Lithology identification method based on spatial texture collaboration

Abstract

The invention discloses a lithology identification method based on spatial texture collaboration, which belongs to the field of image data processing and comprises the following steps: acquiring a lithology identification data set; constructing a spatial texture collaborative enhancement module M1; a space-texture subspace coupling reconstruction module MSRCA is constructed; improving a backbone network of the YOLOv11 module based on the M1 and the MSRCA to obtain an improved YOLOv11 module; and training the rock surface image into a lithology identification model which is used for identifying the rock category of the rock surface image to be identified. According to the invention, the M1 can reinforce multi-scale texture details and direction structure information at the same time, and improves the complex texture expression ability of the rock surface; the MSRCA can improve the complex lithology category discrimination capability and the feature expression integrity. Finally, on the premise of ensuring the detection efficiency, the lithology identification accuracy is remarkably improved, the robustness of rock samples with complex textures and structure changes is enhanced, and the capability of distinguishing small-scale rocks and similar rocks is improved.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, and in particular to a lithology identification method based on spatial texture collaboration. Background Technology

[0002] In recent years, with the development of deep learning technology, lithology identification has gradually shifted from traditional manual discrimination methods to intelligent methods. Specific literature includes: [1] Han Qidi, Zhang Xiaotong, Shen Wei. Lithology identification technology based on gradient boosting decision tree (GBDT) algorithm [J]. Bulletin of Mineralogy, Petrology and Geochemistry, 2018, 37(6):1173-1180. [2] Wu Zhongyuan, Zhang Xin, Zhang Chunlei, et al. Lithology identification method based on LSTM recurrent neural network [J]. Lithology and Oil Reservoirs, 2021, 33(3):120-128. DOI:10.12108 / yxyqc.20210312. [3] Zhang Ye, Li Mingchao, Han Shuai. Automatic lithology identification and classification method based on rock image deep learning [J]. Acta Petrologica Sinica, 2018, 34(2):333-342. Han Qidi et al. used convolutional neural networks to achieve automatic rock image classification; Wu Zhongyuan et al. combined sequence models to learn lithological features; Zhang Ye et al. conducted transfer learning experiments based on classical convolutional architectures. These studies show that deep learning methods can effectively improve the automation level of lithological identification. Despite the progress made in the above research, the following common problems still exist in practical rock surface image processing: First, existing convolutional network structures are mainly designed for general natural images. Although the expressive power can be improved by deepening the network or adding feature fusion layers, there is a lack of specific modeling mechanisms for features such as "multi-scale grain texture", "changes in bedding orientation", and "differences in fracture distribution" on rock surfaces. Differences in rock categories are often reflected in the coupling changes between texture scale and spatial structure, and it is difficult to achieve targeted enhancement by simply relying on conventional convolution stacking.

[0003] Second, some studies have introduced attention mechanisms to enhance the response of key regions, but existing attention structures are mostly based on global average statistical generation of weights and have not established a dynamic modulation relationship between "texture intensity changes" and "spatial attention response". In rock images, densely textured regions and smooth regions have different discriminative significance. If the attention weights cannot adaptively adjust with changes in texture complexity, they may weaken the ability to express key details.

[0004] Third, existing models typically encode spatial structural information and texture statistics in a unified feature space. In rock identification, bedding direction and structural continuity are spatial structural features, while grain density and roughness differences are texture statistics. These two features have different physical meanings in the discrimination process, but existing methods lack explicit subspace separation mechanisms, leading to coupled and mixed feature representations.

[0005] Fourth, even if some methods adopt multi-branch structures or feature fusion mechanisms, they usually only perform simple splicing or weighted fusion, without establishing a bidirectional mapping and collaborative reconstruction mechanism between different feature subspaces, making it difficult to fully explore the intrinsic relationship between spatial structure information and texture statistical information.

