Exploiting On Supervised Hashing Algorithms With Structural Similarity

With the rapid development of science and technology,the data in applications of various industries shows a spurt growth.How to do fast,efficient and accurate retrieval on big data has also become a hot research topic.At the same time,many problems appear in big data retrieval,including the curse of the dimension,the cost of storing and the slow query time.Recently,as an efficient and promising retrieval method,hashing algorithm has received extensive attention.It maps the original high-dimensional data into low-dimensional hamming space to generate a short binary code,which can significantly reduce the storage cost and improve the retrieval speed.After researching on the existing hashing algorithms,we found that most supervised hashing algorithms are only designed for the single-labeled images.In fact,there are many complex structures of labels.Hashing for hierarchical-labeled structure and multi-labeled structure didn’t get much attention.The rich semantic information in these structures can improve the accuracy of hash codes.In order to introduce the hierarchical structural similarity information into hashing algorithm,this work proposes a supervised deep hashing for hierarchical labeled data.This algorithm firstly constructs different intrinsic semantic similarity among hierarchy,and then designs an end-to-end training model based on convolutional neural network,and defines the objective function to keep the balance and similarity of hash codes.Experimental results show that this algorithm improves the accuracy of hash codes.This work also proposes a supervised hashing for multi-label structural similarity with order-preserving features.The algorithm is designed for multi-labeled data.Firstly,the undirected graph is constructed to depict the semantic relevancy between data.Then the convolution neural network is used to extract the order-preserving features,which effectively reduces the information loss.Finally,the algorithm maximizes the likelihood function and optimizes the hash codes.Serveral experiments on multi-labeled datasets show that the proposed algorithm effectively preserves the local order relation of the original data and gets a more accurate sequence of query results in image retrieval tasks.