Data-Driven Multispectral Image Registration

Multispectral imaging is widely used in remote sensing applications from UAVs and ground-based platforms. Multispectral cameras often use a physically different camera for each wavelength causing misalignment in the images for different imaging bands. This misalignment must be corrected prior to concurrent multi-band image analysis. The traditional approach for multispectral image registration process is to select a target channel and register all other image channels to the target. There is no objective evidence-based method to select a target. The possibility of registration to some intermediate channel to the target is not usually considered, but could be beneficial if there is no target channel for which direct registration performs well for every other channel. In this paper, we propose an automatic data-driven multispectral image registration framework that determines a target channel, and possible intermediate registration steps based on the assumptions that 1) some reasonable minimum number of control points correspondences between two channels is needed to ensure a low-error registration; and 2) a greater number of such correspondences generally results in lower registration error. Our prototype is tested on three multispectral datasets captured with UAV-mounted multispectral cameras. The resulting registration schemes had more control point correspondences on average than the traditional register-all-to-one-target-channel approach in all of our experiments. For most channels in our three datasets, our registration schemes produced lower back-projection error than the direct-to-target-channel based registration approach.