L2RT-FIQA: Face Image Quality Assessment via Learning-to-Rank Transformer

Zehao Chen, Hua Yang

Abstract: Face recognition (FR) systems are easily constrained by complex environmental situations in the wild. To ensure the accuracy of FR systems, face image quality assessment (FIQA) is applied to reject low-quality face image unsuitable for recognition. Face quality can be defined as the accuracy or confidence of face images being correctly recognized by FR systems, which is desired to be consistent with recognition results. However, current FIQA methods show more or less inconsistency with face recognition due to the following four biases, including implicit constraint, quality labels, regression models, and backbone networks. In order to reduce such biases and enhance the consistency between FR and FIQA, this paper proposes a FIAQ method based on Learning to rank (L2R) algorithm and vision Transformer named L2RT-FIQA. L2RT-FIQA consists of three parts: relative quality labels, L2R framework, and vision Transformer backbone. Specifically, we utilize normalized intra-class and inter-class angular distance to generate relative quality labels; we employ L2R model to focus more on the quality order rather than the absolute quality value; we apply unpretrained vision transformer as our backbone to improve generalization and global information learning. Experimental results show our L2RT-FIQA effectively reduces the aforementioned four kinds of biases and outperforms other state-of-the-art FIQA methods on several challenging benchmarks.