Fossil pattern recognition and classification method and fossil pattern recognition and classification system based on image recognition

Abstract

The invention discloses a fossil pattern recognition and classification method and system based on image recognition, and relates to the technical field of hydraulic engineering, and the method comprises the steps: obtaining three-dimensional point cloud data of a fossil sample, projecting the three-dimensional point cloud data into a multi-view two-dimensional texture mapping graph, and obtaining a surface element distribution graph of the fossil sample; utilizing an improved GLCM gray level co-occurrence matrix algorithm to calculate the contrast ratio, the correlation and the entropy value of fossil textures, and calculating the curvature and the normal vector distribution of fossils in the three-dimensional point cloud data and geometric parameters of biological parts; analyzing the surface element distribution diagram through a PCA (principal component analysis) method, and extracting a significant element combination related to species; and mapping the feature data to a unified feature space based on a cross-modal attention mechanism, distributing contribution degrees of all modals through learnable weights, and outputting a fossil pattern classification result. Therefore, a large number of dinosaur fossil samples can be quickly and accurately classified.

Description

Technical Field

[0001] The present invention relates to the technical field of water conservancy engineering, and in particular to a fossil pattern recognition and classification method and system based on image recognition. Background Art

[0002] Traditional identification of dinosaur fossils relies primarily on manual identification by paleontologists, who classify them by observing their morphological characteristics and comparing them with existing specimens. However, this method has numerous limitations: it is significantly affected by the integrity and preservation of the fossil, making it difficult to accurately identify broken or severely weathered fossils based on morphological features. Manual identification is also inefficient, requiring significant time and effort. Furthermore, the results are influenced by the subjective experience of the experts, resulting in certain errors and low identification efficiency. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and to design a fossil pattern recognition and classification method and system based on image recognition.

[0004] To achieve the above object, the technical solution of the present invention is, further, in the above-mentioned fossil pattern recognition and classification method based on image recognition, the fossil pattern recognition and classification method comprises the following steps: Acquire three-dimensional point cloud data of the fossil sample, project the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously acquire a surface element distribution map of the fossil sample; The improved GLCM gray-level co-occurrence matrix algorithm is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map, and the LBP local binary pattern is introduced to enhance the sensitivity of edge details to obtain texture feature data; Calculating the curvature, normal vector distribution, and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data; The surface element distribution map was analyzed by PCA principal component analysis, and significant element combinations related to species were extracted to obtain element characteristic data. Based on the cross-modal attention mechanism, texture feature data, geometric feature data and element feature data are mapped to a unified feature space, and the contribution of each modality is distributed through learnable weights to output the fossil pattern classification results.

[0005] Furthermore, in the above-mentioned fossil pattern recognition and classification method based on image recognition, the three-dimensional point cloud data of the fossil sample is obtained, and the three-dimensional point cloud data is projected into a multi-view two-dimensional texture map, and the surface element distribution map of the fossil sample is obtained at the same time, including: A high-precision laser scanner is used to perform an all-around scan of the dinosaur fossil sample to obtain raw point cloud data. The raw point cloud data is subjected to denoising to remove noise points caused by interference from the scanning environment and equipment errors. Missing point cloud data is filled in based on an interpolation algorithm of adjacent point cloud data to obtain three-dimensional point cloud data. Projecting the processed three-dimensional point cloud data at different viewing angles to generate a multi-view two-dimensional texture map, including at least a front view, a side view, and a top view; The elemental composition of the fossil sample surface was analyzed using an X-ray fluorescence spectrometer. The element types and content data in each grid were collected in units of 1 cm, including at least the concentrations of Ca, P, Sr, and Ti, to obtain a surface element distribution map.

[0006] Furthermore, in the above-mentioned fossil pattern recognition and classification method based on image recognition, the improved GLCM gray-level co-occurrence matrix algorithm is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map, and the LBP local binary pattern is introduced to enhance the sensitivity of edge details to obtain texture feature data, including: Based on the GLCM gray-level co-occurrence matrix algorithm, a weighting mechanism is introduced to assign different weights according to the importance of different areas in the image, highlighting the characteristics of key texture areas, and obtaining an improved GLCM gray-level co-occurrence matrix algorithm; The multi-view two-dimensional texture map is converted into a grayscale image, the offset and grayscale levels of the grayscale co-occurrence matrix are set, the contrast, correlation and entropy values ​​are calculated, and feature data are obtained.

