基于分类外形搜索的人脸特征点定位

针对传统由粗糙到精准的人脸外形搜索方法，其每一次外形搜索需要在整个外形搜索空间进行，提出一种基于分类的外形搜索方法。该方法始于一个包含不同人脸形状的外形搜索空间，首先利用基于相关性的特征选择方法对随机森林分类器进行优化，利用训练的随机森林分类器将外形搜索空间分为若干个外形搜索子空间；然后根据输入样本和随机森林分类器确定与当前外形最接近的外形搜索子空间，并计算对应子空间的中心和对应样本的后验概率分布，方便后续阶段更好地进行外形搜索；最后采用级联回归进行人脸特征点定位。在300-W数据集上的实验结果表明，此方法不仅有效降低了外形搜索的时间，同时在无约束环境中具有良好的鲁棒性。

Face alignment based on classified shape searching

Aiming at improving the traditional shape searching which needs to search in the whole shape space each time. A new approach is proposed based on classified face shape searching. This approach begins with a shape space that contains diverse shapes, first we optimize the random forest classifiers by the feature selection method of correlation-based and train the random forest classifiers by training samples, then divides the shape into several sub-spaces by random forest classifiers, and search the sub-space that most similar to the current shape, then estimate the center of the sub-space and probability distribution. Finally, employs cascaded regression to realize face alignment. The approach demonstrates its obvious decreases in searching time and its good robustness in unconstrained environment on the 300-W database.