Super Resolution - by Team Amir & Friends
Copyright (C) 2014 by Amir Shachar ,Oren Renard and Saar Arbel



   What is Super Resolution   back to top
The process of inducing more information out of a set of images (or a single one) onto a high resolution image.
It provides a very helpful tool for image processing of a scene and lets us improve existing data.
Several methods were in use a few years ago, most of which made very big assumptions on the input.
Today the main focus of super-resolution research is the non-parametric SR. One example of such an algorithm is the "by example" algorithm which we implemented and can be activated by the left button in our GUI.
In short we could describe super resolution like so:
Imagine a certain scene which we want to capture, then that scene is the high resolution image.
We capture it using a camera. Our model assumes a few characteristics for each camera:
  • Decimation (down sampling)
  • Geomerical Tanslation
  • Optical Blur
    • Each of the above can be described as a linear operation, which means it is reversable - Our entire roject is to reserve these distortions.




       Our program   back to top
    We provide an application that can run through "Matlab", in which you can select a picture and it will do SR to it.
    Today, pictures are taken in very good resolution, so to find an image in low-resolution is difficult.
    Therefore we take an image and downgrade it - we blur, sample and transform the image supplied to us into a data set consisting of a number of low resolution copies. Then, we apply the algorithm to those images and achieve a reconstruction of the original image (HR).




       How does it work?   back to top
    Firstly, we need to make a clariffiation: there are two main approaches that we implemented, each one of them works diffrently. The "by example" algorithm is an iterative solution which treats the high resolution image as the result of a set of equations (the equations are created from the low-resolution images). In contrast to the iterative solution, the patch-match algorithm modelized the super-resolution problem as a search problem. We treat the given image as a set of patches (a smaller part of the image, its size is predefined) and for each patch we search for its larger counterpart. In order to search successfully the algorithm required a large database of patch-pairs (a small patch and an equivalent enlarged patch).




       Origin of the assumptions   back to top
    Lets start by how a camera works: light (in the form of photons) hits objects and gets deflected in all directions. The photons that reach our eyes make up the world that we see. Due to technological limits the camera can not contain all the photons that it recieves, so the image we get is of lower quality. The low quality is the result of two factors: the camera can not be precise enough to notice each photon so it has to only take some of them. Secondly, light that travels through anything that is not void gets theformed and blurred. These two factors account for two of the opreations we mentioned and the third one (the geometrical translations) is necessary because we use multiple images that were taken from slightly different points of view.


       References   back to top
    The entirely project is follow the "Back Projection" way, mentioned in the papers of M.Irani and her cooperates. link here


       Download   back to top
    Downlaod Our presentation:      
    Download Project: