The Study Of Rock Image Identification Based On Deep Learning

With the rise of computer technology,geological survey has undergone great changes.Rock image recognition is not only an important part of geological research,but also an important content of geological survey.In the current work of lithology identification,it is still necessary for professional person to analyze the thin rock sections,which obtained from the rock samples in the field.And the entire process has a large workload and a long worktime.In this paper,a method of rock image recognition based on depth learning is proposed to explore the automatic recognition of fresh rock section image to optimize this process,.On the one hand,this paper proposes an automatic identification method of rock lithology based on depth learning.By the way of transfer learning,this paper proposes the rock image recognition model for the fresh thin rock sections based on VGG,Res Net and Dense Net and get the best model.The comparative analysis shows that the model using Dense Net structure for feature extraction has the best recognition effect on the fresh thin rock section dataset because of its deeper network and dense connections.The model F1 score is 89.84%,and model accuracy is 94.48%.On the other hand,considering the offline working conditions of geologists,this paper studies the compression of rock recognition model.Based on the idea of light model design,quantization and network pruning,a compression method of rock image recognition model is designed and implemented,which can greatly reduce the size of the model and make it meet the requirements of mobile applications.Experiments show that the method of firstly 10% sparse pruning and then int8 quantization of the whole model can reduce the size of the model and obtain the best model accuracy.The F1 of the model on the mobile test dataset is 89.02%,the accuracy rate is 94.14%,and the size is only 18.7MB.It is suitable for fast and accurate identification of rock images in the field offline.In summary,through the above two aspects of the work,this paper has realized the automatic recognition of the fresh rock section image based on deep learning,obtained the best rock lithology identification model,studied the model compression,and then deployed the model to Android.The automatic identification of rock lithology method provides a brand new solution for rock identification in field geological survey.