A rock identification method, device, and medium based on geological semantic guidance

Abstract

This invention discloses a rock recognition method, device, and medium based on geological semantic guidance, belonging to the field of geological rock detection and recognition technology. It addresses the technical problem in existing rock target detection methods that lack effective modeling of geoscientific features, leading to blindness and redundancy in the feature fusion process and making it difficult to meet the recognition needs of complex geological images. The method includes: performing semantic analysis on effective data from multi-source geological raw data based on rock characteristics to generate a geological semantic map; performing feature fusion processing between the geological semantic map and the feature pyramid based on feature mapping to obtain target fusion features for the current rock image; performing collaborative recognition of rock physical properties in the current multi-source geological data based on visual features to output target detection features for the current rock image; outputting the geological rock classification results of the current rock image; and performing visualization processing on the geological rock classification results within an integrated platform.

Description

Technical Field

[0001] This application relates to the field of geological rock detection and identification, and in particular to a rock identification method, device and medium based on geological semantic guidance. Background Technology

[0002] In critical applications such as tunnel engineering, mineral exploration, and geological disaster monitoring, automatic rock image recognition is crucial for improving construction efficiency, assisting in surrounding rock classification, and ensuring operational safety. However, due to limitations such as complex geological image backgrounds, varying target scales, and blurred textures, traditional deep learning-based target detection models still face challenges in practical applications, including low recognition accuracy and weak generalization ability, especially in scenarios with complex geological structures such as rock fracture zones and weathered areas.

[0003] Current mainstream methods mostly use YOLO series models for rock target detection. Although its feature pyramid structure has a certain multi-scale modeling capability, its feature fusion strategy usually relies on fixed rules or general attention mechanisms, lacking effective modeling of geoscientific features. This results in the feature fusion process being blind and redundant, making it difficult to adapt to the recognition needs of complex geological images. Summary of the Invention

[0004] This application provides a rock recognition method, device, and medium based on geological semantic guidance to solve the following technical problem: In existing rock target detection, there is a lack of effective modeling of geoscientific features, which leads to blindness and redundancy in the feature fusion process, making it difficult to meet the recognition needs of complex geological images.

[0005] The embodiments of this application adopt the following technical solutions: On one hand, this application provides a rock recognition method based on geological semantic guidance, including: performing semantic analysis on effective data in multi-source geological raw data under the relevant rock characteristics to generate a geological semantic map; performing feature fusion processing on the geological semantic map and the feature pyramid under the relevant feature mapping through a preset graph convolutional network to obtain target fusion features under the current rock image; performing collaborative recognition of rock physical attributes in the current multi-source geological data based on visual features through a preset attention residual gating mechanism driven by rock physical attributes to output target detection features under the current rock image; optimizing and inputting the target fusion features and the target detection features into the YOLOv11 algorithm, and outputting the geological rock classification result of the current rock image based on the improved YOLOv11 algorithm; and performing visualization processing on the geological rock classification result under an integrated platform.

[0006] This application constructs a lightweight geological semantic map to extract key geological features such as stratum strike and scale distribution from rock images, embedding these features as prior knowledge into the YOLOv11 feature pyramid network. Building upon this, a graph convolution module is introduced to dynamically adjust the fusion weights of features at different levels based on geological semantic information. Furthermore, to enhance the model's ability to perceive rock properties, a rock physical property-driven attention residual gating mechanism is designed. This mechanism uses physical parameters such as rock density, hardness, and reflectivity as soft attention weights, dynamically adjusting the output residual threshold of the attention module in YOLOv11. This makes the model focus more on areas with specific physical characteristics, significantly improving the recognition accuracy of small-scale and textured areas. The system employs an end-to-end training strategy, combining classification loss, localization loss, and geological semantic consistency constraints to jointly optimize model parameters, enhancing its generalization ability and robustness. The proposed method is applicable to various geological tasks such as stratum fault zone identification, weathering area analysis, and digital core processing, possessing good computational efficiency and deployment adaptability, providing effective support for intelligent geological perception and automated geotechnical analysis.

[0007] In one feasible implementation, semantic analysis of effective data from multi-source geological raw data is performed on rock characteristics to generate a geological semantic map. Specifically, this includes: extracting effective data from the multi-source geological raw data, which includes borehole data, remote sensing images, and lidar data; associating the effective data with acquired rock images to obtain a current rock image with semantic and visual features; performing semantic analysis on the current rock image based on rock characteristics to calculate and determine the strike characteristics, scale distribution characteristics, and lithological probability characteristics of the strata; and calculating and processing a three-dimensional spatial semantic map of the current rock image based on the strike characteristics, scale distribution characteristics, and lithological probability characteristics to construct the geological semantic map.

