SS-TTA : Test-Time Adaption for Self-Supervised Denoising Methods

Even though image denoising has already been studied for decades, recent progress in deep learning has provided novel and considerably better results for this classical signal reconstruction problem. One of the most significant advances in recent years has been relaxing the requirement of having noise-free (clean) images in the training dataset. By leveraging self-supervised learning, recent methods already reach the reconstruction quality of classical and some supervised schemes. In this paper, we propose SS-TTA, a generic test-time adaptation policy that can be applied on top of various self-supervised denoising methods. Taking a pre-trained self-supervised denoising model and a test image as input, our SS-TTA algorithm improves the denoising performance through a proposed ’inference-guided regularization’ process. Based on experiments with three synthetic and three real noise datasets, SS-TTA improves the denoising results of several state-of-the-art self-supervised methods, outperforms recent test-time adaptation approaches, and shows promising performance with supervised models. Finally, SS-TTA also generalizes to cases where the test-time noise distribution differs from the noise distribution of training images.