In this thesis, we propose a novel finger vein recognition system which improves its usability, stability, computational complexity, and recognition accuracy comparing to previous researches. In order to make the system practical in real-world applica...
In this thesis, we propose a novel finger vein recognition system which improves its usability, stability, computational complexity, and recognition accuracy comparing to previous researches. In order to make the system practical in real-world applications, algorithms in the system are also designed to be very robust to external physical factors such as lighting and finger positioning, and to easily accommodate large-sized database which may have frequent changes of entries. The proposed system consists of three major components, each with separate, significant contribution to the performance of the system: finger vein quality assessment, enhancement and feature extraction, and classification. In finger vein quality assessment, we propose an energy model, developed from the structure of veins and corner minutiae, to generate assessment scores and then to discard low-quality vein images for high usability of the system. For the image enhancement and feature extraction component, a novel explicit guided directional filter is proposed to obtain high-quality vein contours from noisy, non-uniform, low-contrast images without introducing any segmentation process. It enhances an input image with an additional supervisor image which strongly instructs to preserve vein patterns and reduces background impacts such as haze and variation of illumination. Veins after guided directional filtering are magnified enough to directly extract average absolute deviation (AAD) features, which represent strengths of directional block information with eight different angles, even from images with thin, vague ridges and non-uniform backgrounds. Three types of classifiers, based on the extreme learning machine (ELM), are designed to best utilize the characteristics of AAD features, to obtain high recognition accuracy and to provide the stability of the system. A typical ensemble ELM (E-ELM) consists of a number of unit ELMs, each trained with whole AAD features, and combines the outputs of unit ELMs for stable classification. A newly-developed component-based ELM (C-ELM) has eight much smaller-sized unit ELMs and an output layer to combine outputs of eight ELMs. For the structured training of vein patterns, the C-ELM is designed to first train small differences between patterns with the same angle and then to aggregate the differences at the output layer. We also design an ensemble C-ELM (EC-ELM) to provide the stability of the component-based matching system. Experimental results show that the proposed system with a C-ELM achieves an excellent matching performance in terms of matching accuracy of 99.89% and speed of 0.87 ms for each image.