TAPE: Task-Agnostic Prior Embedding for Image Restoration

Learning a generalized prior for natural image restoration is an important yet challenging task. Early methods mostly involved handcrafted priors including normalized sparsity, $$\ell _0$$ gradients, dark channel priors, etc. Recently, deep neural networks have been used to learn various but do not guarantee generalize. In this paper, we propose novel approach that embeds task-agnostic into transformer. Our approach, named Task-Agnostic Prior Embedding (TAPE), consists of two stages, namely, pre-training and task-specific fine-tuning, where the first stage knowledge about images transformer second extracts assist downstream restoration. Experiments on types degradation validate effectiveness TAPE. The performance in terms PSNR improved by as much 1.45 dB even outperforms algorithms. More importantly, TAPE shows ability disentangling from degraded images, which enjoys favorable transfer unknown tasks.