Forensic injury condition analysis method and system based on three-dimensional modeling and visual language large model

Abstract

The invention provides a forensic injury condition analysis method and system based on three-dimensional modeling and a visual language large model, and relates to the technical field of data processing.The method comprises the steps that an original three-dimensional model of a wound part is collected through a structured light scanner, the original three-dimensional model is preprocessed, and a standardized wound three-dimensional model is obtained; the method comprises the following steps: identifying three wound edge vertexes of a wound edge in a standardized wound three-dimensional model, establishing a reference topological relation, and carrying out multi-level structure subdivision on the reference topological relation based on a deep feature extraction mechanism of a convolutional neural network to obtain a subdivided topological structure; deep extraction and quantitative analysis of wound deformation features are carried out on the subdivided topological structure, the feature information of the wound tiny deformation and the fine topological structure is accurately captured, and the feature information obtained through extraction and analysis is combined to calculate and obtain proportion calibration parameters. According to the invention, accurate quantitative extraction of wound features, standardized data processing and intelligent wound identification reasoning are realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for forensic injury analysis based on three-dimensional modeling and large-scale visual language models. Background Technology

[0002] In traditional forensic injury assessment scenarios, due to the lack of accurate three-dimensional data support and intelligent analysis tools, the assessment results often rely on the visual observation, manual measurement and subjective experience of the forensic personnel, which has technical limitations.

[0003] For example, if a patient suffers an irregular wound on the left forehead caused by a blunt object, in the traditional identification process, the examiner observes the wound morphology through two-dimensional photographs taken at the scene, manually measures the wound length using a ruler, and determines the time of injury by combining clinical records and personal experience. Even if no organic damage is found, the wound may be confirmed through surgery, and the time of injury may be reconstructed through video surveillance. This leads to serious flaws in the initial identification conclusion due to data bias and experience-based misjudgment. Two-dimensional images cannot fully present the three-dimensional topological structure and depth characteristics of the wound. Manual measurement is easily affected by angle and operation techniques, resulting in insufficient accuracy. Furthermore, key information such as minor deformations and subtle tissue damage are easily overlooked. At the same time, the determination of the time of injury and organic damage lacks standardized feature extraction and quantitative analysis mechanisms, making it difficult to guarantee the accuracy of the identification results. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method and system for forensic injury analysis based on three-dimensional modeling and visual language large model, so as to realize accurate quantitative extraction of trauma features, standardized data processing and intelligent injury identification reasoning.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a forensic injury analysis method based on 3D modeling and a large-scale visual language model, the method comprising: The original 3D model of the trauma site was acquired using a structured light scanner, and the original 3D model was preprocessed to obtain a standardized 3D model of the trauma. By identifying the three wound edge vertices in the standardized 3D trauma model and establishing a baseline topological relationship, the baseline topological relationship is further subdivided into multiple levels based on the deep feature extraction mechanism of the convolutional neural network, resulting in the subdivided topological structure. By performing in-depth extraction and quantitative analysis of the wound deformation features of the subdivided topology, the feature information of the wound's minute deformation and fine topology is accurately captured, and the proportional calibration parameters are calculated by combining the feature information obtained from the extraction and analysis. Based on the proportional calibration parameters, the 3D trauma model is subjected to geometric pose normalization and scale standardization to obtain the processed 3D trauma model; the wound is quantitatively measured and labeled on the processed 3D trauma model to obtain the wound length and area data, thus obtaining the quantitatively labeled 3D model. Based on the quantified and annotated 3D model, a virtual camera is used to take standardized multi-view shots in the world coordinate system to obtain a 2D image containing wound morphology, physical reference points and quantified annotation information. Two-dimensional images and associated quantitatively labeled data are input into a large visual language model fine-tuned by forensic injury data to extract visual and quantitative features of the wound. Based on the extracted features, the time interval of the injury formation and the identification of organic damage are completed.

[0006] Furthermore, a structured light scanner was used to acquire the original 3D model of the trauma site, and the original 3D model was preprocessed to eliminate scanning noise and repair data defects, resulting in a standardized 3D model of the trauma, including: Receive the original 3D model uploaded by the structured light scanner, and separate and extract the point cloud data and the initial mesh model from the original 3D model; By performing statistical filtering on the point cloud data, the mean and standard deviation of the distance of each point in the neighborhood are calculated, and outliers that deviate from the mean by more than a preset multiple of the standard deviation are removed, thus obtaining the point cloud data after statistical filtering. By performing radius filtering on the point cloud data, the number of neighboring points of each point cloud point is counted within the set search radius, and isolated points with fewer than the preset threshold of neighboring points are removed to obtain clean point cloud data after noise reduction. The surface mesh is reconstructed based on pure point cloud data. The reconstructed mesh model is then subjected to Laplacian mesh smoothing to eliminate burrs and jagged distortions on the mesh surface, resulting in a smoothed mesh model. The topological void regions in the smooth mesh model are detected, and the void regions are filled with mesh to obtain a complete, continuous, and standardized 3D trauma model without missing data.

[0007] Furthermore, by identifying the three wound edge vertices in the standardized 3D trauma model and establishing a baseline topological relationship, multi-level structural subdivision of the baseline topological relationship is carried out based on the deep feature extraction mechanism of a convolutional neural network, resulting in a subdivided topological structure, including: Based on the standardized three-dimensional model of trauma, and based on the curvature change characteristics of the wound and the surrounding normal skin, as well as prior knowledge of forensic anatomy, three wound edge vertices with geometric features are automatically located and identified at the edge of the wound. The three wound edge vertices correspond to the starting point, ending point and key point of morphological transition of the wound, respectively, and the vertex set of the spatial coordinates of the three wound edge vertices is obtained. Based on the vertex set, the initial topological skeleton of the wound edge is constructed with three wound edge vertices as the reference. By generating initial connecting edges between the three vertices, a reference topological relationship representing the overall contour of the wound is formed, and the reference topological data is obtained. The baseline topological data is mapped to a two-dimensional parameterized grid and input into a pre-trained convolutional neural network feature extractor. The local geometric features and global morphological features of the topological structure are extracted step by step through the convolutional layers to obtain a feature map of multi-level feature responses. Based on multi-level feature maps, the system adaptively identifies geometrically discontinuous and morphologically changing regions at the wound edge. It then adaptively subdivides the baseline topological relationships based on feature response intensity, increasing the subdivision density in regions with drastic curvature changes and complex morphologies, while maintaining sparse connections in smooth regions, thus obtaining a refined topological structure.

[0008] Furthermore, by performing in-depth extraction and quantitative analysis of wound deformation features on the subdivided topology, the characteristic information of minute wound deformations and subtle topological structures is accurately captured. The proportional calibration parameters are then calculated based on the extracted and analyzed feature information, including: Based on the refined topology, the local geometric deformation features are calculated point by point along the subdivided topological path at the edge of the wound, including the principal curvature, normal vector deflection angle and Gaussian curvature gradient at each subdivided node, to obtain the feature vector sequence. By performing temporal and quantitative analysis on the feature vector sequence, the continuity and discontinuity of the wound edge deformation at the microscale are identified. Based on morphological prior knowledge, an effective deformation feature subset representing the true edge morphology of the wound is selected, and pseudo-deformation interference caused by residual scanning noise is eliminated to obtain the effective deformation feature set. Based on the effective deformation feature set, a mapping relationship between the deformation features of the wound edge and the standard reference scale is constructed. By analyzing the relative deformation distribution law of the wound edge within the reference frame formed by the connection of the three wound edge vertices, the proportional calibration parameters used to compensate for the scale distortion caused by the non-rigid deformation of the wound are calculated. The proportional calibration parameters include the linear scaling factor along the long axis of the wound and the non-uniform deformation compensation coefficient along the normal of the wound edge.

[0009] Furthermore, the 3D trauma model is subjected to geometric pose normalization and scale standardization based on the scaling parameters to obtain a processed 3D trauma model. The wound is then quantified and annotated on the processed 3D trauma model to obtain the wound length and area data, resulting in a quantified and annotated 3D model, including: Based on the proportional calibration parameters, the standardized 3D trauma model is spatially transformed; the model is scaled in the long axis direction of the wound by the linear scaling factor; the local deformation of the model is compensated along the normal direction of the wound edge by the non-uniform deformation compensation coefficient to eliminate the scale distortion caused by non-rigid deformation, and the model is adjusted to the standard anatomical posture to obtain the 3D trauma model after geometric posture normalization and scale standardization. Quantitative measurements are automatically performed on the processed 3D model of the wound. The measurements include calculating the Euclidean distance between the two ends of the longest axis of the wound as the wound length and calculating the surface area of ​​the wound grid as the wound area. Optionally, the maximum wound depth and average wound edge width can be calculated as supplementary quantitative data to obtain the quantitative measurement results. The quantitative measurement results are annotated on the 3D model in a unified visualization style; the annotation content includes at least the length value and measurement path, and the area value and boundary range; the annotated model is stored in association with the quantitative data to form a quantitatively annotated 3D model.

