union area of ​​these two regions can be calculated. The closer the ratio is to 1, the more the shape and position of the visible part matches the corresponding part of the standard contour.

[0050] The area coverage ratio can be calculated as the percentage of the area enclosed by the visible part of the outline to the area enclosed by the entire reference standard outline. This ratio directly reflects the contribution of the visible part to the complete outline in terms of area.

[0051] The system can choose to use one or a combination of these two measures to initially quantify visibility from two dimensions: shape consistency and area ratio.

[0052] Step S204: Identify multiple predefined key geometric features in the reference standard contour and count the proportion of key geometric features covered by the visible part of the contour.

[0053] For each type of hardware, a set of key geometric features crucial to its structural integrity and deformation detection is predefined. These features may include specific corner points, representative long line segments, curvature extrema, or the shape of specific functional parts, such as the edge of an insulator cap or the end edge of a crossarm. The system automatically locates these predefined key features on the reference standard contour generated in step S201, and then determines whether each key feature point or feature segment falls within the area defined by the visible contour extracted in step S202. Finally, the proportion of covered key features to the total number of key features is calculated. This proportion reflects how much of the critical information relied upon for structural health assessment of the material is directly available. For example, even if the visible area is large, if all the key positioning points of the connecting bolts are obscured, their contribution should be reduced.

[0054] Step S205: Based on the geometric overlap metric and the proportion of key geometric features covered by the visible part of the outline, the contribution of the visible area to the complete material outline is obtained by weighted calculation.

[0055] The system performs a weighted fusion calculation by combining the calculated geometric overlap metric and the obtained key feature visibility ratio, according to pre-defined weights. The weighting can reflect the emphasis on different evaluation dimensions; for example, for deformation detection tasks, the weight of the key feature visibility ratio might be set higher. The weighted calculation ultimately outputs a comprehensive scalar value, namely the contribution score. This value, between 0 and 1, comprehensively reflects the value of the current visible area in reconstructing and evaluating the overall outline of the material in terms of both the completeness of geometric coverage and the observability of structural features. This contribution score will serve as the core criterion, directly triggering different subsequent detection strategies.

[0056] In a specific calculation process, the geometric overlap metric uses the area coverage ratio (ACR), which is expressed as:

[0057] Among them, C v For the visible outline, C s For reference to the standard profile, Area(C) v ) and Area (C s ) represent C respectively v With C s The area enclosed by the boundary, Area(C) v ∩C s The intersection of two contour regions is represented by , which can be obtained through polygon Boolean operations. The intersection area is divided by the standard contour area to obtain the ACR. The value range of ACR is [0, 1].

[0058] Key Feature Visibility Ratio (KFVR) is expressed as:

[0059] Wherein, V(k) i ) is feature k i The visibility judgment function, with a predefined key geometric feature set K = {k1,k2, ..., k n}, where n is the total number of features in the key geometric feature set.

[0060] For point features, if point k i Located in C v Inside the polygon, V(k) i =1, otherwise 0.

[0061] For line segment characteristics, if line segment k i The proportion exceeding a certain length is located at Cv Inside the polygon, V(k) i =1, otherwise 0.

[0062] Count the number of times V(ki)=1 among all n key features, and divide by n to obtain KFVR. The value range of KFVR is [0,1].

[0063] By weighted fusion of the two sub-metrics mentioned above, the final contribution CS can be obtained: CS = α·ACR + (1-α)·KFVR Here, α is an adjustable weighting coefficient, and 0≤α≤1.

[0064] If area coverage information is trusted more, α can be set to >0.5, such as 0.7; if key features are considered crucial for defect detection of this type of material, α should be set to <0.5, such as 0.3, to amplify the negative impact of key features not being visible. α can be preset empirically for different material categories or obtained through a small amount of learning.

[0065] In some embodiments, the direct detection strategy is adopted, which performs defect analysis based on the visible area and outputs a first defect detection result, specifically including steps S401-S404. This step is suitable for scenarios where visible information is sufficient and contour reconstruction is not required, and is used to quickly identify and quantify various defects on the surface of the target material.

