Research And Application Of Model Compression Based On Convolutional Neural Network

Deep neural network has always been the focus of artificial intelligence research,it combines relatively low-level features to form more abstract high-level feature representations or category attributes.With the continuous development of deep learning technology and the continuous upgrading of computer hardware systems,in recent years,many problems have already been solved with deep neural networks.Such as image recognition,target detection,semantic segmentation,and image retrieval,etc,and achieved widespread success.All have indeed achieved wide success.Powerful computer hardware makes it easier to train deeper neural networks,and deep neural networks clearly have better recognition accuracy and achieve better results.With the development of the social economy,people use mobile devices more and more frequent,such as mobile phones,tablet computers,and driving recorders.People began thinking about deploying deep neural networks on these mobile devices in order to make mobile devices more powerful.Obviously,running deep neural networks on mobile devices has many benefits,such as better real-time processing.However,due to the large number of complex neural networks,their large number of weight parameters consume a considerable amount of storage space,memory bandwidth,and computing resources,and this prevents deep neural networks from running on mobile applications.For this reason,this paper uses the TensorFlow framework to compress the current mainstream deep neural network model to deploy the depth model to mobile devices.The main tasks include:(1)While maintaining the accuracy of the neural network,pruning network connections and removing redundant connections,and only the effective connection within the threshold is reserved,so as to reduce the storage space required by the neural network.Sparse training of the pruned network is then performed so that the remaining connections can make up for the pruned connections and maintain accuracy.(2)Converting the sparse matrix of the weight parameters to the CSR representation,and the absolute position is replaced with a relative index that stores the effective weights.Then use the K-Means clustering algorithm to quantify the weights of the network and its gradients,and then limit the number of effective weights we need to store by making multiple connections share the same weight for further compression.(3)Huffman coding technology is used to compress the network.For weight parameters in the network,according to the size of their appearance probability,the parameters of different code lengths are represented by fewer bits,thus solving the problem of redundancy caused by different lengths of codes.(4)Implementation of Image Classification APP Based on Android.Transplant the compressed network model to the mobile terminal,and a mobile phone application is developed based on this,so as to achieve the purpose of classifying images on the mobile terminal using a convolutional neural network.