Optimizing contactless palmprint biometrics via Min-Max, Z-Score, and Global Contrast Normalization

This study isolates the impact of pixel-level normalization on deep palmprint recognition under unconstrained conditions. Using a vanilla ResNet-18 on the IITD database, we compare Min-Max scaling, Z-Score standardization, and Global Contrast Normalization, computed on-the-fly in RAM to preserve floating-point precision. Z-Score attained the highest Rank-1 accuracy (95.09%) and Min-Max the lowest Equal Error Rate (1.5877%). As these differences are small and based on single runs, we present them as plausible tendencies rather than proven mechanisms. Overall, proper preprocessing is an important, low-cost factor in biometric optimization.