[0006] Definitions:

[0007] The YOLOv11 module includes a backbone network, a neck network, and a head network. The penultimate layer and the first layer of the backbone network are SPPF and C2PSA modules, respectively.

[0008] Atrous Separable Convolution: This technique uses depthwise convolution with dilation and combines it with pointwise convolution to achieve channel fusion. Summary of the Invention

[0009] The purpose of this invention is to provide a lithology identification method based on spatial texture collaboration that addresses the shortcomings of existing lithology identification technologies, such as the lack of mechanisms for specialized enhancement of multi-scale texture features, texture intensity-driven spatial attention collaborative modulation, explicit subspace decoupling modeling structure of spatial structure information and texture statistical information, and bidirectional cross-reconstruction across subspaces.

[0010] To achieve the above objectives, the technical solution adopted by this invention is as follows: a lithology identification method based on spatial texture synergy, comprising the following steps: S1, Obtain the lithology identification dataset, where the samples are rock surface images containing rock category labels; S2, construct a spatial texture collaborative enhancement module M1, including a normalization layer, a multi-scale local texture enhancement unit, a coordinate attention unit, a texture adaptive collaborative fusion unit, and a residual connection unit; The normalization layer is used to process the initial features F of the input, where the scale is C×H×W. in A biased layer normalization process is performed to obtain the normalized feature F1, where C, H, and W are the features of F1 and F2, respectively. in The number of characteristic channels, height, and width; The multi-scale local texture enhancement unit processes F1 through 3×3 depthwise separable convolution, 5×5 depthwise separable convolution, and dilated separable convolution respectively to obtain three corresponding features, which are then fused into the enhancement feature F2. The coordinate attention unit generates a height direction weight g for F2. h and width direction weight g w ; The texture adaptive collaborative fusion unit is based on the formula F3=Fin ⊗(1+βg h )⊗(1+βg w Generate fusion feature F3, where β is the texture response modulation factor and ⊗ represents element-wise multiplication; The residual connection unit is used according to formula F out =F in +F3 yields the output feature F of M1. out ; S3, construct a space-texture subspace coupled reconstruction module M2, used for F out Structural decomposition and cross-domain reconstruction are performed to obtain the reconstruction feature F. final ; S4. Obtain a YOLOv11 module whose last two layers of the backbone network are SPPF module and C2PSA module respectively. Replace the C2PSA module with M2 and set M1 between SPPF module and M2 to obtain the improved YOLOv11 module. S5. Train the improved YOLOv11 module with the lithology identification dataset until convergence to obtain the lithology identification model; S6: Obtain the surface image of the rock to be identified, and output the rock category through the lithology identification model.

[0011] As preferred, F1, F2, F3, F out The scale is C×H×W.

[0012] Preferably, the multi-scale local texture enhancement unit includes three parallel convolutional branches and a fusion layer; Each convolutional branch consists of a convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in sequence. The convolutional layers in each of the three branches are respectively a 3×3 depthwise separable convolution, a 5×5 depthwise separable convolution, and a dilated separable convolution. The F1 score is then processed through these three convolutional branches to obtain the feature F. 1,1 F 1,2 F 1,3 ; The fusion layer is used according to the formula This generates enhanced feature F2, where Concat(∙) is the concatenation operation along the channel dimension, and Conv 1×1 This is a 1×1 convolution operation.

[0013] Preferably, M2 includes a subspace orthogonal decomposition unit, a spatial structure modeling unit, a texture statistical modeling unit, a bidirectional cross reconstruction unit, and a collaborative fusion output unit; The subspace orthogonal decomposition unit is used to perform F using two learnable convolutional projection matrices. out Subspace decomposition is performed to obtain the spatial structure subspace features F. s and texture statistical subspace features Ft ; The spatial structure modeling unit processes F based on a spatial attention mechanism. s The spatial enhancement feature F is obtained. s '; The texture statistical modeling unit processes F based on the channel attention mechanism. t Statistical enhancement feature F is obtained t '; The bidirectional cross-reconstruction unit uses two convolutional mapping functions Φ(∙) and Ψ(∙) to respectively map F... s 'and F t 'Reconstruction is performed to obtain spatial reconstruction features F' r1 and statistical reconstruction features F r2 F r1 =F s '+Φ(F t '), F r2 =F t '+Ψ(F s '); The collaborative fusion output unit is used to calculate according to formula F final =F out +F r1 +F r2 The reconstructed feature F is obtained. final .