[0007] Furthermore, in the above-mentioned fossil pattern recognition and classification method based on image recognition, the improved GLCM gray-level co-occurrence matrix algorithm is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map, and the LBP local binary pattern is introduced to enhance the sensitivity of edge details to obtain texture feature data, which also includes: The LBP algorithm is applied to the grayscale two-dimensional texture map, and the sensitivity of the image edge details is enhanced by setting different radius and neighborhood point number parameters. A local binary pattern code is generated for each pixel in the image, and the binary pattern code is statistically analyzed to obtain an LBP feature histogram. The LBP feature histogram is fused with feature data to obtain texture feature data.

[0008] Furthermore, in the above-mentioned fossil pattern recognition and classification method based on image recognition, the calculation of the curvature, normal vector distribution and geometric parameters of the biological parts of the fossil in the three-dimensional point cloud data to obtain geometric feature data includes: Fitting a local surface of the three-dimensional point cloud data based on a least squares method to obtain the Gaussian curvature and the mean curvature of each point, and obtaining a normal vector of the plane; Normalize the calculated curvature and normal vector data so that their value range is unified between [0,1] to obtain normalized data; Different biological parts are identified in the three-dimensional point cloud data, and the measurement tools of the point cloud processing software are used to calculate the geometric parameters of each biological part to obtain geometric feature data.

[0009] Furthermore, in the above-mentioned fossil pattern recognition and classification method based on image recognition, the surface element distribution map is analyzed by PCA principal component analysis to extract significant element combinations related to species and obtain element feature data, including: Standardizing the surface element distribution map, calculating the covariance matrix, eigenvalues ​​and eigenvectors of the data, and selecting the principal component according to the size of the eigenvalues; The principal components with a cumulative contribution rate of 89% and the significant element combinations related to species were extracted to obtain element characteristic data.

[0010] Furthermore, in the above-mentioned fossil pattern recognition and classification method based on image recognition, the texture feature data, geometric feature data, and element feature data are mapped to a unified feature space based on the cross-modal attention mechanism, and the contribution of each modality is distributed through learnable weights to output the fossil pattern classification result, including: The MLP multi-layer perceptron is used to map each modal data into a feature space of the same dimension; the softmax function is used to process the fused feature data and convert it into a probability distribution belonging to different dinosaur genera and species. Based on the size of the probability value, the genus with the highest probability is selected as the fossil pattern classification result.

[0011] Furthermore, in a fossil pattern recognition and classification system based on image recognition, the fossil pattern recognition and classification system includes the following modules: A fossil data acquisition module is used to acquire three-dimensional point cloud data of fossil samples, project the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously acquire a surface element distribution map of the fossil samples; A texture feature extraction module is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map using an improved GLCM gray-level co-occurrence matrix algorithm, introduce LBP local binary pattern to enhance edge detail sensitivity, and obtain texture feature data; A geometric feature extraction module is used to calculate the curvature, normal vector distribution and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data; The element feature extraction module is used to analyze the surface element distribution map through the PCA principal component analysis method, extract the significant element combinations related to the species, and obtain element feature data; The fossil pattern classification module is used to map texture feature data, geometric feature data, and element feature data into a unified feature space based on a cross-modal attention mechanism, and to distribute the contribution of each modality through learnable weights to output the fossil pattern classification results.

[0012] Furthermore, in a fossil pattern recognition and classification system based on image recognition, the texture feature extraction module includes the following submodules: The weighted submodule is used to introduce a weighting mechanism based on the GLCM gray-level co-occurrence matrix algorithm, assign different weights according to the importance of different areas in the image, highlight the characteristics of key texture areas, and obtain an improved GLCM gray-level co-occurrence matrix algorithm; The calculation submodule is used to convert the multi-view two-dimensional texture map into a grayscale image, set the offset and grayscale level of the grayscale co-occurrence matrix, calculate the contrast, correlation and entropy value, and obtain feature data.

[0013] Furthermore, in a fossil pattern recognition and classification system based on image recognition, the texture feature extraction module further includes the following submodules: The enhancement submodule is used to apply the LBP algorithm to the grayscaled two-dimensional texture map, and enhance the sensitivity of the image edge details by setting different radius and neighborhood point number parameters; The fusion submodule is used to generate a local binary pattern code for each pixel in the image, perform statistical analysis on the binary pattern code to obtain an LBP feature histogram, and fuse the LBP feature histogram with the feature data to obtain texture feature data.

