SER-FIQ: Unsupervised Estimation of Face Image Quality Based on Stochastic Embedding Robustness

Face image quality is an important factor to enable high performance face\nrecognition systems. Face quality assessment aims at estimating the suitability\nof a face image for recognition. Previous work proposed supervised solutions\nthat require artificially or human labelled quality values. However, both\nlabelling mechanisms are error-prone as they do not rely on a clear definition\nof quality and may not know the best characteristics for the utilized face\nrecognition system. Avoiding the use of inaccurate quality labels, we proposed\na novel concept to measure face quality based on an arbitrary face recognition\nmodel. By determining the embedding variations generated from random\nsubnetworks of a face model, the robustness of a sample representation and\nthus, its quality is estimated. The experiments are conducted in a\ncross-database evaluation setting on three publicly available databases. We\ncompare our proposed solution on two face embeddings against six\nstate-of-the-art approaches from academia and industry. The results show that\nour unsupervised solution outperforms all other approaches in the majority of\nthe investigated scenarios. In contrast to previous works, the proposed\nsolution shows a stable performance over all scenarios. Utilizing the deployed\nface recognition model for our face quality assessment methodology avoids the\ntraining phase completely and further outperforms all baseline approaches by a\nlarge margin. Our solution can be easily integrated into current face\nrecognition systems and can be modified to other tasks beyond face recognition.\n