Research On Hashing Method For Cross-media Retrieval

In recent years,with the arrival of the development of Internet and the popularity of portable devices,there are more and more multimedia(e.g.photos,audio,video and texts)on the web.Large-scale cross-media retrieval has become a new challenge.On the one hand,because of the heterogeneous of low-level feathers for different modalities,the key problem for cross-media retrieval is how to measure the similarity between samples from different modalities.On the other hand,the number of multimedia data is large,and the dimension of data representation is high too.How to accurately and efficiently retrieve samples in heterogeneous data becomes an urgent need.Hashing method,which maps high-dimensional data to a low-dimensional shared Hamming space,provides an effective way for cross-media retrieval.This dissertation focuses on the research of hashing methods for large-scale images and texts cross-media retrieval,and the innovative achievements are as follows:(1)The values of hashing code does not contain any meaning in most existing hashing methods.To address this problem,we propose an unsupervised hashing method,named Projective Dictionary Learning Hashing(PDLH).Firstly,a sharing space is learned by projective dictionary learning.However,the computing complexity of traditional dictionary learning method is too high,because of the sparse item.In this dissertation,we propose to use a substitution of linear mapping for nonlinear sparse coding,which can reduce the computing complexity.And then an iterative optimization method is proposed to achieve a local optimal solution.At last,an orthogonal rotation matrix is learned by minimizing the quantization loss to improve the performance.(2)Most works generally ignores to preserve the intra-model consistency for learning a sharing subspace.To address this problem,we propose a cross-media hashing method,named Semantic Consistency Hashing(SCH).Firstly,we propose to preserve inter-model and intra-model consistency by collective non-negative matrix factorization and neighbor-preserving algorithm respectively to learn a better sharing semantic subspace.Then an efficient optimal method is proposed to achieve low computational cost(O(N)),which has good scalability.Finally,experimental results conducted on two public datasets show the effectiveness of SCH.(3)When labels are available,existing cross-media hashing works only take advantage of label-based similarity to learn hashing functions,and ignores the diversity of similarities among samples.We propose a Supervised Coarse-to-Fine Semantic Hashing(SCFSH)method.Firstly,we formulate to utilize the feature information and the labels to construct a refined similarity matrix.And the the refined similarity matrix is utilized to learn more discriminative hashing functions.The experimental results compared with baseline methods verify the effectiveness of the proposed method.(4)Existing cross-media hashing works ignore the unbalanced semantic gaps between different modalities and high-level semantic concepts.To address this problem,we propose a discrete semantic alignment hashing(DSAH)for cross-media retrieval.First,DSAH formulates to exploit collaborative filtering to mine the relations between semantic labels and hashing codes,which can reduce computational cost and memory consumption.Then the attribute of image modality is employed to align semantic information with the text modality.Finally,to further improve the quality of hashing codes,we propose a discrete optimization algorithm to directly learn hashing codes.Extensive experiments on two public databases show its superior performance over compared cross-media hashing methods.