Abstract
Dorsal hand vein biometrics depend strongly on accurate region-of-interest (ROI) extraction, yet conventional convexity-defect-based methods remain vulnerable to contour noise, skin texture, and shadow-induced artifacts. This paper introduces contour-guided convexity defect geometry (CG-CDG), a geometry-driven ROI extraction framework in which a candidate convexity defect is accepted as an anatomical valley only if it jointly satisfies two independent criteria: chord-contour topological consistency, ensuring anatomical relevance, and a minimum enclosed-area threshold, ensuring physical plausibility. Unlike prior convexity-defect-based methods, which accept a candidate valley on the basis of a single depth- or angle-based criterion evaluated in isolation, CG-CDG applies this joint plausibility test before defining the region of interest, a square ROI centered on the hand centroid and aligned with the dominant valley orientation. Experiments on a public dorsal hand vein database of 1,024 images from 138 subjects show that, combined with HOG descriptors and City Block distance, CG-CDG achieves an equal error rate (EER) of 0.0266, with a 95% bootstrap confidence interval of [0.0241, 0.0289] over the matching scores
(B = 1,000 resamples), the narrowest interval among the four ROI extraction methods compared under the primary evaluation setting. A post-hoc Wilcoxon signed-rank test with Bonferroni correction confirms that this improvement is statistically significant over all competing methods (p < 0.001). On this dataset, geometric validation substantially reduces intra-class variance and yields statistically validated performance without requiring any training data, making CG-CDG a strong candidate both as a standalone, training-free biometric solution and as an anatomically grounded preprocessing stage for hybrid deep-learning pipelines.