Research On Image Retrieval Based On Discrete Supervised Hashing

With the rapid development of Internet and multimedia technology,the number of images available on the Internet is also increasing at an alarming rate.The society has entered the network data era marked by "big data".How to quickly find images similar to specific images from massive databases has become a very challenging task.Hashing has become one of the most popular candidates for ANN search because it can achieve better performance than other methods in real applications.Because the hashing-code learning problem is essentially a discrete optimization problem which is hard to solve,most existing hashing methods try to solve a relaxed continuous optimization problem by dropping the discrete constraints.However,these methods typically suffer from poor performance due to the errors caused by the relaxation.Another part of approaches to hashing apply a single form of hash function,and an optimization process which is typically deeply coupled to this specific form.This tight coupling restricts the flexibility of the method to respond to the data,and can result in complex optimization problems that are difficult to solve.When applying hashing for nearest neighbor retrieval,the integer-valued Hamming distance produces ties(items that share the same distance).If left uncontrolled,different tie-breaking strategies could give drastically different values of the evaluation metric.Unfortunately,the learning to hash literature largely lacks tieawareness,and current evaluation protocols rarely take tie breaking into account.Here we propose a flexible yet simple framework that is able to accommodate different types of hash functions.Our framework decomposes the hashing learning problem into three steps: hash bit learning,hash function learning based on the learned bits and ranking-based evaluation metrics optimizing.In this paper,we propose a novel method,called column sampling based discrete supervised hashing,to directly learn the discrete hashing code from semantic information.The second step can be accomplished by training standard binary classifiers.For out-of-sample extension,we chooses linear classifier,boosted decision trees and deep convolutional network.Finally,we develop learning to rank formulations for hashing,aimed at directly optimizing ranking-based evaluation metrics such as Average Precision and Normalized Discounted Cumulative Gain.We extensively evaluate the retrieval performance on two large scale datasets CIFAR-10 and NUS-WIDE.And the evaluation shows that the more powerful classifiers we use for out-ofsample extension,the better accuracy we can achieve and also the more training time will be consumed.The optimization stage of ranking-based evaluation metrics can further improve the retrieval performance of hashing.