Software:
Image Restoration
 

 

SOFTWARE:
 Image Restoration      
VistaRestoreTools 1.0


The (VI (S) TA) Image Restoration Toolbox ( VistaRestoreTools ) is a Matlab Toolbox for image restoration that includes (1) classical regularization techniques, (2) classical wavelet thresholding techniques, (3) regularization functionals based on non-linear human vision models, and (4) denoising techniques based on Kernel regression in wavelet domains.

This toolbox is based on (and reproduces) the results shown in [Gutierrez03 ,  Gutierrez06 ,  Laparra08 , Laparra10b ].
 

Citation:

We have decided to make the library available to the research community free of charge. If you use VistaRestoreTools in your research, we kindly ask that you reference this website: 

J. Gutiérrez, V. Laparra, G. Camps and J. Malo. "VistaRestoreTools: an image restoration toolbox for Matlab", http://www.uv.es/vista/vistavalencia/software/software.html

 

and the paper(s) associated to each algorithm.

 

Note also that the package contains some previously released public domain wavelet software authored by Eero P. Simoncelli, belonging to his MatlabPyrTools toolbox (http://www.cns.nyu.edu/~lcv/software.php). When using wavelet-based functions in VistaRestoreTools you shall aknowledge the author of MatlabPyrTools as well!.

 

Installation:
  • Download the file VistaRestoreTools1.0.zip (13.5 MBytes).
  • Decompress at your computer and set the Matlab path accordingly.
  • Look at the help of the functions below for instructions on how to use each algorithm.
  • Warning!: VistaRestoreTools1.0 has been tested on Matlab 7.2 (Matlab 2006a).  Posterior Matlab versions for windows may need recompilation of some mex files of MatlabPyrTools.
Basic features:
  • Image degradation: 

  • apply_degradation_2.m    Routine for controlled image degradation. Includes tunable blurring plus white or tunable colored Gaussian noise.

  • image_degradation_demo.m    Shows examples on how to use the image degradation function.
  • Denoising and deconvolution by regularization in Local-Fourier domains:

  • Classical regularization techniques:  restore.m

    • Second derivative regularization functional

    • Auto-regressive (AR) models of different order [Banham&Katssaggelos, IEEE Sig. Proc. Mag. 1997]

    • Inverse of CSF regularization functional [Hunt, Proc. IEEE, 1975]

  • Regularization by perceptually-based non-linear divisive normalization functionals  [Gutiérrez03 ,  Gutiérrez06]:   restore.m

  • image_restoration_demo.m    Shows examples on how to use the different settings of the image restoration function restore.m.
  • Denoising in wavelet domains:

  • Classical thresholding techniques:

    • Hard Thresholding [ Donoho, J. Am. Stat. Assoc. 1995 ]: hardthres.m

    • Soft Thresholding [ Donoho, J. Am. Stat. Assoc. 1995 ]: softthres.m

    • Bayesian denoising assuming Gaussian noise and signal with Gaussian marginals in the wavelet domain [ Figueiredo&Nowak, IEEE Tr.Im.Proc. 2001]: bayesian_gauss_margin.m

  • Mutual Information Kernel Regularization in the steerable wavelet domain  [ Laparra08 , Laparra10b ] more information here:  mi_kernel_denoising.m

  • image_denoising_demo.m    Shows examples on how to use the different image denoising functions in the wavelet domain.

Download   VistaRestoreTools !
 
Results I:
Regularization functionals in Local-Fourier domains
image_restoration_demo.m
 
In these examples classical regularization functionals based in rough spectral image models (L2) linear perception models (CSF) or auto-regressive models of power spectrum (AR) are compared to our perceptually-based regularization functional (Perceptual) that takes into account masking relations among local-Fourier coefficients and is consistent with relations among image coefficients in this domain.
 

Degradations:
Denoising

Blur (cut off freq.=27 cpd)

Gaussian Noise s2=200

 Deblurring and Denoising

Blur (cut off freq.=16 cpd)

Gaussian Noise s2=100

 JPEG Noise

Quality factor=7

 

 Salt and Pepper

1.2% affected

 

PSNR =  25    SSIM = 0.63

S-CIELab =  0.72  d_wav = 0.18

PSNR = 24.6       SSIM = 0.61

S-CIELab = 0.57    d_wav = 0.19

PSNR = 25    SSIM= 0.72

S-CIELab= 1.22   d_wav= 0.25

PSNR= 25.3     SSIM=0.83

S-CIELab= 0.37   d_wav=0.19

 

Regularization Results (Denoising)

 

Regularization Results (Deblurring+Denoising)

 

Regularization Results (JPEG degradation)

 

Regularization Results (Salt and Pepper)

 

 
Results II:
Wavelet-based denoising methods
image_denoising_demo.m
 
In these examples classical thresholding techniques based on image models that assume coefficient independence in the wavelet domain [Soft, Hard, Bayesian Gauss Marginals] are compared to our Kernel regularization technique [Kernel] that includes mutual information relations among wavelets in the kernel.
 

Wavelet Results (Denoising)

Wavelet Results (JPEG noise)


Download   VistaRestoreTools !