When the captured image is blurring, one way of obtaining a clear image is via image deconvolution technique. In order to perform deconvolution successfully, usually we need to pay attention to the following issues:

  1. Image content, which determines what kind of apriori information you will use (Total Variation, Laplacian and so on) during the image reconstruction procedure.
  2. Image noise level, which determines the balance factor between apriori information and data consistence term.
  3. Point Spread Function (PSF), which is used to measure how blurring the image is.

Here my question is related to PSF estimation. The simplest way of finding the right PSF is to use different PSFs to perform deconvolution, and then select the best one visually. Then my question is how we can tell which PSF is the best visually.

  • 1
    $\begingroup$ You've asked about scoring your results visually. Do you mean a machine interpretation or a human interpretation? With humans, you have to deal with subjectivity. Seems like a visit to the eye doctor with a lot of A / B comparisons. There is a good paper that discusses quantitative teniques: "Learning a Bliind Measure of Perceptual Image Quality" $\endgroup$ – user2718 Jan 31 '13 at 14:32
  • $\begingroup$ One way is to not look at visual quality, but to impose prior constraints using properties that you know the image has (for example $1/f$). This paper has a lot of details on priors they use, and the result looks really impressive: cse.cuhk.edu.hk/~leojia/projects/motion_deblurring/… $\endgroup$ – thang Feb 2 '13 at 1:19
  • $\begingroup$ Could you please mark my answer? $\endgroup$ – Royi Apr 8 at 4:44

One way, though not directly visual, is to observe the image statistics. We have a pretty good idea about the statistics of Natural Images, more specifically, their Gradient Distribution (See Statistics of Natural Images and Models by Jinggang Huang, D. Mumford, What Makes a Good Model of Natural Images by Yair Weiss, William T. Freeman, The Statistic Distribution of Image Gradient - Mathematics StackExchnage and UCLA Stat232A-CS266A Statistical Modeling and Learning in Vision and Cognition - Chapter 2 - Empirical Observations: Image Space and Natural Image Statistics).

So what you can do is see how well the statistics of the result matches the models.
Actually many modern (Prior to the Deep Learning era) Deblurring Methods work according to this approach with very nice results.

Look at Blind Motion Deblurring Using Image Statistics by Anat Levin.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.