Multi-spectral Finger based User Verification using Off-the-Shelf Deep Features

Biometrics based person verification is widely employed in several applications like smartphone access, banking, national identification documents and border control. In general, the biometric system employs different biometric characteristics, including physiological (e.g. face, iris, fingerprint, finger vein, palm) and behavioural (e.g. keystrokes, gesture, gait, voice) for verification. Among these biometric characteristics, finger-based biometrics is an efficient biometric characteristic to verify a person considering its unique patterns and accuracy. This work performs person verification using a multi-spectral finger database captured from visible and near-infrared light by employing seven different pre-trained deep Convolutional Neural Network (CNN) models. Further, we perform verification by employing Support Vector Machine classifier (SVM). We employ seven different networks to extract the deep features in this work (AlexNet, DenseNet, GoogleNet, InceptionResNet, InceptionV3, NasNet, ResNet101). Extensive experiments to investigate the performance of the proposed method are performed on the multi-modal finger database (dorsal finger, ventral finger, dorsal finger vein, ventral finger vein) captured from a custom finger capture sensor. Our database consists of 357 unique fingers captured in a span of 3 to 5 days. The experimental results suggest that fusion of all four finger characteristics shows the best performance when compared with individual finger characteristics for person verification.