Today, as the Internet and cellular communication’s influence increases rapidly, there is a need of faster transfer rates of information.
Abstract
Today, as the Internet and cellular communication’s influence increases rapidly, there is a need of faster transfer rates of information.
In the cellular world, for instance, there is bandwidth limit. That makes the transfer of high-resolution images ineffective on the cellular net.
This project suggests a solution for this problem by enabling the transfer of low-resolution images on the net, and by that decreasing the latency
of the transfer. Then when the image arrives to the end station, i.e. the cellular device, it is transformed to a high-resolution image.
This transformation is the Super Resolution Process.
 :
Project Goal
The goal in this project is to implement a Super Resolution System that works on a single image. It gets as input a low-resolution image,
and after processing it using side information (retrieved from the image itself), it produces a high-resolution image.
 :
The Solution
The Super Resolution System is composed out of two parts, as shown in the diagram below.
 :
Interpolation Unit – In this part the image is being interpolated using a Spline interpolation method. Cubic splines interpolate (pass through) the data
with piecewise cubic polynomials. The goal of this interpolation is to derive the smoothest possible cubic spline that simultaneously interpolates the
data and satisfies the monotonicity constraint. This process by itself is not sufficient. It increases the number of pixels in the image but do not increase
the frequency, which results in blurred edges. This is why we need the second part, which is the major part of this project.
Enhancement Unit – This part tries to solve the side effects of the first part. It enhances the edges in the image on one hand, and filters noise on the
other hand. This process is done using backward and forward non-linear diffusion based on the Perona-Malik diffusion equation:
 :
 :
The forward diffusion is used for noise filtering. It is done with a non-negative descending diffusion coefficient:
 :
 :
Two approaches of the selective inverse diffusion were tested.
 :
1. In the first one the image was diffused with the above diffusion coefficient. The diffusion parameters were calculated locally using a window.
Therefore, different thresholds in different locations accomplish the enhancement.
2. In the second approach the edges of the image were calculated using the ‘Canny Edge Detection’ algorithm, first. Then, the image was diffused with selective inverse
diffusion on the edges, and with forward diffusion on the non-edges areas.
 :
Tools
All the algorithms in this project were implemented in Matlab – version 5.
 :
Conclusions
- The interpolation/diffusion algorithms were found effective for building a super-resolution system
- It seems that the edge detection algorithm is slightly better than the windowed selective inverse algorithm
- This system is efficient mostly for images which do not contain texture
 :
Results
The algorithms were tried on several low-resolution images.
 :
 :
 :
Acknowledgments
I would like to thank my supervisor Guy Gilboa for his support and guidance throughout this project. I would also like to thank the VISL laboratory
and the Ollendorff Minerva Center Fund, which supported this project.