Research On Deep Network Compression Method Based On Channel Pruning

With the development of deep learning and GPU technology,computer vision related tasks are becoming more common and better implemented and deployed.Deep convolutional neural networks have achieved great success in various fields,such as classification recognition and target detection tasks,and are also highly concerned neural networks.However,due to the requirement for task recognition accuracy,the number of layers of the convolutional neural network continues to increase,making it deeper and more complex,resulting in a multiplied increase in the amount of parameters and calculations,and deploying the neural network when storage resources and computing resources are insufficient will seriously affect its model performance.This limits the direct application of high-performance deep neural network models to embedded devices with low cost,low power consumption and small memory.In response to this contradiction,this thesis studies channel pruning in deep neural network compression methods,and observes that different filters(also called channels)have different contributions to model performance.The main work content is as follows:(1)A network model compression algorithm based on channel pruning is studied.A channel pruning compression method based on filter elasticity is proposed to reduce the size of the neural network.The importance of the filter in the pruning problem is defined as elasticity and combined with Taylor’s formula to calculate its value,and an iterative pruning framework is proposed to solve the problem.The contradiction between pruning speed and pruning effect.Experiments show that the number of floating-point operations(flops)of the VGG-16 network can be reduced by 80.2% and the number of parameters by 97.0% without significant loss of accuracy.(2)Based on the pruning method proposed in this thesis,a lightweight neural network structure is proposed to build a fire detection unit and deployed on the Raspberry Pi.The network structure has good performance for fire detection and the model occupies a small space.It can be deployed on the Raspberry Pi 4B for real-time monitoring.The data set used is screened and expanded based on the public data set,and a small but diverse model is constructed.Data set;add the CBAM module to the designed network structure,and use the proposed channel pruning compression algorithm to prune the model;the size of the trained model is only 7.1MB,and the accuracy of the test set is 93.7%.The speed of 26 frames per second runs smoothly on a low-cost embedded device Raspberry pi 4B.The channel pruning model compression algorithm based on filter elasticity takes image classification as the basic task and compares different channel pruning compression methods.The experimental results show that the pruned filter subset obtained by the channel pruning algorithm based on filter elasticity can well characterize the full feature information of all channels,and the filters with small elasticity can be safely removed without affecting the model performance.The channel pruning compression method is applied to the fire detection model for compression training,and the trained model is deployed in Raspberry Pi for real-time fire detection,demonstrating the engineering application value of the pruning strategy studied in this thesis.