SSRFD: single shot real-time face detector

Although massive breakthrough has been made in face detection, it is still worth investigating how to achieve a balance between real-time speed and high performance since effective face detectors are computationally restrictive generally. In this paper, we construct a CNN-based face detector for real-time detection, with a novel light-weight Feature Enhance Block proposed to improve the discriminability of features representation of the low-level layer in MobileNet-v3. In addition, to improve the detection of tiny faces, an anchor densification strategy is employed to ensure the same density of different types of anchors on images. Finally, we verify our proposed method by comparative experiments on four prevalent face detection datasets, including PASCAL, AFW, FDDB and WIDER FACE datasets. It is shown that SSRFD achieves good performance with the number of parameters fewer than 0.62 M and a frame-rate of 34 fps on one NVIDIA TITAN X card for 1024 * 1024 high-resolution images.