Cascaded Multispectral Image Fusion with Saliency-Weighted Edge Features

Abstract

Abstract In recent years, UAV inspection technology has been widely applied in infrastructure monitoring and power line inspection. However, significant differences between infrared and visible light images pose challenges for high-precision image registration. To address these challenges, a novel UAV multispectral image registration method based on saliency-weighted edge features and multi-feature cascade matching has been proposed. During feature extraction and description, this method uses a saliency-weighted grayscale window for edge extraction, combined with a multi-scale potential Harris corner selection algorithm to extract significant feature points. These edge features are then described using an eight-direction equal-area sector descriptor. In the feature matching phase, a cascaded matching framework is employed. It comprises an adaptive NNDR pre-screening based on domain priors to filter initial matches, followed by multi-feature fusion matching for fine-grained screening. A final geometric consistency check using the FSC algorithm is applied to effectively reduce the probability of mismatches. Experimental results demonstrate that this algorithm achieves an average high-precision matching rate of 90% on 97 pairs of infrared and visible light power equipment images provided by FLIR, attaining sub-pixel level accuracy. This performance significantly surpasses that of classic algorithms such as SIFT and its derivatives, SURF and ORB.