RegiNet: Gradient guided multispectral image registration using convolutional neural networks

Abstract

Abstract Multispectral image registration suffers from severe inconsistency between reference and target images. In this paper, we propose gradient guided multispectral image registration using convolutional neural networks, called RegiNet. We build an end-to-end network that directly produces the registration result from the input image pair. We use a gradient map of the reference image to guide the target image for registration. RegiNet first encodes the reference image and the gradient map of the target image separately, and then concatenates them to register the target image. For loss function, we use a structure loss to effectively capture gradient information from the reference image. Experimental results demonstrate that the proposed method successfully produces registration results as well as outperforms state-of-the-art ones in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).