[Invited Paper] Overview of Supervised and Self-supervised Image Denoising Using Deep Neural Networks

This paper provides an overview of supervised and self-supervised image denoising techniques, including discussions on practical usages. Through the development of deep network architectures, training strategies, and datasets, supervised deep learning methods outperformed traditional non-learning approaches by substantial margins. However, creating well-registered, real-world noisy-clean image pairs for training is challenging, and thus, there is a limited number of practical training images. Additionally, a denoiser trained with a specific dataset may not perform well for images with different properties. Consequently, there is a growing demand for self-supervised training methods that do not rely on external training sets. In this paper, we first present an overview of supervised techniques and training datasets. We then explore self-supervised methods that effectively remove noise without the need for external datasets and discuss potential future directions for this field. We conclude the paper with a comparison between supervised and self-supervised methods, their selections, and future studies.