[0014] Its beneficial effects are as follows: obtaining three-dimensional point cloud data of fossil samples, projecting the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously obtaining a surface element distribution map of the fossil samples; using the improved GLCM gray-level co-occurrence matrix algorithm to calculate the contrast, correlation, and entropy of the fossil texture in the multi-view two-dimensional texture map, introducing the LBP local binary mode to enhance the sensitivity of edge details, and obtaining texture feature data; calculating the curvature, normal vector distribution, and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data; analyzing the surface element distribution map through the PCA principal component analysis method, extracting significant element combinations related to the species, and obtaining element feature data; mapping the texture feature data, geometric feature data, and element feature data into a unified feature space based on the cross-modal attention mechanism, and distributing the contribution of each modality through learnable weights to output the fossil pattern classification results. 1. Texture features reflect the details of the lines on the fossil surface, geometric features reflect the shape structure of the biological parts, and element features reveal the material composition of the fossil. The three complement each other. 1. Based on the cross-modal attention mechanism, different modal data are mapped to a unified feature space and assigned learnable weights. This can fully utilize the advantages of each modality, significantly improve the accuracy of genus and species identification, and reduce the risk of misjudgment due to incomplete data. 2. The use of scientific curvature and normal vector calculation methods and the measurement of geometric parameters of biological parts can accurately quantify the three-dimensional morphological characteristics of fossils. In terms of element feature extraction, PCA principal component analysis can extract key information from complex element distribution data. The application of these improved algorithms has greatly enhanced the ability to extract fossil features. 3. Through programming and deep learning frameworks, various features are automatically calculated and classified, reducing manual intervention. Compared with traditional manual identification, this greatly shortens the research cycle and reduces labor costs. At the same time, it avoids the influence of subjective errors and fatigue factors in manual operations, allowing a large number of dinosaur fossil samples to be quickly and accurately classified. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0016] Figure 1 Schematic diagram of a first embodiment of a fossil pattern recognition and classification method based on image recognition in an embodiment of the present invention; Figure 2 Schematic diagram of a second embodiment of a fossil pattern recognition and classification method based on image recognition in an embodiment of the present invention; Figure 3 Schematic diagram of a first embodiment of a fossil pattern recognition and classification system based on image recognition in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0018] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 As shown, a fossil pattern recognition and classification method based on image recognition includes the following steps: Step 101: Acquire three-dimensional point cloud data of a fossil sample, project the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously acquire a surface element distribution map of the fossil sample; Specifically, in this embodiment, a high-precision laser scanner is used to perform an omnidirectional scan of a dinosaur fossil sample to obtain raw point cloud data. The raw point cloud data is denoised to remove noise points generated by scanning environment interference and equipment errors. The missing point cloud data is filled in based on an interpolation algorithm of adjacent point cloud data to obtain three-dimensional point cloud data. Projecting the processed three-dimensional point cloud data at different viewing angles to generate a multi-view two-dimensional texture map, including at least a front view, a side view, and a top view; The elemental composition of the fossil sample surface was analyzed using an X-ray fluorescence spectrometer. The element types and content data in each grid were collected in units of 1 cm, including at least the concentrations of Ca, P, Sr, and Ti, to obtain a surface element distribution map.

[0020] Specifically, (1) 3D point cloud data acquisition and multi-view projection Data acquisition equipment Full-surface scans of dinosaur fossil specimens were performed using a high-precision structured light 3D scanner (Artec Eva, with accuracy down to 0.1mm) or a LiDAR scanner (Riegl VZ-400i, suitable for large fossils). Prior to scanning, the fossil surface was cleaned, and high-reflective areas were treated with a non-reflective powder to ensure data integrity.

[0021] Small fossils (≤50cm): Use turntable scanning, collect one viewing angle every 15°, and scan ≥24 viewing angles for a single sample Large fossils (>50cm): Scanning is done using partitions, and data from multiple regions is integrated using marker stitching technology (Artec Studio’s automatic stitching function). Point cloud processing pipeline Scan data is imported into professional software (GeomagicWrap) for processing: Denoising: Use statistical filters (to remove noise points that are more than 3 standard deviations away from the neighborhood mean) and radius filters (to remove isolated points with less than 5 neighborhood points) Hole filling: Using the curvature-based fill algorithm, missing areas smaller than 5cm² are repaired, and damaged areas larger than 5cm² are marked for manual verification. Downsampling: Use voxel grid method (VoxelGrid, resolution 0.5-2mm) to reduce the amount of data and retain geometric features Multi-view texture map generation In Maya or Blender, a coordinate system was established with the center of the fossil as the origin. Using orthogonal projection, six basic viewpoints (positive and negative in the XY / YZ / XZ planes) were generated. Each basic viewpoint was supplemented with ±15° and ±30° tilts, resulting in a total of 18 texture maps. During the mapping process, RGB color information was extracted from the point cloud and generated as 8-bit PNG textures with a uniform resolution of 2048×2048 pixels.

