Abstract: The?traditional?Principal?Component?Analysis（PCA）?requires that training samples are in accordance with Gaussian distribution strictly. However, generating pictures are always influenced by illumination, facial expressions, and postures. In order to solve the problem, a new modular algorithm based on PCA is proposed, which is also a guarantee of the rate of identification. The new algorithm, on the one hand, takes a blocked mode which divides pictures with a same posture into one matrix, so the training sample can be closer to the Gaussian distribution. On the other hand, since the first three characteristics of the principal component are easily affected by light variation, a less than one weighting coefficient is added to reduce the effects of light in the recognition. Thus the improved PCA training matrix is no longer limited to the Gaussian distribution with the combinations of the sub-blocks and the weight coefficients, the recognition rate is improved at the same time. The numerical experiments in the ORL human face databases show that the improved algorithm is superior to the traditional PCA algorithm.

Key words: principal components analysis, face recognition, weight coefficient, improved Principal?Component?Analysis（PCA） method

摘要： 由于传统的PCA要求训练样本符合高斯分布，而现实中得到的图片往往由于光照、表情、姿态的不同，不符合高斯分布。为了使PCA不再局限于高斯分布，并且不影响其识别率，提出一种改进的模块PCA人脸识别新算法。一方面，新算法采取了分块方式，将具有同一姿态的图片划分进同一矩阵，以使训练样本更接近于高斯分布。另一方面，新算法对传统PCA算法中前三个主分量加小于1的权重系数，可以减少光照变化对识别率的影响。利用分块和权重系数的共同作用使得PCA不再局限于高斯分布，同时提高识别率。最后在ORL人脸库上进行实验，结果表明新算法优于传统的PCA算法。

关键词: 主成分分析, 人脸识别, 权重系数, 改进的主成分分析（PCA）算法