Task Adaptive Network for Image Restoration With Combined Degradation Factors

Jingyuan Zhou, Chaktou Leong, Minyi Lin, Wantong Liao, Congduan Li; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2022, pp. 1-8

Existing methods have achieved excellent performance on image restoration, but most of them are designed for one type of degradation. However, the weather is complex in the real world. So networks designed for single tasks are usually difficult to apply. Therefore, we propose a task-adaptive attention module to enable the network to restore images with multiple degradation factors. The task-adaptive attention module mainly includes three parts: Task-Adaptive sub-network, Task Channel Attention, and Task Operation Attention. To evaluate the model, we construct a mixed degradation factors dataset that combines three degradation factors of rain, haze, and raindrop. The experimental results show that our method not only better restores images with mixed degradation factors, but also show competitive results compared to the state-of-the-art models of each task.