Research And Application Of Residual Neural Network In Rock Image Classification

There are certain differences in microstructure characteristics of types of rocks such as porosity,particle size and permeability,which have important influences on the occurrence state of oil and gas resources in rocks.Effective rock classification methods can greatly improve the efficiency of rock microstructure research.Traditional manual classification methods frequently result in empirical errors and subjective problems with expensive costs.The main work of this paper,based on the deep learning approaches,research on the classification task of rock image by residual neural network(hereinafter referred to as ResNet)combined with features of rock image and network structure,and the corresponding improvement of ResNet from the perspective of improving the pertinence of feature extraction and feature fusion,are as follows.(1)This paper proposes a residual network model SEG-ResNet50 with fusion attention mechanism.The model integrates the channel attention SE module to suppress the transmission of unnecessary information during feature extraction,and combines the group convolution and channel shuffling modules at the entrance convolutional layer of the bottleneck structure to balance the amount of model parameters and the amount of computation.The experimental results show that the classification accuracy of SEG-ResNet50 is 92.94%,which is 4% higher than that of the ResNet50 network,and additionally,the amount of parameters and computation is basically the same as that of ResNet50.(2)Combined with the idea of bilinear feature fusion in fine-grained image classification,ResNet models ME-ResNet50 and MSEG-ResNet50 are proposed.The above models use the convolution modules of the ResNet50 and SEG-ResNet50 networks as dual feature extractor to extract features from the same input image,and the Hadamard product method is also used to fuse the features to obtain higher-order features.In the process of feature extraction,the two-way parameters are fully shared to unify the two-way feature dimension.The experimental results show that the proposed model has positive influence on classification accuracy on rock image data set.(3)To verify the effectiveness of the above proposed model in rock image classification task and to provide support for subsequent practical application and development,based on the algorithm research,this paper designs a rock image classification prototype system using the Flask framework,stores and encapsulates the trained model,and reserves the model call interface,which lays a foundation for the follow-up system development.