Sparse Coding Hashing For Cross Modal Multimedia Retrieval

Similarity search methods based on hashing are useful tools for effective cross-modal retrieval, which have drawn extensive attention. The core issue of the cross-model retrieval is how to build relevance between multi-modal data features. The goal of cross-model retrieval is to retrieve data between different models. Cross-Modal Hashing is to map data features of different models to the same low-dimensional Hamming space. Since the data features representation of multimedia of different dimensions are eventually converted into binary representation under the same dimensions, therefore, the efficient retrieval can be implemented by cross-model hashing between the different models. Similar to canonical correlation analysis (CCA), most of the existing cross-modal hashing is used to retrieve data from the same models or within the same angle of view which mapping different models features into a common abstract space. However, this model does not well reduce the semantic gap between different modalities, and extracting a higher level semantic information can be more beneficial. To solve this problem, in this paper, we propose a novel Sparse Coding Hashing for Cross Modal Multimedia Retrieval. Extensive experiments on two different data-sets demonstrate the advantage of our method, and draw the corresponding experimental results. The main contents are as follows:By using sparse coding for cross-modal multimedia retrieval, SCH uses Sparse Coding to learn the latent concepts of text and capture the salient structures from images. Then the learned latent semantic features are mapped to a joint abstraction space. In addition, an iterative strategy is used to find the correlation representation of multi-modalities. Through the high level abstraction space, the unified hash-codes are obtained by quantification. Meanwhile, in order to solve the target function we converted the non convex optimization problem of five matrix variables into four matrix variables. Therefore, the optimization problem can be solved using an iterative strategy with the following listed steps until convergence.In order to verify the effectiveness of the algorithm, evaluate the propose method on the Wiki data-sets and NUS-WIDE data-sets. Experimental results show it’s the superiority of the proposed algorithm.