Self-Supervised Denoising of Single Noisy Images Based on Data Iterative Refinement

Image denoising has improved dramatically in re-cent years due to the rapid growth of neural networks. While some efforts have been made to train image denoising models using only single noisy images, current self-supervised denoising methods encounter challenges such as inefficient network training, loss of valuable information, or reliance on noise modeling. A random neighbor sub-sampler was introduced by Neighbor2Neighbor(NBR2NBR) [2] to create training image pairings, eliminating the requirement for noise-clean pairs or the necessity for multiple noisy observations. However, the training data obtained in this manner introduces a certain level of random noise during sub-sampling. Furthermore, the distribution of these noisy training pairs may not align well with the distribution of actual noisy image pairs, resulting in a data bias between the training and inference distributions. This data bias restricts the denoising performance of the method on real noisy images. This article proposes an effective method to reduce this data bias as much as possible. We introduce iterative data refinement, which trains the model on datasets with higher noise levels and refines noisy images for use in the next round using the learned model. This approach iteratively minimizes the bias between the generated noisier-noisier dataset and the ideal dataset to enhance the training of denoising models. Extensive experiments were carried out using synthetic and real datasets showcase that our approach surpasses previous works in performance.