(A) study on CNN-based PAN-colorization with a locally-adaptive loss function

Abstract

The demand for high resolution satellite image has been increasing quickly in recent days, and there are needs to update hardware and software performance. There are a lot of SR methods for natural images and videos, but not many methods for Remote sensed images. To make satellite image resolution be good quality, the methods can be mainly divided into two categories: Satellite image SR and PAN colorization. While the Satellite image SR methods produce a single high resolution (HR) color output image from the corresponding single low resolution (LR) input image, PAN colorization methods produce high resolution color image from LR multispectral image (MS) and HR gray Panchromatic image (PAN). Because of the HR PAN, PAN colorization methods have much better performance than Satellite image SR. However PAN colorization methods should consider the misalignment between PAN and MS. Since the misalignment, we propose the novel methods with locally adaptive End-to-End learning network. The network takes different loss functions selectively for efficient learning. Our proposed methods do not need any preprocessing for PAN colorization.