多分辨率人脸图像配准方法研究

Multi-resolution Joint Face Alignment

人脸图像配准是人脸检测的后续环节,用于纠正检测结果中存在的空间配准误差。文章提出了一种多分辨率的人脸图像联合配准方法:在粗粒度上,算法通过在较低分配率上处理图像来消除主要的配准误差;在细粒度上,算法采用较高的分辨率对配准结果进行改良。在各个粒度上,将图像配准定义为所有图像信息熵和的最小化问题,并采用牛顿优化算法求解配准参数。文章在AR和Yale B两组测试图像集构造对比实验,结果证明,多分辨率人脸图像配准方法在配准效果和算法效率上优于业界主流的联合配准算法。

Face alignment,usually a following step of face detection,aims at correcting unwanted spatial mis-alignments.This paper proposes a multi-resolution solution to joint face alignment：in coarse levels,images are processed with at low resolutions to remove major mis-alignment errors;in fine levels,alignment is refined using higher resolutions.In each level,the joint alignment problem is defined as the minimization of a sum-ofentropy function calculated over all images,and transformation parameters are simultaneously estimated by a Newton-type optimization method.The article conduct comparison experiments on images from two databases,AR and Yale B,and results prove that the proposed algorithm is robust to large mis-alignment errors,and more computationally efficient than compared methods.