Research On Lithology Recognition Of Rock Slice Image Based On Deep Learning

In view of the complexity of rock structure and the diversity of its components,the determination of rock lithology is difficult.Thin-section identification is to analyze and identify the rock structure and mineral components by making rock thin sections and using polarized light microscope,which requires higher requirements for identification personnel and is easy to be affected by subjective factors.In order to solve these problems,this paper proposes to use convolutional neural network to automatically recognize the images of rock thin sections.On the one hand,this paper proposes a deep learning-based thin section image lithology recognition method.By comparing the performance of classic network models Alex Net,VGG16,Res Ne Xt50 and Goog Le Net on the Image Net dataset,VGG16,Res Ne Xt50 and Goog Le Net are finally selected.By comparing the precision,recall and F1 values of these three network models,the network model more suitable for this paper is selected.The experimental results show that the overall recognition ability of VGG16 network model is better than that of other models,and the feature extraction effect of rock thin section image is better.The precision,recall and F1 values of the model on the test set are 87.84%,83.46%and 85.59%,respectively.In order to improve the inference speed of the model and reduce the parameter and floating point number,two methods,based on deep separable convolution and BN layer channel pruning,are designed to compress the model.Deep separable convolution can effectively compress the model,but it has a certain impact on the accuracy of the model.In order to ensure the performance of the model when compressing the model,the deep separable convolution and SE attention mechanism are combined.By embedding the SE attention mechanism into the convolution layers of the VGG16 network model,the ability of the model to enhance useful features and suppress useless features is enhanced.The experimental results show that the SE attention mechanism can improve the accuracy of the model and avoid the influence of deep separable convolution on the accuracy of the model.The VGG16 network model was pruned based on the BN layer channel pruning and L1 regularization to obtain a sparse solution.In order to maintain the performance of the model,fine-tuning was performed on the pruned model.The parameters of the pruned VGG16 network model were reduced by 50%,and the FLOPS were reduced by 44%.The F1 and precision scores on the test set were 83.06% and 85.30%,respectively.Under the premise of ensuring the performance of the model,the training speed of the model was accelerated,and the rock thin section image lithology could be recognized efficiently.