IDR: Self-Supervised Image Denoising via Iterative Data Refinement (2111.14358v2)

Abstract: The lack of large-scale noisy-clean image pairs restricts supervised denoising methods' deployment in actual applications. While existing unsupervised methods are able to learn image denoising without ground-truth clean images, they either show poor performance or work under impractical settings (e.g., paired noisy images). In this paper, we present a practical unsupervised image denoising method to achieve state-of-the-art denoising performance. Our method only requires single noisy images and a noise model, which is easily accessible in practical raw image denoising. It performs two steps iteratively: (1) Constructing a noisier-noisy dataset with random noise from the noise model; (2) training a model on the noisier-noisy dataset and using the trained model to refine noisy images to obtain the targets used in the next round. We further approximate our full iterative method with a fast algorithm for more efficient training while keeping its original high performance. Experiments on real-world, synthetic, and correlated noise show that our proposed unsupervised denoising approach has superior performances over existing unsupervised methods and competitive performance with supervised methods. In addition, we argue that existing denoising datasets are of low quality and contain only a small number of scenes. To evaluate raw image denoising performance in real-world applications, we build a high-quality raw image dataset SenseNoise-500 that contains 500 real-life scenes. The dataset can serve as a strong benchmark for better evaluating raw image denoising. Code and dataset will be released at https://github.com/zhangyi-3/IDR

Overview of Self-Supervised Image Denoising via Iterative Data Refinement

The paper emphasizes a novel approach in the domain of image denoising through a self-supervised technique termed Iterative Data Refinement (IDR). This technique addresses a significant challenge in supervised denoising methods—the paucity of large-scale noisy-clean image pairs due to prohibitive costs associated with manual annotation. Existing unsupervised methods encounter limitations due to suboptimal performance or impractical requirements (\textit{e.g.,} paired noisy images). The proposed IDR methodology demonstrates a notable advance in this field by only necessitating single noisy images and a readily accessible noise model.

Core Methodology

IDR operates on an innovative iterative mechanism involving two cyclical steps:

1. Data Generation: A noisier-noisy dataset is constructed by applying random noise to single noisy images using a predefined noise model.
2. Model Training and Refinement: The denoising model is trained on this dataset to refine the noisy images, resulting in progressively cleaner target images for the next iteration.

Furthermore, the paper introduces a more computationally efficient fast algorithm approximation of IDR, maintaining high denoising performance with reduced training time.

Empirical Validation

The method's efficacy is showcased through exhaustive experiments on various datasets, exhibiting superior performance against extant unsupervised methods and competitive results compared to supervised models. A significant contribution here is the development of SenseNoise-500, a high-quality dataset featuring 500 real-life scenes, which serves as a robust benchmark to evaluate raw image denoising in real-world conditions.

Theoretical and Practical Implications

Theoretically, the iterative nature of IDR highlights the potential of reducing data bias between synthetically generated data and ideal noisy-clean image pairs, which is pivotal for enhancing the generalization capability of denoising models. Practically, the method demonstrates the viability of employing unsupervised frameworks in real-world applications, reducing dependency on collated annotated datasets. This aligns well with modern, resource-efficient AI paradigms seeking to leverage existing noisy datasets and synthetic augmentation seamlessly.

Future Directions

Speculation on future work points towards refining noise model accuracy and extending IDR's applicability to broader noise types and levels, particularly in environments characterized by complex, dynamic lighting conditions. Additionally, exploring the synergy between IDR and advanced learning architectures, including transformers, may yield further enhancements in denoising quality and efficiency.

In summary, IDR provides a compelling approach to image denoising, notable for its resource-efficient methodology and substantial applicability in both controlled and uncontrolled environments, with promising avenues for future exploration.