Deep Hashing Methods For Image Retrieval

In recent years,with the rapid development of the Internet and the mass popularization of multimedia devices,image has become an important medium for information carrying and dissemination.Faced with massive image data,how to quickly and accurately retrieve the images required by users from them becomes an urgent problem to be solved.Hashing method is a good solution to improve the speed of image retrieval by mapping images to hash codes which are easy to compare.However,the quality of hash codes generated by the hash method depends on the extracted image features.The traditional hash methods based on low-level features do not perform well in retrieval performance because of the limited capacity of low-level features.With the development of deep learning,deep hashing methods have caught the attention of researchers.Deep hashing methods can obtain high quality hash codes by relying on the powerful feature learning ability and hash codes mapping ability of the deep models.For the moment,there are two common types of deep hashing methods:image label-based and image relationship-based methods.The retrieval performance of the image label-based methods depends on the classification ability of the deep models.The image relationship-based methods only consider the semantic relationship between images and ignore the simi-lar relationships between images with the same semantics.In order to solve the above issues,we improve the existing image label based-based and image relationship-based deep hash-ing methods and propose novel deep hashing methods with better performance.In addition,we implement an image retrieval system based on one of the proposed methods.The main work of this thesis is as follows:In order to improve the image label-based deep hashing method,a cascaded deep hash-ing(CDH)method is proposed.CDH constructs three deep hashing mapping sub-models in a cascade way.According to the first sub-model,we can divide an image dataset into two sub-image datasets.The other two sub-models are retrieved on the divided sub-image datasets.When retrieving an image,the specific retrieval model is determined by voting through the three sub-models in a cascade way.Since most of the retrieved images are re-trieved in the sub-image dataset,some images with wrong hash codes are excluded,and the probability including other categories of images in the retrieval results is significantly reduced.Experiments on real image datasets verify the effectiveness of CDH.In order to solve the issue that the image relationship-based deep hashing method can-not compare images with the same semantics,this paper presents a dissimilarity-based deep supervised hashing(DDSH)method.When learning the deep hashing mapping model,DDSH leams dissimilarity between deep features of images to distinguish images with the same semantic,and enhances the learning of the dissimilarity relationship between images with the semantics.Therefore,DDSH can map and generate various hash codes to distin-guish the images with the same semantics.Experimental results show that the proposed method can generate multiple different hash codes for the same semantic images,and the images with the same semantics are more similar in appearance.The fact fully demonstrates the feasibility and effectiveness of the proposed method.In order to quickly retrieve images required by users from the given image dataset,an image retrieval system based on the deep hashing method is designed and implemented.This system is a specific application of DDSH to software system.Since this system is based on the web browser and server architecture,users ean quickly retrieve needed images through the web browser anytime and anywhere.In addition,the Model-View-Controller design pattern is used to separate the interactive view,business processing and data access modules of the system,which reduces the coupling between them and improves the scalability and maintainability of the system.