[0066] Step S401: Perform image analysis on the visible area to extract multi-dimensional visual features for characterizing the surface state of the material. The multi-dimensional visual features include at least two of the following: surface texture statistical features, edge gradient distribution features, and color space anomaly features.

[0067] The system performs in-depth analysis of the acquired visible area images, relying not on a single feature but fusing and extracting multi-dimensional complementary visual features to enhance sensitivity to different types of defects and robustness to environmental changes. Specific extracted features may include surface texture statistical features, edge gradient distribution features, and color space anomaly features.

[0068] The roughness, regularity, and uniformity of a material surface can be quantified by calculating the gray-level co-occurrence matrix, local binary modes, or utilizing Gabor filter bank responses. For example, rusted areas typically exhibit disordered textures and increased roughness.

[0069] Edges are extracted using operators such as Sobel and Canny, and their density, orientation distribution histograms, and gradient magnitudes are statistically analyzed. Defects such as cracks and breaks can introduce abnormal directional edges or local high-gradient regions.

[0070] The image is converted from RGB color space to HSV, Lab, or other color spaces that are less sensitive to changes in lighting. The color mean and standard deviation of the region are calculated, or the Mahalanobis distance from the color of a standard material surface is calculated. Defects such as stains and discoloration will manifest as local deviations in color distribution.

[0071] The system uses at least the two types of features mentioned above in combination to construct a high-dimensional feature vector that can comprehensively characterize the abnormal state of the surface, providing rich and robust input for subsequent model judgment.

[0072] Step S402: Input the extracted multi-dimensional visual features into the pre-trained defect recognition model, and the defect recognition model outputs the judgment results of the category, location and severity level of defects existing in the visible area.

[0073] The pre-trained defect recognition model can employ a deep convolutional neural network, such as Faster R-CNN, YOLO series, or U-Net segmentation networks. This model has been trained on a massive dataset of hardware material images labeled with various defects, such as cracks, corrosion, punctures, and damage. The model receives the feature representation generated in step S401, or uses the original image as input, automatically extracting deeper features internally. After forward propagation computation, the model directly outputs a structured judgment result, which may include the specific category of the defect, its precise location in the image, and the severity level defined based on defect size, contrast, etc.

[0074] Step S403: Optimize the judgment result. The optimization process includes filtering the judgment result based on the confidence level output by the defect recognition model and performing morphological operations on the detected defect area to optimize its boundary.

[0075] To ensure the accuracy and neatness of the output, this step performs post-processing optimization on the model's original output. The optimization mainly involves two aspects: confidence filtering and morphological boundary optimization.

[0076] The model outputs a confidence score for each detected defect instance. The system sets a confidence threshold, such as 0.7. Any judgment result with a confidence score lower than this threshold will be regarded as suspicious or noise and filtered out, thereby effectively reducing the false alarm rate.

[0077] For the defect region mask output by the model, especially the pixel-level segmentation results, its boundaries may have burrs or irregularities. The system adopts morphological operations, such as first performing a closing operation, that is, dilation followed by erosion, to fill the small holes in the region, and then performing an opening operation, that is, erosion followed by dilation, to smooth the boundaries and separate the fine adhesions, thereby obtaining a defect morphology description with clearer boundaries and more complete regions, thus improving the accuracy of defect geometric parameter measurement.

[0078] Step S404: Integrate the optimized judgment results to generate a first defect detection result containing defect type, location, size, and severity.

[0079] The system collects and integrates all optimized and valid defect assessment results to generate a structured first defect detection result report. This report not only lists all defects but also details each defect instance, including its type, location in the image coordinate system and / or world coordinate system, physical dimensions calculated based on the optimized mask, and the severity determined by the model. This result represents the most comprehensive automated assessment of the surface condition of materials under visible area conditions.

[0080] In some embodiments, the step of extracting local geometric features from the visible area and combining these local geometric features with pre-stored three-dimensional geometric priors of the target hardware material's category to reconstruct a complete material outline model of the target hardware material specifically includes steps S511-S514. This step utilizes deep learning and generative models to infer a complete geometric outline that conforms to physical laws and category commonalities from occluded local information.

