Injury-causing tool inference method and system based on point cloud and siamese network

Abstract

The application provides a wounding tool inference method and system based on a point cloud and a Siamese network, and relates to the technical field of computer vision. The wounding tool inference method and system based on the point cloud and the Siamese network comprises the following steps: S1, reconstructing skull three-dimensional data based on a forensic CT scan; S2, based on a PointNet++ multi-scale feature extraction network; S3, collecting tool models of typical blunt objects and sharp objects from a three-dimensional model database such as GrabCAD; S4, based on the TGPM of S3; and S5, jointly inputting damage geometric feature codes of S2 and cross-domain feature mapping models of S4 into a double-channel Siamese matching network. Noise suppression and spatial consistency guarantee of original three-dimensional damage data are realized through point cloud preprocessing and geometric normalization strategies, so that the comparability and reproducibility of geometric description are maintained under the difference of heterogeneous scanning protocols and sampling densities.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method and system for inferring injury-causing tools based on point clouds and Siamese networks. Background Technology

[0002] Computer vision is a crucial branch of artificial intelligence, primarily researching how to enable computers to understand, analyze, and interpret information in images or videos like humans. Its core objective is to automatically extract meaningful features and semantic information from visual data through algorithms and models. Key technologies encompassed in computer vision include image processing, 3D reconstruction, object detection, scene recognition, pose estimation, and point cloud analysis. Among these, the damage-causing tool inference method and system based on point clouds and Siamese networks refers to using 3D point cloud data and a dual-channel contrastive network structure in deep learning to analyze and compare the geometric features of the damaged surface, thereby inferring the type of tool that may have caused the damage.

[0003] Existing technologies employing single-scale geometric descriptions exhibit insufficient stability when faced with noisy clinical scans. This is because they fail to simultaneously preserve local details and global morphology, leading to confusion in the identification of similar tools. For example, a circular blunt object may be misidentified as a small-diameter blunt-ended tool based on edge curvature, thus affecting the qualitative judgment of a case. Comparisons based on synthetic tool models often show distributional mismatches in real-world injury scenarios, resulting in high false positives or low recall in actual operation. This is because synthetic samples and real injuries have systematic differences in sampling noise and material interaction characteristics, leading to decreased retrieval reliability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for inferring injury-causing tools based on point clouds and Siamese networks. This solves the problem that the single-scale geometric description in existing technologies lacks stability when facing noisy clinical scans. This is because the local details and global morphology are not preserved simultaneously, leading to confusion in the identification of similar tools. For example, a circular blunt object and a small-diameter blunt-ended tool may be misjudged based on their edge curvature, which in turn affects the qualitative judgment of the case.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for inferring damage-causing tools based on point clouds and Siamese networks, characterized by comprising the following steps: S1: Based on the three-dimensional skull data reconstructed from forensic CT scans, point cloud resampling and denoising algorithms are used to extract and purify the damaged area; then, the suspected crack and depression areas are separated by curvature analysis and region growing segmentation methods. In order to maintain the consistency of geometric structure, the point cloud is subjected to normal vector estimation and scale normalization to ensure the stability and comparability of subsequent feature extraction and generate a standardized damage point cloud model. S2: Based on the PointNet++ multi-scale feature extraction network, local geometric attributes such as normal vector change rate, curvature gradient, and indentation depth distribution are captured at different spatial scales. The graph convolutional network is further used to construct the point cloud neighborhood topology to realize the continuous encoding of crack morphology. Finally, the local and global information are fused to form a unified morphological feature representation and generate damage geometric feature encoding. S3: Collect tool models of typical blunt and sharp objects from the GrabCAD 3D model database, and obtain standardized input using the same point cloud preprocessing process as S1; based on the self-supervised learning pre-training mechanism, train the geometric structure invariant encoding network, learn the morphological prior knowledge representation of the tool, and generate the tool geometric prior feature model. S4: Based on the tool geometric prior feature model of S3, the tool features are transferred to the damage domain through the domain adaptation layer to reduce the distribution difference between virtual tools and real damage. An adversarial feature alignment training strategy is adopted to enable the network to learn a cross-domain consistent geometric representation and generate a cross-domain feature mapping model. S5: The damage geometric feature encoding of S2 and the cross-domain feature mapping model of S4 are jointly input into the dual-channel Siamese matching network. The two networks share the weight structure. The key morphological regions are focused through the cross-domain feature attention mechanism. The triple loss and contrast loss are jointly optimized to make similar tools and corresponding damage features closely aggregated in the embedding space, generating a damage-tool joint embedding feature space. S6: Based on the damage-tool joint embedding feature space generated by S5, a high-resolution embedding feature index library is constructed through spherical metric learning. Hierarchical clustering is further used to separate different tool types between classes, and t-SNE is used to visualize and verify the interpretability of the embedding space structure. This enables Top-K similar tool retrieval for unknown samples and generates a tool similarity index library. S7: Based on the tool similarity index library of S6, it integrates point cloud geometric features with CT image texture features and finite element simulation data, and uses a multimodal Transformer coding network to achieve cross-modal feature alignment. It comprehensively outputs the prediction results of potential injury-causing tool category, direction of action and force depth. This process integrates visual geometric and mechanical features, enhances the forensic interpretability of the model, and generates a multimodal injury inference model. S8: Based on the multimodal injury inference model output of S7, the three-dimensional visualization rendering and heat map generation algorithm are used to explicitly display the correspondence between injury and prediction tool, and generate a forensic visualization report that includes tool type, injury angle, force direction and injury matching probability, for judicial identification evidence presentation, and generate a forensic visualization injury report.

[0006] Preferably, the generation of a normalized damage point cloud model based on S1 includes the following steps: S101: The generation of a normalized damage point cloud model based on S1 includes the following steps: S101: Based on the reconstruction algorithm of forensic CT scan data, the damaged skull CT slices are converted into a surface mesh model through three-dimensional reconstruction technology; then, a point cloud sampling algorithm is used to generate a dense point cloud representation, generating the original damage point cloud data. S102: Based on the statistical outlier removal algorithm, noise is removed from RIPC to eliminate scan pseudo-points and external interference points; then, the point cloud density is balanced through the uniform resampling algorithm to ensure the consistency of point distribution and generate denoised balanced point cloud data. S103: Based on the curvature estimation algorithm and the region growth segmentation method, crack region separation and depression detection are performed on DBPC to extract the main damage region and generate a point cloud of the damage region; S104: Based on the normal vector estimation algorithm and scale normalization strategy, the IRPC is oriented and geometrically scaled to establish a consistent geometric basis for subsequent feature extraction and generate a standardized damage point cloud model.

[0007] The preferred S2-based generative damage geometry feature encoding includes the following steps: S201: Based on PointNet++ hierarchical sampling and grouping strategy, multi-scale feature extraction is performed on NIPC to capture local geometric information, normal vector change rate and curvature distribution, and generate multi-scale local feature set; S202: Based on graph convolutional networks, topological connections are established in the point cloud neighborhood of MSLFS to explicitly model crack morphology and edge continuity and generate crack topological structure encoding. S203: Based on a global feature aggregation network, CTE and MSLFS are fused to construct a unified descriptor that integrates local details and overall morphology, generating damage geometric feature encoding.

[0008] Preferably, the S3-based generative tool geometric prior feature model includes the following steps: S301: Based on sampling from a 3D model library, blunt and sharp tool models are extracted from databases such as GrabCAD, and data is standardized by surface resampling and denoising algorithms to generate standardized tool point cloud data. S302: Based on data augmentation algorithms, STPC is subjected to occlusion, rotation, erosion, and perturbation operations to form diverse geometric input samples and generate an augmentation tool point cloud; S303: Based on a self-supervised learning mechanism, the geometric encoding network is trained using ATCS to learn the invariant representation of tool morphology and generate a tool geometry pre-trained model.

[0009] Preferably, the S4-based model for generating cross-domain feature maps includes the following steps: S401: Based on the transfer learning mechanism, TGPM is transferred to the damage feature extraction task. The feature space of the input damage data is adaptively adjusted to generate a primary transfer feature model. S402: Based on the domain adaptation layer, PTFM is used to minimize the difference between the distribution of virtual tool features and real damage features to generate a domain-aligned feature model; S403: Based on the adversarial feature alignment algorithm, a discriminator is introduced on the basis of DAFM to optimize cross-domain feature consistency and generate a cross-domain feature mapping model.