[0010] Furthermore, based on the quantized and annotated 3D model, multi-view standardized photography is performed using a virtual camera in the world coordinate system to obtain 2D images containing wound morphology, physical reference points, and quantized annotation information, including: The quantized and annotated 3D model is registered to a unified world coordinate system to obtain the registered 3D model. The geometric center of the wound surface in the registered 3D model is used as the origin of the coordinate system, and the spatial coordinate axes are established based on the main plane of the wound and the depth direction to establish a standardized spatial reference. Based on a standardized spatial reference, the internal and external parameters of the virtual camera are configured to set the shooting conditions; according to the three-dimensional spatial orientation of the wound in the standardized spatial reference, a set of standardized shooting angles and corresponding camera poses covering key anatomical structures are planned and generated. Based on the planned camera pose, the virtual camera is driven to perform multi-view rendering and shooting of the registered 3D model; the model geometry and associated 3D visualization annotation information are projected together onto the 2D imaging plane to generate a set of original view images containing wound morphology, physical reference points and quantitative annotation information. The original perspective image is standardized and post-processed to uniformly adjust its brightness and contrast, and embeds image caption information containing case identification, shooting perspective and key quantitative data, and finally outputs a set of two-dimensional image sequences that meet the preset imaging standards.

[0011] Furthermore, the two-dimensional image and associated quantitatively labeled data are input into a large-scale visual language model fine-tuned from forensic injury data to extract the visual and quantitative features of the wound. Based on the extracted features, the determination of the time interval of the injury formation and the identification of organic damage are completed, including: Standardized two-dimensional image sequences and their associated quantized annotation data are loaded as multimodal inputs into a large visual language model that has been fine-tuned using forensic injury data. Based on the loaded multimodal input, the visual language large model extracts visual features from two-dimensional images and quantitative features from quantized labeled data in parallel, and fuses the visual features and quantitative features to obtain a comprehensive feature vector containing wound color, shape, tissue state and geometric size. By integrating feature vectors, the visual language big model calculates and determines the time interval to which the trauma formation time belongs based on the internally pre-learned trauma healing features and time mapping relationship, and obtains the confidence level of the determination result. Based on the time interval determination, the visual language big data model further identifies and classifies organic damage by using information related to tissue depth and damage range in the comprehensive feature vector. Finally, it outputs a structured identification report that includes the time interval determination, the organic damage identification conclusion, and the corresponding confidence level.

[0012] Secondly, a forensic injury analysis system based on 3D modeling and a large-scale visual language model includes: The acquisition module is used to acquire the original 3D model of the trauma site using a structured light scanner and to preprocess the original 3D model to obtain a standardized 3D model of the trauma. The segmentation module is used to identify the three wound edge vertices in the standardized 3D trauma model and establish a baseline topological relationship. Based on the deep feature extraction mechanism of the convolutional neural network, the baseline topological relationship is further subdivided into multi-level structures to obtain the subdivided topological structure. The extraction module is used to perform in-depth extraction and quantitative analysis of wound deformation features on the subdivided topology, accurately capture the feature information of small deformations and fine topology of the wound, and calculate the proportional calibration parameters by combining the feature information obtained from the extraction and analysis. The quantization module is used to perform geometric pose normalization and scale standardization on the trauma 3D model according to the proportional calibration parameters to obtain the processed trauma 3D model; on the processed trauma 3D model, the wound is quantitatively measured and labeled to obtain the wound length and area data, and the quantized and labeled 3D model is obtained. The shooting module is used to take standardized multi-view shots in the world coordinate system based on the quantized and annotated 3D model, using a virtual camera to obtain a 2D image containing wound morphology, physical reference points and quantified annotation information. The processing module is used to input two-dimensional images and associated quantitatively labeled data into a large visual language model that has been fine-tuned by forensic injury data, extract the visual and quantitative features of the wound, and determine the time interval of the injury formation and identify organic damage based on the extracted features.

[0013] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: Because a standardized 3D model is obtained by using a structured light scanner to acquire the original 3D model of the trauma and preprocessing it through filtering, smoothing, and hole filling, the topological structure is subdivided into multiple levels through wound edge vertex recognition and convolutional neural networks. Combined with wound deformation feature extraction and calculation of proportional calibration parameters, the model standardization and quantitative annotation are completed. Standardized 2D images are generated by taking pictures from multiple perspectives with a virtual camera. Then, a large visual language model fine-tuned by forensic injury data is used to integrate multimodal features to complete the determination of the time of trauma formation and the identification of organic damage. This achieves the effects of accurate quantitative extraction of trauma features, standardized data processing, and intelligent injury reasoning, improving the accuracy of forensic injury identification results and fully meeting the requirements of judicial practice for the rigor of evidence. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the forensic injury analysis method based on three-dimensional modeling and visual language large model provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a forensic injury analysis system based on three-dimensional modeling and visual language large model provided in an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] like Figure 1 As shown, embodiments of the present invention propose a forensic injury analysis method based on 3D modeling and a large visual language model. The method includes the following steps: Step 1: Use a structured light scanner to acquire the original 3D model of the trauma site, and preprocess the original 3D model to obtain a standardized 3D model of the trauma. Step 2: Identify the three wound edge vertices in the standardized 3D trauma model and establish a baseline topological relationship. Then, perform multi-level structural subdivision of the baseline topological relationship based on the deep feature extraction mechanism of the convolutional neural network to obtain the subdivided topological structure. Step 3: By performing in-depth extraction and quantitative analysis of the wound deformation features of the subdivided topology, the feature information of the wound's minute deformation and fine topology is accurately captured, and the proportional calibration parameters are calculated by combining the feature information obtained from the extraction and analysis. Step 4: Perform geometric pose normalization and scale standardization on the trauma 3D model according to the proportional calibration parameters to obtain the processed trauma 3D model; perform quantitative measurement and annotation of the wound on the processed trauma 3D model to obtain the wound length and area data, and obtain the quantified and annotated 3D model. Step 5: Based on the quantized and annotated 3D model, use a virtual camera to perform multi-view standardized shooting in the world coordinate system to obtain a 2D image containing wound morphology, physical reference points and quantified annotation information. Step 6: Input the two-dimensional image and associated quantitative annotation data into the visual language model fine-tuned by forensic injury data, extract the visual and quantitative features of the wound, and determine the time interval of the injury formation and identify organic damage based on the extracted features.

[0020] In this embodiment of the invention, a standardized three-dimensional model of the wound is obtained by preprocessing the original three-dimensional model of the wound site using a structured light scanner. A baseline topological relationship is established by identifying the vertices of the wound edge, and multi-level structural subdivision is carried out based on a convolutional neural network. The deformation features of the wound are extracted, and the proportional calibration parameters are calculated to complete the normalization of the model's geometric posture and the standardization of scale, as well as the quantitative annotation of the wound. A two-dimensional image containing the wound morphology, physical reference points, and quantitative annotation information is generated by taking standardized multi-view shots in the world coordinate system using a virtual camera. The two-dimensional image and the associated quantitative annotation data are then input into a large visual language model that has been fine-tuned by forensic injury data to fuse visual features and quantitative features to complete the determination of the time interval of wound formation and the identification of organic damage. Therefore, this method effectively overcomes the technical problems in traditional forensic injury identification, such as the inability of two-dimensional images to fully present the three-dimensional topological structure and depth features of the wound, the susceptibility of manual measurement to interference leading to insufficient accuracy, the easy neglect of minor deformations and subtle tissue damage of the wound, and the lack of standardized feature extraction and quantitative analysis mechanisms and strong subjectivity in injury identification. This achieves the technical effects of accurate quantitative extraction of wound features, standardized data processing flow, and intelligent injury identification reasoning, thereby improving the accuracy of forensic injury identification results.

[0021] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Receive the original 3D model uploaded by the structured light scanner, and separate and extract the point cloud data and initial mesh model from the original 3D model. Specifically, this includes: First, starting the structured light scanner to perform a comprehensive scan of the trauma area. After the scan is completed, the scanner will automatically upload the collected original 3D model data to the receiving end through the data transmission channel. After receiving the original 3D model, the receiving end starts the data parsing program to classify and identify the data information contained in the original 3D model. Since the original 3D model consists of two parts: point cloud data and initial mesh model, the point cloud data records the 3D spatial coordinate information of a large number of discrete points on the surface of the trauma area, while the initial mesh model is a polygonal connection structure initially constructed based on these point cloud data. By identifying the different data formats and characteristics of the two types of data, the point cloud data and the initial mesh model are separated from the original 3D model.