[0014] Preferably, the subspace orthogonal decomposition unit is based on formula F s =W s *F out F t =W t *F out , to obtain F s and F t ; In the formula, W s W t These are the first learnable convolutional projection matrix and the second learnable convolutional projection matrix, respectively, with * representing the convolution operation.

[0015] Compared with the prior art, the advantages of the present invention are as follows: (1) This invention constructs a spatial texture collaborative enhancement module to enhance the feature expression capability of rock surface images. The multi-scale local texture enhancement unit is used to generate multi-scale texture features by acting on F1 through three convolution branches, and then weighted and fused to obtain the enhanced feature F2. It has the advantages of expanding the receptive field, enhancing the ability to express texture details at different scales, and improving the ability to simultaneously depict fine-grained mineral particles and macroscopic layering structures on the rock surface. The coordinate attention unit is used to perform global pooling operations on F2 along the height and width directions respectively, and generates the height direction weight g by sharing a 1×1 convolution and a Sigmoid function. h and width direction weight g w While preserving accurate location information, the system enhances orientation perception, improves the consistency of layer orientation, and suppresses interference from irrelevant background textures. The texture adaptive collaborative fusion unit realizes a texture-driven spatial enhancement mechanism, improves the adaptability of feature expression, and avoids information bias caused by fixed attention allocation. The residual connection unit prevents over-enhancement from causing feature distortion, improves network training stability, and enhances gradient propagation ability. Finally, the spatial texture collaborative enhancement module can simultaneously enhance multi-scale texture details and orientation structure information, realize texture intensity-driven spatial collaborative modulation, and improve the ability to express complex textures on rock surfaces.

[0016] (2) The present invention also constructs a spatial-texture subspace coupled reconstruction module for structural decomposition and cross-domain reconstruction of features enhanced by spatial texture collaboration, so as to further improve the rock category discrimination ability. The module includes a subspace orthogonal decomposition unit, a spatial structure modeling unit, a texture statistical modeling unit, a bidirectional cross-reconstruction unit, and a collaborative fusion output unit. The subspace orthogonal decomposition unit is used to convert F out F is obtained by subspace projection decomposition. s and F t This reduces the coupling interference between spatial and texture features, improves the interpretability of feature representation, and provides a structural foundation for subsequent targeted modeling; the spatial structure modeling unit for F s Structural continuity enhancement is performed by using local spatial convolution and nonlinear mapping to improve the consistency of layering directions, resulting in F. s This enhances the continuous representation of bedding structure, improves the accuracy of structural rock category identification, and suppresses random texture interference; the texture statistical modeling unit is used to analyze F... t Channel statistical modeling is performed, and the differences in texture distribution are characterized by calculating the channel correlation matrix. Based on the correlation matrix, F is generated. t 'Explicitly modeling texture distribution differences, enhancing fine-grained texture differentiation capabilities, and improving the accuracy of similar lithology classification; while bidirectional cross-reconstruction units are used to establish F s 'and F tThe mapping relationship is obtained by bidirectional feature reconstruction using learnable mapping functions Φ(∙) and Ψ(∙), resulting in F. r1 and F r2 This involves establishing cross-subspace information exchange channels, compensating for information gaps in single-subspace modeling, and enhancing the collaborative expression capabilities of space and texture; finally, a collaborative fusion output unit is used to integrate F... out F r1 F r2 Merging into reconstructed feature F final The final space-texture subspace coupling reconstruction module has the advantages of enabling explicit decoupling modeling of spatial structure information and texture statistical information, constructing a bidirectional cross-reconstruction mechanism across subspaces, and improving the ability to distinguish complex lithology categories and the integrity of feature expression.

