Research On Cross-modal Retrieval Based On Discriminative Discrete Hashing

With the arrival of big data information era,multimedia data has shown explosive growth in all professions and trades,which contains rich social and economic value.Due to the gigantic amount and variety of multimedia data,how to perform efficient cross-modal retrieval from massive data has become a fundamental problem for effectively using these data resources.Hashing-based cross-modal retrieval methods can deal with the problems of high data dimension,heavy storage cost and low retrieval efficiency.Although most existing methods have achieved considerable results,there are still some problems to be solved,such as high algorithm complexity,discarding the discrete constraints of the hash codes and insufficient similarity preserving of origi nal data,which lead to the algorithms unable to achieve satisfactory retrieval performance.This thesis focuses on the above problems,and proposes two novel supervised cross-modal hashing methods,which has the following main contributions:(1)Most existing methods have high training complexity and learn the hash codes without the discrete constraints,making them unscalable to large-scale data applications and bringing up undesired performance.To overcome above shortcomings,this thesis proposes a novel supervised method,named Fast Discrete Matrix Factorization Hashing(FDMFH).Specifically,it utilizes matrix factorization to learn a latent semantic space constructed by labels for each modal.Then,it quantifies the latent semantic repres entation into the hash codes by rotation quantization.Meanwhile,it transforms the hash codes learning procedure into the classification problem between hash codes and labels,which further improves the discriminability of hash codes and maintains the lin ear training complexity.Moreover,an efficient discrete optimization strategy is proposed for one-step learning of the hash codes.Extensive experiments on two benchmark datasets,MIRFlickr and NUS-WIDE,demonstrate that FDMFH has better training efficiency and higher retrieval accuracy than several state-of-the-art methods.(2)Several existing methods can not fully maintain the similarity between the original data and discard the discrete constraints of the hash codes,which leads to suboptimal retrieval performance.To address these problems,this thesis proposes a novel supervised method,named Discrete Cros s-Modal Hashing with Semantic Correlation(DCMH-SC).It utilizes the inner products of the hash codes to reconstruct the pairwise similarity matrix constructed by the class labels non-explicitly.Meanwhile,the linear regression term between the hash codes and the labels is introduced to maintain the pairwise semantic correlation between different modalities,which further improves the discriminability of to-be-learnt hash codes.In addition,an ALM-based discrete optimization strategy can learn the hash codes with the discrete constraints in a single step.Extensive experiments on two benchmark datasets,MIRFlickr and NUS-WIDE,demonstrate that DCMH-SC outperforms several state-of-the-art methods.