[0022] Step 1.2 involves performing statistical filtering on the point cloud data to calculate the mean and standard deviation of the distances of each point within its neighborhood. Outliers deviating from the mean by more than a preset multiple of the standard deviation are removed, resulting in statistically filtered point cloud data. Specifically, this includes: first, setting a fixed neighborhood range for each point in the point cloud data, defined as the 50 neighboring point ranges around each point; then calculating the distances from each point to all points within its neighborhood, summing the distances, and taking the average to obtain the mean neighborhood distance for that point; next, calculating the corresponding standard deviation based on the mean neighborhood distance and each distance value; then setting a judgment criterion: if the distance value of a point deviates from the mean neighborhood distance by more than 3 times the standard deviation, the point is determined to be an outlier. These outliers are usually noise points generated during the scanning process due to environmental interference or equipment errors; finally, removing all outliers from the point cloud data to obtain statistically filtered point cloud data.

[0023] Step 1.3 involves performing radius filtering on the point cloud data. Within a set search radius, the number of neighboring points for each point cloud point is counted, and isolated points with fewer than a preset threshold of neighboring points are removed, resulting in denoised clean point cloud data. Specifically, this includes: initiating the radius filtering process; first, setting a search radius of 2 mm, determined by considering the accuracy requirements of trauma scanning and the distribution characteristics of common noise points; then, for each point cloud point in the statistically filtered data, within the set 2 mm search radius, searching and counting the number of all adjacent point cloud points; simultaneously, setting a threshold of 5 for the number of neighboring points. If a point cloud point has fewer than 5 neighboring points within the 2 mm search radius, it indicates that the point lacks effective connection with the surrounding point cloud and is considered an isolated noise point; finally, removing all isolated points with fewer than 5 neighboring points from the point cloud data. After this step, the noise in the point cloud data is further eliminated, resulting in denoised clean point cloud data.

[0024] Step 1.4: Reconstruct the surface mesh based on clean point cloud data. Perform Laplacian mesh smoothing on the reconstructed mesh model to eliminate burrs and jagged distortions on the mesh surface, resulting in a smoothed mesh model. Specifically, this involves: using clean point cloud data as a basis, constructing a complete surface mesh model representing the surface morphology of the trauma site through the three-dimensional spatial coordinate information of each point in the clean point cloud data and the pre-calculated normal vector information. This reconstruction process automatically generates reasonable polygonal mesh units based on the density and distribution characteristics of the point cloud, forming the initial surface mesh. Then, perform Laplacian mesh smoothing on the reconstructed mesh model. Iteratively calculate the average spatial position of each mesh vertex and its adjacent vertices, and adjust the coordinate position of the current vertex based on this average value. After multiple iterations, gradually eliminate the protruding burrs and irregular jagged distortions on the mesh surface caused by scanning accuracy limitations or residual noise, finally obtaining a smoothed mesh model that accurately reflects the true morphology of the trauma.

[0025] Step 1.5 involves detecting topological voids in the smoothed mesh model and filling these voids with mesh to obtain a complete, continuous, and data-free standardized 3D trauma model. This process includes: comprehensively traversing the smoothed mesh model; analyzing the connectivity of polygonal units to identify unclosed regions, which are the topological voids; after detecting all voids, using a triangular mesh filling algorithm to repair them; automatically generating new triangular mesh units based on the spatial distribution and connectivity of each void's edge vertices, ensuring perfect fit with existing vertices at the void's edge; then gradually filling the voids with these newly generated triangular mesh units from the void's edge to its center until all topological voids are completely filled; finally, a complete, continuous, smooth, and data-free standardized 3D trauma model is obtained.

[0026] In this embodiment of the invention, by using the original 3D model uploaded by the structured light scanner and separating and extracting point cloud data from the initial mesh model, outliers that deviate from the mean by more than a preset multiple of the standard deviation are removed by statistical filtering, isolated points that have fewer than a preset threshold of neighboring points within a set search radius are removed by radius filtering, and Laplacian mesh smoothing is performed after reconstructing the surface mesh based on the clean point cloud data to eliminate burrs and jagged distortions, and then topological void regions in the smooth mesh model are detected and filled, the technical means effectively overcome the problems of scanning noise, mesh surface distortion, and data incompleteness in the original 3D model, thereby achieving the technical effect of obtaining a complete, continuous, noise-free, and data-free standardized trauma 3D model.

[0027] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Based on the standardized 3D trauma model, and using the curvature abrupt change characteristics of the wound surface and surrounding normal skin, as well as prior knowledge of forensic anatomy, three geometrically characteristic wound edge vertices are automatically located and identified at the wound edge. These three vertices correspond to the start point, end point, and key morphological transition point of the wound, respectively, resulting in a set of vertices with spatial coordinates. Specifically, this involves: using the standardized 3D trauma model as the core basis, performing a comprehensive curvature calculation on all data points on the model surface; and accurately determining the curvature value of each point by analyzing the angle between each data point and the tangent of the surface formed by multiple adjacent data points. Combining prior knowledge of forensic anatomy regarding the boundary characteristics of the wound edge and surrounding normal skin, a curvature abrupt change threshold of 0.7 radians is explicitly set. When the curvature value of a data point exceeds this threshold, the point can be identified as such. The boundary points between the wound and normal skin are used to initially screen the boundary points. From all the screened boundary points, key nodes with clear geometric significance are further identified. Based on the natural morphological distribution of the wound, the two points located at the two ends of the wound with the greatest spatial distance and the largest curvature value are selected as the start and end points of the wound, respectively. Then, among all the boundary points between the start and end points, the curvature change amplitude of each point is compared, and the point whose curvature change clearly reflects the turning trend of the wound shape is selected as the key turning point. The three-dimensional spatial coordinates of the three wound edge vertices—the start point, end point, and key turning point—are accurately collected, and the specific coordinate values ​​of each vertex on the X, Y, and Z axes are recorded in detail. Finally, a vertex set containing the complete spatial coordinate information of the three wound edge vertices is formed.

[0028] Step 2.2: Based on the vertex set, construct the initial topological skeleton of the wound edge using the three wound edge vertices as a reference. This is achieved by generating initial connecting edges between the three vertices to form a reference topological relationship representing the overall contour of the wound, thus obtaining reference topological data. Specifically, this includes: using the acquired spatial coordinate set of the three wound edge vertices as the core reference for constructing the initial topological skeleton of the wound edge; firstly, connecting the starting and ending endpoints directly with straight lines to form a basic trunk line segment that represents the length direction of the wound, providing core support for the subsequent topological structure; then, using the basic trunk line segment as a reference, starting from the morphological transition points... A perpendicular line is drawn from the key point to the basic trunk line segment, and the coordinates of the intersection point of the perpendicular line and the basic trunk line segment are accurately calculated. Then, the key points of shape turning are connected to the intersection point with straight lines to form branch lines that support the wound outline. Initial connecting edges are generated by combining trunk lines and branch lines. All initial connecting edges are interconnected and connected in an orderly manner to form a continuous and complete network structure. The network structure comprehensively and accurately represents the overall outline shape of the wound. On this basis, key information such as the spatial coordinates of each vertex, the endpoint relationships of each connecting edge, and the actual length of each connecting edge are further integrated to finally form complete baseline topology data.

[0029] Step 2.3: Map the baseline topology data into a two-dimensional parameterized mesh and input it into a pre-trained convolutional neural network feature extractor. The local geometric features and global morphological features of the topology are extracted step by step through convolutional layers to obtain a feature map of multi-level feature responses. Specifically, the acquired baseline topology data is first subjected to two-dimensional projection transformation. During the projection process, the relative spatial position and connection relationships between each vertex are strictly maintained. The topology in three-dimensional space is accurately mapped to a two-dimensional plane to form a two-dimensional parameterized mesh with a size of 512×512, ensuring that the core features of the topology are completely preserved in two-dimensional space. The two-dimensional parameterized grid is used as input data and loaded into a pre-trained convolutional neural network feature extractor. This extractor contains six convolutional layers, each with a specific kernel size and stride. The first three convolutional layers use smaller kernels to extract local geometric features of the topology, such as small protrusions and depressions at the wound edge. The last three convolutional layers use larger kernels to extract global morphological features of the topology, such as the overall contour and aspect ratio of the wound. Through the progressive processing of each convolutional layer, each layer outputs a corresponding feature response map. The feature response maps output by all convolutional layers are integrated hierarchically to form a feature map that contains both local details and global features.