[0010] Preferably, the S5-based generative damage-tool joint embedding feature space includes the following steps: S501: Based on the dual-channel shared weight network structure, IGFE and CDFMM are input into two encoders respectively to maintain the consistency of feature learning and generate the initial dual-channel feature embedding. S502: Based on the cross-domain feature attention mechanism, the geometric difference region in ISE is weighted and enhanced to highlight the key crack and tool edge morphology, and generate weighted attention embedding features; S503: Based on joint loss function optimization, embedding space clustering training is performed on AWEF to aggregate homologous tools and damage features with high similarity and separate heterologous features, generating a damage-tool joint embedding feature space.

[0011] Preferably, the S6-based similarity index library for generation tools includes the following steps: S601: Based on the spherical metric learning algorithm, a similarity metric function is constructed in JMES to quantify the morphological distance between tools and generate an embedding space similarity matrix; S602: Based on the hierarchical clustering algorithm, ESM is clustered and divided to form feature clusters of different tool categories, generating tool feature clustering structures; S603: Based on the t-SNE visualization dimensionality reduction algorithm, TFCS is mapped in two dimensions and displayed in an interpretable manner; then, the Top-K retrieval strategy is used to output the tool type most similar to the unknown damage, generating a tool similarity index library.

[0012] Preferably, the S7-based generative multimodal injury inference model includes the following steps: S701: Based on a multi-source data fusion framework, the geometric features of the tool point cloud corresponding to TSID are fused with the CT images of the damaged area and the stress data of finite element simulation at the feature level to generate a multimodal input dataset. S702: Based on the multimodal Transformer network, cross-modal feature alignment and deep interaction modeling are performed on MID to generate a fused feature representation model; S703: Based on a multi-task prediction mechanism, a multimodal injury inference model is generated by utilizing key indicators such as FFRM prediction tool category, injury angle, and depth of action.

[0013] Preferably, generating a forensic visualization injury report based on S8 includes the following steps: S801: Based on a 3D rendering algorithm, it generates a 3D damage-to-visualization model by utilizing the geometric and mechanical results output by MIIM and matching the scene with the tool. S802: Based on the heat map generation algorithm, the matching probability of stress intensity and crack corresponding regions is mapped on the 3DVMM surface to generate a damage matching heat map. S803: Based on the automatic forensic report generation module, it integrates IMH, tool prediction results and mechanical parameters to form an interpretable forensic conclusion document and generate a visual forensic injury report.

[0014] A damage-causing tool inference system based on point clouds and Siamese networks, the system comprising: The point cloud preprocessing module, based on 3D reconstruction and point cloud processing technology, uses CT slice 3D reconstruction algorithm and point cloud sampling method to obtain initial point cloud data of skull injury; then, noise removal and point density balancing are completed by statistical outlier removal algorithm and uniform resampling method; then, the core injury region is extracted by curvature estimation and region segmentation algorithm, and geometric standardization is achieved by scale normalization and normal correction strategy, so as to provide a consistent spatial expression for subsequent feature extraction and generate a standardized injury point cloud model. The feature encoding module, based on a point cloud deep learning feature extraction framework, performs multi-scale geometric encoding on the normalized damage point cloud model output by the previous module. It uses hierarchical sampling and local feature aggregation algorithms to extract fine-grained geometric information; it uses graph convolution feature construction methods to establish topological connections between points and capture the continuity of crack edges; and it combines a global feature fusion algorithm to achieve feature integration of local and global morphology. At the same time, the system loads the trained tool geometric encoding model based on self-supervised geometric pre-training and transfer learning strategies to ensure the consistency between the damage and tool feature spaces and generate damage geometric feature encoding. The matching inference module, based on a dual-channel Siamese network structure, inputs the damage geometric feature encoding output from the feature encoding module and the pre-trained tool geometric features into two shared weight encoders to construct a joint embedding space. A cross-domain attention mechanism is used to strengthen the correspondence between key cracks and edges, and the matching discrimination ability is improved by a joint optimization strategy of contrastive loss and triplet loss. Subsequently, the feature similarity matrix is ​​calculated using the spherical metric learning algorithm, and the similarity space is clustered and visualized using hierarchical clustering and t-SNE dimensionality reduction analysis methods to output the tool similarity retrieval results and generate a tool similarity index library. The visualization report module, based on 3D rendering and multimodal fusion technology, calls the tool similarity index library of the matching inference module and integrates the geometric features of the damage point cloud, the structure of the tool point cloud, and the stress data from finite element simulation. It uses a multimodal Transformer interactive algorithm to achieve feature-level fusion and injury mechanism reconstruction. Then, it uses a 3D visualization rendering algorithm to generate a damage-tool matching scene and presents the stress intensity and matching probability distribution through a heat map mapping method. Finally, it automatically outputs an interpretable inference report with the help of a forensic report generation algorithm, generating a forensic visualization injury report.

[0015] This invention provides a method and system for inferring damage-causing tools based on point clouds and Siamese networks. It has the following beneficial effects: This invention achieves noise suppression and spatial consistency assurance of original 3D damage data through point cloud preprocessing and geometric normalization strategies. Therefore, it maintains the comparability and reproducibility of geometric descriptions under heterogeneous scanning protocols and sampling density differences. Multi-scale local geometric encoding and neighborhood topology modeling jointly output richer semantic information about cracks and indentations, allowing subtle edge features and overall morphological information to be expressed simultaneously. This improves similarity discrimination ability and reduces the confusion rate between tools with similar morphologies. Self-supervised pre-training forms robust geometric priors on tool morphological data, maintaining feature representation quality even with few or missing samples, thereby improving the reliability of matching unknown samples. Domain adaptation and adversarial feature alignment shorten the distribution gap between virtual models and real damage, thus achieving higher retrieval accuracy and lower false positive rates in real forensic scenarios. Dual-channel contrastive learning, combined with cross-domain attention focusing mechanism and joint optimization of triples and contrastive loss, jointly promotes the tight aggregation of similar samples and the clear separation of dissimilar samples within the embedding space, significantly improving the usability and ranking stability of Top-K retrieval. Spherical metric learning and hierarchical clustering provide an efficient vector retrieval framework for large-scale indexing and endow the embedding space with a certain interpretable structure, facilitating subsequent visualization and manual review. Feature-level fusion of point cloud morphological information, image texture, and mechanical simulation data can link geometric evidence with force analysis, thereby providing physically meaningful inferences about the direction of action and the depth of force and quantifying the credibility of matching. The combination of 3D rendering and heat map visualization with structured report output enhances the intuitiveness and reviewability of evidence presentation, making the comparison results more readable and explanatory during judicial proceedings. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the main steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 Here is a detailed schematic diagram of S6 of the present invention: Figure 8 This is a detailed schematic diagram of S7 of the present invention; Figure 9 This is a detailed schematic diagram of S8 of the present invention; Figure 10 This is a system block diagram of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example: like Figures 1-10 As shown, this embodiment of the invention provides a method for inferring damage-causing tools based on point clouds and Siamese networks, including the following steps: S1: Based on the three-dimensional skull data reconstructed from forensic CT scans, point cloud resampling and denoising algorithms are used to extract and purify the damaged area; then, the suspected crack and depression areas are separated by curvature analysis and region growing segmentation methods. In order to maintain the consistency of geometric structure, the point cloud is subjected to normal vector estimation and scale normalization to ensure the stability and comparability of subsequent feature extraction and generate a standardized damage point cloud model. S2: Based on the PointNet++ multi-scale feature extraction network, local geometric properties such as normal vector change rate, curvature gradient, and indentation depth distribution are captured at different spatial scales. The graph convolutional network is further used to construct the point cloud neighborhood topology to realize the continuous encoding of crack morphology. Finally, the local and global information are fused to form a unified morphological feature representation and generate damage geometric feature encoding. S3: Collect tool models of typical blunt and sharp objects from 3D model databases such as GrabCAD, and obtain standardized input using the same point cloud preprocessing process as S1; based on a self-supervised learning pre-training mechanism, train a geometric structure invariant encoding network, learn the morphological prior knowledge representation of the tool, and generate a geometric prior feature model of the tool. S4: Based on S3, TGPM transfers tool features to the damage domain through a domain adaptation layer to reduce the distribution difference between virtual tools and real damage. It adopts an adversarial feature alignment training strategy to enable the network to learn cross-domain consistent geometric representations and generate a cross-domain feature mapping model. S5: The damage geometric feature encoding of S2 and the cross-domain feature mapping model of S4 are jointly input into the dual-channel Siamese matching network. The two networks share the weight structure. The key morphological regions are focused through the cross-domain feature attention mechanism. The triple loss and contrast loss are jointly optimized to make similar tools and corresponding damage features closely aggregated in the embedding space, generating a damage-tool joint embedding feature space. S6: Based on the JMES generated by S5, a high-resolution embedding feature index library is constructed through spherical metric learning. Hierarchical clustering is further used to separate different tool types between classes, and t-SNE is used to visualize and verify the interpretability of the embedding space structure. This enables Top-K similar tool retrieval for unknown samples and generates a tool similarity index library. S7: Based on TSID of S6, point cloud geometric features are fused with CT image texture features and finite element simulation data. A multimodal Transformer coding network is used to achieve cross-modal feature alignment and comprehensively output the prediction results of potential injury-causing tool category, direction of action and force depth. This process integrates visual geometric and mechanical features, enhances the forensic interpretability of the model, and generates a multimodal injury inference model. S8: Based on the MIIM output of S7, it uses a 3D visualization rendering and heat map generation algorithm to explicitly display the correspondence between damage and prediction tools, and generates a forensic visualization report that includes tool type, injury angle, force direction and damage matching probability, for judicial identification evidence presentation, and generates a forensic visualization injury report.

