CN-122024079-A - Lithology recognition method based on space texture cooperation

CN122024079ACN 122024079 ACN122024079 ACN 122024079ACN-122024079-A

Abstract

The invention discloses a lithology recognition method based on space texture cooperation, which belongs to the field of image data processing and comprises the steps of obtaining a lithology recognition data set, constructing a space texture cooperation enhancement module M 1 , constructing a space-texture subspace coupling reconstruction module M SRCA , improving a backbone network of a YOLOv module based on M 1 and M SRCA to obtain an improved YOLOv module, training the improved YOLOv module into a lithology recognition model, and recognizing rock types of a rock surface image to be recognized. The M 1 can strengthen multi-scale texture detail and direction structure information simultaneously, improve the complex texture expression capability of the rock surface, and the M SRCA can improve the complex lithology category discrimination capability and the feature expression integrity. Finally, the invention obviously improves the lithology recognition accuracy, enhances the robustness of rock samples with complex textures and structure changes and improves the capability of distinguishing small-scale and similar types of rocks on the premise of ensuring the detection efficiency.

Inventors

LENG XIAOPENG
LUO SIJIA
HU ZHEN
LIN XIANG

Assignees

成都理工大学
成理智源科技(成都)有限公司

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (5)

1. The lithology recognition method based on the space texture cooperation is characterized by comprising the following steps of: S1, acquiring a lithology recognition data set, wherein a sample is a rock surface image containing rock category labels; S2, constructing a spatial texture cooperative enhancement module M 1 which comprises a normalization layer, a multi-scale local texture enhancement unit, a coordinate attention unit, a texture self-adaptive cooperative fusion unit and a residual error connection unit; The normalization layer is used for carrying out layer normalization processing with bias on an initial feature F in input with the scale of C multiplied by H multiplied by W to obtain a normalized feature F 1 , wherein C, H, W is the number, the height and the width of a feature channel of a feature F in respectively; The multi-scale local texture enhancement unit respectively processes F 1 through 3X 3 depth separable convolution, 5X 5 depth separable convolution and cavity separable convolution to obtain 3 corresponding features, and then fuses the 3 corresponding features into enhancement features F 2 ; The coordinate attention unit generates a height direction weight g h and a width direction weight g w for F 2 ; The texture self-adaptive collaborative fusion unit generates fusion characteristics F 3 according to a formula F 3 =F in ⊗(1+βg h )⊗(1+βg w ), beta is a texture response modulation factor, and ⊗ is multiplication element by element; The residual connection unit is used for obtaining an output characteristic F out of M 1 according to a formula F out =F in +F 3 ; S3, constructing a space-texture subspace coupling reconstruction module M 2 which is used for carrying out structural decomposition and cross-domain reconstruction on the F out to obtain a reconstruction characteristic F final ; S4, obtaining a YOLOv module, wherein the last two layers of a backbone network are an SPPF module and a C2PSA module respectively, replacing the C2PSA module with M 2 , and setting M 1 between the SPPF module and M 2 to obtain an improved YOLOv module; S5, training and improving YOLOv the module to convergence by using the lithology recognition data set to obtain a lithology recognition model; s6, acquiring a rock surface image to be identified, and outputting rock types through a lithology identification model.
2. The lithology recognition method based on space texture cooperation of claim 1, wherein the F 1 、F 2 、F 3 、F out scale is c×h×w.
3. The lithology recognition method based on spatial texture coordination according to claim 1, wherein the multi-scale local texture enhancement unit comprises 3 parallel convolution branches and a fusion layer; Each convolution branch comprises a convolution layer, a batch normalization layer and a ReLU activation function layer which are sequentially connected, wherein the convolution layers in the 3 convolution branches are respectively 3 x3 depth separable convolution, 5 x 5 depth separable convolution and cavity separable convolution, and the characteristic F 1,1 、F 1,2 、F 1,3 is respectively obtained by the F 1 through three convolution branches; The fusion layer is used for preparing the fusion layer according to a formula Enhancement features F 2 are generated, where Concat (∙) is a splice operation along the channel dimension and Conv 1×1 is a1 x 1 convolution operation.
4. The lithology recognition method based on space texture cooperation of claim 1, wherein the M 2 comprises a subspace orthogonal decomposition unit, a space structure modeling unit, a texture statistics modeling unit, a bi-directional cross reconstruction unit and a cooperation fusion output unit; The subspace orthogonal decomposition unit is used for carrying out subspace decomposition on the F out through 2 learnable convolution projection matrixes to obtain a space structure subspace characteristic F s and a texture statistics subspace characteristic F t ; The spatial structure modeling unit processes F s based on a spatial attention mechanism to obtain a spatial enhancement feature F s '; the texture statistical modeling unit processes F t based on a channel attention mechanism to obtain a statistical enhancement feature F t '; The bidirectional cross reconstruction unit respectively reconstructs F s ' and F t ' through two convolution mapping functions phi (∙) and ψ (∙) to obtain a spatial reconstruction feature F r1 and a statistical reconstruction feature F r2 ,F r1 =F s '+Φ(F t ')、F r2 =F t '+Ψ(F s '); The collaborative fusion output unit is used for obtaining a reconstruction characteristic F final according to a formula F final =F out +F r1 +F r2 .
5. The lithology recognition method based on space texture cooperation of claim 4, wherein the subspace orthogonal decomposition unit obtains F s and F t according to formula F s =W s *F out 、F t =W t *F out ; Wherein W s 、W t is the first and second learnable convolution projection matrices, respectively, which are convolution operations.

