Image Super-Resolution - Iterative Back Projection Algorithm

A simple maximum-likelihood algorithm for super-resolution.

Description

This project is a simple implementation of the Iterative Back-Projection (IBP)
algorithm for solving the Super-Resolution problem. It was first proposed
by Michal Irani in her 1991 paper "Improving resolution by image
registration". The imaging model being used is described by a paper by
Michael Elad, "Super-Resolution Reconstruction of an image". Both papers
can easily be found through a search in Google Scholar.

I've done two simplifications to the imaging model:
1) The image blur is assumed to be spatially invariant.
2) The spatial transformation model is a global translation.

To run the example code, follow the following steps:
1) Run SRSetup.m
2) Run SRExample.m

The example code operates on a dataset that is generated synthetically from
a reference image. Thus, the exact values for the blur sigma and the
translation offsets are being used.

• Required Products: MATLAB
• MATLAB release: MATLAB 7.11 (2010b)

[imi kh]

Hello,
How to calculate PSNR for comparison purpose, because original image and super-resolved image have different sizes, kindly look into the matter.

[Victor May]

Hi Imi,
the function SynthDataset has an output called croppedOriginal which crops the original image to the size of the super-resolved image. In order to prevent confusion about this, I've updated the example script to include a mean-square error calculation.

[imran khan]

Hello victor,
Where is the updated file, kindly post the updated script.

[Victor May]

I posted it. It might take a few days until it will show up here.

[imi kh]

Hello,
The code only deals with translation motion, how to extend it if the low resolution images are also rotated,
still code is not updated.

[Eric]

Line 41 of SRExample reads:
im = imread('C:\Vic\Work\Src\SuperResolution\Data\typewrtr.jpg');
This doesn't run on somebody else's machine without tweaking.

[imi kh]

Well the problem is now on increase the number of low resolution images, if you increase the numImages =10 and run the code the output is not good whats the problem ?

[Victor May]

Increasing the number of images causes the solution to suffer from overfit. Adding a prior on the image should help. Search the web for papers on priors for super-resolution.

[imi kh]

Can you please increase the images up to 30 and see the results.

[imi kh]

well i do add laplasian term as a prior, but i think the problem lies some where else because if you increase number of low resolution images the output destroyed.

[Yitzi Pfeffer]

In order to avoid divergence when using a large number of LR images. Try to use a coefficient smaller than 1 (I tried 0.25) in the update equation for x:
x = x + 0.25*res; instead of x = x + res; This ensures smaller (and safer) steps in the Gradient Descent procedure.

[imi kh]

Can some one tell me why we model noise as additive Gaussian noise in most of the super-resolution algorithms.