Geometric and Semantic similarity preserved Non-Negative Matrix Factorization for image retrieval

Preserving both geometric and semantic similarities simultaneously is of vital importance to enhance the performance of an image retrieval system. However, most graph-based non-negative matrix factorization methods only preserve the geometric similarity and ignore to hold the semantic similarity in image retrieval. To handle this problem, an algorithm called Geometric and Semantic similarity preserved Non-negative Matrix Factorization (GSNMF) is proposed. GSNMF combines two kinds of regularization, i.e. one is geometric similarity regularization and the other one is semantic similarity regularization, in a united framework. Numerous experiments show the superiority of GSNMF over several current state-of-the-art methods on publicly available dataset.