基于相对索引特征回归的人脸对齐算法

Face alignment algorithm based on relative-indexed features regressing

针对传统人脸对齐算法对于较大人脸姿态鲁棒性较差,并且对于人脸检测结果十分敏感的问题,提出了一种基于改进的局部二值特征的人脸对齐算法。不同于传统的形状索引特征,在一种空间依赖假设下的前提下设计相对索引特征。同时,在这种空间依赖假设下使用半全局线性学习替代全局线性学习。通过人脸对齐敏感性分析比较了算法在不同人脸检测器下的鲁棒性,在300-W基准数据集上的实验表明:算法优于传统的级联回归算法。

Aiming at problem that traditional algorithm is inefficient in large shape variations and different face detectors,a face alignment algorithm based on improved local binary features is proposed.Different from traditional shape-indexed features,a relative-indexed features under a spatial interdependency assumption is designed.Moreover,this algorithm abandon a global linear learning and use a semi-global linear learning under this spatial interdependency assumption. A face alignment sensitivity analysis is used to compare robustness the algorithm under different face detector. Extensive experiments on 300-W benchmark dataset demonstrate that this algorithm outperforms the traditional cascaded regression algorithm.