Cross-modal Image-text Retrieval Based On Two-step Hashing Method

In recent years,hash method has attracted wide attention in cross-modal image-text retrieval due to its small storage space and high retrieval efficiency.The cross-modal image-text retrieval method based on hash learning is to project data instances from different models into Hamming space and learn hash codes for image-text retrieval.Although the cross-modal image-text retrieval hash learning method achieves excellent retrieval performance,it still has some limitations.On the one hand,most of the existing methods learn hash code and hash function at the same time,resulting in more matrix variables in the training process,increasing the computational complexity of training.On the other hand,some methods use linear regression strategy to learn the hash function,which makes the hash function not flexible enough,leading to the retrieval performance can not achieve the best.To solve the above problems,this thesis proposes two cross-modal image-text retrieval hash methods.The main research results are summarized as follows:Aiming at the problem of high computational complexity caused by simultaneous learning of hash code and hash function in cross-modal image-text retrieval hash method,a method is proposed in this thesis,called a fast discrete two-step learning hash method.First,the method uses class label matrix to generate hash code instead of traditional similarity matrix.Then,a two-step learning scheme is proposed to learn two different modal correlation hash functions,that is,generate hash codes first,and then learn hash functions from the learned hash codes.Finally,a discrete optimization strategy is adopted to solve the hash function.Aiming at the problem of cross-modal image-text retrieval hash method using linear regression strategy to learn inflexible hash function,which leads to poor performance.On the basis of the above work,this paper introduces the kernel logistic regression strategy to make the hash function suitable for different prediction models,and proposes a fast discrete two-step learning hash method based on the kernel logistic regression.First,hash codes are generated by class labels to avoid the construction of paired similarity matrices.Second,in the stage of learning hash function,advanced kernel logistic regression algorithm is adopted to enhance the flexibility of hash function,so as to improve the performance of cross-modal image-text retrieval.Finally,a discrete optimization strategy is adopted to solve the hash function.To verify the effectiveness of the proposed method,comparative experiments are carried out on three classical data sets,and the experimental results show that the two proposed methods in this paper have higher retrieval performance and learning efficiency.At the same time,it is verified that the fast discrete two-step learning hash function based on kernel logistic regression is suitable for different prediction models and has higher flexibility.