Test-Time Adaptation for Real-World Denoising Networks via Noise-Aware Image Generation

Dongjin Kim, Donggoo Jung, Dajin Han, Kyeonga Kim, Sungyong Baik, Tae Hyun Kim

Abstract: Image denoising aims for a challenging task of recovering clean images from unseen noise, which can follow different distributions depending on scenes, camera models, ISO settings, etc. Previous works have attempted to handle unseen noise by adapting denoising neural networks to each given noisy image. However, a single noisy image can only provide a limited amount of information for training networks. Therefore, we propose to generate noisy images with diverse yet realistic noise that is similar to noise in a given input image. Such noise generation is difficult to achieve given only a single noisy image. To address the challenge, we propose a normalizing flow (NF) framework that can learn the latent representation of noise, conditioned on noisy images. We also employ the Gaussian mixture model to better handle real-world unseen noise by leveraging multiple noise distributions. Using the proposed NF model, our framework can generate multiple synthetic noisy images to facilitate the adaptation of denoising networks to each given image. To further improve the adaptation to unseen noise, we integrate a meta-learning algorithm into our framework. The experimental results demonstrate that our framework substantially improves the performance of several denoising networks on unseen real-world noise across numerous real-world benchmark datasets.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)