The quality of face images varies due to complex environmental factors, and face images with extremely poor qualities would deteriorate the performance of face recognition. As one of the pre-processing modules of face recognition, face quality assessment needs to consider both environment factors and practical applications. In this paper, we propose a multibranch face quality assessment (MFQA) algorithm considering comprehensive factors acting as a reliable reference for its following recognition. A light-weight convolution neural network (CNN) is used for face image feature extraction, and quality scores corresponding with alignment, visibility, deflection and clarity are predicted by multi-branch layers. Moreover, a score fusion module is implemented by fusing the above scores to obtain a final quality confidence. Compared with other relevant quality assessment works, our method is quite suitable for practical applications because of its better performance, faster speed and smaller model size. Experiments show that our proposed method is able to assess face quality objectively, and the performance of face recognition is significantly improved by introducing our approach into its training and testing procedures.
Paper
Full text
Multi-branch Face Quality Assessment for Face Recognition
Semantic Scholar · Computer Science · 2019
Abstract
The quality of face images varies due to complex environmental factors, and face images with extremely poor qualities would deteriorate the performance of face recognition. As one of the pre-processing modules of face recognition, face quality assessment needs to consider both environment factors and practical applications. In this paper, we propose a multibranch face quality assessment (MFQA) algorithm considering comprehensive factors acting as a reliable reference for its following recognition. A light-weight convolution neural network (CNN) is used for face image feature extraction, and quality scores corresponding with alignment, visibility, deflection and clarity are predicted by multi-branch layers. Moreover, a score fusion module is implemented by fusing the above scores to obtain a final quality confidence. Compared with other relevant quality assessment works, our method is quite suitable for practical applications because of its better performance, faster speed and smaller model size. Experiments show that our proposed method is able to assess face quality objectively, and the performance of face recognition is significantly improved by introducing our approach into its training and testing procedures.