改进的求解独立成分分析参数的算法及应用

Improved method of solving the parameters of independent component analysis model and its application

提出改进的求解独立成分分析参数的算法。该算法采用平均场近似原理,通过改变模型参数的限制条件,解决传统ICA算法中不能解决的问题。在图像特征的提取中,将模型中的混合矩阵和源信号均限制为非负,使提取的特征更为独立。利用语音信号、仿真图形及剑桥大学ORL人脸数据进行实验,从实验结果可以看出,改进的算法能有效地解决源信号数大于观测信号数的问题,同时该算法能分离出更独立的人脸五官特征,对ORL人脸识别结果分析表明,识别率从原来的87%左右提高到95%左右,也说明该算法比传统的算法更有利于进行图像特征提取及识别。

A traditional ICA is a noise-free model and has some rigorous conditions. However, in fact, there are many noises, and people don't know the relations between the number of sources and the number of observations. In the paper, the authors make full use of the theory of mean field approximation (MFA) in statistical physics to estimate the model parameters. In order to obtain more independent features from extraction, a scheme by adding some restrictions to model parameters, such as non-negative mixing matrices and non-negative source signals is proposed. Experiments have been done for several different cases, such as, speech signals, simulated face graphics and ORL face recognition. The over-complete case can also be separated well in speech signals experiment by the proposed method. The simulated graphics experimental results show this proposed method can extracted more independent face features, and the recognition ratio can be improved from 85% to 95% in the ORL face recognition experiment.