关键词: 独立成分分析, 平均场近似, 特征提取

Abstract: A traditional Independent Component Analysis(ICA) is a noise-free model and has some rigorous conditions. However, there are many noises. In this paper, these conditions were changed by the factor analysis in multi-variable statistics and a noisy ICA model was formed. The theory of Mean Field Approximation (MFA) in statistical physics was used to estimate the model parameters. In order to obtain more independent features in feature extraction, some restrictions were added to model parameters, such as non-negative mixing matrices and non-negative source signals. Experiments were done by two different cases that simulated graphics and ORL face databases. The simulated graphics experimental results show this proposed method can extract more independent features than the traditional ICA and the unrestricted MFAICA method do.

Key words: Independent Component Analysis(ICA), Mean Field Approximation(MFA), feature extraction