SwinIA: Self-Supervised Blind-Spot Image Denoising with Zero Convolutions

9 May 2023  ·  Mikhail Papkov, Pavel Chizhov ·

The essence of self-supervised image denoising is to restore the signal from the noisy image alone. State-of-the-art solutions for this task rely on the idea of masking pixels and training a fully-convolutional neural network to impute them. This most often requires multiple forward passes, information about the noise model, and intricate regularization functions. In this paper, we propose a Swin Transformer-based Image Autoencoder (SwinIA), the first convolution-free architecture for self-supervised denoising. It can be trained end-to-end with a simple mean squared error loss without masking and does not require any prior knowledge about clean data or noise distribution. Despite its simplicity, SwinIA establishes state-of-the-art on several common benchmarks.

PDF Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Color Image Denoising BSD300 lambda30 SwinIA SSIM 0.775 # 1
PSNR 27.92 # 1
Color Image Denoising BSD300 lambda5-50 SwinIA PSNR 27.74 # 1
SSIM 0.764 # 1
Color Image Denoising BSD300 sigma25 SwinIA PSNR 28.4 # 1
SSIM 0.789 # 1
Color Image Denoising BSD300 sigma5-50 SwinIA SSIM 0.785 # 1
PSNR 28.4 # 1
Grayscale Image Denoising BSD68 sigma15 SwinIA PSNR 31.07 # 16
SSIM 0.856 # 1
Grayscale Image Denoising BSD68 sigma25 SwinIA PSNR 29.17 # 12
SSIM 0.801 # 1
Grayscale Image Denoising BSD68 sigma50 SwinIA PSNR 26.61 # 3
SSIM 0.706 # 1
Medical Image Denoising FMD Confocal Fish SwinIA SSIM 0.871 # 1
PSNR 31.79 # 1
Medical Image Denoising FMD Confocal Mice SwinIA PSNR 37.65 # 1
SSIM 0.960 # 1
Medical Image Denoising FMD Two-Photon Mice SwinIA SSIM 0.915 # 1
PSNR 33.25 # 1
Grayscale Image Denoising Hanzi SwinIA SSIM 0.556 # 1
PSNR 14.35 # 1
Color Image Denoising Kodak24 lambda30 SwinIA SSIM 0.805 # 1
PSNR 29.51 # 1
Color Image Denoising Kodak24 lambda5-50 SwinIA PSNR 29.06 # 1
SSIM 0.788 # 1
Color Image Denoising Kodak24 sigma25 SwinIA SSIM 0.819 # 1
PSNR 30.12 # 1
Color Image Denoising Kodak24 sigma5-50 SwinIA PSNR 30.3 # 1
SSIM 0.820 # 1
Grayscale Image Denoising Set12 sigma15 SwinIA PSNR 30.37 # 6
SSIM 0.857 # 1
Grayscale Image Denoising Set12 sigma25 SwinIA PSNR 28.72 # 5
SSIM 0.817 # 1
Grayscale Image Denoising Set12 sigma50 SwinIA PSNR 26.03 # 7
SSIM 0.736 # 1
Color Image Denoising Set14 lambda30 SwinIA SSIM 0.799 # 1
PSNR 28.74 # 1
Color Image Denoising Set14 lambda5-50 SwinIA SSIM 0.780 # 1
PSNR 28.27 # 1
Color Image Denoising Set14 sigma25 SwinIA SSIM 0.814 # 1
PSNR 29.54 # 1
Color Image Denoising Set14 sigma5-50 SwinIA SSIM 0.809 # 1
PSNR 29.49 # 1

Methods