Deep Order Hashing For Image Retrieval

With the rapid development of Internet and multimedia,image data has increased dramatically.How to effectively and efficiently retrieve images from large-scale image data has become research hotspots in many fields.Hashing has drawn more and more attention in image retrieval due to its high search speed and low storage cost.Traditional hashing methods project the high-dimensional hand-crafted visual features to compact binary codes by linear or non-linear hashing functions.Such hand-crafted visual features may be not optimal to hash coding,so the produced hash codes may be suboptimal.Deep hashing methods,which integrate image representation learning and hash functions learning into a unified framework,have shown more superior performance.Most of existing supervised deep hashing methods mainly consider the semantic similarities among images by using pair-wise or triplet-wise constraints as supervision information.However,as a kind of crucial information,the order of the retrieval result,is neglected.Consequently,the produced hash codes may be suboptimal.In this paper,a new Deep Hashing with Top Similarity Preserving(DHTSP)method is proposed to optimize the quality of hash codes for image retrieval.Specifically,we utilize Convolutional Neural Networks(CNNs)to extract discriminative image representations directly from the raw image pixels and learn hash functions simultaneously.Then a top similarity preserving loss function is designed to preserve the similarity of returned images at the top of the ranking list.Experimental results on three benchmark datasets show that our proposed method outperforms most of state-of-the-art deep hashing methods and traditional hashing methods.The label of image can be used to guide the learning of network's features.Based on the DHTSP,this paper proposes to combine softmax loss to improve the quality of learned hash codes.Softmax loss is used to improve the distinguishability of the image feature representation,making the features of different classes more separable,which make the quality of hash codes better in some certain.Experimental results on two benchmark datasets show that integrating softmax loss into DHTSP could improve the performance of the learned hash codes.