[0017] In summary, the improved YOLOv11 network based on the above two modules, and the resulting lithology identification model, significantly improves the accuracy of lithology identification, enhances the robustness to rock samples with complex textures and structural variations, and improves the ability to distinguish between small-scale and similar rock categories while ensuring detection efficiency. Attached Figure Description

[0018] Figure 1 This is a flowchart of the present invention; Figure 2 Here is a structural diagram of the spatial texture co-enhancement module M1; Figure 3 The structure diagram of the space-texture subspace coupling reconstruction module M2. Detailed Implementation

[0019] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0020] Example 1: See Figures 1 to 3 A lithological identification method based on spatial texture collaboration includes the following steps: S1, Obtain the lithology identification dataset, where the samples are rock surface images containing rock category labels; S2, construct a spatial texture collaborative enhancement module M1, including a normalization layer, a multi-scale local texture enhancement unit, a coordinate attention unit, a texture adaptive collaborative fusion unit, and a residual connection unit; The normalization layer is used to process the initial features F of the input, where the scale is C×H×W. in A biased layer normalization process is performed to obtain the normalized feature F1, where C, H, and W are the features of F1 and F2, respectively. in The number of characteristic channels, height, and width; The multi-scale local texture enhancement unit processes F1 through 3×3 depthwise separable convolution, 5×5 depthwise separable convolution, and dilated separable convolution respectively to obtain three corresponding features, which are then fused into the enhancement feature F2. The coordinate attention unit generates a height direction weight g for F2. h and width direction weight g w ; The texture adaptive collaborative fusion unit is based on the formula F3=F in ⊗(1+βg h )⊗(1+βg w Generate fusion feature F3, where β is the texture response modulation factor and ⊗ represents element-wise multiplication; The residual connection unit is used according to formula F out =F in +F3 yields the output feature F of M1. out ; S3, construct a space-texture subspace coupled reconstruction module M2, used for F out Structural decomposition and cross-domain reconstruction are performed to obtain the reconstruction feature F. final ; S4. Obtain a YOLOv11 module whose last two layers of the backbone network are SPPF module and C2PSA module respectively. Replace the C2PSA module with M2 and set M1 between SPPF module and M2 to obtain the improved YOLOv11 module. S5. Train the improved YOLOv11 module with the lithology identification dataset until convergence to obtain the lithology identification model; S6: Obtain the surface image of the rock to be identified, and output the rock category through the lithology identification model.

[0021] Example 2: See Figures 1-3 Based on Example 1, more specifically, F1, F2, F3, F out The scale is C×H×W.

[0022] The multi-scale local texture enhancement unit includes three parallel convolutional branches and a fusion layer; each convolutional branch includes a convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in sequence, and the convolutional layers in the three convolutional branches are 3×3 depthwise separable convolution, 5×5 depthwise separable convolution, and dilated separable convolution, respectively. The F1 score is obtained by passing through the three convolutional branches to obtain the feature F. 1,1 F 1,2 F 1,3 The fusion layer is used according to the formula This generates enhanced feature F2, where Concat(∙) is the concatenation operation along the channel dimension, and Conv 1×1 This is a 1×1 convolution operation.

