Hash Code Learning And Its Application In Image Retrieval

With the popularity of digital camera devices such as digital cameras and smartphones,the number of digital images has grown rapidly,and hundreds of millions of photos have been uploaded to the Internet every day.In face of massive image data,how to store it and how to calculate it quickly become the two major problems in image retrieval.In hashing image retrieval,high-dimensional image data is projected into lowdimensional ‘0',‘1' hash code,which greatly reduces the memory cost.At the same time,the distance between image features calculates through Hamming distance,which significantly improves the efficiency of image similarity calculation and makes image real-time retrieval possible.In this paper,we will study image retrieval method based on hash code learning from the four aspects: research background,research purpose,related technology and research status at home and abroad.With the advantages of computational efficiency and low memory feasibility in retrieval of large scale image data,the application of hash code technology in image retrieval has been paid more and more attention by domestic and foreign scholars.However,in hashing image retrieval,the discrete hash code also brings the constraint problem of discrete condition.Under the constraint of discrete condition,the optimization problem of hashing learning becomes a difficult problem to be solved,which is NP hard.In order to solving this problem,a large number of scholars choose to relax the discrete condition,relaxing hash code into real-value,and then optimize it,and finally get a suboptimal result.In recent years,a number of scholars have proposed some discrete hashing optimization algorithm.These algorithms,such as Supervied Discrete Hashing,update each hash bit directly,without any selection.Thus leads algorithm to become time comsuming.This paper proposes a method to solve the discrete hash code optimization problem effectively.The algorithm is designed to improve the efficiency of discrete hashing optimization while maintaining its original performance.In this paper,it is compared with other robust algorithms on three data sets: CIFAR-10,NUS-WIDE and MIRFLickr-25 k and shows that our method achieves speed-up over compared the stateof-the-art methods,while having on-par and in some cases even better performance.