[0022] (2) Surface element distribution collection A portable X-ray fluorescence spectrometer (ThermoScientific Niton XL3t, ppm-level accuracy) was used to scan the fossil surface line by line, with a 2 mm × 2 mm spot diameter and a 5 mm working distance. The scanning step size was set to 1 mm (horizontally) × 2 mm (vertically). A gridded elemental concentration matrix was generated. After baseline correction and peak fitting (using the fundamental parameter method), pseudo-color elemental distribution maps were generated using ArcGIS Pro. Each pixel corresponds to an actual area of ​​1 mm², and element values ​​are normalized to a grayscale value of 0–255 (corresponding to a concentration range of 0–1000 ppm).

[0023] Step 102: using an improved GLCM gray-level co-occurrence matrix algorithm to calculate the contrast, correlation, and entropy of the fossil texture in the multi-view two-dimensional texture map, introducing LBP local binary pattern to enhance edge detail sensitivity, and obtaining texture feature data; Specifically, in this embodiment, a weighting mechanism is introduced based on the GLCM gray-level co-occurrence matrix algorithm, different weights are assigned according to the importance of different regions in the image, the features of key texture regions are highlighted, and an improved GLCM gray-level co-occurrence matrix algorithm is obtained; The multi-view two-dimensional texture map is converted into a grayscale image, the offset and grayscale level of the grayscale co-occurrence matrix are set, the contrast, correlation and entropy values ​​are calculated, and the feature data are obtained.

[0024] The LBP algorithm is applied to the grayscale two-dimensional texture map, and the sensitivity of the image edge details is enhanced by setting different radius and neighborhood point number parameters. A local binary pattern code is generated for each pixel in the image, and the binary pattern code is statistically analyzed to obtain an LBP feature histogram. The LBP feature histogram is fused with the feature data to obtain texture feature data.

[0025] Specifically, (1) Implementation details of the improved GLCM gray-level co-occurrence matrix algorithm 1. Preprocessing stage: image standardization and blocking strategy Grayscale normalization: Convert the 2048×2048 multi-view texture image to an 8-bit grayscale image (0-255 value range), and use adaptive histogram equalization (CLAHE) to enhance local contrast, focusing on improving detail recognition in low-texture areas (entropy value <3). Dynamic blocking mechanism: For smooth surface fossils (sauropod bones): use 64×64 pixel basic blocks with an overlap rate of 30% to reduce edge information loss For highly textured fossils (such as ankylosaur carapaces): 16×16 pixel sub-blocks are enabled, the overlap rate is increased to 70%, and an edge detection algorithm (Canny operator, threshold 0.3-0.7) is used to automatically identify texture-rich areas. Each texture image generates 4096 overlapping sub-blocks (2048 / 32-1=63, 63×63≈4096) to ensure that feature extraction covers the entire image. 2. Weighted GLCM feature calculation process Dynamic allocation of directional weights: Initialize the weight matrix: By default, the four main directions of 0° (horizontal), 45°, 90° (vertical), and 135° are given equal weights (0.25) Adaptive adjustment strategy: Calculate the entropy value of the sub-block texture. When the entropy value is greater than 5 (complex texture), increase the weight of the 45° / 135° direction (+0.15); when the entropy value is less than 3 (simple texture), strengthen the 0° / 90° direction (+0.2). Example: The 45° direction weight of the Ankylosaur skin wrinkle texture (high entropy) is adjusted to 0.4, and the 0° direction weight of the Stegosaurus bone plate smooth surface (low entropy) is increased to 0.45 Refined calculation of characteristic parameters: Contrast: Focuses on pixel pairs with grayscale difference ≥ 16, using a weighted counting method (the greater the grayscale difference, the higher the weight, formula: weight = grayscale difference / 255) to highlight high-frequency textures such as deep grooves and bumps. Correlation: Introduce local mean normalization (sub-block grayscale - sub-block mean), calculate the Pearson correlation coefficient with the global texture, and identify layered / striped periodic texture patterns Entropy value: Using the Shannon entropy formula (non-formula description: the degree of disorder of grayscale distribution within the statistical sub-block, the larger the value, the more complex the texture), set the sub-blocks with entropy values ​​> 6 as key feature areas, and automatically filter flat areas 3. LBP edge detail enhancement implementation Multi-scale LBP parameter configuration: Base layer (R=1, P=8): captures pixel-level edges (crack details) and generates 8-dimensional raw LBP features Enhancement layer (R=2, P=16): covers a 3×3 neighborhood, identifies millimeter-level texture structures (scale edges), and generates 16-dimensional rotation-invariant LBP features Global layer (R=3, P=24): captures centimeter-level texture patterns (bone plate arrangement) and generates 24-dimensional unified pattern LBP features Feature cascade: concatenate the three layers of features by channel to form a 48-dimensional edge-sensitive feature vector. Edge enhancement processing: Non-maximum suppression (NMS): Set the gradient amplitude threshold (take the first 30% of pixels) and retain significant edge points (curvature mutations, texture boundaries) Binarization: Perform Otsu threshold segmentation on the LBP feature map to generate an edge mask (EdgeMask) and mark texture boundary pixels (accounting for about 15%-25%) Feature fusion: Multiply the edge mask and GLCM features point by point to enhance the contribution of texture features in edge areas (weight increased by 30%) (2) Texture feature data integration solution Multi-view feature fusion: Single-view processing: 64×64 sub-block features are extracted from 18 texture images respectively, and 180-dimensional single-view global features are generated through maximum pooling (retaining the maximum value within the sub-block) Cross-view fusion: Use attention mechanism weighted averaging (weights are dynamically allocated according to the view entropy value, high entropy view weight × 1.2), and finally form a 3240-dimensional (180×18) texture feature vector Quality Control Node: Eliminate invalid viewing angles: Automatically filter out viewing angles with entropy values ​​< 2 (completely backlit surfaces), and retain ≥ 12 valid viewing angles Abnormal sub-block detection: Mark sub-blocks with contrast > 200 or entropy > 8 (possibly noise areas), with manual verification accounting for ≤ 5% Step 103: Calculate the curvature, normal vector distribution, and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data; Specifically, in this embodiment, the LBP algorithm is applied to the grayscaled two-dimensional texture map, and the sensitivity of the image edge details is enhanced by setting different radius and neighborhood point number parameters; A local binary pattern code is generated for each pixel in the image, and the binary pattern code is statistically analyzed to obtain an LBP feature histogram. The LBP feature histogram is fused with the feature data to obtain texture feature data.

