Research On Compression Method And Application Of Deep Convolution Neural Network

As an important part of current artificial intelligence technology,Deep convolution network has excellent performance in computer vision.However,the deployment of deep convolution network is limited by its huge model size and computation.The application scope of the deep neural network can be extended if using a suitable algorithm to compress the model.The Xnor-net is a kind of compression algorithm of CNN,which has strong compression model ability and can accelerate the forward propagation of the network model.But Xnor-net can not achieve good accray because of its binary operation.We proposed an improved algorithm for the accuracy of Xnot-net.Firstly,the classical deep convolution network model and the traditional Quantization Compression algorithm would be studied.Then improved Xnor-net algorithm would be explained.Finally,the performance of the model is verified by simulation experiments based on the proposed algorithm.The main work of this paper is as follows:(1)The composition of classical convolution network is studied.Focus on the role of convolution layer,loss function layer,sampling layer,full connection layer,activation function layer and forward propagation process.The composition of data set and the way of data enhancement under supervised learning are studied.The interpretability of convolution layer is expounded and verified by examples.(2)The classical quantized convolution network compression algorithm is studied.This paper mainly studies the propagation characteristics of weights and feature maps,and proposes a targeted binarization improvement method for their different computing processes.In view of the particularity of the binary network,a discrete back propagation process is adopted to match the local optimal solution of the full-precision model with the forward propagation result of the binary model.(3)The algorithm of the same or convolution network is studied,and on this basis,an improved algorithm of accuracy compensation is proposed.Secondary iteration compensation factor is used to compensate the compressed model,so as to improve the accuracy of the model.(4)Using this algorithm,simulated experiments were carried out on the self-made weld image data set.Firstly,the image processing system is calibrated,the camera parameters are obtained,and the conversion relationship between the pixel coordinate system and the world coordinate system is determined.Then,based on the algorithm in this paper,the model is trained on the data set,and the simulation experiment and the performance analysis of the model are carried out.