Deep Learning based Biometric Feature Extraction

Title: Compact and Distinctive Biometric Feature Extraction using Deep Learning

Work under review

Overview: Biometric authentication frameworks rely on various feature generation and classification tasks for which convolutional neural networks (CNN) have demonstrated state-of-the-art effectiveness. However, for resource-constrained devices, e.g., those used in the Internet of Things (IoT), implementation of CNNs remains challenging due to their intensive memory access patterns, operational delays, communication bandwidth, and power consumption. In this work, we propose a CNN-based biometric system where binarized parameters are used to generate compact, yet, meaningful binary biometric features, thereby enabling computation on resource-constrained edge devices and low bandwidth communication with the cloud.

Fig: Operation flow in BinCNN based feature extraction system for face biometrics.

As shown in the figure, the proposed system performs by receiving input image from users, preprocessing them to create normalized images, and then applying the proposed CNN framework on them. Finally, the features for individual users are extracted from the CNN when the learning process is over. The extracted features are found to be compact, secured, and promising for further applications or processing in resource constrained devices.