Research On Hashing For Image Retrieval

With the development of information technology and the rapid growth of data scale,how to search for similar samples efficiently from massive image data has become a research hotspot in the field of image retrieval.As an important representative of approximate near neighbour search,the Hashing(Learning)technology significantly reduces the cost of image storage and query by learning the binary coded representaion of images,therefore resolving the retrieval problem in large-scale image data.The hash algorithm based on deep learning has become the main direction of hash learning.This thesis focuses on the study of deep supervised hashing with its detailed work summarized as follows:1.Considering the shortcomings of existing deep hash methods in training strategy and label semantic utilization,this study proposes a deep semantic asymmetric hashing(DSAH)algorithm.First of all,inspired by the asymmetric deep supervised hashing,DSAH employs this asymmetric technique in training the deep hash network using the query images.The hash codes of all the images in the data base are directly computed by the optimization algorithm.Then,in order to make full use of the association between the image label and the semantic content,we designed an end-to-end deep hash network,which contains an auto-encoding label network and an image network.We use the label autoencoder network to supervise the learning of image network,and embed the semantic information of image label into the learning of deep hash network.2.In view of the influence of the skewness of similarity distribution on deep hash method in the training process,especially on the asymmetric deep hashing method,as well as the shortcomings of the existing deep hash network for image representation,this paper proposes a deep priority local aggregated hashing(DPLAH).First,the Net VLAD and deep hash network are combined to improve the expression ability of the hash network for similar images.Then,by assigning different weights to the image pairs,the influence of the skew of similarity information distribution on the hash network can be reduced,so that the image pairs that are difficult to maintain similarities in the Hamming space are trained in prior.3.An image retrieval system based on deep supervised hash method is developed by using the FLASK framework,which eventaully implements the image retrieval.