Rethinking Image and Video Restoration: An Industrial Perspective

Image and video restoration as a fundamental low-level vision task can significantly improve the visual quality and benefit a lot of downstream computer vision tasks (e.g., video surveillance and satellite imagery). However, early works mainly focus on some ideal settings that strongly limit their applications. Recent years have witnessed increasing interest in designing restoration approaches under real-world scenarios. In this article, we rethink the challenges of restoration deployment from an industrial perspective and share our experiences from three aspects: network design, model training, and deployment environments. According to those thinking and our solutions, we conclude the current progress of restoration tasks and point out some future opportunities that we will focus on.