Research On Image Classification Algorithm Based On Non-negative Matrix Factorization

The “semantic gap” between low-level visual features and high-level semantics becomes the bottleneck of semantic understanding of image content. In order to reduce the semantic gap and improve the image content utilization efficiency, the core problem is how to represent image visual features. Meanwhile, the high dimensionality of image visual features will affect the recognition rate and running time. Therefore, reducing the dimension of image visual features effectively has become one of the hot spots in the field of pattern recognition,computer vision and image processing.As an important method of data representation, matrix decomposition technique has received extensive attention in recent years. The existing matrix decomposition methods include principal component analysis algorithm, linear discriminant analysis algorithm,independent component analysis algorithm, singular value decomposition algorithm and so on.Non-negative matrix decomposition(NMF) algorithm is different from these methods and is required to be non-negative in the decomposition process. The constraint of NMF is consistent with the view of psychology and physiology, namely the human perception of the whole is composed of the components of the perception. NMF is to decompose a non-negative matrix into the product of two non-negative matrices. One is called base matrix,the other one is called coefficient matrix. In the framework of NMF, the column vectors in the decomposition matrix can be interpreted as the weighted sum of all the column vectors in the base matrix. In order to improve the effectiveness of the NMF algorithm, many scholars introduced various constraints, such as sparse, orthogonality, discriminant, manifold etc. They proposed several kinds of improved algorithms which have been applied to face detection and recognition, digital watermarking, gene and cell analysis, recognition of musical instruments,sound source classification, text analysis and clustering, blind signal analysis and so on.Under the framework of NMF, this paper mainly studies how to enhance NMF performance by applying the constraints and online learning. This paper presents three improved NMF methods:(1) graph regularized and constrained NMF with sparseness;(2)incremental NMF with sparseness constraint;(3) graph regularized and incremental NMF with sparseness constraint. After dimensionality reduction and feature extraction, we choose support vector machine(SVM) for image classification. Experimental results show that the proposed algorithms can improve the accuracy and stability of image classification.