Acta Scientiarum http://periodicos.uem.br/ojs ISSN on-line: 1807-8664 Doi: 10.4025/actascitechnol.v45i1.61948 COMPUTER SCIENCE Acta Scientiarum. Technology, v. 45, e61948, 2023 Dorsal hand vein biometrics with a novel deep learning approach for person identification Felix Olanrewaju Babalola, Yıltan Bitirim and Önsen Toygar * Computer Engineering Department, Faculty of Engineering, Eastern Mediterranean University, 99628, Famagusta, North Cyprus, via Mersin 10, Turkey. *Author for correspondence. E-mail: onsen.toygar@emu.edu.tr ABSTRACT. Hand dorsal biometric recognition system proposed in this study combines the strength of information in regions of dorsal vein biometric trait in a deep learning based Convolutional Neural Networks (CNN) model. The approach divides each dorsal image into five overlapping regions; consequently, five different training and test sets are obtained for each image, modeling a multi-modal biometric system while using only one trait. The test outputs are combined by score-level fusion. Experimental results on FYO, Bosphorus and Badawi datasets indicate the efficiency of the proposed method and its comparability with other recognition systems. The results are also compared with the state- of-the-art dorsal hand vein recognition systems to show the ability of the proposed biometric architecture to perform well in different conditions that may affect dorsal vein pattern acquisition and have con-sequent effect on the efficiency of the recognition system. Keywords: dorsal vein recognition; convolutional neural network; score-level fusion; overlapping image regions. Received on November 11, 2021. Accepted on August 22, 2022. Introduction Biometric systems with physical traits such as palmprint, fingerprint, iris, and face are employed for person recognition and secure system authentication. These systems have become widely popular due to enormous success made over the years in biometric research and consequent technological developments for the acquisition of the traits both for research work and real life implementations (Farmanbar & Toygar, 2015). Moreover, hand vein patterns from palm, wrist, fingers and other regions of the hand have also received massive popularity due to inherent attributes of vein pattern, such as stability, uniqueness and spoof-proof properties, as well as the emergence of technologies that can easily capture these traits in a user friendly manner. These properties make vein pattern recognition one of the most reliable way of effectively restricting access to a system and safeguarding data. This study explores the promising attributes of dorsal vein (back of palm) as bio-metric trait and propose a system which exploits the strength of multi-modal systems for enhancing dorsal vein biometric recognition systems by merging outputs from five over-lapping sections of dorsal vein samples. The system is designed with a Convolutional Neural Network (CNN) model which is trained separately for each overlapping region. The overall decision of the system is determined by merging scores from five predictions which correspond to each region. The experiments were conducted on FYO, Bosphorus and Badawi dorsal databases to demonstrate the efficacy of the proffered system. It is generally perceived that when human are subjected to different routine activities or environmental conditions, it bears an effect on sensors' ability to capture vein pattern effectively for recognition, or that captured images may be slightly different in different conditions which will therefore affect the efficiency of the system. The robustness of the system against this perception was also tested by experimenting with data captured under different conditions such as very cold weather, using the device after strenuous exercise or under normal conditions, which are available in Bosphorus database. The contributions of this work can be summed up as follows: i) This study introduces a new dorsal hand vein biometric authentication system that fuses scores from five overlapping sections of dorsal vein samples using a CNN-based architectures. ii) The proposed system takes the advantage of multi-modality within a unimodal system. iii) Texture-based and CNN-based systems are compared with the proposed biometric recognition system.