Manchester Metropolitan University's Research Repository

    From Eyes to Face Synthesis: a New Approach for Human-Centered Smart Surveillance

    Chen, Xiang, Qing, Linbo, He, Xiaohai, Su, Jie and Peng, Yonghong ORCID logoORCID: https://orcid.org/0000-0002-5508-1819 (2018) From Eyes to Face Synthesis: a New Approach for Human-Centered Smart Surveillance. IEEE Access, 6. pp. 14567-14575.

    Published Version
    Download (7MB) | Preview


    With the popularity of surveillance cameras and the development of deep learning, significant progress has been made in the field of smart surveillance. Face recognition is one of the most important yet challenging tasks in human-centered smart surveillance, especially in public security, criminal investigation and anti-terrorism, and so on. Although, the state-of-the-art algorithms for face recognition have achieved dramatically improved results and have been widely applied in authentication scenario, the occlusion problem on face is still one of the critical issues for personal identification in smart surveillance, especially in the occasion of terrorist searching and identification. To address this issue, this paper proposed a new approach for eyes-to-face synthesis and personal identification for human-centered smart surveillance. An end-toend network based on conditional generative adversarial networks (GAN) is designed to generate the face information based only on the available data of eyes region. To obtain photorealistic faces and identitypreserving information, a synthesis loss function based on feature loss, GAN loss, and total variation loss is proposed to guide the training process. Both the subject and objective experimental results demonstrated that the proposed method can preserve the identity based on eyes-only data, and provide a potential solution for the identification of person even in the case of face occlusion.

    Impact and Reach


    Activity Overview
    6 month trend
    6 month trend

    Additional statistics for this dataset are available via IRStats2.


    Repository staff only

    Edit record Edit record