[0081] Step S511: Process the visible region using a deep convolutional neural network to extract deep feature maps, and encode the deep feature maps into a set of feature vectors representing local geometric structures and spatial context relationships.

[0082] The system employs a pre-trained deep convolutional neural network, such as ResNet or VGG, as its backbone. This network has been pre-trained on a large number of natural and industrial images and possesses powerful feature extraction capabilities. The obtained visible region image is used as input and passed through the network's forward propagation. Deeper layers, such as the last convolutional layer or a feature pyramid, are used to obtain deep feature maps. These feature maps retain low-level geometric information such as edges and corners of the visible region and encode high-level semantic information such as component combinations and structural relationships. Subsequently, an encoder module further compresses and encodes this feature map into a fixed-dimensional feature vector. This feature vector serves as a conditional seed for subsequent reconstruction processes, representing the geometric structure of the currently visible portion and its possible spatial context within the whole.

[0083] Step S512: Based on the category information of the target hardware material, retrieve the corresponding basic 3D model from the pre-stored 3D geometric prior library, match the feature vector with the projection features of the basic 3D model under different viewpoints, and estimate the corresponding part and pose of the visible area in the complete 3D model.

[0084] The system maintains a 3D geometric prior library, which stores one or more standard base 3D models for each type of hardware material, such as parametric CAD models or typical point clouds. First, based on the known target material category or obtained through image recognition, the corresponding base 3D model is retrieved from the library. Then, the system pre-calculates or generates online 2D projections of this base 3D model from a series of discrete viewpoints, extracts similar depth features from each projection, and compares the similarity of the feature vector of the currently visible area with these projection features (cosine similarity can be used). By finding the best-matching projection viewpoint, the system can estimate the approximate location of the currently visible part corresponding to the complete 3D model and the coarse spatial pose of the target material relative to the camera.

[0085] Step S513: Using the feature vector and the matched part and pose information as conditions, drive a pre-trained generative model to generate an implicit representation of the complete two-dimensional contour of the target hardware material in pose.

[0086] A conditional generative model, such as a Conditional Variational Autoencoder (CVAE), a Conditional Generative Adversarial Network (cGAN), or a Conditional Diffusion Model, is employed as the core. This generative model has been pre-trained on a large amount of local-to-complete contour data, learning the mapping relationship between local information and the generation of complete contours. In this step, the feature vector extracted in step S511 is concatenated or fused with the estimated part and pose information, serving as the conditional input to the generative model. Additionally, the mask of the occluded region can also be included as a conditional input, explicitly informing the model of the areas to be filled. Driven by these conditions, the generative model operates in the latent space, outputting an implicit representation of a complete two-dimensional contour. This representation is a high-dimensional feature tensor or latent code that contains all the information required for a continuous contour, but has not yet been parsed into a definitive geometry.

[0087] Step S514: Decode the implicit representation of the complete two-dimensional contour into an explicit set of contour points or polygons, and perform geometric normalization on the decoded contour to ensure that it conforms to the overall topological constraints of the category's three-dimensional geometric prior, and output the complete material contour model.

[0088] This step transforms the implicit representation into explicit geometry that can be directly used for comparison. The system uses a decoder paired with the generative model to convert the previously output implicit representation into specific image space coordinates, typically represented as a binary contour mask or a sequence of contour points. However, the initially decoded contour may not be smooth enough in some details or contain minor topological errors, such as small self-intersections or discontinuities. Therefore, geometric normalization is required. This normalization is based on the overall topological constraints implied by the 3D geometric priors of this type of material. For example, insulator strings should be linearly arranged, skirts should be approximately equidistant, and the connection points of hardware should maintain specific symmetries. Normalization operations include applying moving averages or spline curves to smooth the contour; checking and correcting segments that violate basic physical connections; and ensuring that the generated contour maintains overall topological consistency with the projection of the base 3D model in the estimated pose. After normalization, a geometrically reasonable, smooth-boundary, and topologically correct complete material contour model is output, which can be a two-dimensional polygon. The model not only fills in the obscured parts, but its shape also connects naturally with the currently visible parts and conforms to the general geometric laws of this type of material, thus becoming a virtual observation benchmark for subsequent overall deformation comparison.

