Hashing Based On Distance Learning And Its Applications In Image Retrieval

Thanks to the rapid advances of the Internet,we can easily share our pictures and images on the website for various purposes,such as Flickr and You Tube.This in turn results in millions of images online.Therefore there is an emerging need of searching visually relevant images from very huge databases.Recently,hashing method has attracted considerable attentions in computer vision community and has been successfully applied to computer vision,information retrieval,and data mining.These all due to that hashing can not only decrease storage space,but also can significantly improve the efficiency of searching.This paper mainly focuses on the Hashing algorithms.Based on some traditional algorithms and part of the relevant machine learning knowledge,Then it proposes new Hashing algorithms to further improve accuracy and efficiency.This paper combines the idea of distance learning,and the make use of the traditional Linear Discriminant Analysis(LDA)and existing supervise Hashing algorithms,then it proposes an unsupervised Hashing algorithm and applies to a large-scale image retrieval.The main work and innovations of this paper are as follows:· Hashing algorithms can be divided into three categories: unsupervised,semi-supervised and supervised.Because the semi-supervised and supervised algorithms have label informations,so theirs retrieval accuracy is more higher than the unsupervised algorithm.While in most case,the train datasets don't have label informations.In this paper,it first uses the k-means clustering algorithm to classify,then applies the LDA algo-rithm to the model.Thus it successfully introduces the idea of supervised algorithm to solve the problem of having no labeled data.· Most of the proposed Hashing algorithm,often don't take into account the local and global structure of the data points at the same time.In this paper,it not only makes use of the Local structure to makes the average Hamming distance minimized for the intra-cluster pairs,and maximized for the inter-cluster pairs,but also preserve the distance of non-neighbor points in the global structure.· many Hashing algorithms require the model be uniformly barrel and the hashing codes be irrelevant,then the problem become an NP-hard problem,while ignoring the real data distribution characteristics.In this paper,it combines the data distribution and relaxs the strict constraints,then it can not only simplify the model,but also significantly improve the retrieval accuracy.