Research And Application On Deep Hashing Method Based Large-scale Image Index

The arrival of large data age has led to a great deal of attention to the study of large-scale image retrieval methods and efficient indexing.Finding the nearest neighbor to the query image is a fundamental research problem.Approximately Nearest Neighbor(ANN)searching based on hashing technology has become a popular method because of the good performance in terms of efficiency and accuracy.In the large-scale image retrieval application,the hashing technology generate an image to a binary hash code so that there are many information losses and the ability to distinguish the image is not enough,which affect the accuracy of the retrieval results very much.In order to solve the above problems,the paper firstly studies the deep hashing method based largescale image retrieval and proposes a new method,in which hash-codes are fused together.The method not only can preserve the image information and similarity,but also can enhance the ability to express and distinguish images.The method can improve retrieval performance for large-scale image retrieval applications eventually.The thesis of the paper focus on the following aspects:Firstly,the paper introduces the research background of large-scale image retrieval technology,and reviews the history and current situation of large-scale image retrieval and hashing methods.Through the review of large-scale image retrieval and hashing methods,the paper finds out that the deep hashing method can have a better understanding to the content of the image and can provide an effective indexing method for largescale image retrieval.Secondly,the paper studies a deep supervised hashing with pairwise Labels method,which can study the deep feature and the hash code of an image simultaneously.In the paper,the deep neural network models of VGG-F and Alex Net are applied to the method,and two deep hashing networks of different architecture are constructed.The experimental results show that the method using the two different structure's deep neural network can outperform the state of the art performance.Thirdly,the paper proposes a new deep hashing method,which is called Deep Hashing based Fusing Index Method.The method can generate more effective hashcode by fusing hash-codes generated from two different architecture's deep hashing networks.Experiments and analyzes are carried out on two large-scale image benchmarks.From the experiment results,it can be seen that the model proposed in the paper can preserve more image visual information and similarity when compared with the deep hashing model based on single structure's network.Moreover,the fused hash code can enhance the expression ability of hash code so that the large-scale image retrieval application can effectively distinguish the image.The method proposed in the paper has good performance in index construction aspect and image retrieval aspect.Finally,the paper applies the Deep Hashing based Fusing index method to the image retrieval application,and develops a retrieval system.The system realizes the function of "what you see is what you get(WYSIWYG)",and has shown excellent performance on Corel and NUS-WIDE image databases.