Resource-efficient hardware implementation of a neural-based node for automatic fingerprint classification

Modern mobile communication networks and Internet of Things are paving the way to ubiquitous and mobile computing. On the other hand, several new computing paradigms, such as edge computing, demand for high computational capabilities on specific network nodes. Ubiquitous environments require a large number of distributed user identification nodes enabling a secure platform for resources, services and information management. Biometric systems represent a useful option to the typical identification systems. An accurate automatic fingerprint classification module provides a valuable indexing scheme that allows for effective matching in large fingerprint databases. In this work, an efficient embedded fingerprint classification node based on the fusion of a Weightless Neural Network architecture and a technique, namely Virtual Neuron, which efficiently maps a neural network architecture into hardware resources, is presented. The key novelty of the proposed paper is a new neural-based classification methodology that can leverage devices and sensors with limited number of resources, allowing for resource-efficient hardware implementations. Furthermore, the classifier efficiency and the accuracy have been optimized to obtain high classification rate with the best trade-off between minimum area on chip and execution time. The proposed neural-based classifier analyzes a directional image, which is extracted from the original fingerprint image without any enhancement, and classifies the processed item into the five NIST NBIS classes. This approach has been designed for FPGA devices, by exploiting pipeline techniques for execution time reduction. Experimental results, based on a 10-fold cross-validation strategy, show an overall average classification rate of 90.08% on the whole official FVC2002DB2 database.