多尺度光照不变人脸特征图像的提取方法

Multiscale illumination invariant facial feature image extraction method

现实环境下的人脸认证系统性能受光照变化影响很大。本文提出一种基于各向异性扩散算法的多尺度人脸光照不变特征图像提取算法。其特点是针对人脸图像中的光照问题引入新的区间不一致描述子,并提出新的传递系数以消除传统各向异性扩散算法带来的图像光晕效应,进而形成新的各向异性扩散算法。该算法可以在多尺度空间中有效地提取不随光照变化的人脸结构特征图像,不需复杂的光照变化建模,且对训练样本无特殊要求。在Yale B及CMU PIE标准人脸库上进行了实验,结果表明该算法在低频光照域上具有很好的边缘保持能力,即使在光照变化很大的条件下也能获得良好的处理效果,并明显的降低了人脸认证的错误率。

The performance of face verification system is greatly affected by illumination condition.This paper presents an anisotropic diffusion based multiscale illumination invariant facial feature image extraction method.To eliminate the image halo effect of traditional anisotropic diffusion algorithm,a novel anisotropic diffusion algorithm is proposed,within which a new local inhomogeneity and a conduction function are defined.The proposed method can extract illumination invariant facial feature image from multiscale space without the needs of complex illumination modeling and special requirement to training samples.Experiments on Yale B and CMU PIE face databases show that the proposed method can preserve edge information in low frequency illumination fields,and achieve promising results even under harsh illumination conditions.The proposed method also reduces the error rate of face verification evidently.