Conventional image restoration models are difficult to apply efficiently in real-world scenarios because they are designed to handle only specific types and levels of degradation. This study proposes a model that can handle multiple degradations with a single restoration model using prompt learning. Furthermore, we introduce a sub-network, PE (Pixel-level Encoder), that modulates the encoder of the main network and prompts. In this process, the proposed model adaptively intergates features across spatial and channel spaces through DSCA (Deformable Spatial Cross-Attention) and MDTCA (Multi-Dconv head Transposed Cross-Attention). Also, this model exploits PCL (Pixel-wise Contrastive Loss) to capture the style and context information of target images. Experimental evaluation was conducted using a widely-used dataset in the field of all-in-one image restoration, including dehazing, deraining, and denoising. Additionally, this study evaluated the robustness of models for monocular depth estimation and visual odometry using images reconstructed from noise degradation.
- Pretrained Models: Link