Proceedings of the 13th EAI International Conference on Mobile Multimedia Communications, Mobimedia 2020, 27-28 August 2020, Cyberspace

Research Article

A Novel Sample-Enhanced Dataset based on MFF for Large-Angle Face Recognition

Download372 downloads
  • @INPROCEEDINGS{10.4108/eai.27-8-2020.2297864,
        author={He  Wang and Yan  Wang and Jie  Liu and Guisheng  Ying},
        title={A Novel Sample-Enhanced Dataset based on MFF for Large-Angle Face Recognition},
        proceedings={Proceedings of the 13th EAI International Conference on Mobile Multimedia Communications, Mobimedia 2020, 27-28 August 2020, Cyberspace},
        publisher={EAI},
        proceedings_a={MOBIMEDIA},
        year={2020},
        month={11},
        keywords={enhanced-dataset large angle mff face recognition},
        doi={10.4108/eai.27-8-2020.2297864}
    }
    
  • He Wang
    Yan Wang
    Jie Liu
    Guisheng Ying
    Year: 2020
    A Novel Sample-Enhanced Dataset based on MFF for Large-Angle Face Recognition
    MOBIMEDIA
    EAI
    DOI: 10.4108/eai.27-8-2020.2297864
He Wang1, Yan Wang1,*, Jie Liu1, Guisheng Ying1
  • 1: College of Computer Science and Technology, Harbin Engineering University, China
*Contact email: wangyanj@hrbeu.edu.cn

Abstract

Large-angle face recognition has always been a huge challenge due to the scarcity of large-angle dataset. In this paper, a novel sample-enhanced dataset is constructed, which is composed of various angle face picture samples from -90° to 90° relative to the front face. The constructed dataset is obtained by enhancing large-angle face samples of the CASIA-WebFace dataset. The large-angle face samples are generated from small-angle face samples of the CASIA-WebFace dataset, which is based on the multi-task feature framework (MFF). By employing these sample datasets, four trained FaceNets are achieved for face recognition. Finally, to test the effectiveness of the four face recognition networks for the large-angle face, 300 large-angle face pictures of different basketball players are selected as the samples of the experiment. The results demonstrate that the accuracy of large-angle face recognition has been greatly improved when utilizing the FaceNet that is trained by the novel enhanced dataset.