A Method for Identification of Humans from Dorsal Hand Sub-images using Siamese Network Models

Alghamdi, Mona and Pellicer, Alvaro Lopez (2026) A Method for Identification of Humans from Dorsal Hand Sub-images using Siamese Network Models. IEEE Access. ISSN 2169-3536

Abstract

This study proposes a method for person identification utilizing Siamese networks, knuckle crease patterns, and fingernail sub-images. The framework incorporates automatic localization of multiple regions of interest (ROI) within hand images, component recognition and segmentation via bounding boxes, and similarity matching between segmented image sets. Feature extraction is a central element of the framework, implemented within Siamese networks. Various deep learning neural networks (DLNNs), including DenseNet201, ResNet152, and a finetuned DenseNet201, are employed to extract discriminative high-level features. Euclidean distance is used for similarity matching. The approach is validated on established benchmarks, specifically the 11k Hands dataset and the Hong Kong Polytechnic University Contactless Hand Dorsal Images (PolyU). Results indicate that knuckle and fingernail patterns are critical for person identification. Performance on left-hand images in the 11K Hands dataset surpasses that of right-hand images, potentially due to less frequent use of the left hand. Furthermore, finetuning the DenseNet201 model demonstrates high effectiveness across multiple finger regions in the 11K Hands dataset. The right hand achieves notable identification accuracy in the fingernail area, with 95.69% for the thumb and 96.19% for the ring finger. The model demonstrated even higher accuracy in the minor knuckle region, achieving 96.53% for the index finger and 97.68% for the ring finger on the right hand. On the left hand, the model attained 99.18% accuracy for the ring finger, 99.06% for the index finger, and 97.45% for the little finger. In the major knuckle region, the model achieved 97.39% accuracy for the middle finger on the right hand. Collectively, these findings underscore the robustness and discriminative power of the finetuned DenseNet201 model, particularly for major knuckle features, which consistently produced superior results across both hands and multiple fingers.