Research On Supervised Discrete Hashing Method Based On Dictionary Learing For Cross-modal Retrieval

In recent years,with the development of Internet and social media,the scale and type of multimedia data grow rapidly.Given a query in one modal,the ideal return of cross-modal retrieval is the similar instances of another modal,for example,use text to retrieve relevant images or videos.Hashing methods have attracted considerable attention due to their high retrieval speed and low storage costMany cross-modal hashing methods are proposed and achieve considerable results However,these methods still have some limitations.First,most hashing methods di-rectly use linear projection matrices to project multi-modal data to common Hamming space.Due to the complexity of heterogeneous data,linear projection is hard to pre-serve similarity,which is the learning goal of hashing methods.Second,the rule for most supervised cross-modal hashing methods to construct a similarity matrix is that two instances are considered similar if they share at least one label.However,this kind of definition is too simple and loses a lot of useful information.Finally,most hash-ing methods try to relax or drop discrete constraints to solve a continuous optimization problem,and then quantize the real-value solution into binary codes,which causes huge quantization errors.Some papers have proposed discrete optimization strategies,but these methods are usually only suitable for single modeTo overcome these limitations,this paper proposes a novel cross-modal hashing method called Dictionary Learning based Supervised Discrete Hashing(DLSDH).DLS-DH is a two-step hashing method,divided into a hash codes learning process and a hash functions learning process.The pairwise similarity matrix and discrete iterative opti-mization strategy used in DLSDH make the hash codes learning process learn higher quality hash codes.In the learning phase of the hash functions,DLSDH first generates a sparse representation for each instance,and then maps it into the low-dimensional Hamming space.The main contributions of the method proposed in this paper are as follows,·DLSDH proposed a dictionary-based supervised cross-modal hashing method.First,a dictionary is learned for each modality and a sparse representation is gen-erated for each instance.Compared to the complex original feature,sparse repre-sentation is more suitable for mapping to low-dimensional space.In this way,the problem that similar instances are difficult to be close in low-dimensional space is solved.·In order to make full use of the original label information,DLSDH uses cosine similarity to construct pairwise similarity matrix,which can better supervise the hash codes learning process·DLSDH uses a column sampling strategy to solve the discrete optimization prob-lem and avoid errors caused by quantization.At the same time,DLSDH uses the entire data set instead of sampled data during training,which ensures the gener-alization of the hash functions·The results of comparative experiments on three datasets with state-of-the-art cross-modal hashing methods demonstrate that the retrieval performance of DLS-DH is better than or comparable to several state-of-the-art methods.In addition,experiments prove the validity of the cosine similarity matrix.