Hybrid attention network and center-guided non-maximum suppression for occluded face detection

The face detection technique has obtained significant development with the huge application of convolutional neural networks. However, various types of occlusion are widespread in face detection, inevitably destroying the visual features of faces and significantly increasing the difficulty of post-processing. These problems make the occluded face detection a challenging and crucial task. In this paper, we propose a new occlusion-aware face detector (OFDet) to deal with the problem of occluded face detection, which mainly includes a hybrid attention module (HAM) and a center-guided non-maximum suppression (cgNMS) algorithm. Specifically, the HAM consists of three types of attention blocks, i.e., spatial attention block (SAB), channel attention block (CAB), and channel-spatial attention block (CSAB), integrated in a hybrid manner. This module can help the network learn more discriminative and robust feature representation by adaptively highlighting the features of more informative visible facial regions and weakening the features of occluded facial regions, contributing to solving the inter-class occlusion issue. The cgNMS introduces the information of center point distance between detected boxes as a new suppression metric to supplement the traditional intersection over union (IoU) metric. This dual-metric design of cgNMS can ensure that it makes the correct post-processing from highly overlapped detected boxes to deal with the intra-class occlusion problem. Experimental results show that our OFDet achieves state-of-the-art results on the MAFA dataset and obtains competitive results on the WIDER FACE and FDDB datasets, which demonstrate the effectiveness of our method. In addition, HAM and cgNMS are highly efficient, and their cost basically does not affect the efficiency of the model.