Learning low-level vision
read more
Citations
Image Super-Resolution Via Sparse Representation
Learning a Deep Convolutional Network for Image Super-Resolution
Robot vision
Super-resolution through neighbor embedding
SIFT Flow: Dense Correspondence across Scenes and Its Applications
References
Neural networks for pattern recognition
Stochastic Relaxation, Gibbs Distributions, and the Bayesian Restoration of Images
Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
Neural Networks for Pattern Recognition
The Laplacian Pyramid as a Compact Image Code
Related Papers (5)
Frequently Asked Questions (8)
Q2. How many samples are allowed to be within two standard deviations of the observed image patches?
The authors set σi to allow roughly 10 samples at each node to be within two standard deviations of the observed image patches, and set σs to allow roughly 5 or 10 matrix transitions to be appreciably different than zero.
Q3. What is the scene explanation given new image data?
To find the best scene explanation given newimage data, the authors apply belief propagation in the Markov network even though it has loops, an approach supported by experimental and theoretical studies.
Q4. How do the authors compute the compatibility between neighboring patches?
To compute the compatibilities between neighboring patches at different scales, the authors first interpolated the lower-resolution patch by a factor of 2 in each dimension so that it had the same sampling rate as the high resolution patch.
Q5. What is the way to sample the scene and image variables?
for reasonably sized patches, the scene and image variables need to be of a high enough dimensionality that an evenly-spaced discrete sampling of the entire high dimensional space is not feasible.
Q6. What is the first method for finding the compatibility functions of Eqs. 21 and 20?
The first method uses the message-passing rules of Eqs. (21) and (20), based on the joint probability factorization which is not valid for a network with loops.
Q7. What is the way to handle textured areas?
the successes of recent texture synthesis methods (Heeger and Bergen, 1995; DeBonet and Viola, 1998; Zhu and Mumford, 1997; Simoncelli, 1997), gives us hope to handle textured areas well, too.
Q8. What is the compatibility function between patches at the same scale?
Letting dljk be the pixels of the lth candidate in the high resolution patch k, and dmkj be the pixels of the mth candidate in the interpolated lowresolution patch j , the authors take as the compatibility,9 ( xlk, x m j ) = exp−|dljk−dmkj |2/2σ 2s , (30) where the authors scale σs to give the same per pixel variance as for the compatibility function between patches at the same scale.