Joint-task learning to improve perceptually-aware super-resolution of aerial images

Abstract

ABSTRACT Deep neural networks have become very popular for solving many problems in computer vision. Super-resolution (SR) is a particularly challenging task since new information must be created at increased resolution, possibly leading to visual artefacts or incoherent texture. In the context of remote sensing, this image restoration technique has great potential for synthesizing high-resolution (HR) data from low-resolution (LR) images. While there are multiple methods that enhance the perceptual quality of SR images, most of them fail to recover detailed information from aerial imagery. One of the main reasons for that is the difficulty in defining a ‘good-looking’ image from the perspective of the machine. In this work, we propose an end-to-end training procedure that unifies networks related to different tasks: an SR module based on generative adversarial networks (GANs) and a semantic segmentation module. Our claim is that by including a classification loss when estimating the HR image, the GAN generator produces more coherent structures and textural information, synthesizing, therefore, more realistic images according to perception-based scores. Our experimental results show that the proposed method is capable of improving perceptual outputs of deep-learning oriented networks with a small training overhead, surpassing multiple state-of-the-art super-resolution methods. Our code is available at http://github.com/elitonfilho/SegSR.