[0008] In one feasible implementation, before performing feature fusion processing on the geological semantic map and the feature pyramid using a preset graph convolutional network to obtain the target fusion features of the current rock image, the method further includes: extracting graph nodes of graph structure data from the geological semantic map corresponding to the current rock image; according to The graph adjacency matrix of the nodes in the geological semantic map under Euclidean distance and lithological similarity is obtained. ;in, and These are the semantic feature vectors of the i-th and j-th nodes in the geological semantic map, respectively. It is a Gaussian kernel function, and represents the spatial and feature similarity of the map nodes in the geological semantic map; d Let be the spectral coefficient between the i-th and j-th spectral nodes, and represent the lithological similarity between the two spectral nodes i and j; according to The fusion weight matrix is ​​obtained. Wherein, GCN stands for Graph Convolutional Network; These are global features in the geological semantic map; is a learnable matrix for the graph convolutional network, and is used to map the 128-dimensional semantic features in the feature pyramid into weight coefficients.

[0009] In one feasible implementation, a preset graph convolutional network is used to perform feature fusion processing on the geological semantic map and the feature pyramid under relevant feature mapping to obtain the target fusion features under the current rock image. Specifically, this includes: based on... The target fusion features under the current rock image are obtained. ;in, For the fusion weight matrix; This represents the feature map of the i-th layer in the feature pyramid. This indicates that the size of the i-th feature is adjusted.

[0010] In one feasible implementation, before performing collaborative recognition of rock physical attributes in the current multi-source geological data based on visual features through a preset rock physical attribute-driven attention residual gating mechanism, and outputting the target detection features under the current rock image, the method further includes: extracting rock physical attributes from the current multi-source geological data corresponding to the current rock image; wherein, the rock physical attributes include at least: density, Shore hardness, and reflectance frequency spectrum; calculating the attribute dimension of the rock physical attributes to determine the physical attribute tensor of the current rock image; and performing nonlinear mapping calculations on the physical attribute tensor under relevant activation functions through a preset adaptive feature projection network, based on the feature channel number of the feature pyramid and the hierarchical mapping matrix, to obtain the attribute mapping feature vector.

[0011] In one feasible implementation, a pre-defined rock physical property-driven attention residual gating mechanism is used to perform collaborative recognition of rock physical properties in the current multi-source geological data based on visual features, and output target detection features of the current rock image. Specifically, this includes: based on... σ( T(P′ ) S(F)+b) yields the gating tensor of the current rock image under the synergy of attention mechanism and visual features. Where σ is the Sigmoid activation function; T(P′) is the output of the cross-modal transformation operator, and T is the cross-modal transformation operator, P′ is the attribute mapping feature vector; S(F) is the output after visual feature enhancement, and S is the feature enhancement function in the YOLOv11 detection head attention module, including bottleneck structure and residual connection, F is the original feature map of the current rock image in the YOLOv11 detection head attention module; b is the bias tensor; according to The feature map after gating fusion is obtained. ;in, R is the complementary tensor of the gated weights; R is the deformable convolution function. The feature map is based on the previous layer of the YOLOv11 detection head attention module; based on the gated fusion feature map, the target detection feature under the current rock image is output; wherein, the target detection feature is the target task feature identified under the collaborative residual gating mechanism of rock physical properties and visual features in the current rock image.

[0012] In one feasible implementation, the target fusion features and the target detection features are optimized and input into the YOLOv11 algorithm, and the geological rock classification result of the current rock image is output based on the improved YOLOv11 algorithm. Specifically, this includes: adjusting the target feature detection strategy according to the target fusion features and the target detection features to obtain the improved YOLOv11 algorithm; and using the improved YOLOv11 algorithm to perform feature recognition processing on the current rock image based on rock characteristics to obtain the geological rock classification result of the current rock image.

[0013] In one feasible implementation, the geological rock classification results are visualized using an integrated platform, specifically including: performing data structure classification processing on the geological rock classification results to obtain basic rock detection information and enhanced auxiliary information; wherein, the basic rock detection information includes at least: rock category, confidence level, and bounding box coordinates; the enhanced auxiliary information includes at least: feature heatmap path information, information on potential geological hazard risk areas, multimodal data association ID analysis information, and extended semantic knowledge interface information.

