Research And Application Of The Convolutional Neural Network Optimization Model In The Actual Measured Data Classification Of GPR

With the acceleration of highway construction in China,tunnel geological Penetrating Radar(GPR)image interpretation is a key step of tunnel GPR image interpretation.Combined with the advantages of convolutional neural network in image recognition,this paper studies and implements an efficient tunnel GPR image classification model,and studies the model compression of convolutional neural network.In practical engineering,GPR images of tunnels are complex and changeable,and difficult to be recognized.Therefore,traditional machine learning methods are difficult to effectively improve the recognition accuracy of GPR images.Therefore,in this paper,the Visual Geometry group-16 network model is improved by using transfer learning and data amplification techniques,aiming at fully training the convolutional neural network with fewer marked samples,so as to effectively improve the model’s accuracy in GPR image recognition.On this basis,further research found that the vgg-16 network model based on migration learning improvement has problems,such as large model size and long prediction time,which is not conducive to the deployment of the network model in embedded devices.Therefore,after studying relevant model compression technology,this paper adopted channel pruning and weight quantization to optimize the improved vgg-16 network model.First,VisTex texture database was used as the source domain data in the migration learning stage,and the vgg-16 network model was pre-trained under Caffe platform and GPU acceleration calculation.Then,more than 6000 measured tunnel GPR images were used to fine-tune the network model.Then,in the model optimization stage,the improved vgg-16 network model was optimized by using the channel pruning and weighted three-valued quantization method based on Taylor expansion.The experimental results show that the recognition accuracy of the improved vgg-16 network model based on migration learning is as high as 97.54%.After optimization,the volume of the vgg-16 network model is compressed by nearly 77.35%,the prediction time is shortened by 3.6 times,and the recognition accuracy is only reduced by 1.73%.Moreover,the optimized network model can achieve better robustness.