Targeted-instance super-resolution via deep residual networks

Abstract

Super-resolution is the process of restoring detailed information from low-resolution to high-resolution and it is an undetermined computer vision problem. Due to the advent of deep learning, many works on super-resolution has been conducted and the potential has been presented. In particular, surveillance, medical imaging and satellite imaging area demands high-resolution images for their special purposes. Among them, satellite images has difficulties for super-resolution due to immense image size and complex background. In this paper, we proposed 1) targeted-instance dataset by extracting the area of targeted-instance from an entire satellite image and 2) the optimal model for targeted-instance dataset. Targeted-instance dataset which size was shrunk and removed complex backgrounds made learning possible to adopt the state-of-the-art super-resolution models and explore the optimal deep residual model. As the result, targeted-instance dataset presented good or better result than the whole dataset and the proposed deep residual network model by replacing the activation function affected positively on the performance and efficiency.