Single-Image Deblurring verus Multi-Image Denoising on 'Kid Sweater' synthetic dataset

(for all figures, noise standard deviation in the range [0-255])

Ground Truth	one of 100 noisy images, noise std = 20, PSNR=22.5281 dB

The following figures show a comparison between Single-Image Deblurring and Multi-Image Denoising for 4 different synthetic kernels.

The performance of deblurring depends on the kernel and the prior. The performance of multi-image denoising with a weak prior is higher than that for single-image deblurring method with either a weak or stronger prior.

synthetic kernel #1	blurry image	deblurring using weak prior, PSNR=24.7206 dB	deblurring using strong prior, PSNR=31.2855 dB	denoising result using weak prior, PSNR=39.6556 dB	ground truth
synthetic kernel #2	blurry image	deblurring using weak prior, PSNR=24.7948 dB	deblurring using strong prior, PSNR=31.1536 dB	denoising result using weak prior, PSNR=39.6556 dB	ground truth
synthetic kernel #3	blurry image	deblurring using weak prior, PSNR=25.1226 dB	deblurring using strong prior, PSNR=31.2247 dB	denoising result using weak prior, PSNR=39.6556 dB	ground truth
synthetic kernel #4	blurry image	deblurring using weak prior, PSNR=23.8543 dB	deblurring using strong prior, PSNR=30.5894 dB	denoising result using weak prior, PSNR=39.6556 dB	ground truth

The two priors used the above comparison:

The weak prior is the ground truth image corrupted with synthetic Gaussian noise of std deviation 40. The strong prior is the denoised result of a state-of-the-art single image denoising method CBM3D.

weak prior	strong prior