Abstract

The advancement of deep learning in biometric systems, in which face and hand modalities have been widely implemented, leads to significant improvements in terms of speed performance and data confidentiality. Palmprint recognition is the main focus of the proposed approach, which deals with databases that are relatively smaller than other biometric datasets. A large and complex deep learning models may overfit and lose their ability to generalize when applied to such data. This study addresses this challenge by implementing a deep learning model suitable for palmprints, which are characterized by diversity and limited data. Initially, the appropriate Region of Interest (ROI) is extracted using active segmentation, which is fitting for dealing with the difficulty of obtaining palmprints from hand images with closely spaced or connected fingers. In the second stage, a novel customized LOOCV Leave-One-Out Cross Validation (A Modified-LOOCV) technique is integrated with a Siamese deep learning network for palmprint verification. Unlike conventional LOOCV, our modified scheme optimizes the computational cost while achieving a balanced evaluation on three different datasets. The proposed framework rivals the effectiveness of the advanced palmprint recognition systems with a high recognition accuracy of 99.75%, improved equal error rates (EER), reduced to 0.002, and faster matching time, making it highly suitable for field application.

Key words: Palmprint Recognition, Deep Learning, Customized LOOCV, Siamese Network.