[0019] The generation of a normalized damage point cloud model based on S1 includes the following steps: S101: Based on the reconstruction algorithm of forensic CT scan data, the damaged skull CT slices are converted into a surface mesh model through three-dimensional reconstruction technology; then, a point cloud sampling algorithm is used to generate a dense point cloud representation, generating the original damage point cloud data. Based on a forensic CT scan data reconstruction algorithm, the acquired CT slices of the damaged skull are first imported into a 3D reconstruction program. Continuous 2D slices are reconstructed using voxel stacking, with a spacing of 0.5 mm between each slice and a pixel matrix size of 512×512. Threshold extraction is performed on bone tissue regions with grayscale values ​​between 200 and 1200 HU, and missing information between slices is filled in using linear interpolation. For surface extraction of the 3D volume data, the MarchingCubes algorithm is used. The vertex positions of triangular facets are determined by calculating the grayscale distribution of each cube corner, generating a skull mesh model containing approximately 500,000 triangular facets. This model is then imported into a point cloud sampling module, and a uniform sampling strategy is used to sample the mesh surface at an average point spacing of 0.3 mm, resulting in a dense point cloud of approximately 1.6 million points. Each point contains 3D coordinates (x, y, z) and local normal vector information, ultimately forming the original damage point cloud data. S102: Based on the statistical outlier removal algorithm, noise is removed from RIPC to eliminate scan pseudo-points and external interference points; then, the point cloud density is balanced through the uniform resampling algorithm to ensure the consistency of point distribution and generate denoised balanced point cloud data. Based on a statistical outlier removal algorithm, noise is removed from the point cloud data generated in the previous stage. The mean Euclidean distance between each point and its k=20 nearest neighbors is calculated as a local density index. If the mean distance of each point exceeds the global mean distance μ plus twice the standard deviation σ, then the algorithm satisfies the following condition. Points identified as outliers are removed. With μ = 0.35 mm and σ = 0.12 mm, the threshold is 0.59 mm. After removal, the number of points decreases by approximately 5%, leaving 1.51 million points. A uniform resampling process is then performed. The target point spacing is set to r = 0.4 mm. A voxel cube is constructed in space using a voxel grid filtering algorithm, with each voxel having a side length of r. When multiple points exist within the same voxel, only the center point coordinates are retained as a representative point, thus achieving point density balance. The number of points is reduced in densely populated areas, while interpolation points are added to sparse areas to maintain uniformity. The final output is denoised and balanced point cloud data.

[0020] S103: Based on the curvature estimation algorithm and the region growth segmentation method, crack region separation and depression detection are performed on DBPC to extract the main damage region and generate a point cloud of the damage region; Based on curvature estimation algorithms and region growing segmentation methods, damaged regions are extracted from equilibrium point clouds. First, a local plane is fitted using a point set within a neighborhood of radius r = 1 mm. The rate of change of the normal vector direction is calculated, and then the principal curvature is used to extract the damaged region. and The difference defines the curvature. If C > 0.15, it is marked as a high curvature point, forming a set of suspected crack boundaries. Then, seed points are selected based on the region growing method, and points with a neighborhood angle difference of less than 15° are grouped into the same region. The average depth of each region is calculated. The data were sorted, and regions with a depth greater than the global average depth plus 0.3 mm were selected as depression areas. This threshold was defined based on the statistical distribution of the average fracture depth in actual skull injury samples. The average value of the samples was 1.7 mm and the standard deviation was 0.3 mm. Therefore, 0.3 mm can be regarded as a significant depression boundary. After multiple iterations of aggregation and screening, a point cloud of the injury area was generated.

[0021] S104: Based on the normal vector estimation algorithm and scale normalization strategy, the IRPC is oriented and geometrically scaled to establish a consistent geometric basis for subsequent feature extraction and generate a standardized damage point cloud model.

[0022] Based on the normal vector estimation algorithm and scale normalization strategy, orientation correction and geometric standardization are performed on the extracted point cloud of the damaged region. First, the overall normal vector direction is calculated using the least squares plane fitting method, and then the principal axis direction vector of the point cloud is calculated through PCA principal component analysis. Align with the Z-axis in the world coordinate system and adjust the rotation matrix R to make To achieve directional unification, the scaling factor is then calculated based on the bounding box size of the point cloud. ,in Set the standard reference length to 10mm. Given the maximum side length of the current point cloud, if If s = 0.8, then scale all point coordinates according to this ratio to ensure that the geometric dimensions are within the same standard space. Finally, calculate the normal vector normalization for each point's normal vector component ( To unitize, so that Output a standardized damage point cloud model.

[0023] The S2-based generation of damage geometric feature encoding includes the following steps: S201: Based on PointNet++ hierarchical sampling and grouping strategy, multi-scale feature extraction is performed on NIPC to capture local geometric information, such as the rate of change of normal vector and curvature distribution, and generate multi-scale local feature sets. Based on PointNet++'s hierarchical sampling and grouping strategy, the input normalized damaged point cloud model NIPC is first sampled at multiple levels according to its spatial hierarchy. In the initial layer, approximately 1.5 million points are sampled proportionally. Select 75,000 center points, and establish a radius for each center point. The local neighborhood of the points is calculated, and the local center coordinates of the points within the neighborhood are calculated by distance-weighted averaging. These coordinates are used as the input for the first layer of features. Then, in the second layer, the output feature points are downsampled again at a ratio of p2=0.02, with the radius... Extended to 2.0 mm, it captures the geometry of a wider area, calculating the rate of change of the normal vector for each neighborhood. The mean value is obtained by weighting the angle differences between points in the neighborhood to obtain the local directional variation characteristics. For example, when the average value in a certain crack edge region is... When the value is 0.28, compared to the smooth region of 0.05, it is classified as a high-variance region. In curvature analysis, its variance is calculated using the set of curvature values ​​of neighboring points. As a local geometric fluctuation indicator, when Defined as a high curvature region, these local descriptions are concatenated with point coordinates to form a feature vector set. After feature extraction at each layer, local information is aggregated through max pooling to generate a multi-scale local feature set.

