Research And Implementation Of The Object Recognition Based On Convolutional Neural Network

With the extensive use of smart phones,and 4G,Wifi and other high-speed network popularity,more and more people communicate with each other through images,how to get information about something unknown on the image by computers or mobile phones becomes increasingly desired.Earlier object recognition methods need to design manual characteristics according to different classification tasks and can not meet the needs of object recognition of natural images in a complicated environment.Along with the mass production of training data and the development of computer hardware,especially the GPU,it makes the convolutional neural network model as part of the deep machine learning methods be able to cope with more complex scenes and have higher accuracy.This paper introduces the research significance and development status of object recognition based on convolutional neural network,and analyzes the basic theory of object recognition and convolutional neural network in detail,and combined the application development technology,we design a kind of object recognition system based on convolutional neural network in the computer terminal and the Android mobile phone terminal.The main research contents are summarized as follows:(1)The theory of image classification and recognition based on the bag-of-words model,the Fisher vector,the vector of locally aggregated descriptors and support vector machine is discussed.And the object recognition experiments are carried out,then this paper compares and analyzes the three models.(2)It proposes the object recognition model based on convolutional neural network combined with the transfer learning theory,this paper modifies the existing deep convolutional neural network structure,and through the comparison test of two image data sets,it proves that compared with other image classification algorithms in accuracy,the modified image classification model has obvious advantages.Then the convolutional neural network model is simplified and partial convolution blocks are removed to reduce the model parameters,the experimental results show that the modified model is less time-consuming in training and testing and the recognition accuracy of images on multiple categories is not decreased.(3)Adopted the separate architecture of server and client,the simplified convolution neural network model combined transfer learning is applied to the server ofthe object identification system,and a whole set of schemes including PC terminal and mobile phone are designed and implemented.And the cross domain resource sharing and the inner network mapping technology are applied to solve the problem of mutual communication between mobile terminal and the local server,and a system test is carried out on three image data sets,the experimental results show that the system has a good online recognition ability,and this will help people obtain information about unknown items more conveniently and efficiently.Finally,the paper makes a summary and outlook.