[0014] Secondly, embodiments of this application also provide a rock identification device based on geological semantic guidance, the device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to execute a rock identification method based on geological semantic guidance as described in any of the above embodiments.

[0015] Thirdly, embodiments of this application also provide a non-volatile computer storage medium, characterized in that the storage medium is a non-volatile computer-readable storage medium, the non-volatile computer-readable storage medium stores at least one program, each program including instructions, and when the instructions are executed by a terminal, the terminal executes a rock identification method based on geological semantic guidance as described in any of the above embodiments.

[0016] This application provides a rock identification method, device, and medium based on geological semantic guidance. Compared with the prior art, the embodiments of this application have the following beneficial technical effects: 1. This application supports access to multiple data sources, including high-resolution cameras, borehole data (including lithological stratification information and rock physical parameters), hyperspectral / multispectral remote sensing imagery, and LiDAR point cloud data. These data sources provide comprehensive information coverage from the subsurface to the surface, ensuring a complete understanding of the geological environment. The system adopts standardized data interface protocols to ensure effective integration of data from different sources and supports real-time or batch processing modes to adapt to different application scenarios.

[0017] 2. By combining prior geological knowledge with a multi-scale dynamic feature fusion mechanism, the model achieves accurate identification of key attributes such as rock type, texture features, and weathering degree. Furthermore, the model introduces a Graph Convolutional Network (GCN) to generate a spatial semantic map. Combining attributes such as rock mineral composition, texture direction, and grain size distribution, the model adaptively adjusts the fusion weights at different levels of the feature pyramid, significantly improving the identification accuracy for small-scale rocks, blurred areas, and complex backgrounds. The module's inference engine supports ONNX / TensorRT acceleration, achieving end-to-end identification latency of less than 600ms on edge computing devices. Output results are returned in structured JSON format, including rock type, confidence level, bounding box coordinates, feature heatmap path, and potential risk markers, facilitating subsequent visualization and decision analysis. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of a rock identification method based on geological semantic guidance provided in this application embodiment; Figure 2 A schematic diagram of a multi-scale dynamic feature fusion process based on geological semantic guidance is provided for an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a rock identification device based on geological semantic guidance, provided in an embodiment of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0020] It should be noted that this application relates to the fields of computer vision algorithms and geological image processing technology, specifically to a rock recognition method based on geological semantic guidance and multi-scale dynamic feature fusion. This method is applicable to rock target detection tasks in complex geological contexts, such as rock recognition and classification in scenarios involving rock fault zones and weathered areas. Specifically, it proposes a geological semantic guidance and multi-scale dynamic feature fusion mechanism. This mechanism constructs a lightweight geological semantic map of key features such as rock orientation and scale distribution, encodes it as spatial semantic prior knowledge, and dynamically guides the fusion weights of different levels of features in the YOLOv11 feature pyramid using a learnable graph convolution module. This significantly improves the recognition accuracy of small-scale rocks and areas with blurred textures.

[0021] Furthermore, a rock physical property-driven attention residual gating mechanism is introduced, which uses physical parameters such as rock density, hardness, and reflectivity as soft attention weight parameters to dynamically adjust the output residual threshold of the attention module in the YOLOv11 detector head module. This enhances the model's expressive power while improving its adaptability to geological characteristics and its discrimination accuracy.

[0022] This application provides a rock identification method based on geological semantic guidance, such as... Figure 1 As shown, the rock identification method based on geological semantic guidance specifically includes steps S101-S105: S101. Perform semantic analysis on the effective data in the multi-source geological raw data under the relevant rock characteristics to generate a geological semantic map.

[0023] Specifically, the effective data is first extracted from the multi-source geological raw data. The multi-source geological raw data includes: borehole data, remote sensing imagery, and lidar data.

[0024] In one embodiment, the system can support access to multiple data sources, including high-resolution cameras, borehole data (including lithological stratification information and rock physical parameters), hyperspectral / multispectral remote sensing imagery, and LiDAR point cloud data. These data sources provide comprehensive information coverage from the subsurface to the surface, ensuring a complete understanding of the geological environment. The system employs standardized data interface protocols to ensure effective integration of data from different sources and supports real-time or batch processing modes to adapt to different application scenarios.