[0030] Step 2.4: Based on the multi-level feature map, adaptively identify the geometric discontinuities and morphological changes of the wound edge. Adaptively subdivide the baseline topology based on the feature response intensity. Increase the subdivision density in areas with drastic curvature changes and complex morphology, and maintain sparse connections in smooth areas to obtain a refined topology. Specifically, this includes: using the multi-level feature map as the basis for analysis, first setting the feature response intensity threshold to 0.6. By traversing all feature response values ​​in the feature map, areas with response values ​​exceeding the threshold are identified as geometric discontinuities or morphological changes of the wound edge, while areas with response values ​​below the threshold are identified as smooth areas. Based on the distribution of feature response intensity, adaptive subdivision processing is applied to the baseline topology: For regions with drastic curvature changes and complex shapes, i.e., regions where the feature response intensity exceeds the threshold, the subdivision density of the topology is increased, and each existing connecting edge in this region is evenly divided into 3 segments. At the same time, new connecting relationships are added at the segmentation points to form a finer topology network, so as to accurately capture the minute deformations and complex shapes of the wound; For smooth regions, i.e. regions where the feature response intensity is below the threshold, the original sparse connection state is maintained, and each connecting edge is retained as 1 segment to avoid data redundancy caused by excessive subdivision; Through differentiated subdivision processing, a refined topology structure that accurately restores all the detailed features and overall shape of the wound edge is finally obtained.

[0031] In this embodiment of the invention, by employing the curvature mutation features of the wound and surrounding normal skin, as well as prior knowledge of forensic anatomy, three vertices of the wound edge—the starting point, the ending point, and the key point of morphological transition—are automatically located and identified at the wound edge, and a set of spatial coordinates is obtained. Based on this set of vertices, an initial topological skeleton of the wound edge is constructed to obtain benchmark topological data. The benchmark topological data is mapped into a two-dimensional parameterized mesh and input into a pre-trained convolutional neural network feature extractor to extract local and global multi-level feature maps. Then, the benchmark topological relationship is adaptively subdivided according to the feature response intensity. Therefore, this method effectively overcomes the technical problems of traditional methods, such as the reliance on manual and low-precision positioning of key vertices of the wound, the lack of targeted topological structure division, the difficulty in taking into account both complex local morphology and global feature capture, and the inaccurate feature extraction due to coarse subdivision structure. Thus, it achieves the technical effect of accurately constructing a refined topological structure of the wound.

[0032] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: Based on the refined topology, calculate the local geometric deformation features point by point along the subdivided topological path at the wound edge, including the principal curvature, normal vector deflection angle, and Gaussian curvature gradient at each subdivided node, to obtain the feature vector sequence. Specifically, this includes: based on the refined topology, first defining the subdivided topological path at the wound edge, which extends continuously along the refined topological network at the wound edge, covering all subdivided nodes without omission; calculating the local geometric deformation features of each subdivided node along the path in the order from the wound's starting endpoint through key morphological turning points to the ending endpoint; calculating the principal... When calculating curvature, the maximum and minimum curvatures of the surface at each node are determined by analyzing the local morphology of the surface formed by multiple adjacent nodes around that node; these are the principal curvatures of that node. When calculating the normal vector deflection angle, the direction of the normal vector of each node is first obtained, and then the angle between the normal vector of the current node and the normal vector of the previous adjacent node is calculated to obtain the normal vector deflection angle. When calculating the Gaussian curvature gradient, the Gaussian curvature of each node is first solved, and then the difference between the Gaussian curvature of the current node and the Gaussian curvature of adjacent nodes is calculated. The ratio of this difference to the distance between nodes yields the Gaussian curvature gradient. The principal curvature, normal vector deflection angle, and Gaussian curvature gradient of each subdivided node are arranged sequentially according to the calculation order to form a feature vector sequence containing the complete local geometric deformation features of all nodes.

[0033] Step 3.2 involves performing temporal-series quantitative analysis on the feature vector sequence to identify the continuity and discontinuity abrupt change points of the wound edge deformation at the microscale. Based on morphological prior knowledge, a subset of effective deformation features characterizing the true morphology of the wound edge is selected, and pseudo-deformation interference caused by residual scanning noise is eliminated to obtain the effective deformation feature set. Specifically, this includes: performing temporal-series quantitative analysis on the acquired feature vector sequence, treating the feature vectors as temporal data and processing them segment by segment according to their order in the sequence; setting a deformation continuity judgment threshold of 0.4, calculating the similarity between two adjacent feature vectors, and determining that the deformation of the corresponding node is continuous if the similarity is higher than the threshold; otherwise, determining that the node is a deformation discontinuity of the wound edge at the microscale. Continuous mutation points are identified, and their location information is recorded one by one. Based on prior morphological knowledge of the natural laws governing wound edge deformation (i.e., true wound deformation should have a certain continuity and rationality, without irregular and drastic fluctuations), a false deformation threshold of 0.3 is set. Each feature vector in the feature vector sequence is evaluated. If the deformation feature corresponding to a certain feature vector deviates too much from the deformation trend of surrounding continuous feature vectors and exceeds the false deformation threshold, then the feature is determined to be false deformation interference caused by residual scanning noise. All feature vectors that conform to the deformation continuity law and do not exceed the false deformation threshold are selected to form an effective deformation feature subset representing the true edge morphology of the wound. False deformation feature vectors are removed from the sequence, ultimately yielding a pure set of effective deformation features.

[0034] Step 3.3: Based on the effective deformation feature set, construct the mapping relationship between the deformation features of the wound edge and the standard reference scale. By analyzing the relative deformation distribution law of the wound edge within the reference frame formed by the connection of the three wound edge vertices, calculate the proportional calibration parameters used to compensate for the scale distortion caused by the non-rigid deformation of the wound. The proportional calibration parameters include the linear scaling factor along the long axis of the wound and the non-uniform deformation compensation coefficient along the normal of the wound edge. Specifically, based on the effective deformation feature set, first construct the mapping relationship between the deformation features of the wound edge and the standard reference scale. Select a standard geometric model of known size as a reference. The topology of the standard model is consistent with the topology of the wound. Extract the deformation features of the standard model under different deformation states and establish a corresponding database between the standard deformation features and the standard scale. The effective deformation feature set of the wound is compared and matched with features in the standard database to determine the standard scale reference corresponding to the wound deformation features. Using a triangular region formed by connecting the three wound edge vertices as the reference frame, the relative deformation distribution of the wound edge within this frame is comprehensively analyzed, including the stretching, compression, and torsion of each part of the wound edge relative to the reference frame. For deformation along the long axis of the wound, the ratio of the actual long axis length of the wound to the standard long axis length within the reference frame is calculated to obtain a linear scaling factor along the long axis of the wound. This linear scaling factor is used to compensate for the overall scale distortion along the long axis. For deformation along the normal direction of the wound edge, the difference between the actual deformation and the standard deformation at each location is calculated segment by segment according to the deformation differences at different locations of the wound edge, forming a non-uniform deformation compensation coefficient along the normal direction of the wound edge, used to compensate for local scale distortion at different locations in the normal direction. Through the above process, a complete proportional calibration parameter including the linear scaling factor and the non-uniform deformation compensation coefficient is calculated.

[0035] In this embodiment of the invention, because the principal curvature, normal vector deflection angle, and Gaussian curvature change gradient are calculated point by point along the topological path of the wound edge according to the refined topological structure to obtain the feature vector sequence, the feature vector sequence is subjected to temporal quantitative analysis and the effective deformation feature set is screened based on morphological prior knowledge, and then the mapping relationship between the deformation features of the wound edge and the standard reference scale is constructed, and the proportional calibration parameters containing the linear scaling factor along the long axis of the wound and the non-uniform deformation compensation coefficient along the normal direction of the wound edge are calculated in combination with the relative deformation distribution law within the reference frame formed by the connection of the three wound edge vertices, the technical means of effectively overcoming the technical problems of traditional methods that are difficult to capture the micro-scale deformation and subtle topological features of the wound, are easily affected by noise interference leading to feature extraction distortion, cannot compensate for the scale distortion caused by the non-rigid deformation of the wound, and lack scientific proportional calibration basis, the technical effect of accurately obtaining the real and effective deformation features of the wound and obtaining reliable proportional calibration parameters is achieved.

[0036] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Based on the scaling parameters, perform spatial transformation on the standardized 3D trauma model; perform scale correction on the model along the long axis of the wound using a linear scaling factor; compensate for local deformation of the model along the wound edge normal using a non-uniform deformation compensation coefficient to eliminate scale distortion caused by non-rigid deformation; adjust the model to a standard anatomical posture to obtain a 3D trauma model after geometric posture normalization and scale standardization. Specifically, this includes: first, clarifying the linear scaling factor and the non-uniform deformation compensation coefficient along the wound edge normal in the scaling parameters, using these two sets of parameters as the core basis for the spatial transformation of the model; first, locate the long axis of the wound in the standardized 3D trauma model by identifying the spatial distribution of the three wound edge vertices in the model, determining the direction from the starting endpoint to the ending endpoint as the long axis of the wound. The model is scaldredged along its long axis using a linear scaling factor. By uniformly adjusting the coordinates of all relevant nodes in the model, the actual size of the wound along its long axis is made consistent with the standard scale. Next, for local deformation compensation along the wound edge normal, the wound edge is divided into 10 uniform segments, each corresponding to a non-uniform deformation compensation coefficient. Based on the compensation coefficient of each segment, the normal coordinates of the model nodes within the corresponding region are adjusted one by one to accurately compensate for scalar distortions caused by local stretching, compression, or twisting. After completing the scaling correction and deformation compensation, the model is adjusted to a standard anatomical posture, i.e., the wound surface is parallel to the horizontal plane, the long axis is horizontal, and key morphological turning points are facing upwards. This ensures that the model posture conforms to the unified observation standards for forensic identification, ultimately resulting in a 3D trauma model after geometric posture normalization and scalar standardization.

