The Hashing Algorithm Based On Objective Optimization For Large Scale Image Retrieval

With the rapid development of computer technology,digital information of texts,images and video has been rapidly increasing.Due to the richness and intuition of image information,the images have become an important tool for researchers to obtain information.The Image retrieval technology that obtains useful information from massive images has attracted more and more attention.At present,the most popular technology is content-based image retrieval,which finds similar images according to the given image.Hashing methods can map floating-point data with high dimension into binary code,which has low storage cost and fast query speed,and it has been widely used in the approximate neighbor search task of large-scale data sets.Relative to the unsupervised hashing,the supervised hashing has demonstrated better performance in many real applications.This paper mainly studies the supervised hashing algorithm,which generates compact binary code according to the label information,and preserves the similarity based on the original label information.In this paper,a new learning-based supervised discrete hashing algorithm is proposed to generate the best binary codes,which well fit the linear classification,however,the discrete constraint conditions during the optimization process makes the target turn to an NP-hard problem.Essentially,the supervised hashing algorithm is the problem of the discrete optimization.At present,many hashing methods are proposed to solve the discrete optimization problem.Based on the previous research on the supervised discrete hashing,a new algorithm is proposed.The original supervised discrete hashing by introducing an auxiliary variable,and using regularization algorithm to solve binary related optimization problems,the algorithm efficiently solves the problem of mixed integer optimization by alternate optimization and discrete loop gradient descent algorithm,but the operation process which makes the time consumption of the whole process relatively large,is complicated.The proposed method aims to improve the computational efficiency and retrieval performance of the original supervised hashing algorithm.By analyzing the original supervised discrete hashing algorithm,it is found that the solution to optimize the binary code in this method depends on the initial value,and the obtained binary code is not optimal.Moreover,the discrete cyclic gradient descent algorithm is greedy,which makes it easy to generate local minimum value during the iteration.This paper proposes a simplified scheme based on the idea of approximate deviation term approximation and proves that the original supervised discrete hash framework can be simplified.The improved algorithm does not rely on the initial value,and not use the alternate mechanism to optimize the algorithm.In addition,the optimized algorithm only allocates binary codes for each class.The exact mathematical solution of the improved algorithm is obtained by introducing Hadamard matrix.But the new algorithm also has some limitations:(1)The number of bits in a binary code is a power of 2;(2)The number of bits in the data set is greater than the number of classes;(3)It can be used to handle single-label problems;(4)The idea of approximate deviation term approximation.The experiments in this paper are conduced in three large data sets CIFAR-10,MINST and SUN-10.The experimental results show that the proposed algorithm has lower time cost and higher accuracy than the other methods in image retrieval,and the learning time and retrieval performance of the new algorithm are independent of the code length.In addition,the bit scalability and the suitability of processing multi-category data sets of the new algorithm are verified.