[0025] Furthermore, the valid data is correlated with the acquired rock images to obtain the current rock image with semantic and visual features.

[0026] Furthermore, semantic analysis of the current rock images is performed to determine the rock strata strike characteristics, scale distribution characteristics, and lithological probability characteristics.

[0027] Furthermore, based on the characteristics of rock strata strike, scale distribution, and lithological probability, a three-dimensional spatial semantic map was calculated and processed from the current rock image to construct a geological semantic map.

[0028] In one embodiment, Figure 2 This application provides a schematic diagram of a multi-scale dynamic feature fusion process guided by geological semantics, as shown in the embodiments of this application. Figure 2 As shown, the feature pyramid of YOLOv11 is composed of multi-scale feature maps (such as 1 / 8, 1 / 16, and 1 / 32 resolutions) output by the backbone network (such as CSPDarknet). Traditional FPN achieves feature fusion through top-down paths and lateral connections, but the fusion weights are fixed. This patent generates a dynamic weight matrix through graph convolution, realizing adaptive adjustment of the YOLOv11 feature pyramid weights.

[0029] In one embodiment, the data extracted from multi-source data includes: borehole data (containing lithological stratification information, rock physical parameters, etc.), remote sensing images (such as hyperspectral and multispectral images, providing surface spectral characteristics), and lidar data (LiDAR point clouds, providing surface three-dimensional structure information). Then, for the above multi-source data, this application constructs a method for building a three-dimensional spatial semantic map G, namely, first generating a geological semantic map, specifically: utilizing... .in, The Gaussian core of rock strata strikes divides the 360° strike of rock strata into 8 intervals (each interval 45°) to capture the characteristics of geological bodies in the directional dimension and to distinguish geological strike structures with different strikes. The scale distribution fractional logarithmic binning function divides the scale range of 0.1-10m into 10 levels (such as 0.1-0.3m, 0.3-1m, etc.) according to logarithmic intervals, distinguishing between small-scale (such as gravel) and large-scale (such as rock blocks) targets. Lithology Probability Classifier: Based on a borehole lithology database, it outputs the probability that a certain location belongs to lithology such as sandstone, mudstone, or limestone.

[0030] S102. Using a pre-defined graph convolutional network, feature fusion processing is performed between the geological semantic map and the feature pyramid under relevant feature mapping to obtain the target fusion features under the current rock image.

[0031] Specifically, first extract the graph nodes of the graph structure data in the geological semantic graph corresponding to the current rock image.

[0032] Furthermore, according to The Euclidean distance of the nodes in the geological semantic map and the graph adjacency matrix under lithological similarity are obtained. .in, and These are the semantic feature vectors of the i-th and j-th nodes in the geological semantic map, respectively. is a Gaussian kernel function, and represents the spatial and feature similarity of the map nodes in the geological semantic map; d Let be the spectral coefficient between the i-th and j-th spectral nodes, and let represent the lithological similarity between the two spectral nodes i and j.

[0033] In one embodiment, the geological semantic map G is input as graph structure data into a lightweight GCN module to generate the fusion weight matrix W for each layer of the feature pyramid. Pi ~W Pn Through the adjacency matrix Aggregate cross-scale semantic relationships. Construct a graph adjacency matrix based on Euclidean distance and lithological similarity of graph nodes. ,Right now, ,in, , is the semantic feature vector of the i-th and j-th nodes in the graph (including the strike, scale, and lithological probability of the rock strata). This is a Gaussian kernel function used to represent the similarity between the nodes and features of the graph. The Gaussian kernel bandwidth parameter controls the decay rate of spatial distance similarity (usually set to 0.5 times the feature dimension). d The coefficient is used to measure the lithological similarity between two map nodes i and j.

[0034] Furthermore, according to The fusion weight matrix is ​​obtained. GCN stands for Graph Convolutional Network. These are global features in a geological semantic map; is a learnable matrix for the graph convolutional network, used to map 128-dimensional semantic features in the feature pyramid to weight coefficients.

[0035] Furthermore, based on The target fusion features under the current rock image are obtained. .in, For the fusion weight matrix; This represents the feature map of the i-th layer in the feature pyramid. This indicates that the size of the i-th feature is adjusted.

