We model the blur on the observed picture as:
To deblur our observed image, we do several personal image enhancement stages as following:
All the process is done mainly to sharp the face area in the observed image and then by the
estimated blurring kernel we will deblur the whole photo. We make affine transformation to the
base picture so they will be in the same plane as the observed image, then next stage is detecting
and segmenting the faces from the observed image and the base images, called "Priors".
After segmenting every image, priors and observed is divided into four layers and by using the
texture layer we create eigenface basis, as demonstrate here:
Using the eigenface basis found in the previous stage we solve the minimization problem of
Maximum A Poseriori (MAP) Bayesian estimation:
Since the above minimization problem is not a quadratic problem, we to use an iterative solution to
solve this problem, we use Iterative Reweighted Least Square (IRLS) method:
The result of our algorithm:
Conclusions and suggestions:
About the project
| • | The entire solution is using in canonical loops, this kind of algorithm is better to implement in C/C++ instead of matlab. Iterative implementation in matlab is very slow |
| • | There is a need to find a good stopping condition to the IRLS |
| • | There is a need to figure out how to choose the best In a generic way |
| • | As it shown above we didn’t succeed to get to good result, Many approximations has been done in the way, some of them are probably the reason for the current results |