[0089] In some embodiments, comparing the complete material outline model with a standard outline to obtain outline deformation defects, and fusing the outline deformation defects with visible surface defects to obtain the overall defect detection result, specifically includes steps S521-S524. This step intelligently fuses and comprehensively evaluates directly observed surface information and indirectly inferred overall deformation information to generate a structured final detection report that includes risk inference.

[0090] Step S521: Spatially align the complete material outline model with the standard outline, calculate the geometric difference field between the two, and identify and quantify the location, deformation direction, and deformation amount of the outline deformation area based on the geometric difference field.

[0091] The generated complete material outline model is then finely spatially aligned with a standard outline already aligned to the current viewpoint. This alignment can be achieved using feature-point-based registration or iterative nearest-point algorithms, ensuring optimal alignment at the pixel level. After alignment, the system calculates a dense geometric difference field, which can be a distance transformation map. Each pixel value represents the shortest distance from that point to another outline, with a sign (positive for outward expansion, negative for inward collapse). By analyzing this difference field, the system can accurately identify localized deformation regions, determine the deformation direction of each region, and quantify the deformation amount, such as the maximum deformation depth or average offset expressed in pixels or actual physical units. This field-based analysis method can capture subtle local deformations on the outline.

[0092] Step S522: Within the visible area, perform anomaly detection of surface texture and structure to identify and locate surface defects.

[0093] Within a defined visible area, the system runs a highly sensitive defect detection algorithm. This can be a deep learning-based classification / segmentation model specifically designed to identify microscopic or macroscopic surface anomalies such as cracks, corrosion, peeling, and electrolytic corrosion. Alternatively, it can be combined with traditional image processing techniques, such as texture analysis and spot detection, to discover imperfections like color unevenness and stains. The detection output is a collection of all identified surface defects within the visible area, with each defect including its category's pixel-level location and a severity estimate.

[0094] Step S523: Establish the spatial relationship between the contour deformation region and the surface defect. Based on the predefined fusion rules, perform information fusion on the contour deformation defect and the surface defect to obtain defect information. The fusion rules are used to handle the defect priority of the spatially overlapping region and the merging logic of different types of defects.

[0095] The identified contour deformation regions and located surface defects are overlaid and analyzed in the image coordinate system to determine whether they overlap, are adjacent, or contain each other in space. For example, a piece of rust might be located inside a local depression. Then, the system calls a predefined fusion rule base to process these relationships.

[0096] For example, the rules could stipulate that when both macroscopic contour deformation and microscopic surface defects exist at the same location, macroscopic deformation should be reported as the primary defect, while surface defects should be labeled as an accompanying phenomenon or possible cause of the deformation, thereby avoiding duplicate alarms and improving the logic of the reports.

[0097] For example, a rule could stipulate that if multiple short cracks are distributed along a profile deformation band and the spacing is less than a threshold, they should be merged and reported as a single continuous crack with deformation.

[0098] Step S524: Based on the fused defect information, generate a structured second defect detection result, which includes at least an overall contour deformation description, a list of local surface defects, and inferred defect hints for occluded areas.

[0099] The system will integrate the defect information and organize it into a structured report of the second defect detection results that is both machine-readable and easy for humans to interpret. This report may include an overall contour deformation description, a list of local surface defects, and inferred defect hints for occluded areas.

[0100] In some embodiments, based on predefined fusion rules, information fusion is performed on contour deformation defects and surface defects to obtain defect information. Then, through an interpretable rule system, defect evidence from different sources and of different types is rationally integrated and its credibility is assessed, thereby generating high-quality comprehensive defect information.

[0101] First, based on the category of the target hardware material and the type of the contour deformation defect, the corresponding rule group is retrieved from a predefined defect fusion rule base. This rule group at least includes defect conflict resolution rules and defect merging conditions. The system maintains a structured defect fusion rule base. The rules in this base are indexed and organized according to material categories (e.g., insulators, fittings, conductors) and contour deformation types (e.g., overall bending, local depressions, asymmetric expansion). When the system needs to process defect fusion for the current target, it first precisely retrieves and retrieves the matching rule group from the rule base based on the target's known category and main contour deformation type. This rule group is tailored for this type of material and deformation scenario, and necessarily includes defect conflict resolution rules for handling information contradictions and defect merging conditions for integrating related evidence.