[0024] S202: Based on graph convolutional networks, topological connections are established in the point cloud neighborhood of MSLFS to explicitly model crack morphology and edge continuity and generate crack topological structure encoding. Based on graph convolutional networks, neighborhood topological relationships are constructed within the multi-scale local feature set generated in the previous stage. First, k=10 nearest neighbors of each point in Euclidean space are selected to form a local subgraph, and edge weights are defined. It is a weighted combination of the distance between points and the cosine of the angle between the normal vectors. When the distance between points... Angle At that time, calculate The weighting coefficients of α and β are set to 0.6 and 0.4, respectively. This setting is derived from the fitting statistics of 20 skull crack samples, making the influence of distance slightly higher than that of orientation consistency. Subsequently, the aggregated features of each node are calculated in the graph convolutional layer, and the node feature orientation is updated after weighted summation of the neighborhood features. When the average edge weight of a node is lower than the global mean minus 0.15, it is marked as a crack boundary point; if it is higher than the global mean plus 0.10, it is marked as a continuous structure point. The division interval is determined by the statistical distribution of edge weights in the sample. The global mean is about 0.72 and the standard deviation is 0.09. Therefore, the thresholds for judging cracks and continuous boundaries are 0.57 and 0.82, respectively. The crack topology structure code is generated by iterative aggregation and updating through multi-layer graph convolution.

[0025] S203: Based on a global feature aggregation network, CTE and MSLFS are fused to construct a unified descriptor that integrates local details and overall morphology, generating damage geometric feature encoding.

[0026] Based on a global feature aggregation network, the crack topology structure encoding obtained in the previous stage is fused with the multi-scale local feature set. First, each local feature vector is weighted and normalized, and the local feature weights are set. Global topology weights This ratio is derived from empirical parameter adjustments. The optimal ratio is determined by calculating the average decrease in reconstruction error across 5 groups of samples, and the fused features are calculated as follows: each It includes 12-dimensional geometric descriptions (such as curvature, orientation difference, local density, etc.). It includes 8-dimensional structural encoding (such as node aggregation degree, edge weight mean, etc.), which are fused to form a 20-dimensional unified feature description. Then, global pooling is performed on all point cloud features, and the mean and variance of each feature dimension are calculated. The global mean vector is then... Sum of variance vectors The splicing forms the final damage geometric feature encoding, in a typical sample. The value range is [-0.4, 0.7]. Between [0.05, 0.25], the complete geometric feature code is generated through this global feature aggregation process.

[0027] The S3-based generative tool geometric prior feature model includes the following steps: S301: Based on sampling from a 3D model library, blunt and sharp tool models are extracted from databases such as GrabCAD, and data is standardized by surface resampling and denoising algorithms to generate standardized tool point cloud data. Blunt and sharp weapon models were extracted from databases such as GrabCAD based on sampling from 3D model libraries. First, the databases were filtered using keywords such as "blunt weapon," "knife," and "hammer." The found 3D STEP and STL models were downloaded, and their file size, face count, and format attributes were recorded. For each model, a face count M>10 was considered optimal. 5 Considered a high-precision model, if M < 10 4 This model is then marked as low-precision and a surface subdivision operation is performed. A uniform resampling algorithm is used to reset the point spacing on the model surface to 0.5 mm. The model surface area A is then compared with the target point density. Calculate the number of sampling points For example, the surface area of ​​a knife is A = 2100 mm². 2 The number of sampling points is approximately 8400. Then, bilateral filtering is performed for noise reduction, and the rate of change of the normal vector direction at each point is calculated. Rate of change in direction of the mean in the neighborhood When comparing, Anomalies are defined as points that are outliers and their coordinates are smoothly adjusted. The denoising process is completed by iterating point by point three times. Then, the three-dimensional coordinates of all points are normalized to the range of [-1, 1] and adjusted according to their maximum side length. Compared with the reference length Calculate the scaling factor ,when When s=0.8, the coordinate data is uniformly scaled to form standardized tool point cloud data.

[0028] S302: Based on data augmentation algorithms, STPC is subjected to occlusion, rotation, erosion, and perturbation operations to form diverse geometric input samples and generate an augmentation tool point cloud; Based on data augmentation algorithms, the standardized tool point cloud generated in the previous stage is subjected to various geometric transformations. First, each point cloud sample is processed according to the occlusion ratio. 15% of the points were randomly deleted to simulate occlusion, and then the overall model was rotated. The rotation angle θ was randomly generated within the range of [-20°, 20°]. At that time, a three-axis rotation transformation was performed according to the Euler rotation sequence. A morphological erosion operation was then performed on the rotated point cloud, dividing the point cloud into cubes with a side length of 1.0 mm according to the spatial mesh. The local point density was then... Below the global average minus Density compensation is performed in real time, and complementary points are generated through neighborhood interpolation, while for Higher than add Random perturbation is performed on the region, and the coordinate offset δ is randomly assigned between [-0.2, 0.2] mm to form a set of morphological change instances. In each enhancement operation, 10 derivative samples are generated, so that each tool model corresponds to a point cloud set with a total of 10 × the number of original samples, thereby constructing the enhancement tool point cloud set.

[0029] S303: Based on a self-supervised learning mechanism, the geometric encoding network is trained using ATCS to learn the invariant representation of tool morphology and generate a tool geometry pre-trained model.

[0030] Based on a self-supervised learning mechanism, a geometric coding network is trained using the augmented point cloud dataset ATCS. ​​First, an encoder-decoder network framework is constructed. The input point cloud is randomly divided into a training set and a validation set in an 8:2 ratio, with an input data dimension of N×3, where N is the number of points and 3 represents the coordinate components. Feature embedding is performed on each sample to extract point features. And through point feature mean With variance Normalization is used to suppress scale differences, and reconstruction loss is employed. Calculate the average distance between the network output point cloud and the input point cloud. At that time, a loss threshold ε=0.4mm was set as the stopping condition, the network training iteration batch was set to 200 epochs, the learning rate η=0.001, and the batch size B=16. During training, the loss change rate of adjacent batches was compared. ,when The learning rate is then reduced to 0.5η, and the training parameters are stabilized through this layer-by-layer adjustment method, ultimately generating a tool geometry pre-trained model.

[0031] The S4-based model for generating cross-domain feature maps includes the following steps: S401: Based on the transfer learning mechanism, TGPM is transferred to the damage feature extraction task. The feature space of the input damage data is adaptively adjusted to generate a primary transfer feature model. Based on the transfer learning mechanism, TGPM is transferred to the damage feature extraction task. The feature space of the input damage data is adaptively adjusted. First, the damage data is standardized using mean and variance to ensure data transfer at the same scale. Next, key geometric features, such as surface normal vector changes and curvature distribution, are extracted from the damage point cloud data and learned through a multi-layer convolutional neural network (CNN). During the transfer process, a shared layer between the target damage features and the source task features allows the network to retain the learning capabilities of the source task in the new task. For example, if the source task is crack detection of mechanical parts and the target task is damage identification, transfer learning can effectively reduce model training time and decrease the data requirements for the new task. Finally, by minimizing the target task loss function, a primary transfer feature model is generated for subsequent damage identification tasks.

[0032] S402: Based on the domain adaptation layer, PTFM is used to minimize the difference between the distribution of virtual tool features and real damage features to generate a domain-aligned feature model; Based on a domain-adaptive layer, this study utilizes PTFM to minimize the difference between the distributions of virtual tool features and real damage features. First, features are extracted from both virtual tool and real damage data to construct their respective feature spaces. Then, a distribution difference metric, such as Maximum Mean Difference (MMD), is calculated to determine if there is a significant difference in the feature distributions. A large MMD value indicates a significant distribution difference between the virtual tool and real damage data, requiring further alignment. To minimize this distribution difference, PTFM introduces a domain-adaptive layer to adjust feature alignment during training, ensuring that the features of the virtual tool remain consistent with those of the real damage. During training, if the difference of a feature exceeds a predetermined threshold, the system dynamically adjusts its domain alignment weights based on the degree of distribution matching, making the features more similar. Through this method, the resulting domain-aligned feature model effectively integrates features from both virtual tool and damage data.

[0033] S403: Based on the adversarial feature alignment algorithm, a discriminator is introduced on the basis of DAFM to optimize cross-domain feature consistency and generate a cross-domain feature mapping model.

[0034] Based on an adversarial feature alignment algorithm, a discriminator is introduced on top of DAFM to optimize cross-domain feature consistency. First, an adversarial discriminator is added to the DAFM model to distinguish features between the source and target domains. During training, the discriminator determines whether the input features originate from the source or target domain, while the feature extractor generates features that can deceive the discriminator, making the source and target domain features indistinguishable. Through adversarial training, the game between the feature extractor and the discriminator is continuously optimized, ultimately achieving feature alignment between the source and target domains. Assume the source domain features are... The target domain features are The discriminator D classifies the input features f and ultimately optimizes the feature alignment process by minimizing the loss function. For example, if the training data contains virtual tool features and real damage features, through adversarial training, the discriminator will have difficulty distinguishing between the two, indicating that the feature distribution has become consistent. Finally, a cross-domain feature mapping model is generated, which can be used for practical damage detection tasks.