Description

Lithology recognition method based on space texture cooperation Technical Field The invention relates to the technical field of image data processing, in particular to a lithology recognition method based on space texture cooperation. Background In recent years, with the development of deep learning technology, lithology recognition is gradually changed from a traditional manual discrimination method to an intelligent direction. Specific documents include [1] Korean enlisting, zhang Xiaotong, shen Wei ] lithology recognition technology based on gradient lifting decision tree (GBDT) algorithm [ J ]. Mineral rock geochemistry report, 2018,37 (6): 1173-1180 ] [2] Wu Zhongyuan, zhang Xin, zhang Chunlei, etc.. Lithology recognition method based on LSTM recurrent neural network [ J ]. Lithology hydrocarbon reservoir, 2021,33 (3): 120-128. DOI:10.12108/yxyqc.20210312 ] [3] Zhang Ye, li Mingchao, han Shuai. Lithology automatic recognition and classification method based on rock image depth learning [ J ]. Lithology report, 2018,34 (2): 333-342. Korean enlightenment et al uses convolutional neural network to realize automatic classification of rock images, wuzhong et al combines sequence model to learn lithology characteristics, zhang Ye et al carries out migration learning experiments based on classical convolutional architecture, and the like, which shows that the deep learning method can effectively improve the automation level of lithology recognition. Despite the advances made in the above studies, the following commonalities remain in the actual rock surface image processing: First, existing convolutional network structures are designed mainly for general natural images, and although the expression capacity is improved by deepening the network or adding a feature fusion layer, special modeling mechanisms are lacked for the features such as multi-scale grain texture, variation of bedding trend, and crack distribution difference of the rock surface. Rock class differences often manifest themselves in texture scale and spatial structure coupling variations, which are difficult to develop with targeted enhancements relying only on conventional convolution stacks. Second, some studies introduce attention mechanisms to strengthen key regional responses, but existing attention structures mostly generate weights based on global average statistics, and do not establish dynamic modulation relations between "texture intensity changes" and "spatial attention responses". In the rock image, the discrimination meaning of the texture dense region and the discrimination meaning of the smooth region are different, and if the attention weight cannot be adaptively adjusted along with the change of the texture complexity, the key detail expression capability can be weakened. Third, existing models typically mix coded spatial structure information with texture statistics in a unified feature space. In the rock identification process, the bedding direction, structural continuity and the like belong to spatial structural characteristics, and the particle distribution density, roughness difference and the like belong to texture statistical characteristics. The two have different physical meanings in the distinguishing process, but the existing method lacks an explicit subspace separation mechanism, so that the coupling mixing phenomenon exists in the feature expression. Fourth, even if a multi-branch structure or a feature fusion mechanism is adopted in some methods, only simple splicing or weighted fusion is usually performed, a bidirectional mapping and collaborative reconstruction mechanism between different feature subspaces is not established, and it is difficult to fully mine the internal correlation between the spatial structure information and the texture statistical information. Noun interpretation: YOLOv11 modules, including a backbone network, a neck network, and a head network, wherein the penultimate layer and the first layer of the backbone network are SPPF modules and C2PSA modules, respectively. The cavity separable convolution is Atrous Separable Convolution, the depth convolution with expansion rate is adopted, and the channel fusion is realized by combining point-by-point convolution. Disclosure of Invention The invention aims to provide a lithology recognition method based on spatial texture cooperation, which solves the problems that in the existing lithology recognition technology, the mechanisms of special enhancement for multi-scale texture features, spatial attention cooperative modulation driven by texture intensity, explicit subspace decoupling modeling structure of spatial structure information and texture statistical information, cross subspace bidirectional cross reconstruction and the like are lacked. In order to achieve the purpose, the technical scheme adopted by the invention is that the lithology recognition method based on space texture cooperation comprises the following steps: S1, acquiring a lith