[0036] In one embodiment, such as Figure 2 As shown, the weights are obtained by fusing the layers of the feature pyramid module in the original YOLOv11 network with the Graph Convolutional Network (GCN). Generation. This network mainly utilizes the adjacency matrix. The features between nodes are passed along, and combined with the global graph features G, the weights of nodes and edges are learned. Finally, The learnable matrix of the GCN is used to map the 128-dimensional semantic features into weight coefficients. Then, feature fusion is performed to obtain the target fused features under the current rock image. , This is the feature map of the i-th layer of the YOLOv11 feature pyramid. This indicates that the size of the i-th feature is adjusted (e.g., by interpolation, downsampling, etc.), and then compared with the weights. Multiplication is performed to achieve "dynamic weighting." Finally, the fused features are output. .

[0037] As a feasible implementation method, this application constructs a three-dimensional spatial semantic map (geological semantic map) to realize the semantic representation of multi-source geological data, deeply fusing the strike, scale characteristics, and lithological probability of rock strata. Based on a lightweight GCN, a dynamic weight adjustment mechanism is designed to construct an adjacency matrix using spatial distance and lithological similarity, adaptively generating multi-scale feature fusion weights and adjusting the YOLOv11 feature pyramid. This technology effectively solves the problem of blind feature fusion in traditional methods, significantly improving the accuracy and robustness of small-scale rock detection in complex scenarios.

[0038] S103. Through a preset attention residual gating mechanism driven by rock physical properties, collaborative recognition of rock physical properties in the current multi-source geological data is performed based on visual features, and target detection features under the current rock image are output.

[0039] Specifically, it is also necessary to first extract the rock physical properties from the current multi-source geological data corresponding to the current rock image. These rock physical properties include at least: density, Shore hardness, and reflectance frequency spectrum.

[0040] Furthermore, the attribute dimensions of the rock's physical properties are calculated to determine the physical property tensor of the current rock image.

[0041] Furthermore, a pre-defined adaptive feature projection network is used, and based on the feature channel number of the feature pyramid and the hierarchical mapping matrix, nonlinear mapping calculations are performed on the physical property tensor under relevant activation functions to obtain the property mapping feature vector.

[0042] In one embodiment, the YOLOv11 detector head includes a path aggregation network (PAN) following the feature pyramid and a detection layer. Traditional attention modules (such as SPP and SAM) are based solely on visual features. The attention residual gating mechanism designed in this patent is embedded in the feature fusion node following the PAN module, directly acting on the input features of the detector head, and dynamically adjusting the attention residual threshold through rock physical properties (density, hardness, etc.).

[0043] As a feasible implementation, in the neural network mapping module for physical property features, this mechanism uses rock physical property parameters as the core source of soft attention weights. First, a multi-source data fusion module is constructed to extract density ρ and Shore hardness from borehole data. Reflection frequency spectrum, etc., form a physical attribute tensor P∈R1×D (D is the attribute dimension). Dimension alignment is achieved through an Adaptive Feature Projector (AFP), and the attribute mapping feature vector is shown as: P′=AFP(P)= Where P is a physical tensor (e.g., P = [density, hardness, reflectivity] ∈ R1 × 3). ∈RC1×D、 ∈RC×C1 is the hierarchical mapping matrix (D is the attribute dimension, C is the number of YOLOv11 feature channels, here C=256). The LeakyReLU activation function is used to enhance nonlinear mapping capabilities.

[0044] Furthermore, according to σ( T(P′ ) S(F)+b) yields the gating tensor of the current rock image under the synergy of attention mechanism and visual features. Where σ is the Sigmoid activation function; T(P′) is the output of the cross-modal transformation operator, where T is the cross-modal transformation operator and P′ is the attribute mapping feature vector; S(F) is the output after visual feature enhancement, where S is the feature enhancement function in the YOLOv11 detection head attention module, including bottleneck structures and residual connections, and F is the original feature map of the current rock image in the YOLOv11 detection head attention module, usually the output feature of the PAN module; b is the bias tensor, used to adjust the activation threshold of the gating generation to improve the model's fitting ability; M is the gating tensor with dimension R. H×W×C σ is the Sigmoid activation function, which generates probabilistic gating weights.

