Advanced Variants of Feature Level Fusion for Finger Vein Recognition

Abstract

Authentication based on vein patterns is a very promising biometric technique. The most important step is the accurate extraction of the vein pattern from sometimes low quality input images. A single feature extraction technique may fail to correctly extract the vein pattern, entailing bad recognition performance. One of the solutions that can be used to improve recognition results is biometric fusion. A possible fusion strategy is feature level fusion, that is the fusion of several feature extractors' outputs. In our work, we exploited the feature level fusion to improve the quality of the extracted vein patterns and thus the feature extraction accuracy. An experimental study involving different feature extraction techniques (maximum curvature, repeated line tracking, wide line detector, ...) and different fusion techniques (majority voting, weighted average, STAPLE, ...) is conducted on the UTFVP finger-vein data set. The results show that feature level fusion is able to improve the recognition accuracy in terms of the EER over the single feature extraction techniques.