Deep Super-resolution Based Hashing for Image Retrieval
Image retrieval based on deep convolutional neural networks(CNNs)has achieved promising performance in recent years.However,there are many low-resolution images in real-world retrieval tasks,and they would result in inaccurate hash representations for the CNN,which has only trained on high-resolution images.In this paper,we propose a novel framework,which is called deep super-resolution based hashing(DSR-Hashing),to solve the problem.DSR-Hashing is constructed by two components:a super-resolution net-work and an encoding network.For low-resolution images,the super-resolution network can upscale them to their high-resolution versions,so as to provide more semantic information from pixel-wise reconstruction.Then,the up-scaled images are fed into the encoding network which consists many residual blocks.The en-coding network is deep enough and adopts transfer learning strat-egy for better training.Extensive experiments conducted on two benchmark datasets demonstrate the state-of-the-art performance of DSR-Hashing on low-resolution image retrieval.
Image retrieval Super-resolution Convolutional neural network Deep learning
Xianbo Li Naiguang Zhang Qian Wang Feng Dai Ziyu Xue
School of Computer and Information Engineering,Beijing Technology and Business University,Beijing,Ch Information Technology Institute,Academy of Broadcasting Science,SART,Beijing,China Beijing Goldwind Science&Creation Windpower Equipment Co.,Ltd
国际会议
2019国图灵大会(ACM Turing Celebration conference-China 2019 )
成都
英文
835-841
2019-05-17（万方平台首次上网日期，不代表论文的发表时间）