Periocular Recognition using CNN based Feature Extraction and Classification

Due to the COVID-19 pandemic, there has been an increased focus on contactless biometric systems for authentication. With face masks covering half of the face, it becomes significantly difficult for existing face recognition algorithms to work with the same accuracy (as for non-occluded faces). The main objective of this paper is to successfully build an authentication system that can detect and recognize a user based only on the periocular region of their face. The UBIPr - ubiris dataset has been analyzed and by using transfer learning a CNN model based on VGG-16 has been built for feature extraction. Later, these features have been passed through an ANN for fine-tuned classification. Two such ANN models have been maintained depending on the eye position i.e., separate models for left and right eyes and their outputs have then been combined for classification. A traditional split of 70-30 for training and testing has been followed. The accuracy for the test set is found to be 99.4%. The overall reduction in the loss for the test set is 0.09%. The average confidence for true positives is 0.78 and false positives is 0.11. This provides security to the system from false positives by setting the confidence threshold higher as well as minimizes the chances of physical contact. Currently, the system works only on still images.