[0102] Then, for areas where the contour deformation region and surface defect spatially overlap or are adjacent, analysis is performed according to defect conflict resolution rules. These rules include prioritizing the contour deformation defect when both macroscopic contour deformation and microscopic surface defects exist simultaneously at the same location, and labeling the surface defect as an accompanying or derivative phenomenon of the contour deformation. The system uses spatial overlay analysis to locate areas where the contour deformation region and surface defect region spatially overlap or are closely adjacent. For these areas, the invoked defect conflict resolution rules are applied for analysis. When both macroscopic contour deformation and microscopic surface defects exist simultaneously at the same spatial location, the system prioritizes recording and reporting the macroscopic contour deformation as the primary defect in that region. Simultaneously, the semantic labels of the microscopic surface defects are updated or appended as "accompanying phenomena" or "possible derivative results" of the macroscopic deformation.

[0103] For non-conflicting defects, information aggregation is performed based on defect merging conditions. These conditions include a weighted calculation based on deformation and surface defect severity levels to determine whether multiple discrete defects are caused by the same mechanical reason and whether to merge them into a single comprehensive defect description. Specifically, for spatially separated defects or remaining defects that have been processed by conflict resolution rules, the system determines whether logical merging is necessary based on the defect merging conditions. For example, if multiple short cracks are distributed along a straight line, and the direction of this line coincides with the principal axis of a significant bending deformation, the system calculates the overall severity and bending deformation of these cracks, performs a weighted comprehensive evaluation, and if the weighted comprehensive index exceeds a threshold, it is inferred that these discrete defects are likely caused by the same mechanical reason. Therefore, they are merged with the profile deformation and described as a single comprehensive defect, such as bending deformation on the northeast side causing local microcracks.

[0104] Finally, a confidence level is assigned to each defect information generated after fusion. The confidence level is determined based on the confidence level of the original evidence in the visible area on which the defect is based, as well as the integrity of the logical links it undergoes during the reconstruction and fusion process.

[0105] This step assigns a confidence level to each defect information in the final output. It is not a single value, but rather a comprehensive assessment of the confidence level of the original evidence and the integrity of the logical link.

[0106] For surface defects originating directly from the visible area, their confidence level primarily inherits the confidence score output by the previous defect detection model. For deformation defects inferred entirely from the reconstructed contour, their confidence level is strongly correlated with the confidence level of the generated complete contour model.

[0107] Then, the reliability of the reasoning chain that led to the conclusion about the defect is assessed. For example, a surface defect labeled as an accompanying phenomenon may have its credibility appropriately lowered due to its long reasoning chain; while a directly observed surface defect that does not conflict with any deformation has a short and direct logical chain and is therefore more reliable.

[0108] The system combines the two dimensions mentioned above into a final credibility level based on a preset mapping table or calculation formula, and marks it next to the corresponding defect information.

[0109] The present invention provides a hardware material deformation defect detection system for partially occluded environments, which is used to implement the above-mentioned method and includes an image acquisition and parsing module, a visibility assessment module, a strategy decision module, a direct detection module, and a reconstruction-assisted detection module.

[0110] The image acquisition and parsing module is used to acquire and parse the image to be detected in order to obtain the visible area of ​​the target hardware material.

[0111] The visibility assessment module is used to calculate the contribution of the visible area to the complete material outline based on the visible area and the pre-stored standard outline.

[0112] The strategy decision module is used to compare the contribution with a preset selection threshold and select either a direct detection strategy or a reconstruction-assisted detection strategy accordingly.

[0113] The direct detection module is used to perform defect analysis based on the visible area and output a first defect detection result when the direct detection strategy is selected.

[0114] The reconstruction-assisted detection module is used to perform detection when the reconstruction-assisted detection strategy is selected, and it includes a contour reconstruction unit and a fusion detection unit.

