Study On The Lithology Identification Method Of Rock Image Based On Deep Learning

With the deepening of geological exploration work,automated analysis of rock images has become an important research field.Traditional methods for analyzing rock images rely on manual feature extraction and classifier design,but this approach is subjective and limited.With the emergence of deep learning,research on using convolutional neural networks for rock image classification is increasing,and the approach has high levels of automation and accuracy.This paper first studies the problem of classifying five types of rock images using four different deep learning network models(Goog Le Net,Res Net,Mobile Net,Shuffle Net).In the data preprocessing stage,data augmentation techniques are used to increase the diversity of the data,and the images are scaled to the same size.In the model training and validation phase,the loss function is set to cross-entropy,and stochastic gradient descent(SGD)optimizer is used for model optimization.The experimental results show that all four models can effectively classify the five types of rock images,with the Shuffle Net V2 model performing the best,followed by Goog Le Net V1 and Res Net34.Among all the models,Mobile Net V2 has the lowest classification accuracy,but still achieves acceptable results.Based on the above network comparison,this paper studies the improvement methods of introducing an effective channel attention mechanism and sawtooth dilated convolution in the Shuffle Net V2 network to improve its performance in rock image classification tasks.The improvement methods include replacing the Re LU activation function with the Mish activation function,reducing the number of Shuffle units used,using the ECA attention mechanism to enhance information at different spatial positions in the feature map,and using sawtooth dilated convolution to expand the receptive field.The experimental results show that the Shuffle Net V2 network with attention mechanism and sawtooth dilated convolution performs better on the rock dataset,with an accuracy improvement of 4.36 percentage points compared to the original model.In addition,the improvement methods also make the model perform better in processing images with texture and complex structures.This method has the characteristics of significantly improving the model classification accuracy,better adaptability and generalization ability,which is important for the research and application in the field of image classification.