Image Semantic Information Mining Algorithm by Non-negative Matrix Factorization

This paper studies on mine the image semantic information through Non-negative Matrix Factorization, which is a powerful computing tools in many applications. Non-negative matrix factorization is a low-rank matrix approximation method for finding two low-rank nonnegative matrices and the product of which can provide a good approximation to the original non-negative matrix. Firstly, a multimodal training matrix is constructed according to the ground truth annotations of the training images dataset. Secondly, the matrix constructed in the first step is decomposed by non-negative matrix factorization. Afterwards, the untagged images with only visual features are represented as the matrix, and then semantic terms can be extracted from the matrix which represents the similarity between test and training images. Experimental results demonstrate the effectiveness of the proposed method.