Research On Image Retrieval Based On Deep Graph Laplacian Hashing

In recent years,with the continuous development of big data technology,large-scale image retrieval research has received considerable attention.In the big data area,the main problem of image retrieval research is how to efficiently retrieve the results according to the query.To this end,the hashing image algorithm is widely used to solve the approximate nearest neighbor image retrieval problem.Hashing approach can transform the high dimensional visual features into compact binary codes.And the similarity between images can be formulated by the Hamming distance between the hash codes.Hash not only improves the retrieval speed,but also reduces the storage space for the features of image data.With the rapid development of deep learning,deep hashing has become a hot topic in hashing research.Furthermore,deep hashing has better performance than the traditional hashing methods.However,most existing deep hashing methods are based on the supervised scenario,and images do not have supervision information in many real applications.For this reason,in this thesis,we proposed a novel unsupervised deep hashing image retrieval system for more real applications.Specifically,the research results and innovations of this thesis include the following:1.We designed an end-to-end deep hashing image retrieval system to simultaneously learn robust image features and compact hash codes.We further utilized back-propagation and gradient descent algorithms to learn the deep network.2.We combined the graph Laplacian constraint and deep neural network to achieve an unsupervised deep hashing image retrieval algorithm,which means we learn the deep network without the label information.And in order to improve the performance of the hash codes,we used other three constraints to reduce the quantization errors,and make the hash codes more balanced,independent and uncorrelated.3.We evaluated the proposed method on CIFAR-10 and CIFAR-100 datasets,and the experimental results demonstrated that our method achieves better performance than other unsupervised hashing methods in image retrieval tasks.