[0115] The contour reconstruction unit is used to reconstruct a complete material contour model of the target hardware material based on the local geometric features of the visible area and the pre-stored category 3D geometric priors.

[0116] The fusion detection unit is used to compare the complete material contour model with the standard contour to obtain contour deformation defects, and fuse them with the surface defects of the visible area to output a second defect detection result.

[0117] The electronic device of the present invention includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the above-described method.

[0118] The present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0119] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0120] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0122] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0124] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting a deformation defect of a hardware material in a local occlusion environment, characterized in that, Includes the following steps: Acquire an image containing the target hardware material to be inspected, and parse the image to be inspected to obtain the visible area of ​​the target hardware material; Based on the visible area and the standard outline of the pre-stored target hardware material, calculate the contribution of the visible area to the complete material outline. A preset selection threshold is set, the contribution is compared with the selection threshold, and one of the direct detection strategy and reconstruction-assisted detection strategy is selected. If the contribution is higher than the selection threshold, a direct detection strategy is adopted, which performs defect analysis based on the visible area and outputs the first defect detection result. If the contribution is not higher than the selection threshold, a reconstruction-assisted detection strategy is adopted, and a second defect detection result is output. The reconstruction-assisted detection strategy includes: Local geometric features are extracted from the visible area and combined with the pre-stored three-dimensional geometric priors of the target hardware material category to reconstruct a complete material outline model of the target hardware material. The complete material outline model is compared with the standard outline to obtain the outline deformation defect. The outline deformation defect is then fused with the visible surface defect to obtain the second defect detection result.

2. The method for hard goods material deformation defect detection under local occlusion environment of claim 1, wherein, The process of acquiring an image to be detected containing the target hardware material and parsing the image to obtain the visible area of ​​the target hardware material specifically includes: Semantic segmentation is performed on the image to be detected to segment the foreground region of the material and the occlusion region, and the depth information of the image to be detected is obtained. By integrating the results of semantic segmentation with depth information, the part of the unobstructed area in the foreground region of the material is determined to be occluded in the depth direction and is taken as the initially visible area. The boundary of the initially visible area is optimized to output the visible area of ​​the target hardware material.

3. The method for hard goods material deformation defect detection under local occlusion environment of claim 1, wherein, The calculation of the contribution of the visible area to the complete material outline based on the visible area and the pre-stored standard contour of the target hardware material specifically includes: The standard outline of the pre-stored target hardware material is projected in two dimensions according to the posture of the target hardware material in the image to be detected, generating a reference standard outline aligned with the image space. Extract the visible outline of the target hardware material from the visible area; Calculate the geometric overlap measure between the visible portion of the contour and the reference standard contour, wherein the geometric overlap measure includes at least one of the contour intersection union ratio and the region area coverage ratio. Identify multiple predefined key geometric features in the reference standard contour and count the proportion of key geometric features covered by the visible part of the contour; Based on the geometric overlap metric and the proportion of key geometric features covered by the visible outline, a weighted calculation is performed to obtain the contribution of the visible area to the complete material outline.

4. The method for hardware material deformation defect detection under local shielding environment of claim 1, wherein, The method employs a direct detection strategy, performs defect analysis based on the visible area, and outputs a first defect detection result, specifically including: Image analysis is performed on the visible area to extract multi-dimensional visual features that characterize the surface state of the material. The multi-dimensional visual features include at least two of the following: surface texture statistical features, edge gradient distribution features, and color space anomaly features. The extracted multi-dimensional visual features are input into a pre-trained defect recognition model, which outputs a judgment result on the category, location, and severity level of defects existing in the visible area. The judgment result is optimized, which includes filtering the judgment result based on the confidence level output by the defect recognition model and performing morphological operations on the detected defect area to optimize its boundary. The integrated and optimized judgment results generate a first defect detection result that includes the defect type, location, size, and severity.