[0045] In one embodiment, such as Figure 2 As shown, in the dynamic residual gating engine based on physical properties, a Physical-Sensitive Neural Gate (PSNG) needs to be designed and embedded in the input node of the attention module of the YOLOv11 detector head. Through cross-modal interaction between rock physical property features and visual features, the weight allocation logic of the attention module is dynamically adjusted. Traditional detector head attention mechanisms calculate weights based solely on visual features, while this engine introduces physical properties such as density and hardness as soft attention sources, achieving collaborative guidance of "geological properties-visual features" to solve the false detection problem of rocks with similar textures (such as sandstone and siltstone). The cross-modal gating weight generation process is essentially the gating tensor of the current rock image under the collaboration of the attention mechanism and visual features. The generation process.

[0046] Furthermore, finally based on The feature map after gating fusion is obtained. .in, R is the complementary tensor of the gate weights, used to preserve historical feature information of the suppressed regions in the current layer; R is a deformable convolution function. This is a feature map from the previous layer in the attention module of the YOLOv11 detection head. Historical characteristics of residual path reconstruction.

[0047] Furthermore, based on the feature map after gating fusion, the target detection features under the current rock image are output; where the target detection features are the target task features identified under the collaborative residual gating mechanism of rock physical properties and visual features in the current rock image.

[0048] As a feasible implementation method, this application, for the first time, uses rock physical property parameters (density, hardness, reflectivity, etc.) as input variables for the attention module of the detection head. Through the cross-modal transformation operator T and residual gating mechanism, a collaborative learning framework of "geological properties-visual features" is constructed. This engine requires no additional labeled data; it can adaptively adjust the attention weights in the YOLOv11 detection head module using only the physical parameters from the borehole data, providing a lightweight and high-precision solution for intelligent rock detection in complex geological scenarios.

[0049] S104. Input the target fusion features and target detection features optimization into the YOLOv11 algorithm, and output the geological rock classification results of the current rock image based on the improved YOLOv11 algorithm.

[0050] Specifically, an improved YOLOv11 algorithm is obtained by utilizing the target fusion features and the target detection features corresponding to the target feature detection adjustment strategy. Then, the improved YOLOv11 algorithm is used to perform feature recognition processing on the current rock image based on rock characteristics, thereby obtaining the geological rock classification result of the current rock image.

[0051] As a feasible implementation method, the YOLOv11 algorithm is improved by combining the two mechanisms mentioned above (target fusion features and target feature detection adjustment strategies corresponding to target detection features) to develop an intelligent detection and classification platform for geological rocks. The platform supports multiple data input methods (such as hyperspectral imagery and point cloud data) to achieve rapid detection and accurate classification of rock targets. A modular and scalable system architecture can also be designed to ensure seamless integration with existing geological survey and resource exploration systems.

[0052] In one embodiment, a lightweight rock recognition model can be built based on an improved YOLOv11 architecture. This model combines prior geological knowledge with a multi-scale dynamic feature fusion mechanism to achieve accurate identification of key attributes such as rock type, texture features, and weathering degree. Furthermore, the model introduces a Graph Convolutional Network (GCN) to generate a spatial semantic map. This map, combined with attributes such as rock mineral composition, texture direction, and grain size distribution, adaptively adjusts the fusion weights at different levels of the feature pyramid, significantly improving recognition accuracy for small-scale rocks, blurred areas, and complex backgrounds. The module's inference engine supports ONNX / TensorRT acceleration, achieving end-to-end recognition latency of less than 600ms on edge computing devices. Output results are returned in structured JSON format, including rock type, confidence level, bounding box coordinates, feature heatmap path, and potential risk markers, facilitating subsequent visualization and decision analysis.

[0053] S105. Visualize the geological and rock classification results using an integrated platform.

[0054] Specifically, the geological rock classification results need to be processed first to obtain basic rock detection information and enhanced auxiliary information. The basic rock detection information includes at least: rock category, confidence level, and bounding box coordinates. The enhanced auxiliary information includes at least: feature heatmap path information, information on potential geological hazard risk areas, multimodal data association ID analysis information, and extended semantic knowledge interface information.

[0055] As a feasible implementation method, a modular end-to-end deployment architecture can be utilized to achieve efficient integration and application of a geological rock intelligent detection and classification platform. The identification method of this application is particularly suitable for scenarios such as tunnel engineering, mineral exploration, and geological disaster monitoring. By seamlessly integrating multi-source data inputs and combining it with geological semantic-guided feature fusion technology, it provides an accurate solution for rock identification in complex geological contexts. To achieve an intelligent closed loop of "data acquisition—semantic analysis—decision feedback," this application proposes a modular and lightweight end-to-end deployment architecture that can seamlessly integrate multi-source equipment such as borehole drilling, remote sensing, and lidar. Combined with a geological semantic-guided deep learning model, it provides a real-time and accurate rock detection solution for complex geological scenarios.

