Research On Object Recognition Based On Deep Convolution Neural Network

Recently,with the fast development of science and technology,we are now in the era of artificial intelligence.A number of classical artificial intelligent commercial products have already appeared in our daily life,such as intelligent security,intelligent robots,unmanned aerial vehicles and automatic driving technology.There are a lot of intelligent algorithms supporting these products,and object recognition algorithm is one of them.Among object detection algorithms,deep neural network is the most popular one proposed in the recent years.The convolutional neural network based approach is the most widely adopted deep learning technology in this field.Scholars found that there are many factors affecting the recognition performance of convolutional neural network model.Among them,the network structure is one of the most important factors.Therefore,it is very important to improve the network performance by modifying the network structure.This dissertation is proposed based on the research performed on deep neural networks and convolution neural network.We have performed a comprehensive investigation on network structure,parameter settings,and training skills of the classical convolution neural network models recently proposed.We then summarized that methods of improving the effect of network can be divided into two kinds,by modifying network structure according to their priorities,i.e.,increase the depth of network and enhance the performance of convolution unit,both methods have two forms Then,this dissertation obtains four improved models and four improvement method of Alex Net network based on aforementioned methods.Accordingly,we design rigorous experiments on several real data sets like Image Net-1000 data sets,and the performance of our approaches is superior to basic network Alex Net which demonstrates the effectiveness of four kinds of improved model.In order to further improve the performance of network model,this dissertation first analyzes the results of experiments on modified models as well as experiments on convolution parameters control study of classical convolutional neural network.We redesign the network structure based on Alex Net,and obtain the deep convolution neural network model.At last,we performed a number of experiments on the data sets including Image Net-1000 and cloth data sets ACS and CAPB.The results of our proposed deep convolutional neural network are superior to other models in the experiment on these three data sets in terms of accuracy,and this verifies the effectiveness of our proposed model.Improved models generally achieve higher accuracy than the underlying model in experiment,but may not perform well on datasets with high similarity between some classes.In this paper,based on tree-based convolutional neural network model,a double-layer tree structure model based on Alex Net network is proposed.The experiment is carried out on the ACS data set to verify the effectiveness of the model;the results show that the model is valid.