Research And Application Of Deep Learning In Rock Image Recognition Based On Hardware Acceleration

With the rapid growth of data scale and accessibility,deep learning algorithms based on hardware acceleration have been widely used in image recognition.In geological work,however,the rock image recognition is completed by artificial.It not only needs experienced professionals,but also needs long time-consuming and high error rate.To solve the problem,rock slice images are experimental data.The preprocessed rock images are recognized by VGGNet which is typical deep convolution neural network model.In the case of guarantee accuracy,the VGGNet model is optimized.What’s more,the convolution operation in the process of image recognition is accelerated by FPGA.The main work and contribution of this dissertation are as follows:(1)In the rock image recognition experiment based on deep learning,the test sets are divided into mixed test sets and classification test sets.The pre-processed rock image data sets are recognized by the optimized VGGNet model to verify the accuracy of the model and measure the error recognition rate of various rock images.The optimization of VGGNet model includes stmcture improvement and parameter optimization.Experimental identification results show that the accuracy of mixed test set and three classification test sets can reach 0.976,1,0.9675 and 0.94875 respectively.By comparing the experimental results from CPU and GPU,it is found that the recognition accuracy and the error recognition rate between various rock images are relatively consistent.However,but the training efficiency of GPU is 8 times that of CPU under the corresponding number of iterations.In addition,the network model AlexNet is also used to recognize rock images.The comparison results show that the optimized VGGNet model is more suitable for rock image recognition.Cpared with other scholars’ different recognition methods,the optimized VGGNet model has higher recogmition rate and better performance.(2)In the accelerated implementation of the convolution operation based on FPGA,parallel acceleration design is caried out for the convolutional layer,pooling layer,activation layer and global average pooling layer of the optimized VGGNet model.On this basis,the advantages of FPGA are applied to the design of the parallel acceleration module of the convolutional layer,which plays a key role in feature extraction during the training process.And different parallel acceleration modes are adopted for each core module.By comparing the analysis of the experimental results of different platforms,the experimental results under different parallelism and the performance evaluation of the parallel acceleration module designed,it is shown that the parallel acceleration module designed in FPGA has realized the parallel convolution layer,including parallel in convolution kernels,parallel between convolution kernels and parallel in input and output channels.Its performance can reach 233.52GOPS,which has certain practical value and practical significance.