Face detection in a complex background using cascaded conventional networks

Abstract

Although significant achievements have been achieved in the field of face detection recently, face detection under complex background is still a challenge issue. Especially, face detection has wide applications in real life, such as face recognition attendance system and crowd size estimation. In this paper, we propose a novel cascaded framework to tackle the challenges based on: blur, illumination, pose, expression and occlusion. Our framework adopt the localization of facial landmarks to boost up their performance. In addition, our detector extracts features from different layers of a deep residual network for complementary information of low-dimensional and high-dimensional features. Our method achieves notable results over the state-of-the-art techniques on the challenging WIDER FACE benchmark for face detection and our results show that average precision of 89.2%. Importantly, we demonstrate superior performance and robustness in a challenging environment.