[0035] The S5-based generative damage-tool joint embedding of the feature space includes the following steps: S501: Based on the dual-channel shared weight network structure, IGFE and CDFMM are input into two encoders respectively to maintain the consistency of feature learning and generate the initial dual-channel feature embedding. Based on a dual-channel shared weight network structure, IGFE and CDFMM are input into two encoders respectively. During execution, the geometric representation of the tool point cloud in IGFE is first aligned by scaling its coordinate range to a ratio of 100 mm between the maximum edge length and the reference length. For example, if the maximum edge length of the input tool point cloud is 125 mm, it is scaled by a factor of 0.8 before being fed into the first encoder channel. Simultaneously, the crack edge morphology data extracted from the damaged surface in CDFMM is spatially sampled, and the density of nearest neighbor points in the crack region is scaled by a radius of 1 mm. The number of points within a 1-meter neighborhood is calculated. When the number of neighborhood points is greater than 20, it is marked as a high-density crack area; when it is less than 8, it is marked as a low-density area. The input data is then locally resampled according to the density level to keep the crack area representation clear before being fed into the second channel encoder. The encoder weights of the two channels are shared. Therefore, during training, the difference between the output embedding vectors of the two channels is calculated and their parameters are synchronized. When the difference is higher than the preset baseline difference value of 0.1, a weight adjustment operation is performed. If the difference is lower than 0.1, the current parameters are kept unchanged. The initial dual-channel feature embedding corresponding to the structure is obtained through multiple batches of training.

[0036] S502: Based on the cross-domain feature attention mechanism, the geometric difference region in ISE is weighted and enhanced to highlight the key crack and tool edge morphology, and generate weighted attention embedding features; Critical cracks refer to crack regions on the damaged surface (or in CT volume data) that simultaneously satisfy three types of indices: geometric abruptness, elongation correlation, and significant mechanical response. They manifest as morphological abnormalities in the geometric domain and as stress or strain concentration points in the mechanical domain, and are the main characteristic response points of tool-induced damage structures. Based on a cross-domain feature attention mechanism, geometric difference regions in ISE are weighted and enhanced. In the specific execution process, the local curvature changes of the corresponding crack surface points and tool edge points in the initial dual-channel embedding are first compared. Regions with a local curvature change greater than 0.3 are marked as geometrically significant feature regions; those between 0.1 and 0.3 are marked as moderately variable regions; and those below 0.1 are marked as smooth regions. These interval judgments allow for the localization of geometric feature difference regions. Weighting coefficients are then assigned to the located regions, with geometrically significant feature regions receiving a weighting factor. The weighting coefficient is 1.5. The medium variation region is assigned a weighting coefficient of 1, and the smooth region is assigned a weighting coefficient of 0.7. For example, when the curvature change value of the crack tip region is 0.35, the region will be multiplied by a weighting value of 1.5. During the weighting process, all coordinate and normal description vectors are uniformly weighted. At the same time, difference comparison is performed on cross-domain comparison regions. When the angle between the tool edge point and the crack region point in the vector direction is less than 20 degrees, it is determined to be a corresponding associated region. The same weighting coefficient is synchronously adjusted in the corresponding region, and finally the formation of weighted attention embedding features is completed.

[0037] S503: Based on joint loss function optimization, embedding space clustering training is performed on AWEF to aggregate homologous tools and damage features with high similarity and separate heterologous features, generating a damage-tool joint embedding feature space.

[0038] Based on joint loss function optimization, embedding space clustering training is performed on AWEF. In the specific execution process, a similarity metric between homologous tool features and corresponding damage features is first constructed, using the Euclidean distance between embedding vectors as the criterion. When the distance is less than 0.5, it is determined to be a high similarity region; when it is between 0.5 and 1.5, it is determined to be a medium similarity region; and when it is greater than 1.5, it is determined to be a low similarity region. During training, in order to cluster homologous features, a similarity aggregation weight of 1 is assigned to high similarity regions, 0.6 to medium similarity regions, and 0.2 to low similarity regions. For example, when the distance between a tool edge embedding and a damage crack embedding is 0.3, its corresponding aggregation weight will be 1. At the same time, a separation term is added between heterologous features to keep the embedding vectors with a distance greater than 1.5 at a larger interval during training. By continuously performing clustering iterations, training stops when the feature clustering radius changes to less than 0.05, thereby generating a joint embedding feature space of damage and tools.

[0039] The generation tool similarity index library based on S6 includes the following steps: S601: Based on the spherical metric learning algorithm, a similarity metric function is constructed in JMES to quantify the morphological distance between tools and generate an embedding space similarity matrix; Based on the spherical metric learning algorithm, a similarity metric function is constructed in JMES to quantify the morphological distance between tools. During execution, the input 3D point cloud data of the tools is first converted to spherical coordinates, transforming the spatial position of each point into a triplet of radius r, azimuth angle θ, and polar angle φ. Normalization is then applied to limit the value of r to between zero and one. When r exceeds one, it is scaled by a coefficient of 0.98. Subsequently, the morphological distance between any two tools is calculated using a weighted average of spherical feature differences. The distance weight for each point is determined based on the local curvature value: a weight of 1 is taken when the curvature value is greater than 0.3, 0.8 is taken when the curvature value is between 0.1 and 0.3, and 0.98 is taken when the curvature value is less than zero. The initial weight is set to 0.5. For example, when the mean radius of the spherical point distribution of tool A is 0.75 and that of tool B is 0.82, the average radius difference between the two is 0.07. Combining the average angle difference of 20 degrees and the weighted weight of 0.8, the calculated morphological distance is 0.56. To determine the similarity threshold, the average distance of all tool sample pairs is taken as the baseline value D_base. When D_base is 0.6, those less than 0.4 are defined as high similarity regions, those between 0.4 and 0.8 are defined as medium similarity regions, and those greater than 0.8 are defined as low similarity regions. In the entire calculation loop, all tool pairs are traversed sequentially and the distance results are stored to finally form the embedding space similarity matrix.

[0040] S602: Based on the hierarchical clustering algorithm, ESM is clustered and divided to form feature clusters of different tool categories, generating tool feature clustering structures; Based on hierarchical clustering, ESM is clustered to form feature clusters for different tool categories. During execution, all distance values ​​in the embedding space similarity matrix are first read and sorted in ascending order. Initial aggregation is performed on any pair of tools with high similarity. If the distance between two tools is less than 0.4, they are grouped into the same primary cluster. After aggregation, the average distance between clusters is calculated, and a new distance matrix is ​​generated. Then, the iterative merging criterion of hierarchical clustering is followed. If the distance between clusters is less than 0.5, merging continues. If the distance is greater than 0.5 but less than 0.8, the current level is retained without further merging. If the distance exceeds 0.8, the tool is removed from the current cluster. The clusters are labeled as independent clusters. In the example, if the sample set contains ten tool models, the similarity matrix shows that the distance between tool one and tool two is 0.38 and the distance between tool two and tool three is 0.45. Then, tool one, two and three are merged into the same level cluster. Then, the merging order is determined according to the average distance between this cluster and other clusters. The whole process is to form a three-layer clustering tree through layer-by-layer aggregation and layer-by-layer calculation. In the parameter settings, the minimum number of clusters threshold is set to three, the maximum level depth is set to five, and the clustering stopping condition is that the average distance change between two adjacent layers is less than 0.05. The final output is the tool feature clustering structure.

[0041] S603: Based on the t-SNE visualization dimensionality reduction algorithm, TFCS is mapped in two dimensions and displayed in an interpretable manner; then, the Top-K retrieval strategy is used to output the tool type most similar to the unknown damage, generating a tool similarity index library.

[0042] Unknown damage refers to damage samples that do not yet have a corresponding tool label or a matching confidence level below a set threshold in the current forensic damage database. Essentially, it is a set of high-dimensional damage features to be identified, which needs to be compared with known tool samples through t-SNE dimensionality reduction mapping and Top-K similarity retrieval to infer the potential tool type.

