Image Restoration - Image Denoiser

What it does

SwinIR Denoiser is a unified image restoration framework based on Swin Transformer that removes noise from degraded images. Supports grayscale and color images at configurable noise levels.

Problem it solves

• Remove noise from photographs and images
• Grayscale image denoising (noise levels 15, 25, 50)
• Color image denoising (noise levels 15, 25, 50)
• Real-world noisy image restoration
• Batch processing of noisy image collections
• Memory-efficient processing of large images (tiling support)

Input/Output

Accuracy & Speed

• Performance:
• Accuracy: State-of-the-art denoising on standard benchmarks
• Noise levels: Optimized for σ = 15, 25, 50
• Metrics:
• Competitive performance on Set12, BSD68, Urban100 benchmarks
• Qualitative visual quality restoration
• Efficient inference relative to other Transformer-based models

Technical Details

• Architecture: Swin Transformer-based restoration framework
• Core Components:
• Shallow Feature Extraction Module
• Extracts initial features from noisy input
• Prepares features for deep processing
• Deep Feature Extraction (Residual Swin Transformer Blocks - RSTB)
• Multiple Residual Swin Transformer Blocks for deep feature learning
• Swin Transformer attention for efficient multi-scale feature extraction
• Skip connections preserve information flow
• High-Quality Image Reconstruction Module
• Reconstructs clean image from extracted features
• Restores spatial resolution and color fidelity
• Key Features:
• Tiling support: Process large images without memory issues
• Multi-task capable: Same framework for denoising, SR, artifact removal
• Residual learning: Skip connections enable training of deeper networks
• Efficient attention: Window-based attention reduces computational cost
• Training:
• Multiple datasets per task (BSD, Set12, Urban100, DIV2K, etc.)
• Practical degradation modeling (BSRGAN for real-world scenarios)
• Supports classical, lightweight, and real-world denoising modes
• Parameters: Vary by model variant (SwinIR-M is middle baseline)

Compliance & Provenance

For more on how we classify models and mark outputs, see our AI Policy.

Model Source

• GitHub Repository: https://github.com/JingyunLiang/SwinIR
• Paper: "SwinIR: Image Restoration Using Swin Transformer" ICCV 2021 Workshops (AIM) https://openaccess.thecvf.com/content/ICCV2021W/AIM/papers/Liang_SwinIR_Image_Restoration_Using_Swin_Transformer_ICCVW_2021_paper.pdf
• License: apache-2.0