A generalized Gaussian image model for edge-preserving MAP estimation
read more
Citations
Proximal Splitting Methods in Signal Processing
Proximal Splitting Methods in Signal Processing
Markov Random Field Modeling in Image Analysis
Deterministic edge-preserving regularization in computed imaging
Markov Random Field Modeling in Computer Vision
References
Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images
Determining optical flow
On the statistical analysis of dirty pictures
Related Papers (5)
Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images
Nonlinear total variation based noise removal algorithms
Frequently Asked Questions (8)
Q2. What is the function that determines the tendency of neighbors in X to be attracted?
The derivative of p determines the tendency of neighbors in X to be attracted and plays a role analogous to the influence function of robust statistics[ll, 271.Â
Q3. What is the resulting form of the probability density function for X?
The resulting form of the probability density function for X is similar to the generalized Gaussian distribution commonly used as a noise model in robust detection and estimation[l4].Â
Q4. How do the authors compute the global MAP estimate?
In practice, the authors have found that the global MAP estimate may be computed by alternating a complete pass of local minimization with a single iteration of a gradient-based method.Â
Q5. What is the significance of the median filtering prior?
Since median filtering has been shown to be of broad practical importance in image filtering, the authors believe this suggests that methods based on the GGMRF prior can also be practically useful in a variety of image estimation applications.Â
Q6. What is the function that allows sharp edges to form in the reconstructed image?
Notice that the function is quadratic near zero, but the flat region beyond the value T allows sharp edges to form in the reconstructed image.Â
Q7. How many iterations did the phantom take to get to the correct estimate?
Initial stages of convergence proceeded rapidly in each case, but the approximately 450 iterations required for convergence of the estimate when p = 1 was over an order of magnitude higher than in the Gaussian case.Â
Q8. What is the connection between the median filter and the global MAP estimate?
This connection is of interest since median filters are a useful class ofhomogeneous edge preserving nonlinear filters for image processing.Â