[0043] Based on the t-SNE visualization dimensionality reduction algorithm, a two-dimensional spatial mapping and interpretable display of TFCS are performed. Then, a Top-K retrieval strategy is used to output the tool type most similar to the unknown damage. During execution, the three-dimensional embedded features are first input into the t-SNE algorithm. The similarity probability distribution of each feature point is calculated, and iterative mapping is performed based on a perplexity parameter set to 30. After mapping, the two-dimensional coordinate range is limited to the interval between -5 and 5. Regions with a Euclidean distance less than 0.5 are defined as tightly clustered regions, those between 0.5 and 1 are defined as moderately dense regions, and those greater than 1 are defined as discrete regions. Then, a system is established within the mapping results. The index table records the coordinates of the center point of each tool category as the retrieval reference point. When an unknown damage sample is input, its distance to each reference point in two-dimensional space is calculated and sorted. The tool categories corresponding to the five smallest distances are selected as the Top-5 output. For example, when the distances of the unknown sample to categories A, B, C, D, and E are 0.3, 0.45, 0.58, 0.62, and 0.71, respectively, then categories A to E are output as candidate tool types. To control the stability of the retrieval, the Top-K threshold K is set to five, and the minimum similarity screening threshold is set to 0.35. Items below these values ​​are not included in the retrieval results, and finally, a tool similarity index library is generated.

[0044] The S7-based generative multimodal injury inference model includes the following steps: S701: Based on a multi-source data fusion framework, the geometric features of the tool point cloud corresponding to TSID are fused with the CT images of the damaged area and the stress data of finite element simulation at the feature level to generate a multimodal input dataset. Based on a multi-source data fusion framework, the geometric features of the tool point cloud corresponding to TSID are fused with the CT images of the damaged area and the stress data from finite element simulation at the feature level. During execution, the tool point cloud data is first morphologically sampled, unifying the number of 3D coordinate points for each tool model to 50,000 points. These points are evenly distributed using the farthest point sampling method, and the curvature, normal vector, and surface area ratio features of each point are extracted. The curvature value is calculated by averaging the angles between the normals of adjacent points. Points with an angle greater than 20 degrees are defined as high curvature points, those between 10 and 20 degrees as medium curvature points, and those below 10 degrees as low curvature points. The curvature intervals are assigned weights of 1, 0.8, and 0.5, respectively. Simultaneously, the CT images of the damaged area are read, and voxels within the grayscale range of 100 to 1200 are converted into density data, and then... The data is resampled by mean averaging every ten voxels. When the mean gray value exceeds 700, it is defined as a high-density damage area. Then, the stress and displacement information of the nodes in the finite element simulation results are extracted. Nodes with a maximum principal stress greater than 120 MPa are marked as high-stress points, those between 60 and 120 MPa are marked as medium-stress points, and those less than 60 MPa are marked as low-stress points. The weights of each interval are set to 1.5, 1, and 0.6, respectively. In the fusion stage, the three types of data are normalized and superimposed according to their weights. The weighted values ​​of the corresponding positions of the point cloud features, CT density features, and stress features are combined into a unified feature vector. For example, if the weights of a certain point cloud point are 1, 0.8, and 0.6, the weighted average of the fused feature values ​​is 0.85. By performing fusion calculations on all samples, a multimodal input dataset is generated.

[0045] S702: Based on the multimodal Transformer network, cross-modal feature alignment and deep interaction modeling are performed on MID to generate a fused feature representation model; Based on a multimodal Transformer network, cross-modal feature alignment and deep interaction modeling are performed on MID. During execution, the feature vectors of each sample in the multimodal input dataset are first divided into three groups: geometric feature group, CT density feature group, and stress feature group. A linear embedding layer unifies the dimensions of each group to 128 dimensions before inputting them into the Transformer encoding module. In the alignment phase, modal attention weights are set, with the geometric feature channel weight set to 0.4, the CT density channel to 0.35, and the stress channel to 0.25. When the average alignment error calculated on the validation set during the early training phase exceeds 0.12, the weight distribution is adjusted, tilting the weight ratio 10% towards the channel with the larger error. For example, if the geometric feature channel weight is higher than 0.12, the weight distribution is adjusted. If the geometric channel error is 0.15 and the CT error is 0.09, then the weight of the geometric channel is adjusted to 0.44, and the weight of the CT channel is reduced to 0.33. In the interaction modeling stage, the number of heads in the multi-head attention mechanism is set to eight, and the single-head dimension is sixteen. By calculating the cross-modal attention matrix, the interaction features between different modalities are compared. When the interaction intensity value is greater than 0.6, it is judged as a highly correlated feature pair. The value between 0.3 and 0.6 is considered medium-correlation, and the value less than 0.3 is considered low-correlation. The weight of the highly correlated feature pair in the feature fusion layer is doubled. For example, in a certain sample, the interaction intensity between the CT density change area and the stress distribution gradient change area is 0.65. Then, the interaction output of the two is weighted and amplified by one time, and finally a stable fusion feature representation model is formed in the multi-layer Transformer structure stack.

[0046] S703: Based on a multi-task prediction mechanism, it generates a multimodal injury inference model by utilizing key indicators such as FFRM prediction tool category, injury angle, and depth of action.

[0047] Based on a multi-task prediction mechanism, this study utilizes FFRM to predict key indicators such as tool category, wound angle, and depth of impact. During execution, the vector output from the fused feature representation model is first input into the multi-task branch network. Each branch corresponds to a different prediction target. The tool category branch employs a Softmax classification structure with 128 input dimensions and 10 output categories. The category threshold is set at a 70% confidence level; an output probability higher than 0.7 is considered a predicted category. The wound angle branch uses a regression structure with a prediction range of 0 to 90 degrees, divided into intervals of five degrees. When a predicted value falls within a five-degree interval, the representative value for that interval is output. For example, the input... If the error exceeds 43 degrees, it falls into the 45-degree range. The depth branch uses a linear regression structure, with a depth prediction range of 0 to 50 millimeters. A baseline depth error threshold of 2 millimeters is set. When the error exceeds this threshold, the gradient is backpropagated to update the parameters. During training, joint loss weights are set for the three tasks: 0.5 for class prediction, 0.3 for angle prediction, and 0.2 for depth prediction. If the average angle error in a batch is greater than 5 degrees, its weight is increased to 0.4, the class weight is reduced to 0.45, and the depth weight remains at 0.2. Through this dynamic weight adjustment process, the parameters are continuously updated in the multi-task iteration. When the model training converges and the prediction is stable, a multimodal injury inference model is generated.

[0048] The generation of forensic visualization injury reports based on S8 includes the following steps: S801: Based on a 3D rendering algorithm, it generates a 3D damage-to-visualization model by utilizing the geometric and mechanical results output by MIIM and matching the scene with the tool. Based on a 3D rendering algorithm, a 3D damage-to-visualization tool is generated using the geometric and mechanical results output by MIIM. During execution, the tool first reads the geometric mesh data and corresponding mechanical distribution data output by the MIIM model. The geometric mesh consists of approximately 100,000 triangular faces, with each vertex containing 3D coordinates (x, y, z) and a normal vector. The mechanical data includes the magnitude of the forces acting on the nodes. With direction vector ,in The value range is 0 to 200 N. The two types of data are registered according to node numbers. When the node index matching error exceeds 0.001, interpolation is performed again to ensure a one-to-one correspondence between the geometric and mechanical data. Subsequently, Phong lighting rendering is applied to the geometric model, setting the ambient light intensity A=0.3, diffuse reflection coefficient D=0.6, and specular coefficient S=0.1. Three-channel color coding maps the force magnitude to a red-blue gradient range, where areas below 50 N are displayed as blue, 50 to 120 N as yellow, and above 120 N as red. The average force at each vertex is calculated during rendering. For example, if the forces on four neighboring points are 30, 45, 40, and 35 N respectively, then F_avg is 37.5 N, and the corresponding mapped color is a blue-green transition color; then, a matching judgment is made based on the relationship between the tool's incident angle θ and the angle between the normal to the damaged surface, when... When a node is identified as a match, it is marked as a green dot and displayed with a transparency of α=0.5 to highlight the matching area. If the angle difference is between 10° and 25°, it is marked as a semi-transparent yellow area. If it exceeds 25°, it is marked as a gray unmatched area. Through point-by-point rendering and angle filtering, a three-dimensional visualization matching model is generated.