[0026] Specifically, (1) Calculation details of surface geometric properties 1. Curvature and normal vector calculation process Local surface fitting: Point cloud preprocessing: Use the K-nearest neighbor algorithm (K=20) to construct the local neighborhood of each point, ensuring that the neighborhood points are evenly distributed (deviation <15%) Quadratic polynomial fitting: The fitting equation is \(z=ax²+by²+cxy+dx+ey+f\), and the coefficients are solved by the least squares method. The fitting error threshold is set to 0.3mm (dynamically adjusted according to the scanner accuracy) Curvature parameter calculation: Gaussian curvature (K): Positive curvature (K>0): identifies convex structures (joint head, crown top), and the threshold K>0.005mm⁻² is marked as a strong convex area Negative curvature (K<0): identifies concave structures (joint sockets, bone grooves), and the threshold K<-0.003mm⁻² is marked as a deep concave area Mean curvature (H): Absolute value > 0.01mm⁻¹: defined as high curvature area (phalangeal articular surface, rib arc surface) Absolute value ≤ 0.01 mm⁻¹: considered as flat area (long bone shaft, nail plate plane) Normal vector estimation: Covariance matrix method: Calculate the covariance matrix of the local neighborhood point cloud and take the eigenvector corresponding to the minimum eigenvalue as the normal vector Direction consistency check: adjacent points with normal vector angles greater than 20° are considered edge points (bone fracture surface, weathering boundary) and marked as geometric feature boundary points. 2. Specifications for extracting geometric parameters of biological parts Anatomical landmark positioning: Establish a standard coordinate system: Taking the skull fossil as an example, set the front end of the nasal bone as the origin, the sagittal plane as the XZ plane, and the horizontal plane as the XY plane Automatic identification of key points: Based on a deep learning model (PointNet++), 21 anatomical landmarks (dentation origin, articular process vertex) are detected, with manual calibration error ≤ 1mm (2) Geometric feature data processing flow Standardization and dimensionality reduction: Z-score standardization: normalize linear / angular parameters to zero mean, the formula is described as (parameter value - mean) / standard deviation Logarithmic transformation: The curvature / normal vector data is compressed by ln(x+1) to ensure that the value is distributed in the range [-3,3]. PCA dimensionality reduction: retain the first 5 principal components (cumulative variance > 95%), the physical meaning of typical principal components: PC1: Overall size feature (contribution 40%-50%) PC2: Surface curvature (contribution 20%-25%) PC3: Symmetry index (contribution 10%-15%) Feature quality control: Missing value processing: KNN interpolation is used for small-scale missing values ​​(<5%), and large-scale missing samples are marked as to be supplemented. Outlier detection: Identify outliers (measured incorrect joint angles) based on the DBSCAN algorithm, with an automatic labeling rate of ≤3%. Step 104: Analyze the surface element distribution map using the principal component analysis method (PCA) to extract significant element combinations related to the species and obtain element signature data. Specifically, in this embodiment, the LBP algorithm is applied to the grayscaled two-dimensional texture map, and the sensitivity of the image edge details is enhanced by setting different radius and neighborhood point number parameters; A local binary pattern code is generated for each pixel in the image, and the binary pattern code is statistically analyzed to obtain an LBP feature histogram. The LBP feature histogram is fused with the feature data to obtain texture feature data.

