Real-Time Finger-Video Analysis for Accurate Identity Verification in Mobile Devices

The rapid advancement of smartphone technology, driven by sophisticated cameras and accelerated computing capabilities, has fueled the demand for reliable and secure identity verification solutions. Finger photo recognition has emerged as a popular and touchless alternative for secure smartphone unlocking. However, systems reliant solely on single RGB images face substantial challenges, including device variability, inconsistent backgrounds, and varying lighting conditions, which can compromise their effectiveness. This research addresses the limitations of current finger-based biometric systems by introducing a novel identity verification system that leverages the rich and dynamic information captured in finger-video data. The proposed approach, VIDVerify, combines binary classification, reduced information redundancy, and loss of cosine embedding to create a robust identity verification pipeline. VIDVerify employs a Siamese architecture with joint embeddings and three complementary loss functions: VICReg loss for redundancy reduction, focal loss for self-class balancing, and cosine embedding loss for video data optimization. The system is evaluated using the Multi-Movement Finger-Video (MMFV) database, which captures finger movements across multiple axes. By comparing convolutional backbones (ResNets, MobileNets) with transformer-based backbones (Swin Transformers), this research establishes a new benchmark for finger-video-based identity verification, paving the way for future advancements in this field. The code is available on GitHub: www.github.com/sulabh-shr/fingerprint.