5. The method for hardware material deformation defect detection under local shielding environment of claim 1, wherein, The process of extracting local geometric features from the visible area and combining these features with pre-stored 3D geometric priors regarding the target hardware material's category to reconstruct a complete material outline model of the target hardware material specifically includes: The visible region is processed by a deep convolutional neural network to extract deep feature maps, and the deep feature maps are encoded into a set of feature vectors that represent local geometric structure and spatial context relationships. Based on the category information of the target hardware material, the corresponding basic 3D model is retrieved from the pre-stored 3D geometric prior library. The feature vector is matched with the projection features of the basic 3D model under different viewpoints to estimate the corresponding part and pose of the visible area in the complete 3D model. Using feature vectors and matched part and pose information as conditions, a pre-trained generative model is driven to generate an implicit representation of the complete two-dimensional contour of the target hardware material in pose. The implicit representation of the complete two-dimensional contour is decoded into an explicit set of contour points or polygons, and the decoded contour is geometrically normalized to ensure that it conforms to the overall topological constraints of the category's three-dimensional geometric prior, thus outputting a complete material contour model.

6. The method for detecting deformation defects in hardware materials under partially obscured environments as described in claim 1, characterized in that, The process of comparing the complete material outline model with a standard outline to obtain outline deformation defects, and fusing the outline deformation defects with visible surface defects to obtain the second defect detection result, specifically includes: The complete material outline model is spatially aligned with the standard outline, and the geometric difference field between the two is calculated. Based on the geometric difference field, the location, deformation direction and deformation amount of the outline deformation area are identified and quantified. Within the visible area, perform anomaly detection of surface texture and structure to identify and locate surface defects; Establish the spatial relationship between contour deformation regions and surface defects. Based on predefined fusion rules, perform information fusion on contour deformation defects and surface defects to obtain defect information. The fusion rules are used to handle the defect priority of spatially overlapping regions and the merging logic of different types of defects. Based on the fused defect information, a structured second defect detection result is generated, which includes at least a description of the overall contour deformation, a list of local surface defects, and inferred defect hints for occluded areas.

7. The method for detecting deformation defects in hardware materials under partially obscured environments as described in claim 6, characterized in that, The method, based on predefined fusion rules, fuses information from contour deformation defects and surface defects to obtain defect information, specifically including: Based on the category of the target hardware material and the type of the contour deformation defect, the corresponding rule group is called from the predefined defect fusion rule library. The rule group contains at least defect conflict resolution rules and defect merging conditions. For regions where contour deformation areas and surface defects overlap or are adjacent in space, analysis is performed based on defect conflict resolution rules. These rules include prioritizing the use of contour deformation defects when macroscopic contour deformation and microscopic surface defects coexist at the same location, and labeling surface defects as accompanying or derived phenomena of contour deformation. For non-conflicting defects, information is aggregated according to defect merging conditions. These conditions include a weighted calculation based on deformation and the severity level of surface defects to determine whether multiple discrete defects are caused by the same mechanical cause and to decide whether to merge them into a comprehensive defect description. Each defect information generated after fusion is labeled with a confidence level. The confidence level is determined based on the confidence level of the original evidence in the visible area on which the defect is based, as well as the integrity of the logical links it undergoes during the reconstruction and fusion process.

8. A system for detecting deformation defects in hardware materials under partially obscured environments, characterized in that, The method for detecting deformation defects of hardware materials under partially obscured environments according to any one of claims 1-7 includes: The image acquisition and parsing module is used to acquire and parse the image to be detected in order to obtain the visible area of ​​the target hardware material; The visibility assessment module is used to calculate the contribution of the visible area to the complete material outline based on the visible area and the pre-stored standard outline. The strategy decision module is used to compare the contribution with a preset selection threshold and select a direct detection strategy or a reconstruction-assisted detection strategy accordingly. The direct detection module is used to perform defect analysis based on the visible area and output a first defect detection result when the direct detection strategy is selected; A reconstruction-assisted detection module, used to perform detection when the reconstruction-assisted detection strategy is selected, includes: The contour reconstruction unit is used to reconstruct and generate a complete material contour model of the target hardware material based on the local geometric features of the visible area and the pre-stored category three-dimensional geometric priors. The fusion detection unit is used to compare the complete material outline model with the standard outline to obtain the outline deformation defect, and fuse it with the surface defect of the visible area to output a second defect detection result.

9. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1-7.

Images