[0027] Specifically, Application of PCA principal component analysis The surface element distribution data was imported into Python data analysis libraries (NumPy and SciPy). Principal component analysis (PCA) was used to reduce the data dimension and extract features. The data was first normalized to eliminate the influence of dimensional differences in elemental content. The covariance matrix, eigenvalues, and eigenvectors of the data were then calculated. The principal components were selected based on the eigenvalues ​​(typically, those with a cumulative contribution of 80%-90%). Principal component analysis was used to extract significant elemental combinations associated with the species, generating elemental signature data.

[0028] Step 105: Map the texture feature data, geometric feature data, and element feature data to a unified feature space based on a cross-modal attention mechanism, distribute the contribution of each modality through learnable weights, and output the fossil pattern classification result.

[0029] Specifically, in this embodiment, each modal data is mapped to a feature space of the same dimension through an MLP multi-layer perceptron; the fused feature data is processed using a softmax function and converted into a probability distribution belonging to different dinosaur genera and species, and the genus with the largest probability is selected as the fossil pattern classification result based on the size of the probability value.

[0030] Specifically, 1. Implementation of Cross-modal Attention Mechanism A network model based on the cross-modal attention mechanism is constructed within a deep learning framework (TensorFlow or PyTorch). Texture, geometry, and element features are used as inputs, and a multi-layer perceptron (MLP) is used to map each modality into a feature space of the same dimension. A cross-modal attention module is introduced into the network. This module learns the correlations between different modal data and assigns a learnable weight to each modality, thereby highlighting the modal information that contributes most to the classification result.

[0031] (2) Classification result output At the output layer of the network model, a softmax function is used to process the fused feature data, converting it into a probability distribution of belonging to different dinosaur genera and species. Based on the probability values, the species with the highest probability is selected as the final fossil pattern classification result. This classification result and the corresponding probability value are then output to assess the accuracy of the classification.

[0032] Its beneficial effects are as follows: 1. Texture features reflect the details of the fossil's surface textures, geometric features reveal the shape and structure of biological parts, and elemental features reveal the fossil's material composition; these three complement each other. Using a cross-modal attention mechanism, data from different modalities are mapped into a unified feature space and assigned learnable weights. This fully leverages the strengths of each modality, significantly improving the accuracy of genus and species identification and reducing the risk of misclassification due to incomplete data. 2. Scientific methods for calculating curvature and normal vectors, as well as measuring the geometric parameters of biological parts, accurately quantify the three-dimensional morphological characteristics of fossils. For elemental feature extraction, principal component analysis (PCA) extracts key information from complex elemental distribution data. The application of these improved algorithms significantly enhances the ability to extract fossil features. 3. Through programming and a deep learning framework, various features are automatically calculated and classified, reducing manual intervention. Compared with traditional manual identification, this significantly shortens research cycles and reduces labor costs. It also avoids subjective errors and fatigue associated with manual operation, enabling the rapid and accurate classification of large numbers of dinosaur fossil samples.

[0033] See also Figure 2 In a fossil pattern recognition and classification method based on image recognition, calculating the curvature, normal vector distribution, and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data includes the following steps: Step 201: Fitting a local surface of the three-dimensional point cloud data based on the least squares method to obtain the Gaussian curvature and mean curvature of each point, and obtaining the normal vector of the plane; Step 202: normalize the calculated curvature and normal vector data so that their value range is unified between [0, 1] to obtain normalized data; Step 203: Identify different biological parts in the three-dimensional point cloud data, and use the measurement tools of the point cloud processing software to calculate the geometric parameters of each biological part to obtain geometric feature data.

