Content-based image retrieval using student's t-mixture model and constrained multiview nonnegative matrix factorization

Hongqing Zhu, Qunyi Xie

Abstract: The expensive and time-consuming effort required for archiving images is the main motive for developing an effective retrieval system. This paper addresses a competitive scheme for Content-Based Image Retrieval (CBIR) based on a constrained multiview Nonnegative Matrix Factorization (NMF) that has the ability to generate a sparse representation. The scheme blends multiple visual features, which can together reflect the content of images in terms of similarity metrics and the Frobenius norm. Then, the proposed method constructs a similarity-preserving matrix factorization via an improved NMF, where the structural constraint, L 1/2-sparse constraint and farness-preserving constraint are integrated into the objective function of conventional NMF. In this way, the structure and content of high-dimensional feature data source can be preserved in low-dimensional space. Another critical part of the proposed system is to establish Student’s t-Mixture Model (SMM) based on a Markov Random Field (MRF), which can best manipulate the clustering of sparse representations according to the statistical properties of the image features. With this method, the task of image retrieval of the whole dataset is reduced to a nearest-neighbour search in a specific category containing the query image. Convergence of the proposed update rule, investigated in this study, is also verified by numerical simulations. Lastly, we conduct experiments on public datasets to compare the performance of the proposed algorithm with existing works in terms of Precision and Recall Rates. The encouraging results indicate the effectiveness of the proposed technique.