[0037] Step 4.2 involves automatically performing quantitative measurements on the processed 3D trauma model. These measurements include calculating the Euclidean distance between the two endpoints of the longest axis of the wound as the wound length, and calculating the surface area of ​​the wound mesh as the wound area. Optionally, the maximum wound depth and average wound edge width can be calculated as supplementary quantitative data to obtain the quantitative measurement results. Specifically, using the geometrically normalized and scale-standardized 3D trauma model as the measurement object, an automatic quantitative measurement program is initiated to achieve accurate multi-dimensional quantization of the wound, adapting to complex wound morphologies while ensuring measurement efficiency. When measuring the wound length, the surface area of ​​the wound is first extracted... The 3D coordinates of the nodes form a wound point set. This point set is recursively divided into two subsets using a divide-and-conquer strategy, with each subset containing no more than 50 nodes. A divide-and-conquer threshold of 50 nodes is set. The farthest pair of points within each subset is found, and the Euclidean distance between these two pairs of farthest points is calculated. Then, the farthest pair of points spanning the two subsets is found, i.e., a combination of nodes from one subset and nodes from the other subset, and their Euclidean distance is calculated. The distance values ​​of the three farthest pairs are compared, and the pair corresponding to the maximum value is selected. The Euclidean distance between this pair is the longest axis length of the wound, which is then determined as the wound length. The divide-and-conquer method reduces computation through recursive partitioning, avoiding the inefficiency of traversing all point pairs while ensuring that no truly farthest point pairs are missed, thus balancing measurement accuracy and speed. When measuring the wound area, the surface mesh data of the model wound is read. This mesh consists of multiple polygonal units. Each polygonal unit is first validated, eliminating invalid units with collinear vertices or those exceeding the wound boundary. For valid polygonal units, if it is a triangle, the area is directly calculated using the shoelace formula; if it is a quadrilateral, its diagonal is divided into two triangles, and the area of ​​each triangle is calculated using the shoelace formula and then summed. When calculating the area of ​​a single polygon, its vertex coordinates are recorded in clockwise order. The precise area of ​​a single unit is obtained by summing the products of the coordinate differences and taking half of the absolute value. This process is repeated for all valid polygonal units, accumulating their areas one by one, while simultaneously setting an area error correction threshold. The area is 0.003 square centimeters. Error correction is applied to the accumulated results to obtain the accurate total wound area. If it is necessary to measure the maximum wound depth, three non-collinear nodes in the undamaged area are selected from the model, and the spatial plane equation is solved as the reference plane for normal skin. Then, the vertical distance from each node on the wound surface to the reference plane is calculated to obtain the initial damage depth of each node. To eliminate the error caused by the fitting deviation of the reference plane, normal skin nodes within a 10 mm range of the reference plane are extracted, and a distance threshold of 10 mm is set to form a normal skin reference point set. For each wound node, the nearest point to all nodes in the reference point set is found by brute force, and the Euclidean distance between the two points is calculated as the corrected depth of the wound node. The corrected depth values ​​of all wound nodes are traversed, and outliers less than 0.02 mm are removed, with an outlier threshold of 0.0.02 mm, the maximum value among the remaining values ​​is selected as the maximum wound depth, and double calculation ensures the accuracy of depth measurement; when measuring the average wound edge width, the vertices of the wound edge contour are sorted: all boundary nodes of the wound edge are extracted, the centroid coordinates of the nodes are calculated, and all nodes are sorted clockwise with the centroid as the center to form an ordered wound edge contour polygon; along the sorted wound edge contour, 20 measurement points are evenly selected according to the principle of equal arc length, the arc length spacing between adjacent measurement points is 0.8 mm, and the arc length threshold is set to 0.8 mm; for each measurement point, the tangent direction of the wound edge is fitted by the coordinates of the two adjacent nodes, and a normal vector perpendicular to the tangent direction is constructed. A line segment is drawn along the normal vector towards the opposite edge of the wound. The opposite edge is considered a polyline composed of ordered nodes. The shortest distance from the measurement point to each line segment in this polyline is calculated; this distance represents the wound width at the corresponding measurement point. The average wound width is obtained by summing the wound width values ​​of 20 measurement points and averaging them. Clockwise sorting ensures a uniform distribution of measurement points, and the shortest distance from each point to a line segment accurately captures the wound width, avoiding deviations caused by local irregularities in the wound edge. Finally, wound length and wound area are used as core quantitative data, while maximum wound depth and average wound width are used as supplementary quantitative data, integrating them to form a complete quantitative measurement result suitable for complex wound morphologies.

[0038] Step 4.3: Annotate the quantified measurement results on the 3D model using a unified visualization style. The annotations should include at least the length values ​​and measurement paths, and the area values ​​and boundary ranges. The annotated model is then stored in association with the quantified data to form a quantified and annotated 3D model. Specifically, this includes: first, setting a unified visualization annotation style specification, where length values ​​are displayed in bold black font, size 12; measurement paths are annotated with solid red lines, 2 pixels wide; area values ​​are displayed in bold blue font, size 12; boundary ranges are annotated with dashed blue lines, 2 pixels wide; the maximum depth from the supplementary quantified data is annotated in bold green font, and the average edge width... The measurements are marked in bold orange font and placed at the corresponding measurement areas. Following this standard, the quantitative measurement results are marked one by one on the processed 3D model of the wound, ensuring that the length values ​​correspond one-to-one with the measurement path and the area values ​​correspond one-to-one with the boundary range. The marking positions of the supplementary quantitative data are accurately matched with the measurement areas. After the marking is completed, the association mapping relationship between the model and the quantitative data is established. The 3D model file is bound to a text file containing data such as wound length, wound area, maximum depth, and average wound edge width. At the same time, auxiliary information such as measurement time, calibration parameters, and measurement program version are recorded and stored together in the designated database to form the quantified and annotated 3D model.

[0039] In this embodiment of the invention, the technical means of using linear scaling factors in the proportional calibration parameters to correct the scale of the model wound along its long axis, compensating for local deformation along the wound edge normal using non-uniform deformation compensation coefficients to eliminate scale distortion caused by non-rigid deformation and adjusting the model to a standard anatomical posture, and automatically quantifying and measuring parameters such as wound length, area, selectable maximum depth, and average wound edge width on the processed model, and then labeling the length values ​​and measurement paths, area values ​​and boundary ranges in a unified visual style, and storing the labeled model and quantified data together, effectively overcomes the technical problems of traditional manual measurement being susceptible to the influence of angle and operation methods, insufficient accuracy, scale distortion and inconsistent posture of the model, single measurement parameters, chaotic labeling styles, and scattered data storage that is difficult to reuse. Thus, it achieves the technical effect of obtaining a geometrically normalized and scale-standardized 3D trauma model, realizing accurate automatic quantification and unified standardized labeling of wound parameters, and ensuring that the quantified data and model are linked and traceable.

[0040] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Register the quantized and annotated 3D model to a unified world coordinate system to obtain the registered 3D model. Using the geometric center of the wound surface in the registered 3D model as the origin, and establishing spatial coordinate axes based on the main plane and depth direction of the wound, a standardized spatial reference is created. This includes: first, acquiring the quantized and annotated 3D model; starting the coordinate system registration program; selecting a standard reference point for the world coordinate system, which is set based on internationally accepted 3D modeling coordinate standards; calculating the spatial positional deviations between multiple feature points in the model and the standard reference point; adjusting the translation, rotation, and scaling of the quantized and annotated 3D model to accurately register it to a unified world coordinate system to obtain the registered 3D model; then, calculating the geometric center of the wound surface; traversing the 3D coordinates of all nodes on the wound surface; calculating the average values ​​of the X-axis, Y-axis, and Z-axis coordinates; and using the spatial point corresponding to this average value as the origin. Next, establish the spatial coordinate axes with the main plane of the wound as the XY plane, where the direction of the longest axis of the wound is the X-axis, the direction perpendicular to the X-axis and within the main plane of the wound is the Y-axis, and the direction perpendicular to the main plane of the wound and pointing outwards from the skin is the Z-axis. A standardized spatial reference is established through the origin and the three coordinate axes.