[0034] Specifically, (1) Curvature and normal vector calculation In 3D point cloud processing software, the built-in geometric calculation module is used to calculate the curvature and normal vector of each point in the 3D point cloud data. For curvature calculation, a method based on least squares fitting of local surfaces is used to obtain the Gaussian curvature and mean curvature of each point. Normal vector calculation is performed by fitting a local plane to the point cloud data and obtaining the plane's normal vector as the point's normal vector. The calculated curvature and normal vector data are normalized to a uniform range of [0, 1] to facilitate subsequent analysis and comparison.

[0035] (2) Calculation of geometric parameters of biological parts Based on knowledge of dinosaur fossil anatomy, different biological parts, such as bones and joints, are identified in the 3D point cloud data. The measurement tools in the point cloud processing software are used to calculate the geometric parameters of each biological part, such as length, width, thickness, and angles. For irregularly shaped biological parts, point cloud-based volume calculation methods can be used to estimate their volume parameters. These geometric parameters are then organized to form geometric feature data.

[0036] The above is an introduction to an embodiment of a fossil pattern recognition and classification method based on image recognition of the present invention. Figure 3 In a fossil pattern recognition and classification system based on image recognition, the fossil pattern recognition and classification system includes the following modules: A fossil data acquisition module is used to obtain three-dimensional point cloud data of fossil samples, project the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously obtain a surface element distribution map of the fossil samples; Texture feature extraction module, which uses the improved GLCM gray-level co-occurrence matrix algorithm to calculate the contrast, correlation, and entropy of fossil textures in multi-view two-dimensional texture maps, introduces LBP local binary pattern to enhance edge detail sensitivity, and obtains texture feature data; The geometric feature extraction module is used to calculate the curvature, normal vector distribution and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data; The element feature extraction module is used to analyze the surface element distribution map through the PCA principal component analysis method, extract the significant element combinations related to the species, and obtain element feature data; The fossil pattern classification module is used to map texture feature data, geometric feature data, and element feature data into a unified feature space based on a cross-modal attention mechanism, and to distribute the contribution of each modality through learnable weights to output the fossil pattern classification results.

[0037] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A fossil pattern recognition and classification method based on image recognition, characterized in that: The fossil pattern recognition and classification method comprises the following steps: Acquire three-dimensional point cloud data of the fossil sample, project the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously acquire a surface element distribution map of the fossil sample; The improved GLCM gray-level co-occurrence matrix algorithm is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map, and the LBP local binary pattern is introduced to enhance the sensitivity of edge details to obtain texture feature data; Calculating the curvature, normal vector distribution, and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data; The surface element distribution map was analyzed by PCA principal component analysis, and significant element combinations related to species were extracted to obtain element characteristic data. Based on the cross-modal attention mechanism, texture feature data, geometric feature data and element feature data are mapped to a unified feature space, and the contribution of each modality is distributed through learnable weights to output the fossil pattern classification results.

2. A fossil pattern recognition and classification method based on image recognition as claimed in claim 1, characterized in that: The method of obtaining three-dimensional point cloud data of a fossil sample, projecting the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously obtaining a surface element distribution map of the fossil sample includes: A high-precision laser scanner is used to perform an all-around scan of the dinosaur fossil sample to obtain raw point cloud data. The raw point cloud data is subjected to denoising to remove noise points caused by interference from the scanning environment and equipment errors. Missing point cloud data is filled in based on an interpolation algorithm of adjacent point cloud data to obtain three-dimensional point cloud data. Projecting the processed three-dimensional point cloud data at different viewing angles to generate a multi-view two-dimensional texture map, including at least a front view, a side view, and a top view; The elemental composition of the fossil sample surface was analyzed using an X-ray fluorescence spectrometer. The element types and content data in each grid were collected in units of 1 cm, including at least the concentrations of Ca, P, Sr, and Ti, to obtain a surface element distribution map.

3. A fossil pattern recognition and classification method based on image recognition as claimed in claim 1, characterized in that: The improved GLCM gray-level co-occurrence matrix algorithm is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map, and the LBP local binary pattern is introduced to enhance the sensitivity of edge details to obtain texture feature data, including: Based on the GLCM gray-level co-occurrence matrix algorithm, a weighting mechanism is introduced to assign different weights according to the importance of different areas in the image, highlighting the characteristics of key texture areas, and obtaining an improved GLCM gray-level co-occurrence matrix algorithm; The multi-view two-dimensional texture map is converted into a grayscale image, the offset and grayscale levels of the grayscale co-occurrence matrix are set, the contrast, correlation and entropy values ​​are calculated, and feature data are obtained.

