Multispectral hand features for secure biometric authentication systems

With rapid growth of mobile-based computing, reliable security and user authentication methods are necessary. This article proposes a proof-of-concept biometric authentication method utilizing hand images collected in different light spectra. An analysis of similarity between pattern of blood vessels extracted from near-infrared images and thermal images and assessment of the correlation between individual biometric features contained in each image type is also performed. Results indicate a large potential of using thermal images in biometrics more extensively than now. Also proposed and evaluated are biometric recognition methods based on images of the hand acquired in visible light, near infrared, and using thermal infrared sensors. Two approaches were used to assess information content in images of each type: one based on texture descriptor and one employing convolutional neural networks. In an evaluation gathering data from 104 subjects, the former yielded the lowest equal error rate (EER) of , whereas the latter approach gave for thermal images. Finally, fusion of different-spectra modalities increases accuracy and further reduces EER to . This is, to the authors' best knowledge, the first study exploring the concept of fusing different spectral representations of the human hand for the purpose of biometric recognition.