Research On Image Classification Method Based On Deep Learning

Image classification is an important task in the field of computer vision.With the rapid development of the Internet and mobile terminals in recent years,more and more pictures have been appeared in various social websites.However,due to the diversity and disorder of these images,it is difficult to obtain effective information from them.Therefore,how to use image classification technology to divide the images into specified categories is significantly important for both the enterprises and the users.In recent years,image classification technology has made great progresses with the advent of deep convolutional neural networks,which still has many drawbacks.This dissertation focuses on the major problems existed in image classification task,i.e.,low classification accuracy,heavy time complexity and space complexity,and security risks.The content of this dissertation is as follows: 1)For the problem of large intra-class dif-ferences and small inter-class differences in the fine-grained image classification task,a corresponding pre-processing algorithm is proposed to reduce the impact of the prob-lem,to improve the classification accuracy? 2)To address the problem of large intra-class differences of local features brought by the diversity of target poses in the end-to-end fine-grained image classification framework,a new classification framework based on semantic part alignment(ASP-CNN)is proposed? 3)To address the problem of large time complexity and space complexity of fine-grained image classification networks,a lightweight fine-grained image classification network based on parameter dimensionality reduction and network compression technology is proposed,to reduce the storage space and computing resources required by the network? 4)To address the issue of security risks that brought by adversarial examples in the field of image classification in recent years,an adversarial example detection algorithm is proposed to improve the security of the clas-sification networks.From the shallower to the deeper,we have studied many challenges in the field of image classification.The main research contents and innovations of this dissertation are listed as follows:1.An image preprocessing algorithm termed as “target alignment” is proposed for fine-grained image classification.In fine grained image databases,the differences between categories sometimes are very subtle,while images of the same category have large intra class differences due to their different poses.This problem brings great difficulties to the training of the network.In this dissertation,we propose to divide the database into three main categories and successfully perform the prepro-cessing algorithm for fine grained image classification by the four steps,i.e.,po-sitioning,rotation,cropping and scale normalization,helping multiple image clas-sification algorithms achieve higher classification accuracy in fine grained image classification tasks.2.An end-to-end fine-grained image classification network based on semantic parts alignment(ASP-CNN)is constructed.In current fine grained image classification network,the information transmitted by the detection sub network to the classifica-tion sub network only contains the location information of the local feature points.Therefore,the extraction results of the local features by the classification sub net-work will be affected by the diversity of the target posture in the picture.This dis-sertation proposes the ASP-CNN,a fine-grained image classification framework,which adopts a detection sub-network and a classification sub-network to estab-lish a basic fine-grained image classification framework.In this basic classification framework,this dissertation adds pose alignment based on the spatial positional relationship of feature points in the detection sub-network,as well as the RRo I(Ro-tation Region of Interest)pooling,that has been commonly used in text recognition tasks,in the classification sub-network.Therefore,the local regions around the feature points are “aligned”,so that the classification sub-network can obtain more discriminative features in the process of feature extraction,to achieve better classi-fication accuracy.3.A lightweight fine-grained image classification network based on parameter dimen-sionality reduction and deep compression technology is proposed.Deep neural networks often bring extremely high time complexity and space complexity,this problem is particularly serious in tasks such as fine-grained image classification that require multiple neural networks for cooperation.In this dissertation,based on the proposed K nearest neighbor based “region” proposal algorithm and the clas-sification sub network structure based on the two stream structure and the 1 × 1convolutional layer,the dimensionality of the network parameters is reduced,on the other hand,the deep compression technology used for the convolutional layer and the fully connected layer realizes the compression of network parameter storage costs.Through the combined use of the above two methods,the double compres-sion on the total amount of network parameters and the storage method is completed,and the time complexity and space complexity required by the fine grained image classification algorithm are greatly reduced.4.A defense framework with two stream structure is proposed for detecting the adver-sarial examples.The existence of adversarial examples brings potential dangers to the safe use of neural networks,how to deal with the threat of adversarial examples is an important topic in the field of computer vision.This dissertation first analyzes the working principle of adversarial examples through the method of neural network visualization,then combines it with the phenomenon that the defense systems pro-posed by many researchers are always cracked quickly,and analyzes the hypothesis of “insufficient data”,and points out that: in current data volume environment,it is impossible to achieve a highly robust defense system in a complete Image Net database,finally,depending on the different types of features that neural networks of different architectures rely on in image classification,a two stream structure is proposed to accurately detect the adversarial examples in the image classification tasks,to enhance the security performance of the neural network.