Face detection: A deep convolutional network method based on grouped facial part

In this paper, a novel method is proposed for face detection, which is of simple structure but robust to severe occlusion. In detail, the size-free images are firstly segmented to a series of candidate windows. Then these candidate windows are further filtered by grouped facial part networks to generate a set of windows which may contain faces. Finally, the set of face proposals are input to a multi-task deep convolutional network (DCN) for further classification and calibration. Importantly, we take the spatial position relations of local facial parts into consideration and find it helpful to handle the severe occlusion. Our method achieves outstanding performance on the widely used datasets FDDB and AFW, compared to the other proposed face detectors. Especially on FDDB, our method achieves a high recall rate of 90.13%.