Feature-Aligned Feature Pyramid Network and Center-Assisted Anchor Matching for Small Face Detection

The performance of face detection has made great progress with the rapid development of deep learning. However, small faces only contain an extremely limited number of pixels, which makes them suffer from poor feature representation and low-quality positive anchors. These problems make small face detection still a highly challenging task In this paper, we conduct an in-depth study from the perspective of feature learning and anchor matching and present a small-scale aware face detector (SFDet), which can accurately detect small faces. Specifically, we first propose a feature alignment module (FAM), which learns the semantic transformation offset and adaptively upsamples features using it as a guide. FAM can construct a powerful feature pyramid to learn more discriminative and robust feature representation of small faces. Then, we design an IoU-and-center based anchor matching strategy (ICAMS), which introduces center point information as a new matching metric to complement the traditional intersection over union (IoU) metric and simultaneously uses these two metrics in the anchor matching stage. ICAMS can ensure the generation of high-quality positive anchors of small faces. Comprehensive experiments are conducted on two challenging face detection datasets, and the results demonstrate the effectiveness of our method.