Geometry based hand vein biometry

Abstract

This thesis documents a study in which new algorithms are developed for geometry based hand vein biometry. Hand vein patterns are assumed not to change over time except in their size, and they are unique to each individual, hence researchers aim to construct a biometric control system based on hand vein patterns. The approach proposed here is using free-posture captured near infrared hand vein images for both enrollment and test. We utilize Gabor lters banks to enhance the visibility of hand vein segments which is a new thing in hand vein biometry literature. Furthermore, it is robust against wrongly aligned hand vein features. The region of interests (ROIs) are extracted from hand vein images. In order to increase the visibility of hand veins in these ROIs, Gabor lter bank approach is applied. Enhanced ROIs facilitate to extract vein line segments as geometric features. To extract these line segments; thresholding, skeletonization and line edge map (LEM) extraction methods are applied on enhanced ROIs, respectively. These methods yield the LEM and the graph version of a hand vein structure. Before identity matching, a preprocessing stage is con gured to alleviate the e ects of wrong registrations. There are two di erent methods for alignment correction. The rst one is based on keypoint matching, whereas the other is based on translating the compared LEMs iteratively. Last but not least, identity matching is done by several distance measurement metrics, namely, line segment Hausdor distance (LHD), weighted line segment Hausdor distance (WLHD), modi ed line segment Hausdor distance (MLHD) and graph edit distance (GED). WLHD and MLHD are di erent versions of LHD, that we specialized for hand vein biometry. Additionally, an algorithm is developed to make a measurement on Graph Edit Instance (GED) metric. GED is de ned as the least cost graph edit operation sequence which is used to transform one graph to another. By examining each metric, we notice that line segment matching based methods give more promising results than graph matching.