[0056] In one embodiment, after rock identification and classification are completed, the results are output in structured JSON format, including basic detection information (such as rock category, confidence level, and bounding box coordinates) and enhanced auxiliary information. Enhanced information may include: feature heatmap paths generated using methods such as Grad-CAM, intuitively displaying the model's judgment criteria; automatically marking potential geological hazard risk areas based on rock weathering degree, compressive strength, and other attributes; providing multimodal data association IDs, supporting backtracking to original borehole logs, remote sensing images, or LiDAR point clouds to achieve cross-modal data linkage analysis; and reserving standard API interfaces for accessing third-party geological databases or expert systems to expand the semantic interpretation and knowledge reasoning capabilities of the identification results. This design facilitates subsequent processing and invocation, improving the system's practicality and flexibility.

[0057] In addition, this application also provides a rock identification device based on geological semantic guidance, such as... Figure 3 As shown, the rock recognition device 300 based on geological semantic guidance specifically includes: At least one processor 301; and a memory 302 communicatively connected to the at least one processor 301; wherein the memory 302 stores instructions executable by the at least one processor 301 to enable the at least one processor 301 to execute: Semantic analysis of rock properties is performed on effective data from multi-source geological raw data to generate geological semantic maps. By using a pre-defined graph convolutional network, feature fusion processing is performed between the geological semantic map and the feature pyramid under relevant feature mapping to obtain the target fusion features under the current rock image. By using a pre-defined rock physical property-driven attention residual gating mechanism, collaborative recognition of rock physical properties in the current multi-source geological data is performed based on visual features, and target detection features under the current rock image are output. The target fusion features and target detection features are optimized and input into the YOLOv11 algorithm, and the geological rock classification results of the current rock image are output based on the improved YOLOv11 algorithm. Visualization processing of geological rock classification results is performed on an integrated platform.

[0058] This application constructs a lightweight geological semantic map to extract key geological features such as stratum strike and scale distribution from rock images, embedding these features as prior knowledge into the YOLOv11 feature pyramid network. Building upon this, a graph convolution module is introduced to dynamically adjust the fusion weights of features at different levels based on geological semantic information. Furthermore, to enhance the model's ability to perceive rock properties, a rock physical property-driven attention residual gating mechanism is designed. This mechanism uses physical parameters such as rock density, hardness, and reflectivity as soft attention weights, dynamically adjusting the output residual threshold of the attention module in YOLOv11. This makes the model focus more on areas with specific physical characteristics, significantly improving the recognition accuracy of small-scale and textured areas. The system employs an end-to-end training strategy, combining classification loss, localization loss, and geological semantic consistency constraints to jointly optimize model parameters, enhancing its generalization ability and robustness. The proposed method is applicable to various geological tasks such as stratum fault zone identification, weathering area analysis, and digital core processing, possessing good computational efficiency and deployment adaptability, providing effective support for intelligent geological perception and automated geotechnical analysis.

[0059] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for devices and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0060] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0061] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0064] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0065] The foregoing has described specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired results. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0066] The above description is merely an embodiment of this application and is not intended to limit this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the embodiments of this application should be included within the scope of the claims of this application.

