Research On Cross-modal Hashing Retrieval Algorithms Based On Latent Semantic Learning

Cross-modal retrieval is a special retrieval way for two modalities in the field of multimedia retrieval.It achieves accurate cross-matching by constructing semantic correlations between different modalities.In recent years,cross-modal hashing retrieval has become the mainstream research direction in cross-modal retrieval due to its fast retrieval speed and low storage cost.The existing cross-modal hashing retrieval methods are designed to transform heterogeneous features into compact binary codes,and use supervised information to improve performance.However,most of methods discard the discrete constraints in the process of hash code generation,resulting in the accumulated quantization errors.And the general use forms of supervision information ignore the semantic corrections among categories.To address the above problems,this paper studies two cross-modal hashing retrieval algorithms based on latent semantic learning,which effectively narrows the semantic gap between heterogeneous modalities.(1)The Semantic-Guided Hashing retrieval algorithm is proposed to solve the problem of neglecting the semantic corrections among categories.Firstly,it generates the semantic representation of category names by word embedding model,and uses the semantic representation to construct the class-level semantic space.Then,according to the encoder-decoder paradigm,a bimodal autoencoder model based on class semantics is designed to learn the projection from original feature space to common space that retains all the information in the original feature,which extends the algorithm to the unseen domain.(2)The Self-Taught Cross-Modal Hashing retrieval algorithm based on minimal semantic loss is presented to focus on large quantization error in the process of hash code generation.It adopts collective matrix factorization technology to learn the common semantic representation of different modalities,and maps the common features to the low-dimensional Hamming space by binary quantization process with orthogonal constrains,while ensuring that the semantic loss is minimized.In addition,the algorithm regards the hashing function learning as a binary classification problem,and further decreases the quantization error by the advantage of Support Vector Machine.This paper designed a large number of experiments on public datasets,and the proposed two algorithms were evaluated by various performance metrics.The experimental results show that the semantic-guided hashing retrieval algorithm can achieve superior performance in retrieval tasks and also demonstrate the effectiveness and applicability of the algorithm in the unseen domain retrieval problem.Furthermore,the self-taught cross-modal hashing retrieval algorithm shows the significant performance advantage in the experiments,which proves the importance of minimizing semantic encoding loss.To sum up the above two experimental results,two algorithms proposed in this paper show excellent performance,and achieve fast and accurate retrieval results in cross-modal retrieval tasks.