Rock Image Classification And Image Feature Visualization Based On Deep Learning

Automatic classification of rock images is of great significance of geological analysis,and lithology is the basis for classification.Recently,although researchers use deep learning technology to improve the efficiency and accuracy of the classification of rock images,the impact on local features of the classification results is still neglected,which makes the classification performance encounter a bottleneck.In this thesis,considering the influence of the local characteristics of rock lithology and its distribution of the classification results,a classification model RockNet for rock images is proposed;Futhermore,the decisionmaking basis of RockNet is explained and the reliability of RockNet is verified through the characteristics visualization of rock lithology.The main work are as follows:1.Making rock data sets.Due to the lack of open rock data sets,this thesis collected more than 4000 rock images from Chongqing Museum of Nature,Chengdu University of Technology and Wuhan University digital cloud platform,and made rock data sets for classification.In order to enhance the persuasion of the experiment,the rock data set is expanded in various ways.2.RockNet,a model for rock lithology classification,is proposed.Considering the local characteristics and distribution law of rock lithology,this thesis integrates and optimizes the Ghost Bottlenecks module in GhostNet and the Inception-A module in Inception V4,and proposes a new rock lithology classification model RockNet.By comparing RockNet with the classical classification model and the classification model in recent three years,the effectiveness of RockNet in rock lithology classification is verified.3.Feature visualization.Based on class activation map algorithm and weighted gradient-based class activation map algorithm,the feature visualization of rock lithology is realized.It not only directly proves the important role of local features in rock lithology classification by neural network,but also provides reliable basis for RockNet model proposed in this thesis.