Semi-supervised Metric Learning Based Anchor Graph Hashing For Large Scale Image Retrieval

With the rapid growth of Internet and smart terminal,the approximate nearest neighbor search of the massive amount of high dimensional data has aroused wide attention.In recent years,the hashing based search method has been proposed to map the high dimensional data into the low dimensional binary codes,while preserving the similarity in the original feature space.These characteristics of the hashing methods can result in the high computational complexity and low storage overhead.Hence,the hashing method has been widely applied into the fields of computer vision and multimedia retrieval.In this paper,we propose the Semi-supervised Metric Learning based Anchor Graph Hashing(MLAGH)method,in which the optimal distance metric can be learned to preserve the semantic and feature similarity.Furthermore,the Anchor Graph Hashing method can be used to convert the similar images into the similar codes for image retrieval.Specifically,first,the similarity graph is constructed on feature space of the training data,and K-means algorithm is performed to obtain the anchor points.Consequently,the triplet relationship between images and the anchors can be exploited in the anchor graph.Then,the objective function can be formulated based on the label smoothness and the triplet constraints in the feature space.The SGD algorithm is employed to minimize the objective function to learn the optimal distance metric.Furthermore,the penalty is introduced to reduce the number of the iterations of SGD to expedite the execute speed of our proposed method.Finally,we conduct multiple image retrieval experiments on two public large scale data sets,and the comparative analysis with several state-of-the-art hashing methods show the superiority of our proposed methods in terms of accuracy and cost.