Cross-illumination Evaluation of Hand Crafted and Deep Features for Fusion of Selfie Face and Ocular Biometrics

This paper addresses the implementation of a multiunit biometric system. Results are shown for multi-unit classification with VISible light mobile Ocular Biometric (VISOB) dataset using feature descriptors such as Local Binary Patterns (LBP) and Histogram of oriented gradients (HOG). We also evaluate the pre-trained deep learning models such as VGG16, ResNet18, MobileNetV1, MobileNetV2, and LightCNN9. Experimental evaluation on large scale VISOB dataset suggests that feature-level fusion followed by score-level fusion of left ocular region, right ocular region and face region in office light condition, daylight and dims condition has provided Equal Error Rates (EER) of 9.3%, 8.0% and 10.6% respectively. Also, combining the pretrained models using feature fusion decreased the EER even further.