|M.Sc Student||Stern Ori|
|Subject||Super Resolution Techniques - Comprehensive Review|
|Department||Department of Electrical and Computer Engineering||Supervisor||PROFESSOR EMERITUS Arie Feuer|
In a wide range of applications like medical, military, space research, terrain mapping and others, it is required to acquire high resolution images. High resolution images are expected to enhance the capability of detection and identification of details in the image, improve performances of pattern recognition algorithms, as well as the performances of automatic classification algorithms in computerized systems.
Super-Resolution algorithms are aimed to reconstruct real high frequency information up to a new Nyquist frequency that was aliased due to low sampling rate.
Since in many cases and applications there is an infinite set of solutions that fits the measured low-resolution images, one can relate the problem of super-resolution as an estimation problem. Given one low resolution, and without essential a prior knowledge regarding the solution, there is no way to remove the aliasing due to the low sampling rate, and one can not reconstruct information at frequencies above the Nyquist frequency of the detector. Reconstructing high frequency information is the purpose of image super resolution. Thus, in many super resolution algorithms and applications, in order to produce improved resolution images and remove the aliasing effect, a set of measurements of the scene is required. This fundamental requirement is achieved by acquiring a set of low resolution images with a relative sub-pixel shift between the acquisition system and the scene, or by using multiple cameras. The motion and the aliasing caused due to under sampling are essential properties for resolution improvement.
On the other hand, there are super resolution methods such as example based methods, which may be able to produce improved resolution image using one low resolution image, based on prior learning from a set of training sequence of images.
The paper concentrates in super-resolution of gray level stills images and deals with the following three issues of super-resolution: (a) Limitations and conditions for resolution improvement. (b) Methods and algorithms for image super-resolution. (c) Does the resolved image contain new high frequency information that was lost during the images acquisition, or does the resolved image just looks better?
The paper includes a comprehensive review of super resolution methods and algorithms for gray level still images. The major pros and cons of each method are discussed. The paper also provides simulation results of two space domain methods - MAP and POCS.