[0049] S802: Based on the heat map generation algorithm, the matching probability of stress intensity and crack corresponding regions is mapped on the 3DVMM surface to generate a damage matching heat map. Based on a heatmap generation algorithm, the matching probability of stress intensity and corresponding crack regions is mapped on the 3DVMM surface. During execution, the 3DVMM model surface is first divided into uniform mesh elements, with each element having an area of ​​1 mm². 2 There are approximately 20,000 units in total, each containing an average force value. With crack density Two characteristics were used: the stress range was defined as low (0–60 MPa), medium (60–120 MPa), and high (120–200 MPa); the crack density range was defined as low (0–0.2), medium (0.2–0.5), and high (0.5). The matching probability was calculated. When using the force weighting coefficient With crack weight coefficient The weighted average, Set it to 0.6. Set to 0.4, the mapping probability is determined by judging the combination of stress and crack level in each region. When the stress level and crack level are consistent, Take 0.85, when there is a difference of one level When the value is 0.55, and the difference is two levels... Take 0.25, for example, for a certain area. , If all corresponding values ​​are medium level, then the matching probability is 0.85; then all... Visualization is achieved by projecting 3D surface coordinates onto a heatmap plane and using a color gradient from red (high match) to blue (low match). The color range is set as follows: Displayed in red, Displayed in yellow, It is green. The color is blue. Finally, the probabilities of adjacent units are made spatially continuous through smooth interpolation. When the difference between the probabilities of adjacent units is greater than 0.3, linear interpolation is performed to take the median value, thus forming a damage matching heatmap.

[0050] S803: Based on the automatic forensic report generation module, it integrates IMH, tool prediction results and mechanical parameters to form an interpretable forensic conclusion document and generate a visual forensic injury report.

[0051] Based on the automatic forensic report generation module, integrating IMH, tool prediction results, and mechanical parameters, the process first imports damage depth data from the IMH. The category of the output of the tool prediction model and angle of injury and the forces calculated by the finite element simulation module ,stress Deformation Isomechanical parameters are stored, with each parameter corresponding to a time step t_k. During the data integration phase, an index table is created for each set of parameters, and average indicators are statistically analyzed according to tool category, for example... The average injury angle of the tool was 42°, the average force was 130 N, and the average damage depth was 18 mm. Then, a structured descriptive paragraph was generated for each sample, and key fields were filled using a preset text template, such as "Sample number A103, inferred tool type is..." The damage was classified as "Class, with an action angle of approximately 42°, corresponding to a damage depth of 18 mm, and a main force of 130 N; the damage characteristics match the corresponding category characteristics"; a text generation threshold was set during the report generation process. When the model generates a confidence level higher than 0.8, it is automatically written into the conclusion area; if it is lower than 0.8, it is marked as "needs review". For outlier handling, if a parameter exceeds the empirical benchmark value by ±20%, such as mean stress... The baseline value of 80 MPa is recorded as "stress is too high" in the report notes. Finally, the conclusion paragraphs, table data and key frame images generated from all samples are integrated and output as a forensic visual injury report.

[0052] A damage-causing tool inference system based on point clouds and Siamese networks, the system comprising: The point cloud preprocessing module, based on 3D reconstruction and point cloud processing technology, uses CT slice 3D reconstruction algorithm and point cloud sampling method to obtain initial point cloud data of skull injury; then, noise removal and point density balancing are completed by statistical outlier removal algorithm and uniform resampling method; then, the core injury region is extracted by curvature estimation and region segmentation algorithm, and geometric standardization is achieved by scale normalization and normal correction strategy, so as to provide a consistent spatial expression for subsequent feature extraction and generate a standardized injury point cloud model. The feature encoding module, based on a point cloud deep learning feature extraction framework, performs multi-scale geometric encoding on the normalized damage point cloud model output by the previous module. It uses hierarchical sampling and local feature aggregation algorithms to extract fine-grained geometric information; it uses graph convolution feature construction methods to establish topological connections between points and capture the continuity of crack edges; and it combines a global feature fusion algorithm to achieve feature integration of local and global morphology. At the same time, the system loads the trained tool geometric encoding model based on self-supervised geometric pre-training and transfer learning strategies to ensure the consistency between the damage and tool feature spaces and generate damage geometric feature encoding. The matching inference module, based on a dual-channel Siamese network structure, inputs the damage geometric feature encoding output from the upper module and the pre-trained tool geometric features into two shared weight encoders to construct a joint embedding space. A cross-domain attention mechanism is used to strengthen the correspondence between key cracks and edges, and the matching discrimination ability is improved by a joint optimization strategy of contrastive loss and triplet loss. Subsequently, the feature similarity matrix is ​​calculated using the spherical metric learning algorithm, and the similarity space is clustered and visualized using hierarchical clustering and t-SNE dimensionality reduction analysis methods. The tool similarity retrieval results are output, and a tool similarity index library is generated. The visualization report module, based on 3D rendering and multimodal fusion technology, calls the tool similarity index library of the previous module and integrates the geometric features of the damage point cloud, the structure of the tool point cloud, and the stress data from finite element simulation. It uses a multimodal Transformer interactive algorithm to achieve feature-level fusion and injury mechanism reconstruction. Then, it uses a 3D visualization rendering algorithm to generate a damage-tool matching scene and presents the stress intensity and matching probability distribution through a heat map mapping method. Finally, it automatically outputs an interpretable inference report with the help of a forensic report generation algorithm, generating a forensic visualization injury report.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for inferring damage-causing tools based on point clouds and Siamese networks, characterized in that: Includes the following steps: S1: Based on the three-dimensional skull data reconstructed from forensic CT scans, point cloud resampling and denoising algorithms are used to extract and purify the damaged area; Then, the suspected crack and depression areas are separated by curvature analysis and region growth segmentation. In order to maintain the consistency of geometric structure, the point cloud is subjected to normal vector estimation and scale normalization to ensure the stability and comparability of subsequent feature extraction and generate a normalized damage point cloud model. S2: Based on the PointNet++ multi-scale feature extraction network, local geometric attributes such as normal vector change rate, curvature gradient, and indentation depth distribution are captured at different spatial scales. The graph convolutional network is further used to construct the point cloud neighborhood topology to realize the continuous encoding of crack morphology. Finally, the local and global information are fused to form a unified morphological feature representation and generate damage geometric feature encoding. S3: Collect tool models of typical blunt and sharp objects from the GrabCAD 3D model database, and obtain standardized input using the same point cloud preprocessing process as S1; based on the self-supervised learning pre-training mechanism, train the geometric structure invariant encoding network, learn the morphological prior knowledge representation of the tool, and generate the tool geometric prior feature model. S4: Based on the tool geometric prior feature model of S3, the tool features are transferred to the damage domain through the domain adaptive layer to reduce the distribution difference between virtual tools and real damage. An adversarial feature alignment training strategy is adopted to enable the network to learn a cross-domain consistent geometric representation and generate a cross-domain feature mapping model.

2. The method for inferring damage-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: The generation of a normalized damage point cloud model based on S1 includes the following steps: S101: Based on the reconstruction algorithm of forensic CT scan data, the damaged skull CT slices are converted into a surface mesh model through three-dimensional reconstruction technology; then, a point cloud sampling algorithm is used to generate a dense point cloud representation, generating the original damage point cloud data. S102: Based on the statistical outlier removal algorithm, noise is removed from RIPC to eliminate scan pseudo-points and external interference points; then, the point cloud density is balanced through the uniform resampling algorithm to ensure the consistency of point distribution and generate denoised balanced point cloud data. S103: Based on the curvature estimation algorithm and the region growth segmentation method, crack region separation and depression detection are performed on DBPC to extract the main damage region and generate a point cloud of the damage region; S104: Based on the normal vector estimation algorithm and scale normalization strategy, the IRPC is oriented and geometrically scaled to establish a consistent geometric basis for subsequent feature extraction and generate a standardized damage point cloud model.

