Multiple Task-driven Face Detection Based on Super-resolution Pyramid Network

Abstract

Although research in face detection and recognition has achieved tremendous progress through the various frameworks that are being put forward every year, face detection under complex circumstances is still a challenging issue. Multiple task-driven face detection has wide applications, such as crowd number estimation, face recognition attendance and so on. In this paper, we propose a multiple task-driven cascade detection networks based on super-resolution Pyramid, to effectively tackle the following challenges in face detection: low-resolution faces under the lens; faces from blur, illumination, scale, pose, expression and occlusion. Our method integrates the advantages of the super-resolution technology and an efficient image pyramid structure. The design of this structure not only recover high frequency information lost in the sampling process, but also can handle multi-scale invariants. Also, facial landmarks play non-negligible roles during detection. Our method achieves state-of-the-art results over prior arts on both the WIDER FACE dataset and the Face Detection Dataset and Benchmark (FDDB), and our results show a higher average detection precision of 90%. Notably, we demonstrate superior performance and robustness in a challenging environment.