A Lightweight Super-resolution for Compressed Image

In recent years, there are significant improvements in super-resolution task using deep neural networks. However, these models have a problem that artifacts become stronger when upscaling lossily compressed images such as JPEG. Conventional super-resolution algorithms are trained on uncompressed images, but the images mainly dealt with on-device such as consumer electronics are compressed ones. Therefore, to perform on-device super-resolution, a lightweight model is required, which can reduce compression artifacts. We propose a novel loss and training scheme that can produce high-resolution images while minimizing artifacts for compressed ones with a single network. Experimental results show that our method can reduce artifacts significantly with less computation than existing approaches.