Deep Learning Based Feature Representation And Advanced Classifier Design For Image Recognition

With the development of science and technology,human beings are no longer satisfied w ith relying solely on their organs to perceive and process external information.It is an ambitio us goal of human beings to realize an image recognition system that can replace human beings.Image recognition technology imitates how humans analyze and understand target images to recognize various targets and objects with different patterns.After more than half a century of development,image recognition objects have transitioned from simple text and digital symbo ls to recognition scenarios such as autonomous driving.With the increasing complexity of the scenes,the image recognition task has also ushered in new difficulties and challenges.How to extract discriminative image features from a large amount of image information? How to desi gn a robust classification system? These problems have been the focus of research in the field of image recognition.In this paper,I explore and innovate on two critical aspects of image rec ognition tasks,feature representation and image classifier design,in order to improve the accu racy and generalization ability of image recognition algorithms.The main research work of th is paper includes.1.Elastic net regularized class specific dictionary learning algorithm(ENR-CSDL)is proposed in terms of both how to improve the discriminative power of image feature expressions and how to learn to perform effective dictionary learning.The algorithm combines the advantages of the sparse representation algorithm and the collaborative representation algorithm,resulting in a final expression of the sample that is both highly discriminable and has a low fit error.At the same time,the algorithm incorporates the advantages of class-specific dictionary learning algorithms to express sample features in a fine-grained manner to achieve better classification results.Finally,ENR-CSDL algorithm extends the algorithm to the nuclear space and improves its ability to represent samples of nonlinear structures.Experiments on the face classification datasets show the effectiveness of the ENR-CSDL algorithm.The classification accuracy of the KENR-CSDL reaches 91.14% on the Extended Yale B dataset.Compared with SVM and KCSDL,KENR-CSDL has higher accuracy and more generalization ability.2.To solve the problem of how to mine the inherent topology of samples in the feature space and how to preserve information about the nearest neighbor relationships between samples in the feature mapping process,the laplace graph embedding class specific dictionary learning algorithm(LGECSDL)is proposed.LGECSDL introduces regularization constraints based on Laplacian graph embedding and applies them to the sparse representation algorithm so that the final expression of the samples maintains both local nearest-neighbor relationships between the original samples and strong discriminability.Simultaneously,this paper extends the algorithm to class-specific dictionary learning algorithms under kernel space,which further improves the algorithm’s ability to handle nonlinear samples and improves the classification accuracy.3.For how to design and improve the performance of convolutional neural networks,this paper proposes an enhanced residual convolutional network(ERNet).First,this paper analyzes the basic design principles of convolutional neural networks,and optimizes Res Net networks on the basis of these principles by designing new stem modules and adding dropout layers to enhance the receptive field and generalization ability of the network.Second,the image classification accuracy is further improved by extracting and fusing multiple convolutional neural network features and classifying the fused features using an ensemble learning algorithm.The classification accuracy of the proposed method reaches 97.26% and 98.27% on UC-Merced and WHU-RS remote sensing classification datasets,respectively,which is significantly better than the results of Res Net18,Inception-V2.4.To obtain rich multi-scale image features and to filter redundant features,this paper proposes a contextual multi-scale learning network(CMSNet)for person re-identification task.CMSNet has designed a forward hierarchical connection group to enable the information flow from small-scale features to large-scale features.And the CMSNet has designed backward hierarchical connection groups for feature fusion between scales.At the same time,the CMSNet filtering the redundant features by developing a novel channel attention selection mechanism.Compared with the results published recently,the CMSNet has higher recognition accuracy.The experimental results show that the Rank-1 accuracy of the CMSNet is 95.6%,91.7%,78.9%,and 83.4% on Market1501,Duke,CUHK03,and MSMT17.Compared with the results of the latest papers,the proposed algorithm has higher retrieval accuracy.