Quality measure for Blind deblurring methods

(Matlab code)

The following reconstruction measure was specially developed for the Blind Image Deblurring (BID) problem, and it was used for the first time in [1]. The measure is based on the Increased Signal to Noise Ration (ISNR). However, the computation of a meaningful ISNR in blind deblurring situations raises some special issues that we now address.

These special issues have to do with the infinite number of solutions existing in the blind deblurring problem. There are two different kinds of variability of solutions that we need to distinguish, here the BID problem. One corresponds to changes in the shape of the estimated blurring filter’s Point Spread Function (PSF), compensated by matching changes in the estimated image. In this case, different estimated images will, in general, exhibit different amounts of residual blur and/or different artifacts (e.g. ringing), which affect their quality. These degradations should be taken into account by the quality measure. However, other two forms of variability are of a different kind and should not be account by the quality measure:

(1) affine transformations of the intensity scale of the filter, compensated by affine transformations of the estimated image, and

(2) small translations of the blurring filter’s PSF, compensated by opposite translations of the estimated image. These degradations do not affect the quality of the deblurred image, and the restoration measure should be insensitive to them.

To address these invariance issues, we have performed an image adjustment (spatial alignment and intensity rescaling) before comparing the images with the original sharp one. The estimated image was spatially aligned, and the pixel intensities were rescaled by an affine transformation, so as to minimize the image’s squared error relative to the original sharp image. In [1] we used this demo with a spatial alignment with a maximum shift of 3 pixels in each direction and with a resolution of a ¼ of a pixel. For more details on this quality measure, see [1].

Reference:

[1] M. S. C. Almeida and L. B. Almeida, "Blind and Semi-Blind Deblurring of Natural Images", accepted in IEEE Trans. Image Processing. ( Preprint PDF , Blind Deblurring Quality Measure )

Matlab Code: Blind Deblurring Quality Measure (.rar)

If you find any bug, please report it to me: M. S. C. Almeida. Thank you!

License:  This code and these data sets are copyright of Luis B. Almeida and Mariana S.C. Almeida. Free permission is given for their use for nonprofit research purposes. Any other use is prohibited, unless a license is previously obtained. To obtain a license please contact Luis B. Almeida or Mariana S.C. Almedia. 

This package is compressed with win-rar. Download it here.