[0023] The M2 includes a subspace orthogonal decomposition unit, a spatial structure modeling unit, a texture statistical modeling unit, a bidirectional cross-reconstruction unit, and a collaborative fusion output unit; the subspace orthogonal decomposition unit is used to perform F through two learnable convolutional projection matrices. out Subspace decomposition is performed to obtain the spatial structure subspace features F. s and texture statistical subspace features F t The spatial structure modeling unit processes F based on a spatial attention mechanism. s The spatial enhancement feature F is obtained. s The texture statistical modeling unit processes F based on a channel attention mechanism. t Statistical enhancement feature F is obtained t The bidirectional cross-reconstruction unit uses two convolutional mapping functions Φ(∙) and Ψ(∙) to respectively map F... s 'and F t 'Reconstruction is performed to obtain spatial reconstruction features F' r1 and statistical reconstruction features F r2 F r1 =F s '+Φ(F t '), F r2 =F t '+Ψ(F s The collaborative fusion output unit is used to calculate the output according to formula F. final =F out +F r1 +F r2 The reconstructed feature F is obtained. final .

[0024] The subspace orthogonal decomposition unit is based on formula F. s =W s *F out F t =W t *F out , to obtain F s and F t In the formula, W s W t These are the first learnable convolutional projection matrix and the second learnable convolutional projection matrix, respectively, with * representing the convolution operation.

[0025] Example 3: See Figures 1 to 3 To illustrate the effectiveness of this invention, a comparative experiment was designed as follows: Dataset: Rock Hand Specimen Dataset, containing 3190 color images in 10 categories, divided into training and test sets.

[0026] Experimental environment: The experimental platform is equipped with an NVIDIA RTX 3090 GPU (24GB VRAM) that supports CUDA 12.2; the processor is an AMD EPYC 7601 with 8 CPU cores and 63GB of memory.

[0027] Experimental groups: The experimental group consisted of the present invention, while the control group consisted of the original YOLOv11 model, ResNet network, and ViT network.

[0028] The performance metrics of the above networks are compared and presented in Table 1 below:

[0029] Table 1. Comparison of Performance Indicators of Different Models Model Name Cross-entropy loss Accuracy (%) ViT Network 0.06 0.85 ResNet network 0.004 0.72 YOLOv11 model 0.046 0.88 This invention 0.033 0.91 Data Analysis: Experimental results show that the proposed model comprehensively outperforms the comparison group in terms of performance. Compared to the original YOLOv11 model's accuracy of 88.0% and cross-entropy loss of 0.046, this invention significantly improves the accuracy to 91.0% while reducing the cross-entropy loss to 0.033. This demonstrates that the proposed invention effectively enhances the discriminative power of feature extraction and accelerates model convergence through synergistic local texture enhancement and bidirectional spatial attention mechanisms.

[0030] In contrast, although the ResNet network has an extremely low cross-entropy loss of 0.004, its accuracy is only 72.0%, showing serious overfitting or undergeneralization problems; while the ViT network has the highest cross-entropy loss due to convergence difficulties and its accuracy is lower than that of the YOLO series.

[0031] In summary, this method successfully overcomes the limitations of traditional CNNs, such as limited receptive fields and the difficulty in optimizing Transformers, achieving the highest detection accuracy with the lowest training error. This verifies the superior effectiveness of the spatial-texture co-enhancement module M1 and the spatial-texture subspace coupling reconstruction module M2 in balancing local detail and long-range dependency modeling. M1 and M2 successfully find the optimal balance between model complexity and feature expressiveness. It avoids the generalization bottleneck caused by the simple structure of ResNet networks and circumvents the high computational cost and convergence difficulties brought about by the global self-attention mechanism of ViT networks. Through lightweight depthwise separable convolution and coordinate attention mechanisms, M1 and M2 endow YOLOv11 with the ability to capture long-range spatial dependencies at minimal parameter cost, enabling it to achieve a leap from "local perception" to "global co-enhancement" while maintaining real-time detection speed. This result not only confirms the scientific validity of the improvement strategy but also provides a valuable solution for high-precision object detection in resource-constrained scenarios.