Claims

1. A rock identification method based on geologic semantic guidance, characterized in that, The method includes: Semantic analysis of rock properties is performed on valid data from multi-source geological raw data to generate geological semantic maps, specifically including: Extract the valid data from the multi-source geological raw data; The valid data is correlated with the acquired rock images to obtain a current rock image with semantic and visual features; Semantic analysis of the current rock image is performed on the rock characteristics to calculate and determine the strike characteristics, scale distribution characteristics, and lithological probability characteristics of the rock strata. Based on the strike characteristics of the rock strata, the scale distribution characteristics, and the lithological probability characteristics, the current rock image is processed to calculate a three-dimensional spatial semantic map, thereby constructing the geological semantic map. Extract the graph nodes of the graph structure data from the geological semantic graph corresponding to the current rock image; according to The graph adjacency matrix of the nodes in the geological semantic map under Euclidean distance and lithological similarity is obtained. ;in, and These are the semantic feature vectors of the i-th and j-th nodes in the geological semantic map, respectively. Let Jaccard(k) be a Gaussian kernel function, and let Jaccard(k) represent the spatial and feature similarity of the map nodes in the geological semantic map. i ,k j ) represents the spectral coefficients between the i-th and j-th spectral nodes, and indicates the lithological similarity between the two spectral nodes i and j; according to The fusion weight matrix is ​​obtained. Wherein, GCN stands for Graph Convolutional Network; These are global features in the geological semantic map; is a learnable matrix for a graph convolutional network, used to map 128-dimensional semantic features in the feature pyramid to weight coefficients; Using a pre-defined graph convolutional network, feature fusion processing is performed between the geological semantic map and the feature pyramid under relevant feature mapping to obtain the target fusion features of the current rock image, specifically including: according to The target fusion features under the current rock image are obtained. ;in, For the fusion weight matrix; This represents the feature map of the i-th layer in the feature pyramid. This indicates that the size of the i-th feature is adjusted; Extract the rock physical properties from the current multi-source geological data corresponding to the current rock image; calculate the attribute dimension of the rock physical properties to determine the physical property tensor of the current rock image; By using a pre-defined adaptive feature projection network and based on the feature pyramid's feature channel number and hierarchical mapping matrix, nonlinear mapping calculations are performed on the physical property tensor under relevant activation functions to obtain the property mapping feature vector. Through a pre-defined rock physical property-driven attention residual gating mechanism, collaborative recognition of rock physical properties in the current multi-source geological data is performed based on visual features, outputting target detection features in the current rock image, specifically including: according to σ( T(P′ ) S(F)+b) yields the gating tensor of the current rock image under the synergy of attention mechanism and visual features. Where σ is the Sigmoid activation function; T(P′) is the output of the cross-modal transformation operator, and T is the cross-modal transformation operator, P′ is the attribute mapping feature vector; S(F) is the output after visual feature enhancement, and S is the feature enhancement function in the YOLOv11 detection head attention module, which includes bottleneck structure and residual connection, F is the original feature map of the current rock image in the YOLOv11 detection head attention module; b is the bias tensor; according to The feature map after gating fusion is obtained. ;in, R is the complementary tensor of the gated weights; R is the deformable convolution function. This is the feature map of the previous layer in the attention module of the YOLOv11 detection head; Based on the gated fused feature map, the target detection features of the current rock image are output; Based on the target fusion features and the target feature detection adjustment strategy corresponding to the target detection features, an improved YOLOv11 algorithm is obtained, and based on the improved YOLOv11 algorithm, the geological rock classification result of the current rock image is output. The geological rock classification results are then visualized using an integrated platform.

2. The rock identification method based on geological semantics guidance according to claim 1, characterized in that, The multi-source geological raw data includes: borehole data, remote sensing images, and lidar data.

3. The rock identification method based on geological semantics guidance according to claim 1, characterized in that, The physical properties of the rock include at least: density, Shore hardness, and reflectance frequency spectrum.

4. The rock identification method based on geological semantics guidance according to claim 1, characterized in that, The target detection features are the target task features identified by the collaborative residual gating mechanism of rock physical properties and visual features in the current rock image.

5. The rock identification method based on geological semantics guidance according to claim 1, characterized in that, Based on the improved YOLOv11 algorithm, the geological rock classification results of the current rock image are output, specifically including: The improved YOLOv11 algorithm is used to perform feature recognition processing on the current rock image based on rock characteristics, thereby obtaining the geological rock classification result of the current rock image.

6. The rock identification method based on geological semantics guidance according to claim 1, characterized in that, The geological rock classification results are visualized using an integrated platform, specifically including: The geological rock classification results are processed by data structure classification to obtain basic rock detection information and enhanced auxiliary information; The basic rock detection information includes at least: rock type, confidence level, and bounding box coordinates; the enhanced auxiliary information includes at least: feature heatmap path information, information on potential geological hazard risk areas, multimodal data association ID analysis information, and extended semantic knowledge interface information. 7.A rock identification device based on geosemantic guidance, characterized in that, The device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform a geological semantic-guided rock identification method according to any one of claims 1-6.

8. A non-transitory computer storage medium, comprising, The storage medium is a non-volatile computer-readable storage medium that stores at least one program, each program including instructions that, when executed by a terminal, cause the terminal to perform a rock identification method based on geological semantic guidance according to any one of claims 1-6.

Images