Options
2021
Bachelor Thesis
Titel
Towards Quality-Aware Face Recognition
Alternative
In Richtung Qualitätsabhängiger Gesichtserkennung
Abstract
Recently, face recognition systems have reached near-perfect levels of performance on face verification tasks with images of high face image quality, which is the utility of a face image for recognition purposes. However, in the real world, images can vary greatly in quality. Some images may have the face partially occluded, other images might depict the subject in profile, and some images might be of low resolution. Due to these inherent variabilities, face recognition systems perform much worse in uncontrolled environments. In order to alleviate this issue, previous works focused on verification with sets of images for each person. By using the quality of each image in a weighted average, the output of each image is combined into one, placing a larger emphasis on high-quality images. However, not much research has been done into the effectiveness of directly incorporating image quality into the similarity metric, which would also allow quality-aware comparisons with single images. In this work, we propose a quality-aware similarity score, which uses MagFace [34] features and qualities as a basis. We construct a quality-aware similarity score by adapting the standard cosine similarity score with a model-specific quality value. This adaptation is based on a linear weight function of the cosine similarity score, which controls how the quality influences the quality-aware similarity score. We show that training the linear weight function is an optimization problem, which can be robustly solved with a computationally low-effort brute force approach. By incorporating the quality into our similarity score, we can improve the system's capability to handle low-quality images. We prove the effectiveness of this approach by testing on a variety of face recognition benchmarks, including cross-age, cross-pose, cross-quality, and general unconstrained datasets. Additionally, we show robustness in training by providing a comparison of results when training on different datasets. The results show an improvement over the baseline in 14 out of 16 benchmarks. Moreover, the proposed approach beats state-of-the-art on several single image face recognition benchmarks such as AgeDB [35], CALFW [60], CFP-FP [42], CPLFW [59], and XQLFW [27] with a verification accuracy of 98.50%, 96.12%, 98.74%, 93.50% and 83.95% respectively. Additionally, we also achieve state-of-the-art performance on the video-based multi-media benchmarks IJB-B [51] and IJB-C [33], for the different FMRs 10-2, 10-3, 10-4 and 10-2, 10-3.
ThesisNote
Darmstadt, TU, Bachelor Thesis, 2021
Verlagsort
Darmstadt