Research On Image Classification And Retrieval Method Based On Deep Learning And Sparse Representation

Image classification and image retrieval technology are the hot topics in the field of intelligent information processing under the rapid development of computer science and the explosive growth of digital image resources.Image representation is a key problem in image classification and image retrieval.At present,feature description is the main form of image representation in image classification and image retrieval technology,which is accompanied by problems such as high dimension and complexity.As an emerging feature representation method,the sparse representation model of images can effectively solve the problems of storage quantity,computational complexity and image interpretation in practical applications.In addition,the emergence of deep learning technology makes the image feature extraction no longer rely on the artificial design method affected by subj ective factors,but focus on exploring the deep relationship hidden in a large number of data,so that the image feature description ability can be further improved.This paper focuses on the image feature representation in image classification and image retrieval.The image representation method based on sparse representation and the deep learning model based on hierarchical feature extraction are studied to improve the image classification and retrieval performance.The main work of this paper includes:(1)An image retrieval method based on sparse representation and feature fusion is proposed.Aiming at the problems of high dimensional and complexity in the local features,a sparse feature learning model is constructed through the dictionary learning,coefficient learning and feature pooling.The obtained feature description is more sparse and distinguishable than the local features.At the same time,a phased image retrieval algorithm based on sparse representation and feature fusion is constructed.The main advantage of this algorithm are as follows:First,the traditional local feature description has the problems of unequal quantity and high dimension,which have great limitations in the application field of image retrieval.Therefore,sparse feature learning model can effectively solve this problem.Second,a phased retrieval structure combining global features and local features is adopted,so that the image can be comprehensively described,thus the performance of image retrieval can be improved.The experimental results on Coil20 and the improved Caltech256 datasets verify the effectiveness of the proposed algorithm.(2)An image classification and retrieval method based on convolution deep belief network is proposed.Firstly,combining with the idea of "local receptive field" in CNN,the convolution operation is introduced into the DBN model to construct a convolutional deep belief network(CDBN)stacked by three CRBMS.Therefore,the spatial information of the neighborhood of the image data can be well obtained,and the feature description with good local invariance and high level can be obtained.Secondly,in order to explore the effectiveness of CDBN model in the application of image classification and retrieval,an image classification framework is constructed by combining the algorithm of CDBN and softmax classifier.Then,the class information of the query image is obtained based on image classification,the HOG feature is used for retrieval within the class.The image retrieval framework based on convolution deep belief network is further constructed,which optimizes the performance of image retrieval through image classification.Experiments on three standard image datasets(Coil20,UCM,and improved Caltech256)verify that the convolutional deep belief network is used as a feature extractor,which can effectively improve the expressive ability of image features and exhibit certain advantages in image classification and retrieval applications.