Research On Convolutional Neural Network Based On Lightweight Computing Platform

The 21 st century is an age of information explosion.Every day,a large amount of information such as video,voice,text,and images are produced.Deep learning solves the problem of mining hidden rules or features in large data sets.In the past decade,the rapid development of artificial intelligence has largely been the result of deep learning theory and engineering research.Deep learning has become the most popular technology in artificial intelligence.Convolutional neural network is the core component module of deep learning.It is exactly its application that makes deep learning back to life.The convolutional neural network omits the process of human participation in feature extraction and is robust to some deformations.The breakthrough research results of convolutional neural network to a certain extent stimulates the fiery development of current deep learning and artificial intelligence.However,there are some key problems in the convolutional neural network,such as imperfect foundation theory support,as well as complex network structure,too many parameters,and high hardware level requirements.The research content of this paper is mainly the application of convolutional neural network on the lightweight computing platform.The main work is as follows:Firstly,it introduces and analyzes the latest research progress on convolutional neural networks abroad and domestically,and its application on lightweight computing platforms.Then,to solve the problem of occupying large storage space in the convolutional neural network,a clustering algorithm based on weight distribution property is proposed to compress the neural network model.Then the direct migration of the compressed network to the lightweight computing platform was explored,and the weights clustered by block of the convolutional layer and the full connect layer was proposed,as well as the error correction algorithm.In the experiment,the compression algorithm achieved a6%~10% increase in the compression rate over the conventional algorithm,and it can maintain the network accuracy at a certain level.In addition,the algorithm proposed is mathematically deduced and the error upper bound of the algorithm is given.Aiming at the repeatability of filters in traditional convolutional neural networks,a cascaded activation function with parameters--PCReLU,is proposed.The activationfunction has two outstanding points.One is to avoid the linear repetition of the filter in the network,and the other is to increase the activity of the neuron in the network.Eliminating repeatability makes the extraction of the network's shallow features more efficient.Enhancing activity is reflected in the fact that neurons do not “dead” during training.Finally,an optimization algorithm is proposed for PCReLU.Simulation results also show that the algorithm can accelerate convergence and make the training process more stable.This paper finally studies the combination of super-resolution reconstruction and convolutional neural networks.In order to improve the reconstruction performance of the network,a deconvolution module is used to reconstruct the details of the input image.In order to reduce the network structure,a reconstruction model based on the residual network and the deconvolution network is constructed.Experiment results show that the proposed algorithm can reduce the network while improving the reconstruction quality.Finally,the application of convolutional neural network to mobile phones is realized.