Research On Lithology Recognition Method Of Rock Thin Section Image Based On Deep Learning

Lithology identification is a crucial technique in the field of geology,as it helps geologists determine the origin,geological age,and physical and mechanical properties of rocks.Lithology identification is the foundation of geological research,providing important evidence for the exploration and development of mineral resources such as petroleum,coal,and metallic ores.Methods for identifying lithology in rocks include field observations,physical and chemical analyses,and identification through thin section analysis.Thin section images of rocks contain a wealth of geological information that is often invisible to the naked eye,and the identification of lithology in these images is a cornerstone for the subsequent exploration and exploitation of oil and gas resources.Traditional methods for identifying lithology in thin section images primarily rely on the observations of geological experts,which can be costly,time-consuming,and heavily dependent on specialized knowledge.However,with the continuous development of computer science,big data and machine learning are changing the field of geology,and machine learning algorithms can be used to model and analyze thin section datasets to achieve automatic identification of lithology in thin section images.This paper aims to design a lightweight model that can achieve fast,accurate,and fully automatic identification of lithology in thin section images.Firstly,low-resolution thin section images of rocks are collected and reconstructed into high-resolution images using Res-SRCNN.All clear thin section images of rocks are then added to the Nanjing University rock teaching thin section micro-image dataset,and the dataset is augmented with random rotations,scalings,and adjustments to brightness and saturation to increase the quantity of the dataset.Thin section images of rocks are then labeled and classified by category to establish a complete thin section image dataset.Secondly,a deep learning model is used as a feature extractor to extract features from the established thin section image dataset.Since deep convolutional neural networks use fixed-size convolution kernels to extract features from images,the model’s perception of the image is limited to the local region.Therefore,this paper uses a light weight MobileViT network that can capture global image information as the feature extractor for the thin section image dataset,and further improves it.To reduce the problem of neuron death,the GELU function is used to replace Re LU6 as the activation function of the MV2 module in the MobileViT model.To further increase the model’s ability to capture channel and positional information in images,a coordinate attention mechanism is introduced to increase the model’s receptive field,resulting in the CA-MobileViT model.Experimental results demonstrate that this identification model has good overall performance.Thirdly,a knowledge distillation method is used to train the CA-MobileViT model with a larger receptive field Swin Transformer V2 model as the teacher network,thereby improving the accuracy of the automatic lithology identification model for thin section images of rocks.Experimental results show that the overall performance of the model has been further improved.Finally,thin section images of rocks are used from the Guizhou Geological Survey,the Experimental Teaching Database of China University of Geosciences,and the Classic Metamorphic Rock Image Set to verify the reliability and feasibility of the automatic lithology identification model for thin section images of rocks.The results show that the model can correctly identify the lithology of all verification images and has good robustness.