本文为杨建超CVPR08上文章Image Super-Resolution as Sparse Representation of Raw Image Patches的读书笔记,针对如何运动压缩感知的理论、稀疏表示来进行超分辨重建。
Image Super-Resolution as Sparse Representation of Raw Image Patches
To_捭阖_youth
一、Theoretical basis and Motivation
1、Previous methods
①Conventional approaches
require multiple low-resolution images of the same scene,typically aligned with sub-pixel accuracy
The SR task—recover the original high-resolution image by fusing the low-resolution images.
Problem —much information is lost in the high-to-low generation process,the reconstruction problem is severely undetermined,
the solution is not unique.
methods —MAP、BTV、MRF(【2】new problems:the performance degrades rapidly if the magnification factor is large or if the is not enough low-resolution images to constrain the solution)
===>overcome(【12】an example-based learning(require enormous databases)、【22】LEE from manifolds learning(result in blurring effects,due to over-or under-fitting))
②Our method
not require any learning on high-resolution patches,insteadworking directly with the low-resolution training patches or features.
Two steps:
①a local model from the sparse prior to recover lost high-frequency for local details
(using the sparse prior,find the sparse representation for each local patch)
②the global model from the reconstruction constraint to remove possible artifact
(regularize and refine the entire image and global optimization to eliminate the reconstruction errors,suppressing noise and ensuring consistency with the low-resolution input)
The advantage:
①require a much smaller database
②superior performance,both qualitatively and quantitatively(the couputed sparse representation adaptively selects the most relevant patches in the dictionary to the best represent each patch of the given low-resolution image)
二、Super-resolution from Sparisity
1、Reconstruction constraint
Y=DHX (1)
(Y:a low-resolution image;X:a higher-resolution image X of the same scene;
H represent a burring filter;D the downsampling operator.)
2、Sparse representation prior
the patches x of the high-resolution image X:
2.1、Local Model from Sparse Representation
【Remarks】
①Subtract the mean pixel value for each patch,so that the dictionary represents image textures rather than abolute intensities
②(3)式F penalize visually salient high-frequency errors;F is chaosen as some kind of high-pass filter.
(more sensitive to high-frequency content image;the high-frequency components of low-resolution image are the most important for predicting the lost high-frequency content)
2.2、Enforcing Global Reconstruction Constraint
Notice do not demand exact equality between y and Dl*a;and noise<