Intelligent Recognition And Classification Of Rock Slice Image Based On Deep Learning

The microscopic data provided by rock slices can help researchers accurately determine the structure,type,fracture development and control of the rock.The method of using traditional digital image processing technology to identify microscopic images of mineral thin slices requires manual extraction of features,which has the problem of difficulty in feature extraction.The method using convolutional neural network does not require manual extraction of features,and can automatically extract mineral features during model training and learning.At the same time,the convolutional neural network has high accuracy and generalization ability.This research mainly includes the following three aspects:(1)Determining the backbone network for extracting the features of the rock slice image.Using the filtered and preprocessed rock slice image data,the recognition results of the rock slice image by the improved Alex Net、VGG16 and Res Net50 network are compared and analyzed.The experimental results show that the improved VGG16 has superior average accuracy and cross-entropy loss when recognizing rock slice images,and has better classification results.Finally,the improved VGG16 is selected as the backbone network for feature extraction.(2)Proposing an improved VGG16 network based on multi-channel features fusion.The improved VGG16 backbone network is used to extract the features of the rock slice image,and the feature maps extracted from different convolutional layers are visualized.Experimental results show that the shallow network mainly extracts texture detail features,and the deep network mainly extracts rich semantic features.Considering that with the deepening of the network,the image size is getting smaller and smaller,and the loss of feature information is more serious.Therefore,this research combines the shallow features and deep features extracted by the improved VGG16.Specific method: Before inputting the fully connected layer,first retain the feature maps extracted by the five groups of convolution kernels;secondly,the size of the feature maps after convolution is continuously reduced,and the bilinear interpolation method in the deconvolution is required to be used for the five The feature map after group convolution is restored to a certain size to ensure that the first three dimensions of the five groups of feature maps are the same;finally,the concatenate operation is used to merge on the last dimension of the five groups of feature maps to increase the features of the image itself.But the characteristic information of each layer has not increased.Through experimental comparison,when the initial value of the learning rate is set to 0.0001 and the number of training images is 11,000,the model classification result is optimal.(3)Verifying the generalization ability of the improved VGG16 multi-channel features fusion network.The crawler technology is used to obtain different types of rock slice images to verify the construction of an improved VGG16 multi-path feature fusion network.According to the recognition average accuracy rate is about 98.6%,the cross entropy loss value is 0.3,the model training time is about 13.64 h,and the test time for a rock slice is about 80 ms.The experimental results show that the improved VGG16 multichannel feature fusion network has good generalization ability.