MACNet: A Lightweight Face Detector

Abstract

In order to meet the requirements of face detection between the accuracy and efficiency, a lightweight face detector named MAC-Net, based on fusing context information and extracting feature channel attention is proposed. This network has designed a new block to construct a network. In this new block, we add the context information module and feature channel attention module to enrich features. The image feature pyramid is used to achieve multi-scale features learning. By adding a facial landmark module, the quality of face detection is inproved. Two models of different sizes are trained and tested on datasets such as AFW, FDDB, and PASCAL face. Experiments show that our lightweight algorithms can reach state-of-the-art with an accuracy of 99.28%, 97.0%, and 97.91% on these three datasets respectively. Our lightweight MAC-Net can run at 18 FPS on a CPU device for VGA-resolution images.