PKLNet: A Non-Contact Biometric Recognition Neural Network Based on Palmprint Key Point Localization

PKLNet is a deep learning architecture based on palmprint key point localization, enabling non-contact palmprint recognition through precise localization of palm key points. Its core innovation lies in transforming palmprint recognition into a key point localization problem, addressing traditional challenges such as pose variation and lighting effects. It combines efficiency, hygiene, and accuracy, providing a new solution for the biometric authentication field.

Compared to fingerprint recognition, palmprint features are more abundant (with a capacity several times that of fingerprints), which can theoretically reduce error rates; non-contact collection avoids fingerprint residue and hygiene issues. Application scenarios include high-security access control (data centers, vaults) and public places during pandemics (airports, hospitals), etc.

PKLNet focuses on key point localization: it uses deep convolution + Feature Pyramid Network (FPN) to process multi-scale key points; leverages pose estimation as an auxiliary task to alleviate the problem of palm pose diversity; applies adaptive contrast enhancement and data augmentation to cope with lighting changes; and uses multi-spectral fusion (visible light + near-infrared) to adapt to changes in skin conditions.

On standard datasets, the key point localization error is controlled within a few pixels; the closed-set Rank-1 recognition accuracy exceeds 99%, and the open-set equal error rate (EER) is less than 1%; after optimization via model quantization and knowledge distillation, it can perform real-time inference on ARM processors, making it suitable for resource-constrained scenarios.

PKLNet supports on-device feature extraction and comparison, with original images not leaving the local device; it uses revocable biometric technology, so even if the template is compromised, the original information cannot be restored; it integrates a liveness detection module to distinguish real palms from attack media (photos, silicone, etc.).

Palmprint recognition is expected to be applied in scenarios such as mobile payment and smart home; PKLNet's key point-driven paradigm can provide a reference for iris, vein recognition, etc.; with the popularization of hardware and optimization of algorithms, it will gradually move towards daily applications.