关键词: 非负矩阵分解, 特征提取, 稀疏表示, 梯度下降法, 特征降维

Abstract: In order to improve the universality of image feature extraction, an image feature extraction method based on improved Nonnegative Matrix Factorization (NMF) is proposed. Firstly, considering the practical significance of extracted image feature, NMF model was used to reduce the dimension of image feature vector; Secondly, in order to represent image by a small number of coefficients, sparse constraint was added to NMF model as one of regular terms; Thirdly, to make the optimized feature have better inter-class differentiation, clustering property constraint would be another regular term of NMF model; Finally, through optimizing model using gradient descent method, the best feature basis vectors and image feature vectors could be acquired. Experiment shows that for two image databases, the acquired feature which is extracted by improved NMF will be more conducive to correct image classification or identification, and the feature extraction method is more universal.

Key words: Nonnegative Matrix Factorization (NMF), feature extraction, sparse representation, gradient descent method, feature dimension reduction