Finger Vein Image Quality Assessment by Mated Comparison Score Prediction

Assessing the quality of finger vein images can help improve the biometric recognition performance. Current methods achieve quality assessment by training a machine learning model against predefined quality labels. The model approximates the target labels and is thereby limited by the labels in use. Therefore, the assignment of quality labels plays an important role for the performance of the quality assessment and provides an additional source of error. The absence of a commonly agreed definition for biometric sample quality makes the assignment of quality labels a research topic on its own. While many of the related studies focus on the machine-learning part, unsophisticated quality assignment practices such as the manual annotation of images are still in use. This paper presents an alternative to the explicit assignment of quality labels. The proposed method uses mated comparison scores as training targets. Without explicit quality labels, the model freely adapts to the underlying data and produces continuous quality scores.