4. A fossil pattern recognition and classification method based on image recognition as claimed in claim 1, characterized in that: The improved GLCM gray level co-occurrence matrix algorithm is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map, and the LBP local binary pattern is introduced to enhance the sensitivity of edge details to obtain texture feature data, which also includes: The LBP algorithm is applied to the grayscale two-dimensional texture map, and the sensitivity of the image edge details is enhanced by setting different radius and neighborhood point number parameters. A local binary pattern code is generated for each pixel in the image, and the binary pattern code is statistically analyzed to obtain an LBP feature histogram. The LBP feature histogram is fused with feature data to obtain texture feature data.

5. A fossil pattern recognition and classification method based on image recognition as claimed in claim 1, characterized in that: The calculation of the curvature, normal vector distribution and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data includes: Fitting a local surface of the three-dimensional point cloud data based on a least squares method to obtain the Gaussian curvature and the mean curvature of each point, and obtaining a normal vector of the plane; Normalize the calculated curvature and normal vector data so that their value range is unified between [0,1] to obtain normalized data; Different biological parts are identified in the three-dimensional point cloud data, and the measurement tools of the point cloud processing software are used to calculate the geometric parameters of each biological part to obtain geometric feature data.

6. A fossil pattern recognition and classification method based on image recognition as claimed in claim 1, characterized in that: The surface element distribution map is analyzed by the PCA principal component analysis method to extract significant element combinations related to species and obtain element characteristic data, including: Standardizing the surface element distribution map, calculating the covariance matrix, eigenvalues ​​and eigenvectors of the data, and selecting the principal component according to the size of the eigenvalues; The principal components with a cumulative contribution rate of 89% and the significant element combinations related to species were extracted to obtain element characteristic data.

7. A fossil pattern recognition and classification method based on image recognition as claimed in claim 1, characterized in that: The cross-modal attention mechanism maps texture feature data, geometric feature data, and element feature data to a unified feature space, and distributes the contribution of each modality through learnable weights to output the fossil pattern classification results, including: The MLP multi-layer perceptron is used to map each modal data into a feature space of the same dimension; the softmax function is used to process the fused feature data and convert it into a probability distribution belonging to different dinosaur genera and species. Based on the size of the probability value, the genus with the highest probability is selected as the fossil pattern classification result.

8. A fossil pattern recognition and classification system based on image recognition, characterized in that: The fossil pattern recognition and classification system comprises the following steps: A fossil data acquisition module is used to acquire three-dimensional point cloud data of fossil samples, project the three-dimensional point cloud data into a multi-view two-dimensional texture map, and simultaneously acquire a surface element distribution map of the fossil samples; A texture feature extraction module is used to calculate the contrast, correlation and entropy of the fossil texture in the multi-view two-dimensional texture map using an improved GLCM gray-level co-occurrence matrix algorithm, introduce LBP local binary pattern to enhance edge detail sensitivity, and obtain texture feature data; A geometric feature extraction module is used to calculate the curvature, normal vector distribution and geometric parameters of the fossil parts in the three-dimensional point cloud data to obtain geometric feature data; The element feature extraction module is used to analyze the surface element distribution map through the PCA principal component analysis method, extract the significant element combinations related to the species, and obtain element feature data; The fossil pattern classification module is used to map texture feature data, geometric feature data, and element feature data into a unified feature space based on a cross-modal attention mechanism, and to distribute the contribution of each modality through learnable weights to output the fossil pattern classification results.

9. The fossil pattern recognition and classification system based on image recognition according to claim 8, characterized in that: The texture feature extraction module includes the following submodules: The weighted submodule is used to introduce a weighting mechanism based on the GLCM gray-level co-occurrence matrix algorithm, assign different weights according to the importance of different areas in the image, highlight the characteristics of key texture areas, and obtain an improved GLCM gray-level co-occurrence matrix algorithm; The calculation submodule is used to convert the multi-view two-dimensional texture map into a grayscale image, set the offset and grayscale level of the grayscale co-occurrence matrix, calculate the contrast, correlation and entropy value, and obtain feature data.

10. The fossil pattern recognition and classification system based on image recognition according to claim 8, characterized in that: The texture feature extraction module also includes the following submodules: The enhancement submodule is used to apply the LBP algorithm to the grayscaled two-dimensional texture map, and enhance the sensitivity of the image edge details by setting different radius and neighborhood point number parameters; The fusion submodule is used to generate a local binary pattern code for each pixel in the image, perform statistical analysis on the binary pattern code to obtain an LBP feature histogram, and fuse the LBP feature histogram with the feature data to obtain texture feature data.

Images