Implement iterative blind deblurring algorithm

[Andrey1994]

This algorithm iterates between updating the restored image and PSF, based on error function minimization

[windyuan]

looking forward to see this feature. thank you.

[windyuan]

@Andrey1994 is selected from GSoC (Google Summer of Code https://summerofcode.withgoogle.com) and now is working on this feature. More details from https://github.com/Andrey1994/ImageDeblurring/wiki
Some steps to enable.

1. investigation on deblurring algorithms (done)
2. simulate by matlab or OpenCV (WIP) and tune a good quality on synthetic and real blurring images.
3. porting and clean up code in libxcam with OpenCL.
4. quality and performance tuning.
   Thanks,
   Wind

[Andrey1994]

The first version of the IBD algorithm is merged(#450). Performance and quality improvements required(WIP)

[windyuan]

Cool. Thanks for @Andrey1994 efforts on IBD enabling.
This feature initiated from GSoC
libxcam Enable a debluring feature with OpenCL Design

[windyuan]

#454
#453
merged for debluring framework refine.

[excllent123]

 The result of function edgetaper don't  agree with matlab‘s,  and  I have tested  this function, It doesn't work. Can you update your code? or tell me how it works.  
Thank  you.

[windyuan]

@excllent123, could you post the picture from both matlab and libxcam. this can help us to figure out if it's a bug. about the pictures, if not convenient in public, you can also send mail to us (feng.yuan@intel.com and a1994ndrey@gmail.com)
@Andrey1994, could you help check whether it's a bug or something wrong in usage?

[Andrey1994]

Yep, I will check it. This function just blur the edges with some weights, could you send me how do you use it?
For now I see one mismatch, I used correlation instead convolution, so for non-symmetric psf it will be differ. I will compare it with matlab results more detailed and fix a code

[excllent123]

My test code is like this:
Firstly, I pad the image by using copyMakeBorder in opencv,
cv::copyMakeBorder(Img,Img,size,size,size,size,cv::BORDER_REPLICATE);
and then  use  edgetaper
Img = edgetaper(Img,BlurKernel); 
and the BlurKernel  is motion-blur-kernel,  
[0,0,0,0,0
 0 0 0 0 0
 1/5 1/5 1/5 1/5 1/5
 0 0 0 0 0 
 0 0 0 0 0]

I have compared with the output of this function with matlab, but the result is different,  
and It does'nt work,the result of  image I will send to your email.
Thank you for your help

[Andrey1994]

why do you pad image? edgetaper saves the initial sizes. Also Img has to be a single channel image.

I've checked matlab version and libxcam version, there are a little difference in weights calculation, but both version are good for ringing artifacts removing. @windyuan Have we fix weights initialization?

[windyuan]

@Andrey1994 If there's only one rule for weights(matlab version). we need fix. Else, keep the current version.

[Andrey1994]

@excllent123 Matlab calculates the autocorrelation of PSF in another and in my opinion rather strange way. They use 1d autocorrelation of vector which is a sum of each PSF rows\columns. In libxcam we use 2d autocorrealtion of PSF. That's why their weights are differ

Maybe their implementation outperform our, but I think both of them can be used. So, I think at least for a while we will keep current version.

[excllent123]

@Andrey1994 Thanks for your help.

[windyuan]

close this feature since already released in 1.1.0