Deep Learning Of Multi-scale Rock Images For Intelligent Lithology Identification

Multiscale rock images are used as the research object,and the method of lithology intelligent identification based on images is studied.We use deep learning methods such as target detection,image classification and feature fusion,and the intelligent identification models are built through image identification technology,which are suitable for different applications.The convolution neural network is used to extract rock features automatically,thus the lithology intelligent identification combined with"macroscopic-mesoscopic-microcosmic" three levels can be realized.The main work and research results can be drawn as follows:(1)An improved rock detection model based on Faster R-CNN is proposed in the macroscopic scale.The Fast R-CNN detector is built through the RPN regional proposal algorithms to predict the lithology information and location information in rock images.The transfer learning method is used to train the rock detection model,which improves the training speed and identification accuracy of the model.The macroscopic rock images data set is established,and the target detection of rock can be realized.(2)A mesoscopic rock image classification model based on Resnet-101,Inception_v3 and Densenetl21 is built in the mesoscopic scale,and the structures of three different convolutional neural networks are compared.Using the image classification technology,the intelligent lithology identification can be realized.The proposed method is verified by establishing the mesoscopic rock image data set.(3)A lithology intelligent identification method based on deep learning of rock orthogonal polarization image is proposed in the microcosmic scale.The Xception,mobilenet_v2,Inception_Resnet_v2,Inception_v3,Densenet 121,Resnet101_v2 and Resnet-101 seven convolutional neural networks are used to build the rock orthogonal polarization image classification model.The rock orthogonal polarization image data set is established,and automatic identification can be realized.(4)A rock single polarization-orthogonal polarization image fusion model is proposed.The rock feature of single polarization and orthogonal polarization are extracted through the residual network of ResNet-18/34/50、ResNeXT-50/101,respectively.The feature fusion method is used to further expand the difference between classes.The rock single polarization-orthogonal polarization image data set is established,and refinement identification of lithology is realized.Compared with the identification method based on orthogonal polarization images,this method has better identification effect and generalization performance.(5)The intelligent lithology identification method based on image fusion is explored from three levels:pixel-level,feature-level and decision-level.We analyze advantages of different fusion methods,and discuss the future development trend.The preliminary research ideas of multi-scale rock image fusion method is formed,and the development of integration technology is prospected.