Efficient Image Restoration Using Deep Learning: A Survey

Image restoration constitutes a pivotal task within various domains of computer vision, such as photography enhancement or autonomous driving. Recent advancements have predominantly concentrated on deep learning networks, which have demonstrated superior performance relative to traditional algorithms. The field of image restoration has witnessed substantial growth in recent years, as evidenced by the numerous surveys published in this domain. Nevertheless, the efficiency of these methods has received limited attention. Certain applications necessitating image restoration require real-time processing or deployment in edge devices, achievable only through the implementation of efficient and well-designed methods. In this survey, we compile a collection of image restoration methods that prioritize efficiency. The primary strategies employed by these methods are thoroughly examined and analyzed. Finally, a comprehensive analysis of the efficiency of the reported methods is presented, culminating in the proposal of efficiency upper bounds for trainable parameters, computational operations, and inference runtime.