Noisy-As-Clean

Noisy-As-Clean implements a self-supervised image denoising strategy that trains denoising networks using the corrupted test image itself to approximate supervised parameters for removing synthetic and realistic noise.

Key Features:

• Self-supervised framework: Treats the corrupted test image as its own clean target during training.
• Synthetic training-pair generation: Generates synthetic images from the corrupted image paired with an additional, similar corruption for training.
• Weak-noise assumption: Leverages the principle that sufficiently weak noise permits training from a single corrupted image.
• Compatibility with DnCNN and ResNet: Demonstrates that networks such as DnCNN and ResNet trained with NAC achieve comparable or superior performance.
• Domain-gap mitigation: Approximates parameters learned by supervised methods, reducing the domain gap between supervised and unsupervised approaches.
• Applicable to synthetic and realistic noise: Targets both synthetic noise and realistic, real-world corruption scenarios.

Scientific Applications:

• Image denoising: Removal of synthetic and realistic noise from images using self-supervised training.
• Real-world imaging where clean references are unavailable: Training denoisers directly from corrupted test images when clean ground truth is impractical or impossible to obtain.
• Comparative evaluation of denoising methods: Benchmarking performance of DnCNN, ResNet, and other denoising approaches under domain-gap conditions.

Methodology:

Train networks in a self-supervised manner by generating synthetic images from the corrupted image paired with an additional similar corruption, using the corrupted image as the training target and relying on the weak-noise assumption to approximate supervised optimal parameters.