[0032] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A lithological identification method based on spatial texture synergy, characterized in that, Includes the following steps: S1, Obtain the lithology identification dataset, where the samples are rock surface images containing rock category labels; S2, construct a spatial texture collaborative enhancement module M1, including a normalization layer, a multi-scale local texture enhancement unit, a coordinate attention unit, a texture adaptive collaborative fusion unit, and a residual connection unit; The normalization layer is used to process the initial features F of the input, where the scale is C×H×W. in Perform layer normalization with bias to obtain normalized features F1, where C, H, and W are the features of F1 and F2, respectively. in The number of characteristic channels, height, and width; The multi-scale local texture enhancement unit processes F1 through 3×3 depthwise separable convolution, 5×5 depthwise separable convolution, and dilated separable convolution respectively to obtain three corresponding features, which are then fused into the enhancement feature F2. The coordinate attention unit generates a height direction weight g for F2. h and width direction weight g w ; The texture adaptive collaborative fusion unit is based on the formula F3=F in ⊗(1+βg h )⊗(1+βg w Generate fusion feature F3, where β is the texture response modulation factor and ⊗ represents element-wise multiplication; The residual connection unit is used according to formula F out =F in +F3 yields the output feature F of M1. out ; S3, construct a space-texture subspace coupled reconstruction module M2, used for F out Structural decomposition and cross-domain reconstruction are performed to obtain the reconstruction feature F. final ; S4. Obtain a YOLOv11 module whose last two layers of the backbone network are SPPF module and C2PSA module respectively. Replace the C2PSA module with M2 and set M1 between SPPF module and M2 to obtain the improved YOLOv11 module. S5. Train the improved YOLOv11 module with the lithology identification dataset until convergence to obtain the lithology identification model; S6: Obtain the surface image of the rock to be identified, and output the rock category through the lithology identification model.

2. The lithological identification method based on spatial texture synergy according to claim 1, characterized in that, F1, F2, F3, F out The scale is C×H×W.

3. The lithological identification method based on spatial texture synergy according to claim 1, characterized in that, The multi-scale local texture enhancement unit includes three parallel convolutional branches and a fusion layer; Each convolutional branch consists of a convolutional layer, a batch normalization layer, and a ReLU activation function layer connected in sequence. The convolutional layers in each of the three branches are respectively a 3×3 depthwise separable convolution, a 5×5 depthwise separable convolution, and a dilated separable convolution. The F1 score is then processed through these three convolutional branches to obtain the feature F. 1,1 F 1,2 F 1,3 ; The fusion layer is used according to the formula This generates enhanced feature F2, where Concat(∙) is the concatenation operation along the channel dimension, and Conv 1×1 This is a 1×1 convolution operation.

4. The lithological identification method based on spatial texture synergy according to claim 1, characterized in that, The M2 includes a subspace orthogonal decomposition unit, a spatial structure modeling unit, a texture statistical modeling unit, a bidirectional cross reconstruction unit, and a collaborative fusion output unit; The subspace orthogonal decomposition unit is used to perform F using two learnable convolutional projection matrices. out Subspace decomposition is performed to obtain the spatial structure subspace features F. s and texture statistical subspace features F t ; The spatial structure modeling unit processes F based on a spatial attention mechanism. s The spatial enhancement feature F is obtained. s '; The texture statistical modeling unit processes F based on the channel attention mechanism. t Statistical enhancement feature F is obtained t '; The bidirectional cross-reconstruction unit uses two convolutional mapping functions Φ(∙) and Ψ(∙) to respectively map F... s 'and F t 'Reconstruction is performed to obtain spatial reconstruction features F' r1 and statistical reconstruction features F r2 F r1 =F s '+Φ(F t '), F r2 =F t '+Ψ(F s '); The collaborative fusion output unit is used to calculate according to formula F final =F out +F r1 +F r2 The reconstructed feature F is obtained. final .

5. The lithological identification method based on spatial texture synergy according to claim 4, characterized in that, The subspace orthogonal decomposition unit is based on formula F. s =W s *F out F t =W t *F out , to obtain F s and F t ; In the formula, W s W t These are the first learnable convolutional projection matrix and the second learnable convolutional projection matrix, respectively, with * representing the convolution operation.

Images