Lighting normalization is a kind of widely used approach for achieving illumination invariant face recognition. Lighting normalization approaches try to regularize various lighting conditions in different face images into ideal illumination before face recognition. However, many existing methods perform lighting normalization by treating face images as natural images, and neglect the particular properties of faces, e. g. face symmetry. As a result, for the face images with side lighting, many existing methods cannot recover the facial features in shadow regions. To resolve this problem, in this paper, a novel lighting normalization approach exploiting face symmetry priori is proposed for illumination invariant face recognition. In the proposed approach, lighting normalization for a shadow region is performed by referring to the face structure information of a symmetric non-shadow region. The symmetry priori of face structure is modeled via an energy minimization framework. In addition, a shadow-free reliability map is further proposed to simplify the original bivariate optimization problem into a univariate one in order to reduce the computation cost. Experiments on face images with synthetic and real shadows show that the proposed lighting normalization approach is effective in recovering facial features in shadow regions of a face, and also robust to face misalignment and asymmetric face geometric normalization.%光照归一化在光照鲁棒的人脸识别中被广泛使用.许多现有光照归一化方法将人脸图像视为自然图像,而忽略了人脸这一类特定物体的先验属性,因此很难从一幅具有侧光的人脸图像中恢复阴影区域中的人脸信息.提出了利用人脸对称性先验的光照归一化方法,在能量最小化框架下,对人脸图像的阴影区域进行光照归一化时参考其对称非阴影区域中的人脸结构信息,同时提出了无阴影信度图将二元最优化问题简化为一元最优化问题,以降低光照归一化方法的计算代价.在合成阴影和真实阴影人脸图像上的实验表明,利用人脸对称性的光照归一化方法能有效恢复图像阴影区域中的人脸特征,并对人脸误配准和非对称几何归一化具有一定的鲁棒性.
展开▼
