Research On Network Pruning Algorithm Based On Vehicle Classification Mode

The deeper the depth and the more complex the structure of the deep convolutional neural network vehicle classification model,the more accurate the analysis and inference ability,which must have high requirements on the storage and computing capabilities of the carrying equipment.Vehicle classification systems are usually loaded on small and medium-sized equipment in key road sections,and are difficult to deploy due to limited hardware conditions.Studying neural network model pruning is the most direct and effective way to reduce model size and computational complexity.This article mainly studies the compression method of vehicle neural network model.The focus of the research is to use the norm and geometric median pruning algorithm under sensitivity analysis to study its performance in Cifar10 public data model and vehicle classification.Optimized computing methods to reduce the amount of computation and storage space when running the model.The main work of this article is divided into the following four points:(1)Establish dataset Car-class and model acquisition.First,referring to the Cifar10 dataset,a Car-class dataset with a size of 6151 was established using web crawler technology and manual classification.Design the data import class Car Dataset to process the data into the format required by the neural network.MobileNet,ResNet56 and ResNet110 are trained with the precision-controlled epoch method,and the average recognition accuracy of Cifar100 and Car-class models reaches 93.41%and 92.98%.(2)The importance of each convolutional layer is different.The pruning rate of the convolutional layer cannot be effectively set.A sensitivity norm filter evaluation criterion pruning algorithm is proposed.This method uses the norm criterion to analyze the influence of the pruning rate of each convolutional layer on the entire network layer by layer,and then obtains the pruning ratio of each layer for pruning.The experimental results show that the model pruning effect of this method is better,and the accuracy is higher than that of equal pruning at the same pruning rate.(3)According to the norm evaluation standard,the offset and possible failure of some filters are calculated.This article presents a sensitivity geometric median filter pruning algorithm.This method uses the filter with the shortest Euclidean distance from all filters in the convolutional layer as the geometric center.Sensitivity analysis and pruning are performed by measuring the importance of the filter by the Euclidean distance from the filter to the geometric center.The experimental results show that the method has a good effect on medium-depth neural network pruning.(4)Use the algorithm in this article to optimize the pruning of the vehicle classification model.The sensitivity analysis of the three methods in this article is carried out respectively,and then the influence of the pruning rate on the model accuracy is analyzed according to the sensitivity.The vehicle classification model is pruned with the best pruning rate.The experimental results show that the method in this article performs well in the vehicle classification model.The MobileNet pruning rate reaches 40%,and the accuracy drops by less than 1%,while ResNet56 can even perform 50% pruning,and the accuracy loss is controlled within 5%.