Matlab code and test results for the paper: Mingyuan Zhou, Haojun Chen, John Paisley, Lu Ren, Lingbo Li, Zhengming Xing, David Dunson, Guillermo Sapiro, and Lawrence Carin, "Nonparametric Bayesian Dictionary Learning for Analysis of Noisy and Incomplete Images," submitted to IEEE Trans. Image Process. PDF
1 Softwares
- BPFA Gray-scale and RGB image denoising code (last update 04/15/2010) can be found HERE.
-
BPFA Gray-scale, RGB and Hyperspectral image inpainting & denoising code (last update 04/15/2010) can be found HERE.
2 Nonparametric
No matter the missing data ratio are 95%, 80%, 20% or 5%, no matter the noise standard deviation are 100, 75, 25 or 5, no matter the images are the gray-scale Barbara, Lena, Peppers or House, the RGB Castle, Mushroom, Kangaroo or Horses, or the hyperspectral Urban or Hill, the code listed above, with exactly the same parameters, can be used to produce state-of-the-art results for all cases.
The results reported in each table below are obtained with the same initialization and parameter setting, the detail can be found in the code listed above:
- Results in Table 1 can be reproduced by Demo_Denoise_Gray.m
- Results in Table 2 can be reproduced by Demo_Denoise_RGB.m
- Results in Table 3 can be reproduced by Demo_Inpainting_GRAY.m
- Results in Table 4 can be reproduced by Demo_Inpainting_RGB.m
- Results in Table 5 can be reproduced by Demo_Denoise_Inpainting_GRAY.m
- Results in Table 6 can be reproduced by Demo_Inpainting_HyperSpecImage.m
3 Test Results
Table 1: BPFA gray-scale image denoising PSNR results, using patch size 8×8. The test images, denoising results and learned dictionaries can be downloaded HERE.
Table 2: BPFA RGB image denoising PSNR results, using patch size 7×7. The test images, denoising results and learned dictionaries can be downloaded HERE.
Table 3: BPFA gray-scale image interpolation PSNR results, using patch size 8×8. The test images, interpolation results and learned dictionaries can be downloaded HERE.
Table 4: BPFA RGB image interpolation PSNR results, using patch size 7×7. The test images, interpolation results and learned dictionaries can be downloaded HERE.
Table 5: Simultaneous image denoising and interpolation
PSNR results for BPFA, considering the Barbara256
image and using patch size 8×8. The test images, denoising & interpolation results and learned dictionaries can be downloaded HERE.
Table 6: BPFA hyperspectral image interpolation PSNR results, considering a 150 × 150 hyperspectral urban image with 210 spectral bands. Test results can be downloaded from HERE
Observed data (%) | 1×1 | 2×2 | 3×3 | 4×4 |
2 | 15.34 | 21.09 | 22.72 | 23.46 |
5 | 17.98 | 23.58 | 25.30 | 25.88 |
10 | 20.41 | 25.27 | 26.36 | 26.68 |
20 | 22.22 | 26.50 | 27.02 | 27.16 |
File translated from
TEX
by
TTH,
version 3.87.
On 26 Apr 2010, 13:03.