Abstract: Non-negative matrix factorization ( NMF ) is a method of parts based feature extraction, according to this characteristic. An improved algorithm dealing with NMF basis was introduced to enhance partial occlusion face recogni- lion rate. Firstly, discrete wavelet transformation was used to produce a representation in the low frectuency domain, and basic matrix was got according to the NMF method. Secondly,parts of face features which possess outstanding perfor- mance were extracted by threshold value judgments,and they were used to form optimized facial subspace feature. The training and testing images were projected to the optimized subspace feature. Finally support vector machine was used for classification. Experiments show that the improved algorithm's computing time is short, and this method achieves remarkable effects under the partially occluded circumstance.

Key words: NMF, DWT，Face recognition, Basic matrix