Supervised Hashing Methods For Information Retrieval

Hashing is one popular type of approximate nearest neighbor approach.Hashing algo-rithms utilize hash functions to transform feature vectors to binary codes,thus improving search efficiency and reducing memory cost.Learning-based hashing methods learn hash functions from training data.Learning-based hashing can produce more compact hash codes and obtain better retrieval results.Learning-based hashing can be divided into three categories:unsuper-vised hashing,supervised hashing and semi-supervised hashing.This paper proposes a new supervised hashing method,i.e.,Natural Supervised Hashing(NSH),which treats label vectors as ideal semantic binary codes and wants to learn hash codes with similar structure as label vectors.The idea can be directly formulated into an inner product fitting problem,and after approximating the discrete binary value with differentiable hyperbolic tangent function,gradients of the loss function can be computed and gradient descent can be used to solve the problem.However,the above method has a long training time and the exper-imental result is not very good.To deal with it,NSH exploits one type of structure-preserving transformation to implicitly bring close inner products of hash codes and label vectors,and ends up with the final optimization problem.To solve the problem,NSH adopts an alternating itera-tion paradigm,and after simple relaxation,every step of the iteration has close-form solutions,which reduces the training time to a large degree.Besides,due to the fact that NSH needs not to explicitly retain Gram matrices in the memory,NSH scales well and can use all available data for training when running on NUS-WIDE.Experimental results on four multi-label datasets in-dicate that NSH achieves superior performance over compared methods with different bits and search procedures.To deal with multimodal scenarios where samples have multiple views,multimodal hashing methods are developed.Different views of multimodal data have different feature spaces.And multimodal hashing learns corresponding hash functions for each view,so as to transform fea-tures in different views into a common Hamming space,where difference between modalities can be ignored and hamming distance can be computed to perform intra-modality and cross-modality nearest neighbor search.In this paper,a new supervised multimodal hashing method,i.e.,Natural Multimodal Hashing(NMMH),is proposed.NMMH adopts the two-step frame-work to learn hash functions.First,to learn hash codes,NMMH follows NSH and utilizes a structure-preserving transformation to implicitly bring close inner products of hash codes and label vectors.After acquiring hash codes,NMMH learns corresponding hash functions for each modality.Besides,NMMH extends iterative quantization(ITQ)to multimodal data and after projecting features from every modality into a common space,a rotation matrix is learned to jointly minimize the quantization loss of all modalities.Experiments in different search scenar-ios are conducted to compare several multimodal hashing methods.And when unified codes are used,NMMH attains obvious superiority over other approaches.