Universal Image Restoration Pre-training via Degradation Classification

Conclusions : l Randomly initialized models demonstrate an inherent capability to classify degradation. l Models trained on the all-in-one task exhibit the ability to discern unkown degradation. l There is a degradation understanding step in the early training of the restoration model. Our codes and trained model weights are available at https://github.com/MILab-PKU/dcpt . 1. Applicable to various network structures. DCPT consistently achieves average PSNR improvements of 2.08 dB and above. 2. Archiving SoTA performance in all-in-one restoration. 3. More efficient than other degradation embedding, e.g., physical degradation models (IDR) and human instructs (InstructIR).