Supervised Hashing For Image-Text Cross-Modal Retrieval

With the development of science and technology,the multimedia information existing in different media types on the Internet increases day by day.To adapt to the development trend,cross-modal retrieval becomes an important problem.Hashing methods have attracted great attention for cross-modal retrieval due to the low memory requirement and fast computation.Cross-modal hashing methods aim to embed data from different modalities into compact binary hash codes.In this way,we can measure the similarity between different modalities via Hamming distance,and then we can perform cross-modal retrieval.In this dissertation,we mainly focus on the research of cross-modal hashing methods.And we propose two novel supervised cross-modal hashing methods.The main contributions are as follows:First,how to better preserve inter-modality and intra-modality similarity is still a challenge during the phase of mapping features in different original spaces into a common Hamming space.Most existing cross-modal hashing methods ignore the restrictions on dissimilar instances.Besides,most cross-modal hashing methods relax discrete constraints and then optimize the relaxed objective functions functions,followed by quantization to obtain hash codes.This relaxation causes quantization error and low retrieval performance.To address above problems,we propose a novel supervised cross-modal hashing method,termed Discrete Similarity Preserving Hashing(DSPH).DSPH simultaneously preserves inter-modality and intra-modality similarity.Specifically,DSPH puts restrictions on both the similar and dissimilar instances to learn more discriminative hash codes.Moreover,we design a discrete optimization method to learn discrete hash codes.Extensive experiments conducted on three datasets verify the superior performance of the proposed DSPH.Second,using labels is beneficial to improve the performance of cross-modal retrieval.And although some methods apply collective matrix factorization in cross-modal retrieval,how to better combine labels with collective matrix factorization in the training phase deserves concern.In addition,most existing cross-modal hashing methods drop the discrete constraints and optimize the relaxed objective functions,which brings retrieval performance degradation.To address above problems,we propose a novel supervised cross-modal hashing method,termed Semantic Associated Discrete Hashing(SADH).SADH regresses class labels of training instances to hash codes,and utilizes collective matrix factorization to learn hash codes directly.Moreover,in order to solve the discrete optimization problem,we use a discrete optimization algorithm to learn discrete hash codes when optimizing the objective function.Extensive experiment results on three datasets validate the effectiveness and superiority of the proposed method.