Weighted Histogram Matching for Improved Automated Synthetic Aperture Radar (SAR)-optical Image Registration

Automated SAR-optical image registration is the first step in many downstream tasks including automatic target recognition (ATR), remote sensing, and visual navigation for unmanned aerial vehicles (UAVs). SAR-optical image registration is particularly challenging due to significant nonlinear intensity and geometric differences between SAR and optical images and speckle noise present in SAR images. Siamese Convolutional Neural Networks and Generative Adversarial Networks are well-suited for multi-modal feature extraction and are the current state-of-the-art for SAR-optical image registration. However, the high memory and computational requirements of such models pose challenges for real-time, automated SAR-optical image registration. Lightweight, efficient models, like XFeat and SuperPoint, have been investigated for SAR-optical image registration but struggle with the differences between SAR and optical images. In this work, we present a weighted histogram matching pre-processing step to create SAR-like optical images. We demonstrate that weighted histogram matching can improve the registration rate of XFeat by 2.5% and the registration rate of SuperPoint by 11.8%. Further, we introduce a novel method to predict image alignment without the use of a ground truth or reference image. Using a threshold of 25 or more inliers required for alignment, we achieve 98% average accuracy, 98.9% average precision, and 94.5% average recall.