[0041] Step 5.2: Based on the standardized spatial reference, configure the internal and external parameters of the virtual camera to set the shooting conditions; according to the three-dimensional spatial orientation of the wound in the standardized spatial reference, plan and generate a set of standardized shooting angles and corresponding camera poses covering key anatomical structures. Specifically, based on the established standardized spatial reference, configure the internal and external parameters of the virtual camera. The internal parameters are set as follows: focal length 50 mm, pixel size 3 micrometers, and image resolution 3000×2000 pixels to ensure that the image clarity and pixel accuracy meet the identification requirements; the external parameters are set as follows: shooting distance 30 cm, that is, the straight-line distance between the camera lens and the coordinate origin is fixed at 30 cm, and the initial values ​​of the camera's pitch angle and roll angle are set to 0 degrees. Based on the three-dimensional spatial orientation of the wound in the standardized spatial reference, a set of standardized shooting angles covering key anatomical structures is planned and generated, including four core angles: frontal view, side view, 45-degree oblique view, and top view. For each angle, the corresponding camera pose parameters are calculated, including the camera's spatial coordinates, pitch angle, roll angle, and yaw angle, to ensure that each angle can clearly present the key anatomical features of the wound, forming a complete set of camera poses.

[0042] Step 5.3: Based on the planned camera poses, drive the virtual camera to perform multi-view rendering and shooting of the registered 3D model; project the model geometry and associated 3D visualization annotation information onto the 2D imaging plane to generate a set of original view images containing wound morphology, physical reference points, and quantized annotation information. Specifically, this includes: according to the planned set of camera poses, sequentially drive the virtual camera to switch to each preset pose, and perform multi-view rendering and shooting of the registered 3D model. During the rendering process, a ray tracing algorithm is used to simulate the imaging principle of a real camera, accurately reproducing the texture, lighting, and 3D shape of the original wound surface. Simultaneously, the associated 3D visualization annotation information of the model is included... The process includes length measurement paths, area boundary ranges, and supplementary quantitative data annotations, all of which are simultaneously projected onto a two-dimensional imaging plane to ensure that the annotation information and wound morphology accurately correspond in the two-dimensional image. To provide a physical reference, standard reference markers of known dimensions are added to the edge area of ​​the three-dimensional model. These markers are cubes with a side length of 1 cm, and their three-dimensional coordinates are fixed at a specific position on a standardized spatial benchmark. After projection, they form clear physical reference points in the two-dimensional image. Through the rendering projection process, a set of original viewpoint images containing the complete wound morphology, standard physical reference points, and all quantitative annotation information are generated. Each shooting viewpoint corresponds to one original image, ensuring that the image information is comprehensive and without omission.

[0043] Step 5.4 involves standardizing the original viewpoint images, uniformly adjusting their brightness and contrast, and embedding caption information containing case identification, shooting angle, and key quantitative data. The final output is a set of two-dimensional image sequences that conform to the preset imaging standards. Specifically, this includes: starting the image standardization post-processing workflow, firstly uniformly adjusting the brightness and contrast of all original viewpoint images, adjusting the average brightness value of each image to 120 (range 0-255), and adjusting the contrast to 1.2 times. By calculating and correcting the brightness and grayscale values ​​pixel by pixel, the brightness differences caused by light and shadow simulation during imaging from different viewpoints are eliminated, ensuring that the visual effect of all images is consistent. Next, standardized image captions are embedded. The case identifier uses a 16-digit numerical code, including the year, region code, and case number, and is marked in the upper left corner of the image. The shooting perspective is labeled with text, such as frontal view and side view, and is marked in the upper right corner of the image. Key quantitative data includes specific values ​​for wound length and wound area, and supplementary quantitative data includes maximum depth and average wound edge width, marked in the lower left corner of the image. All image captions use bold white font, size 14, and a semi-transparent black mask is added to the background to ensure readability. Finally, the processed images are standardized in format and uniformly saved as lossless PNG format, with a resolution of 3000×2000 pixels. The final output is a set of two-dimensional image sequences that meet the preset imaging standards, have complete information, uniform format, and can be directly used for subsequent injury analysis.

[0044] In this embodiment of the invention, by registering the quantified and annotated 3D model to a unified world coordinate system, establishing a standardized spatial reference with the geometric center of the wound surface as the origin and based on the main plane and depth direction of the wound surface, configuring the internal and external parameters of the virtual camera based on this reference, and planning a standardized shooting angle and camera pose covering key anatomical structures, the virtual camera is driven to render and shoot from multiple perspectives, and the model geometry and 3D annotation information are projected to generate an original image containing wound morphology, physical reference points, and quantified annotations. Then, the brightness and contrast of the original image are uniformly adjusted, and case identification, shooting angle, and key quantitative data caption information are embedded. Therefore, this invention effectively overcomes the technical problems of traditional 2D shooting coordinate system inconsistency, arbitrary perspective, lack of standardized image specifications, incomplete presentation of wound morphology and quantitative information, difficulty in comparing images of different cases, and poor traceability due to lack of key related information. Thus, it achieves the technical effect of outputting a set of standardized, complete, traceable, and pre-defined imaging standard 2D image sequences.

[0045] In a preferred embodiment of the present invention, step 6 above may include: Step 6.1 involves loading the standardized two-dimensional image sequence and its associated quantized annotation data as multimodal input into a large-scale visual language model that has been fine-tuned using forensic injury data. Specifically, this includes: first, organizing the standardized two-dimensional image sequence and its associated quantized annotation data, ensuring the two-dimensional image sequence is uniformly formatted as lossless PNG with a resolution of 3000×2000 pixels; the quantized annotation data includes core and supplementary parameters such as wound length, wound area, maximum depth, and average wound edge width, along with basic case information such as the injured person's gender, age, and wound location; the large-scale visual language model used is optimized based on a general visual language pre-trained model and has been tested with large-scale forensic injury data. The data was fine-tuned, encompassing 15,000 real forensic injury cases with different wound types (sharp instrument wounds, blunt instrument wounds, abrasions, etc.) and different healing stages (0-24 hours, 24-72 hours, etc.) and different degrees of injury. Each case included standardized two-dimensional image quantitative annotation information and authoritative identification conclusions. The standardized two-dimensional image sequence was arranged according to the shooting perspective, and a one-to-one correspondence mapping was established between the quantitative annotation data and the corresponding images. The data was uniformly loaded into the visual language big model in a multimodal data format. During the loading process, the data integrity was checked. If any image was missing or the data was abnormal, it was automatically marked and the processing was paused, prompting staff to check and supplement the data to ensure the accuracy and completeness of the input data.

[0046] The large-scale visual language model used employs a four-module architecture: a visual encoder, a text encoder, a cross-modal fusion module, and an output adaptation layer. The first module, the visual encoder, is built based on VisionTransformer and contains 12 encoder layers. It transforms two-dimensional images into sequential feature vectors through 16×16 pixel image patching, while embedding positional encoding and global image features to enhance the extraction of visual features such as wound morphology, physical reference points, and quantized annotation visualization information. The second module, the text encoder, uses a 6-layer Transformer structure for quantized annotation data. The first module is a dedicated vocabulary and text encoding rules for basic case information, which transforms numerical data and categorical information into structured text feature vectors to ensure that machines can accurately understand the semantics of the data. The second module is a cross-modal fusion module, built on a multi-head attention mechanism with eight attention heads. By calculating the mutual information weights of visual and text features, it achieves deep alignment and fusion of the two types of features, eliminating the information fragmentation caused by modal differences. The third module is an output adaptation layer, which contains two fully connected network layers. It fine-tunes and optimizes parameters using forensic injury data to adapt the model output to specific task requirements such as trauma time determination and organic injury identification.

[0047] Step 6.2: Based on the loaded multimodal input, the visual language large model extracts visual features from the two-dimensional image and quantitative features from the quantized labeled data in parallel, and fuses the visual features and quantized features to obtain a comprehensive feature vector containing wound color, shape, tissue state, and geometric dimensions. Specifically, the visual language large model initiates multimodal feature extraction, and performs parallel extraction of visual features from the two-dimensional image and quantized features from the quantized labeled data. When extracting visual features from the two-dimensional image, each image in the image sequence is analyzed at the pixel level to identify and extract wound color features such as color type (red, dark red, purplish red, etc.), color depth, and color uniformity; wound shape features such as wound contour regularity, edge neatness, and whether there are branches or depressions; and tissue state features such as the degree of swelling of the skin around the wound, whether there is bleeding or exudation, and tissue loss. The model extracts visual features such as the wound condition, the growth status of healing tissue, etc., and also extracts auxiliary visual features such as the ratio between physical reference points and the wound in the image. When extracting quantitatively labeled data features, it directly reads specific values ​​such as wound length, wound area, maximum depth, and average wound edge width, and converts the values ​​into a feature vector form that the model can recognize. At the same time, it calculates the relative proportion of each quantitative parameter, such as the ratio of depth to length, as supplementary quantitative features. In the feature fusion stage, an attention mechanism is used to assign weights to visual features and quantitative features respectively, with visual features accounting for 0.45 and quantitative features accounting for 0.55. The two types of features are deeply integrated through a hierarchical feature fusion network to eliminate feature redundancy, strengthen key information, and finally form a comprehensive feature vector that includes wound color, morphology, tissue status, geometric dimensions, and various feature relationships.

