Table 2 Statistical comparison of restoration performance among CycleGAN, DDRM (binarization), and our proposed method on the The Pillow Book dataset with degradation intensity = 1.0

From: Restoration of ancient Japanese manuscripts via the diffusion denoising restoration model and color space-based masking

Method

PSNR ↑

SSIM ↑

LPIPS ↓

 

Min

Max

Mean

Var

Min

Max

Mean

Var

Min

Max

Mean

Var

CycleGAN

12.35

20.39

17.16

1.3887

0.3894

0.9536

0.8369

0.0049

0.0160

0.4231

0.0893

0.0036

DDRM(Bin)

8.96

18.34

13.86

2.7422

0.5176

0.8369

0.7440

0.0028

0.3711

0.7014

0.4877

0.0033

Ours

19.20

27.60

24.83

1.2752

0.8384

0.8884

0.8736

0.0001

0.2498

0.4784

0.2789

0.0003

  1. Quantitative statistics (minimum, maximum, mean, and variance) of PSNR, SSIM, and LPIPS are reported.
  2. The proposed method achieves the highest mean PSNR and SSIM and the lowest LPIPS variance, demonstrating both superior fidelity and stability compared with the other methods.
Back to article page