自适应特征提取的光照鲁棒性人脸识别

Adaptive feature extraction for illumination invariant face recognition

针对光照对人脸特征提取的影响,提出了一种基于多尺度Curvelet变换的自适应局部熵的光照鲁棒性人脸特征提取方法。采用特殊局部对比增强算法对光照不均衡图像进行光照补偿,同时使图像局部特征显著;通过对增强后的图像进行Curvelet多尺度分解,得到的分解系数进行分块求熵从而构成候选特征向量;通过特征鉴别能力分析和评估,对候选特征值进行最优选择。在ORL,Yale,YaleB,AR四个人脸数据库中的实验结果表明,该方法与传统的PCA,LDA方法相比,避免小样本和特征分解问题,同时具有环境适应性和抗光照影响的特点。

An adaptive local entropy feature extraction method based on the multi-resolution Curvelet transform is proposed for face recognition according to illumination.It compensates uneven illumination and makes local details notable through local contrast enhancement.Then the enhanced images are fed into the Curvelet transform.The block entropy of Curvelet coefficients consists of the candidate feature vectors.Feature discrimination power analysis and evaluations can adaptively select the most important features among all candidate features.Experimental results in ORL,Yale,YaleB,and AR face databases,show that the proposed method avoids suffering from the small sample size problem and SVD（Singular Value Decomposition）problem such as PCA and LDA.Meanwhile,it has light resistance and environmental adaptability.