[0048] Step 6.3: By integrating the feature vector, the visual language big data model calculates and determines the time interval to which the injury formation time belongs based on the internally pre-learned wound healing features and time mapping relationship, and obtains the confidence level of the determination result. Specifically, this includes: a database of wound healing features and time mapping relationship constructed based on a large number of forensic injury cases. This database records in detail the feature changes of different types of wounds under different environmental conditions, from injury occurrence to healing, corresponding to each time interval. For example, from 0 to 6 hours, the wound color is bright red with obvious bleeding and tissue swelling; from 6 to 12 hours, the color is slightly darker, bleeding decreases, and swelling continues; from 12 to 24 hours, the color turns dark red and swelling gradually subsides. Each time interval corresponds to a clear feature combination. After receiving the integrated feature vector, the model... The model compares and matches the standard feature vectors of each time interval in the database one by one to calculate the similarity value. A similarity threshold of 0.75 is set for time determination. If the similarity between the comprehensive feature vector and the standard feature vector of a certain time interval is higher than this threshold, the trauma formation time is initially determined to belong to that interval. Then, the model combines factors such as the quantity distribution feature vector matching details within the case data of that time interval to calculate the confidence level of the determination result. The confidence level calculation uses a weighted summation method, with similarity accounting for 0.6, case data support accounting for 0.3, and detail matching accounting for 0.1. The final output confidence level value is rounded to two decimal places. If the confidence level is lower than 0.6, it is marked in the result as requiring further verification in conjunction with clinical medical records to ensure the reliability of the time determination.

[0049] Step 6.4: Based on the time interval determination, the visual language model further identifies and classifies organic injuries based on information related to tissue depth and injury extent in the comprehensive feature vector. Finally, it outputs a structured identification report containing the time interval determination, organic injury identification conclusion, and corresponding confidence level. Specifically, after determining the time interval of trauma formation, the visual language model initiates the organic injury identification and classification process based on information related to tissue depth and injury extent in the comprehensive feature vector. First, it extracts features representing tissue depth from the comprehensive feature vector, such as the maximum wound depth and the degree of tissue damage at the wound edge, and features representing the injury extent, such as the wound area and the anatomical structures involved. It sets tissue depth thresholds: 0.5 cm for superficial injuries, 1.5 cm for intermediate injuries, and 3 cm for deep injuries. Based on the comparison between the maximum wound depth and the thresholds, it initially classifies the injury depth level. Simultaneously, it combines tissue state information from visual features, such as whether subcutaneous fat is involved. The model precisely defines the extent of damage by examining whether there is neurovascular damage in the adipose tissue, muscle tissue, or bone. It predefines an organic injury classification system, including categories such as epidermal injury, dermal injury, subcutaneous tissue injury, muscle injury, bone injury, and neurovascular injury, each with specific feature criteria. The extracted features related to tissue depth and injury extent are compared and matched with the criteria for each category. If all criteria for a category are met and the similarity is higher than 0.8, that category is used as the identification result. Finally, the model integrates the trauma formation time interval determination results, corresponding confidence levels, organic injury identification classification results, and corresponding confidence levels, along with key feature matching basis, case data support, and a description of the model's determination process, generating a structured identification report. The report has a standardized format, divided into modules such as basic case information, identification data sources, feature extraction results, time determination conclusions, organic injury conclusions, confidence levels, explanations, and review suggestions, ensuring comprehensive, logically clear, and traceable report content that can be directly used as evidence in judicial practice.

[0050] In this embodiment of the invention, a standardized two-dimensional image sequence and associated quantized labeled data are used as multimodal input. This input is loaded into a large visual language model fine-tuned from forensic injury data. The model extracts visual features from the two-dimensional images and quantized features from the quantized labeled data in parallel, fusing them to form a comprehensive feature vector containing wound color, shape, tissue state, and geometric dimensions. Then, based on pre-learned wound healing features and time mapping relationships, the time interval and confidence level of the injury formation are determined. Combining the tissue depth and damage extent information in the comprehensive feature vector, organic injury identification and classification are completed. Finally, a judgment conclusion and corresponding settings are output. The structured forensic report technology effectively overcomes the technical problems of traditional forensic injury identification, such as the separation of visual features and quantitative data, reliance on human experience leading to strong subjectivity, lack of standardized analysis mechanisms for determining the time of injury formation and organic damage, low reliability of results, and non-standardized reporting. It achieves the technical effect of deep integration of multimodal features and intelligent injury identification reasoning, improving the accuracy, objectivity, and reliability of determining the time of injury formation and identifying organic damage. The structured report also ensures the standardization and traceability of the injury identification results, fully meeting the requirements of judicial practice for the rigor of evidence.

[0051] like Figure 2 As shown, embodiments of the present invention also provide a forensic injury analysis system based on three-dimensional modeling and a large visual language model, including: The acquisition module is used to acquire the original 3D model of the trauma site using a structured light scanner and to preprocess the original 3D model to obtain a standardized 3D model of the trauma. The segmentation module is used to identify the three wound edge vertices in the standardized 3D trauma model and establish a baseline topological relationship. Based on the deep feature extraction mechanism of the convolutional neural network, the baseline topological relationship is further subdivided into multi-level structures to obtain the subdivided topological structure. The extraction module is used to perform in-depth extraction and quantitative analysis of wound deformation features on the subdivided topology, accurately capture the feature information of small deformations and fine topology of the wound, and calculate the proportional calibration parameters by combining the feature information obtained from the extraction and analysis. The quantization module is used to perform geometric pose normalization and scale standardization on the trauma 3D model according to the proportional calibration parameters to obtain the processed trauma 3D model; on the processed trauma 3D model, the wound is quantitatively measured and labeled to obtain the wound length and area data, and the quantized and labeled 3D model is obtained. The shooting module is used to take standardized multi-view shots in the world coordinate system based on the quantized and annotated 3D model, using a virtual camera to obtain a 2D image containing wound morphology, physical reference points and quantified annotation information. The processing module is used to input two-dimensional images and associated quantitatively labeled data into a large visual language model that has been fine-tuned by forensic injury data, extract the visual and quantitative features of the wound, and determine the time interval of the injury formation and identify organic damage based on the extracted features.

[0052] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A forensic injury analysis method based on 3D modeling and large-scale visual language models, characterized in that: The method includes: The original 3D model of the trauma site was acquired using a structured light scanner, and the original 3D model was preprocessed to obtain a standardized 3D model of the trauma. By identifying the three wound edge vertices in the standardized 3D trauma model and establishing a baseline topological relationship, the baseline topological relationship is further subdivided into multiple levels based on the deep feature extraction mechanism of the convolutional neural network, resulting in the subdivided topological structure. By performing in-depth extraction and quantitative analysis of the wound deformation features of the subdivided topology, the feature information of the wound's minute deformation and fine topology is accurately captured, and the proportional calibration parameters are calculated by combining the feature information obtained from the extraction and analysis. Based on the proportional calibration parameters, the 3D trauma model is subjected to geometric pose normalization and scale standardization to obtain the processed 3D trauma model; the wound is quantitatively measured and labeled on the processed 3D trauma model to obtain the wound length and area data, thus obtaining the quantitatively labeled 3D model. Based on the quantified and annotated 3D model, a virtual camera is used to take standardized multi-view shots in the world coordinate system to obtain a 2D image containing wound morphology, physical reference points and quantified annotation information. Two-dimensional images and associated quantitatively labeled data are input into a large visual language model fine-tuned by forensic injury data to extract visual and quantitative features of the wound. Based on the extracted features, the time interval of the injury formation and the identification of organic damage are completed.

2. The forensic injury analysis method based on three-dimensional modeling and large-scale visual language model according to claim 1, characterized in that, The original 3D model of the trauma site was acquired using a structured light scanner. The original 3D model was then preprocessed to eliminate scanning noise and repair data defects, resulting in a standardized 3D model of the trauma, including: Receive the original 3D model uploaded by the structured light scanner, and separate and extract the point cloud data and the initial mesh model from the original 3D model; By performing statistical filtering on the point cloud data, the mean and standard deviation of the distance of each point in the neighborhood are calculated, and outliers that deviate from the mean by more than a preset multiple of the standard deviation are removed, thus obtaining the point cloud data after statistical filtering. By performing radius filtering on the point cloud data, the number of neighboring points of each point cloud point is counted within the set search radius, and isolated points with fewer than the preset threshold of neighboring points are removed to obtain clean point cloud data after noise reduction. The surface mesh is reconstructed based on pure point cloud data. The reconstructed mesh model is then subjected to Laplacian mesh smoothing to eliminate burrs and jagged distortions on the mesh surface, resulting in a smoothed mesh model. The topological void regions in the smooth mesh model are detected, and the void regions are filled with mesh to obtain a complete, continuous, and standardized 3D trauma model without missing data.

