基于差分特征的多视角人脸检测

Differential features based multi-view face detection

摘要:

针对真实环境下的多视角人脸检测问题，提出了一种基于差分特征的多视角人脸检测算法。它综合运用2种不同的差分特征：一阶NPD特征与二阶Laplace特征，结合Gentle Adaboost算法与分类回归树(CART)，分别训练基于一阶和二阶差分特征的人脸检测器，再将这2种差分特征的检测结果进行融合，得到最终的人脸检测结果。本文的差分人脸检测器充分利用了2种差分特征的互补性，结合了一阶特征对光照的鲁棒性和二阶特征对旋转的鲁棒性，从而更好地实现了复杂环境下的多视角人脸检测。在CMU-MIT和FDDB两大公开人脸检测数据集中对提出的方法进行验证，结果证明了本文提出的差分人脸检测器的有效性，能够较好地检测复杂环境下的多视角人脸。

Abstract:

Pose and illumination variations are two major challenges in face detection. Therefore，a novel face detection method based on differential features is proposed. This method extracts first order and second order differential features from images, which are respectively used to train two face detectors using the Gentle Adaboost algorithm with the Classification And Regression Trees(CART) as weak classifiers. Given a new image, the two face detectors are first separately applied to detect candidate faces in the image, and then their detected face regions are combined to give the final face detection results. Thanks to the illumination invariance of first order derivative features and to the rotation invariance of second order derivative features, the proposed differential features based face detection method can better handle the detection of multi-view faces in complex background. The proposed method has been evaluated on the CMU-MIT and FDDB datasets and the results demonstrate its effectiveness.