3. The method for inferring damage-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: The S2-based generation of damage geometric feature encoding includes the following steps: S201: Based on PointNet++ hierarchical sampling and grouping strategy, multi-scale feature extraction is performed on NIPC to capture local geometric information, normal vector change rate and curvature distribution, and generate multi-scale local feature set; S202: Based on graph convolutional networks, topological connections are established in the point cloud neighborhood of MSLFS to explicitly model crack morphology and edge continuity and generate crack topological structure encoding. S203: Based on a global feature aggregation network, CTE and MSLFS are fused to construct a unified descriptor that integrates local details and overall morphology, generating damage geometric feature encoding.

4. The method for inferring damage-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: The S3-based generative tool geometric prior feature model includes the following steps: S301: Based on sampling from a 3D model library, blunt and sharp tool models are extracted from databases such as GrabCAD, and data is standardized by surface resampling and denoising algorithms to generate standardized tool point cloud data. S302: Based on data augmentation algorithms, STPC is subjected to occlusion, rotation, erosion, and perturbation operations to form diverse geometric input samples and generate an augmentation tool point cloud; S303: Based on a self-supervised learning mechanism, the geometric encoding network is trained using ATCS to learn the invariant representation of tool morphology and generate a tool geometry pre-trained model.

5. The method for inferring damage-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: The S4-based model for generating cross-domain feature maps includes the following steps: S401: Based on the transfer learning mechanism, TGPM is transferred to the damage feature extraction task. The feature space of the input damage data is adaptively adjusted to generate a primary transfer feature model. S402: Based on the domain adaptation layer, PTFM is used to minimize the difference between the distribution of virtual tool features and real damage features to generate a domain-aligned feature model. S403: Based on the adversarial feature alignment algorithm, a discriminator is introduced on the basis of DAFM to optimize cross-domain feature consistency and generate a cross-domain feature mapping model.

6. The method for inferring damage-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: Also includes: S5: The damage geometric feature encoding of S2 and the cross-domain feature mapping model of S4 are jointly input into the dual-channel Siamese matching network. The two networks share the weight structure. The key morphological regions are focused through the cross-domain feature attention mechanism. The triple loss and contrast loss are jointly optimized to make similar tools and corresponding damage features closely aggregated in the embedding space, generating a damage-tool joint embedding feature space. The S5-based generative damage-tool joint embedding of the feature space includes the following steps: S501: Based on a dual-channel shared weight network structure, IGFE and CDFMM are input into two encoders respectively to maintain feature learning consistency and generate initial dual-channel feature embeddings. S502: Based on the cross-domain feature attention mechanism, the geometric difference region in ISE is weighted and enhanced to highlight the key crack and tool edge morphology, and generate weighted attention embedding features; S503: Based on joint loss function optimization, embedding space clustering training is performed on AWEF to aggregate homologous tools and damage features with high similarity and separate heterologous features, generating a damage-tool joint embedding feature space.

7. The method for inferring damage-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: Also includes: S6: Based on the damage-tool joint embedding feature space generated by S5, a high-resolution embedding feature index library is constructed through spherical metric learning. Hierarchical clustering is further used to separate different tool types between classes, and t-SNE is used to visualize and verify the interpretability of the embedding space structure. This enables Top-K similar tool retrieval for unknown samples and generates a tool similarity index library. The generation tool similarity index library based on S6 includes the following steps: S601: Based on the spherical metric learning algorithm, a similarity metric function is constructed in JMES to quantify the morphological distance between tools and generate an embedding space similarity matrix; S602: Based on the hierarchical clustering algorithm, ESM is clustered and divided into feature clusters of different tool categories, generating tool feature clustering structures; S603: Based on the t-SNE visualization dimensionality reduction algorithm, TFCS is mapped in two dimensions and displayed in an interpretable manner; then, the Top-K retrieval strategy is used to output the tool type most similar to the unknown damage, generating a tool similarity index library.

8. The method for inferring damage-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: Also includes: S7: Based on the tool similarity index library of S6, it integrates point cloud geometric features with CT image texture features and finite element simulation data, and uses a multimodal Transformer coding network to achieve cross-modal feature alignment. It comprehensively outputs the prediction results of potential injury-causing tool category, direction of action and force depth. This process integrates visual geometric and mechanical features, enhances the forensic interpretability of the model, and generates a multimodal injury inference model. The S7-based generative multimodal injury inference model includes the following steps: S701: Based on a multi-source data fusion framework, the geometric features of the tool point cloud corresponding to TSID are fused with the CT images of the damaged area and the stress data of finite element simulation at the feature level to generate a multimodal input dataset. S702: Based on the multimodal Transformer network, cross-modal feature alignment and deep interaction modeling are performed on MID to generate a fused feature representation model; S703: Based on a multi-task prediction mechanism, a multimodal injury inference model is generated by utilizing key indicators such as FFRM prediction tool category, injury angle, and depth of action.

9. The method for inferring injury-causing tools based on point clouds and Siamese networks according to claim 1, characterized in that: Also includes: S8: Based on the multimodal injury inference model output of S7, the system employs 3D visualization rendering and heatmap generation algorithms to explicitly display the correspondence between injury and prediction tools. It generates a forensic visualization report containing tool type, injury angle, force direction, and injury matching probability, used for judicial evidence presentation, and generates a forensic visualization injury report. The generation of forensic visualization injury reports based on S8 includes the following steps: S801: Based on a 3D rendering algorithm, it generates a 3D damage-to-visualization model by utilizing the geometric and mechanical results output by MIIM and matching the scene with the tool. S802: Based on the heat map generation algorithm, the matching probability of stress intensity and crack corresponding area is mapped on the 3DVMM surface to generate a damage matching heat map. S803: Based on the automatic forensic report generation module, it integrates IMH, tool prediction results and mechanical parameters to form an interpretable forensic conclusion document and generate a visual forensic injury report.

10. A wound-causing tool inference system based on point cloud and Siamese network, characterized in that: The system comprising the injury tool inference method based on point cloud and Siamese network as described in any one of claims 1-9, wherein the system includes: The point cloud preprocessing module, based on 3D reconstruction and point cloud processing technology, uses CT slice 3D reconstruction algorithm and point cloud sampling method to obtain initial point cloud data of skull injury; then, noise removal and point density balancing are completed by statistical outlier removal algorithm and uniform resampling method; then, the core injury region is extracted by curvature estimation and region segmentation algorithm, and geometric standardization is achieved by scale normalization and normal correction strategy, so as to provide a consistent spatial expression for subsequent feature extraction and generate a standardized injury point cloud model. The feature encoding module, based on a point cloud deep learning feature extraction framework, performs multi-scale geometric encoding on the normalized damage point cloud model output by the previous module. It uses hierarchical sampling and local feature aggregation algorithms to extract fine-grained geometric information; it uses graph convolution feature construction methods to establish topological connections between points and capture the continuity of crack edges; and it combines a global feature fusion algorithm to achieve feature integration of local and global morphology. At the same time, the system loads the trained tool geometric encoding model based on self-supervised geometric pre-training and transfer learning strategies to ensure the consistency between the damage and tool feature spaces and generate damage geometric feature encoding. The matching inference module, based on a dual-channel Siamese network structure, inputs the damage geometric feature encoding output from the feature encoding module and the pre-trained tool geometric features into two shared weight encoders to construct a joint embedding space. A cross-domain attention mechanism is used to strengthen the correspondence between key cracks and edges, and the matching discrimination ability is improved by a joint optimization strategy of contrastive loss and triplet loss. Subsequently, the feature similarity matrix is ​​calculated using the spherical metric learning algorithm, and the similarity space is clustered and visualized using hierarchical clustering and t-SNE dimensionality reduction analysis methods to output the tool similarity retrieval results and generate a tool similarity index library. The visualization report module, based on 3D rendering and multimodal fusion technology, calls the tool similarity index library of the matching inference module and integrates the geometric features of the damage point cloud, the structure of the tool point cloud, and the stress data from finite element simulation. It uses a multimodal Transformer interactive algorithm to achieve feature-level fusion and injury mechanism reconstruction. Then, it uses a 3D visualization rendering algorithm to generate a damage-tool matching scene and presents the stress intensity and matching probability distribution through a heat map mapping method. Finally, it automatically outputs an interpretable inference report with the help of a forensic report generation algorithm, generating a forensic visualization injury report.

Images