A Hardware-Friendly Anomaly Detector for Real-Time Applications with Push-Broom Scanners

Abstract

In this work, we propose a hardware-friendly anomaly detector as a modified version of the LbL-FAD, named HW-LbL-FAD, specifically designed for being executed into parallel hardware devices with limited computational resources. As the original LbL-FAD, the HW-LbL-FAD makes use of the orthogonal projection concept for identifying the subspace in which the background pixels are contained, thus identifying the anomalous pixels as those not totally contained in this subspace. However, the HW-LbL-FAD makes a more efficient use of the Modified Gram-Schmidt orthogonalization process, resulting in a mathematically equivalent algorithm that requires less computational and memory resources. Additionally, the HW-LbL-FAD algorithm can be totally executed using integer arithmetic at different levels of precision that can be adapted for achieving the best relation between detection accuracy and computational burden. The HW-LbL-FAD has been tested in this work using 3 real hyperspectral images typically employed for validating anomaly detection application. These images have been analyzed using both floating-point arithmetic and integer arithmetic with different levels of precision. The obtained results demonstrate the goodness of this proposal.