Deep Learning To Hash For Large-Scale Multimedia Retrieval

Retrieving from the increasing scale of multimedia data on the Internet efficiently and accurately is an important problem.Due to the storage and retrieval efficiency,hashing methods are receiving attention recently,and have been widely adopted to approximate nearest neighbor search systems for large-scale multimedia retrieval.For most existing hashing methods,the multimedia data is firstly represented as vectors of hand-crafted or machine-learned features,followed by a separate step of learning to hash that generates binary codes.Nevertheless,these two-step methods have several disadvantages.First,because of the discrete nature of hash codes,these methods usually learn continuous low dimensional representations and then convert them into binary codes.However,the quantization error between continuous representations and binary codes is not statistically minimized during binarization,which results in an uncontrolled gap between the learned representations and final codes.Second,separating feature extraction and learning stage,the quality of feature representation itself might limit the quality of learned hash codes.To solve these inherent problems,this paper presents a deep hashing framework for multimedia data retrieval.Deep hashing framework goes beyond traditional methods in three aspects.First,inheriting from deep learning,deep hashing can convert raw multimedia data to hash codes end-to-end,which solves the synergistic optimization problem between feature representations and hash codes.Second,benefiting from the powerful fitting ability of deep neural network,deep hashing could help learn high quality and nonlinear hash function that converts continuous representations to hash codes,which also preserves the similarity relationship.Third,a novel quantization loss is proposed in deep hashing framework to reduce the gap between continuous representations and binarized codes,which improves hashing quality and compression ratio significantly.In addition,a method called Deep Hashing Network(DHN)is proposed using Bayesian estimation in the deep hashing framework,to solve image hashing and retrieval problem.The proposed deep hashing framework and DHN method is implemented in Caffe and TensorFlow,and comprehensive experiments are conducted to prove the substantial multimedia retrieval accuracy boosts over the state-of-the-art hashing methods.