3. The forensic injury analysis method based on three-dimensional modeling and large-scale visual language model according to claim 2, characterized in that, By identifying the three vertices of the wound edge in a standardized 3D trauma model and establishing a baseline topological relationship, multi-level structural subdivision of the baseline topological relationship is carried out based on the deep feature extraction mechanism of a convolutional neural network, resulting in a subdivided topological structure, including: Based on the standardized three-dimensional model of trauma, and based on the curvature change characteristics of the wound and the surrounding normal skin, as well as prior knowledge of forensic anatomy, three wound edge vertices with geometric features are automatically located and identified at the edge of the wound. The three wound edge vertices correspond to the starting point, ending point and key point of morphological transition of the wound, respectively, and the vertex set of the spatial coordinates of the three wound edge vertices is obtained. Based on the vertex set, the initial topological skeleton of the wound edge is constructed with three wound edge vertices as the reference. By generating initial connecting edges between the three vertices, a reference topological relationship representing the overall contour of the wound is formed, and the reference topological data is obtained. The baseline topological data is mapped to a two-dimensional parameterized grid and input into a pre-trained convolutional neural network feature extractor. The local geometric features and global morphological features of the topological structure are extracted step by step through the convolutional layers to obtain a feature map of multi-level feature responses. Based on multi-level feature maps, the system adaptively identifies geometrically discontinuous and morphologically changing regions at the wound edge. It then adaptively subdivides the baseline topological relationships based on feature response intensity, increasing the subdivision density in regions with drastic curvature changes and complex morphologies, while maintaining sparse connections in smooth regions, thus obtaining a refined topological structure.

4. The forensic injury analysis method based on three-dimensional modeling and large-scale visual language model according to claim 3, characterized in that, By performing in-depth extraction and quantitative analysis of wound deformation features from the subdivided topology, the characteristic information of minute wound deformations and subtle topological structures is accurately captured. The proportional calibration parameters are then calculated based on the extracted and analyzed feature information, including: Based on the refined topology, the local geometric deformation features are calculated point by point along the subdivided topological path at the edge of the wound, including the principal curvature, normal vector deflection angle and Gaussian curvature gradient at each subdivided node, to obtain the feature vector sequence. By performing temporal and quantitative analysis on the feature vector sequence, the continuity and discontinuity of the wound edge deformation at the microscale are identified. Based on morphological prior knowledge, an effective deformation feature subset representing the true edge morphology of the wound is selected, and pseudo-deformation interference caused by residual scanning noise is eliminated to obtain the effective deformation feature set. Based on the effective deformation feature set, a mapping relationship between the deformation features of the wound edge and the standard reference scale is constructed. By analyzing the relative deformation distribution law of the wound edge within the reference frame formed by the connection of the three wound edge vertices, the proportional calibration parameters used to compensate for the scale distortion caused by the non-rigid deformation of the wound are calculated. The proportional calibration parameters include the linear scaling factor along the long axis of the wound and the non-uniform deformation compensation coefficient along the normal of the wound edge.

5. The forensic injury analysis method based on three-dimensional modeling and large-scale visual language model according to claim 4, characterized in that, The trauma 3D model is subjected to geometric pose normalization and scale standardization based on the proportional calibration parameters to obtain the processed trauma 3D model. The wound is then quantitatively measured and labeled on the processed trauma 3D model to obtain the wound length and area data, resulting in a quantitatively labeled 3D model, including: Based on the proportional calibration parameters, the standardized 3D trauma model is spatially transformed; the model is scaled in the long axis direction of the wound by the linear scaling factor; the local deformation of the model is compensated along the normal direction of the wound edge by the non-uniform deformation compensation coefficient to eliminate the scale distortion caused by non-rigid deformation, and the model is adjusted to the standard anatomical posture to obtain the 3D trauma model after geometric posture normalization and scale standardization. Quantitative measurements are automatically performed on the processed 3D model of the wound. The measurements include calculating the Euclidean distance between the two ends of the longest axis of the wound as the wound length and calculating the surface area of ​​the wound grid as the wound area. Optionally, the maximum wound depth and average wound edge width can be calculated as supplementary quantitative data to obtain the quantitative measurement results. The quantitative measurement results are annotated on the 3D model in a unified visualization style; the annotation content includes at least the length value and measurement path, and the area value and boundary range; the annotated model is stored in association with the quantitative data to form a quantitatively annotated 3D model.

6. The forensic injury analysis method based on three-dimensional modeling and large-scale visual language model according to claim 5, characterized in that, Based on the quantified and annotated 3D model, multi-view standardized photography is performed using a virtual camera in the world coordinate system to obtain 2D images containing wound morphology, physical reference points, and quantified annotation information, including: The quantized and annotated 3D model is registered to a unified world coordinate system to obtain the registered 3D model. The geometric center of the wound surface in the registered 3D model is used as the origin of the coordinate system, and the spatial coordinate axes are established based on the main plane of the wound and the depth direction to establish a standardized spatial reference. Based on a standardized spatial reference, the internal and external parameters of the virtual camera are configured to set the shooting conditions; according to the three-dimensional spatial orientation of the wound in the standardized spatial reference, a set of standardized shooting angles and corresponding camera poses covering key anatomical structures are planned and generated. Based on the planned camera pose, the virtual camera is driven to perform multi-view rendering and shooting of the registered 3D model; the model geometry and associated 3D visualization annotation information are projected together onto the 2D imaging plane to generate a set of original view images containing wound morphology, physical reference points and quantitative annotation information. The original perspective image is standardized and post-processed to uniformly adjust its brightness and contrast, and embeds image caption information containing case identification, shooting perspective and key quantitative data, and finally outputs a set of two-dimensional image sequences that meet the preset imaging standards.

7. The forensic injury analysis method based on three-dimensional modeling and large-scale visual language model according to claim 6, characterized in that, Two-dimensional images and associated quantitatively labeled data are input into a large-scale visual language model fine-tuned from forensic injury data. Visual and quantitative features of the wound are extracted, and based on these extracted features, the time interval of the injury formation and the identification of organic damage are determined, including: Standardized two-dimensional image sequences and their associated quantized annotation data are loaded as multimodal inputs into a large visual language model that has been fine-tuned using forensic injury data. Based on the loaded multimodal input, the visual language large model extracts visual features from two-dimensional images and quantitative features from quantized labeled data in parallel, and fuses the visual features and quantitative features to obtain a comprehensive feature vector containing wound color, shape, tissue state and geometric size. By integrating feature vectors, the visual language big model calculates and determines the time interval to which the trauma formation time belongs based on the internally pre-learned trauma healing features and time mapping relationship, and obtains the confidence level of the determination result. Based on the time interval determination, the visual language big data model further identifies and classifies organic damage by using information related to tissue depth and damage range in the comprehensive feature vector. Finally, it outputs a structured identification report that includes the time interval determination, the organic damage identification conclusion, and the corresponding confidence level.

8. A forensic injury analysis system based on 3D modeling and visual language large model, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to acquire the original 3D model of the trauma site using a structured light scanner and to preprocess the original 3D model to obtain a standardized 3D model of the trauma. The segmentation module is used to identify the three wound edge vertices in the standardized 3D trauma model and establish a baseline topological relationship. Based on the deep feature extraction mechanism of the convolutional neural network, the baseline topological relationship is further subdivided into multi-level structures to obtain the subdivided topological structure. The extraction module is used to perform in-depth extraction and quantitative analysis of wound deformation features on the subdivided topology, accurately capture the feature information of small deformations and fine topology of the wound, and calculate the proportional calibration parameters by combining the feature information obtained from the extraction and analysis. The quantization module is used to perform geometric pose normalization and scale standardization on the trauma 3D model according to the proportional calibration parameters to obtain the processed trauma 3D model; on the processed trauma 3D model, the wound is quantitatively measured and labeled to obtain the wound length and area data, and the quantized and labeled 3D model is obtained. The shooting module is used to take standardized multi-view shots in the world coordinate system based on the quantized and annotated 3D model, using a virtual camera to obtain a 2D image containing wound morphology, physical reference points and quantified annotation information. The processing module is used to input two-dimensional images and associated quantitatively labeled data into a large visual language model that has been fine-tuned by forensic injury data, extract the visual and quantitative features of the wound, and determine the time interval of the injury formation and identify organic damage based on the extracted features.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images