基于视频的无监督人脸质量评估方法

Unsupervised Face Image Quality Assessment Method Based on Video

为解决视频场景下人脸图像因光照、姿态、模糊等因素变化导致的人脸质量评估方法评估效果不佳的问题,提出了一种基于视频的无监督人脸质量评估方法。考虑到当前研究社区未有完备的视频人脸数据集,且人脸图像质量评估有较强的主观性,构建了一个基于视频帧的人脸数据集,并设计了人脸质量伪标签生成方法。在此基础上,提出了基于EMD(Earth Mover’s Distance)损失的视频人脸质量评估模型,通过EMD损失函数优化模型学习,完成了人脸图像质量分布的预测。在公开和构建的测试集上的实验结果表明,该方法在视频场景下的人脸质量评估任务中优于其他人脸质量评估方法,有更好的评估性能和应用前景。

In order to solve the problem that the existing face image quality assessment methods are not effective due to the changes of illumination, posture, blur and other factors, a video-based unsupervised face image quality assessment method is proposed in this paper. Given the lack of video face dataset and the image quality assessment is somewhat subjective, a video frame face dataset was constructed and a method for generating pseudo label of face image quality is designed. Moreover, a video face image quality assessment model based on EMD(Earth Mover’s Distance) loss is proposed. By learning the EMD loss function optimization model, the face image quality distribution is predicted. Experimental results on both public and constructed datasets show that the proposed method is superior compared to other face image quality assessment methods in video scenarios with promising application prospects.