Deblurring shaken and partially saturated images
read more
Citations
Unnatural L0 Sparse Representation for Natural Image Deblurring
Deep Convolutional Neural Network for Image Deconvolution
Deblurring Text Images via L0-Regularized Intensity and Gradient Prior
Recording and playback of camera shake: benchmarking blind deconvolution with a real-world database
A Comparative Study for Single Image Blind Deblurring
References
Image quality assessment: from error visibility to structural similarity
Maximum Likelihood Reconstruction for Emission Tomography
Bayesian-Based Iterative Method of Image Restoration
An iterative technique for the rectification of observed distributions
Related Papers (5)
Frequently Asked Questions (18)
Q2. What are the future works mentioned in the paper "Deblurring shaken and partially saturated images" ?
As future work, their algorithm could potentially be extended to handle other sources of ringing, such as moving objects, impulse noise, or post-capture non-linearities ( such as JPEG compression ). One alternative direction for future work is the investigation of new regularisers that are targeted at suppressing ringing. Whenever it is possible to identify poorly-estimated latent pixels, their approach has the potential to reduce artefacts by decoupling these pixels from the rest of the image.
Q3. What is the classical algorithm for non-blind deblurring with Poisson noise?
The classical algorithm for non-blind deblurring with Poisson noise is the Richardson-Lucy algorithm, which is described by the multiplicative update equation:f̂ t+1 = f̂ t ◦A⊤ ( gAf̂ t), (17)where the division is performed element-wise.
Q4. How did the authors find that dilating the masked regions reduced ringing?
The authors found that dilating the masked regions using a9×9 square window further reduced ringing, at the expense of leaving more blur around the saturated regions.
Q5. What is the only modification required to estimate PSFs for saturated images using this blind algorithm?
The only modification required to estimate PSFs for saturated images using this blind algorithm is to discard potentially-saturated regions of the blurred image using a threshold.
Q6. How can the authors perform non-blind deblurring with missing data?
In order to perform non-blind deblurring with missing data, the authors can define a mask m of binary weights, where mi = 0 if pixel i is missing, and mi = 1 otherwise.
Q7. What is the effect of a poor decision at a neighbouring pixel?
Chou and Brown point out that due to the coupling of pixels in a Markov Random Field image model, the estimate for a pixel with strong observations may be negatively impacted by a poor decision at a neighbouring pixel with weak observations.
Q8. What is the effect of initialising with the bright image?
When the authors initialise with the blurred image, or random values in [0, 1], the algorithm is forced to attempt to estimate the bright values, and in doing so causes ringing.
Q9. What is the reason that poorly-constrained directions appear as ringing?
The reason that poorly-constrained directions appear as ringing, and not as some other kind of visual artefact, is that the nullspace of a PSF tends to be spanned by vectors having small support in the frequency domain, but large spatial support.
Q10. What is the main factor causing poorly-constrained spatial frequencies to become amplified?
In this work, the authors posit that for the case of saturation, the main factor causing poorly-constrained spatial frequencies to become amplified is that there exist pixels in the sharp image that exceed the image’s range, i.e. ∃ j : fj >
Q11. Why do the authors have less ringing in deblured images?
In most cases their results exhibit less ringing than those of Cho et al. (2011), which is due to the fact that the authors explicitly decouple the estimates of bright pixels from other pixels, in addition to removing saturated blurred pixels.
Q12. What is the main reason why the authors have proposed a non-blind deblurring?
Based on this analysis, the authors have proposed a non-blind deblurring algorithm, derived from the Richardson-Lucy algorithm, which is able to deblur images containing saturated regions without introducing ringing, and without sacrificing detail in the result.
Q13. What is the difficulty of inverting the true non-linear forward model?
Given the difficulty of inverting the true non-linear forward model, alternative approaches to handling saturated pixels are needed, in order to prevent ringing artefacts from appearing.
Q14. What are the famous algorithms for non-blind deblurring?
Many algorithms exist for non-blind deblurring in the linear (non-saturated) case, perhaps most famously the Wiener filter (Wiener 1949) and the Richardson-Lucy algorithm (Richardson 1972; Lucy 1974).
Q15. What is the second way that ringing can arise?
The second, and harder to tackle, way that ringing can arise is if there are some unmodelled / outlier pixels in the blurred image g, e.g. due to saturation or impulse noise.
Q16. How do the authors denote the set of blurred pixels that are not affected by fS?
(14)The authors denote by V the set of blurred pixels that are not affected by fS , i.e. V = { i ∣ ∣(AfS)i = 0 } , and construct the corresponding binary mask v (where vi = 1 if i ∈ V , and vi = 0 otherwise).
Q17. What is the main reason why the authors have developed a non-blind deblurring?
In addition, this underlying principle of decoupling poorly-estimated latent pixels could also be applied within other non-blind deblurring algorithms, that are faster or more suitable for different noise models than the Richardson-Lucy algorithm, e.g. (Levin et al.
Q18. What is the shape of the smooth versions of R and R′?
Figure 8 shows the shape of these smooth versions of R and R′.Equation (20) can be roughly interpreted as weighting the blurry pixels according to the value of R′: in the linear (unsaturated) portion where x < 1, R(x) ≃ x and R′(x) ≃ 1, so that the term in parentheses is the same as for the standard RL algorithm.