8/6/2019 Kolte Arora Image Super Resolution

1/2

Image Super-Resolution

Members:

Ritesh Kolte (rkolte@stanford.edu), Abhishek Arora (arorabhi@stanford.edu)

Super-resolution is the task of obtaining a high-resolution image of a scene given low resolution image(s) of the

scene. Applications of super-resolution include forensic, satellite, medical imaging, surveillance, displaying video

on large screens, etc [1]. Obtaining high-resolution images directly via better hardware (better image sensors, larger

chip size) is quite costly. Furthermore, most of the smart-phones today would be hard-pressed to incorporate

hardware enhancements for achieving high resolution images. Hence, a signal processing approach to increase the

resolution of images obtained from the phone camera is desirable.

Image interpolation methods are not considered as super-resolution methods since even the ideal sinc interpolation

cannot recover the high frequency components that are lost in the low-resolution sampling process. Most of the

super-resolution approaches work on the principle of combining multiple slightly-shifted low-resolution images of

the scene [2]. This involves image registration, interpolation and deblurring as the basic operations. However, this

technique is numerically limited to small increases in resolution. There also exist methods based on techniques like

gradient profile priors [3] and sparse representation of images in an over-complete dictionary [4]. A recent approach

based on a single image was proposed in [5] that exploits the redundancy of patches within the image and combines

this information with example-based techniques such as [6]. This technique overcomes the limitation on increase in

resolution incurred in e.g. [2].

In this project, we aim to explore the techniques outlined in [5]. Many a times, one wants to zoom into an image

while preserving image quality, edge effects and texture. This algorithm provides a higher resolution image of the

low resolution input image while maintaining the image quality. It exploits the fact that a small patch in an image

(say 5x5) tends to occur many times in exact or similar forms in the same image. Such similar patches are obtained

by the approximate nearest neighbor search algorithm described in [7]. This patch redundancy is exploited toconstruct the high-resolution image. The process is highly computationally intensive and may take up to a few

minutes. Hence, the system will be implemented in MATLAB. Time permitting; we can extend this application for

use in smart phones (despite the long processing time, since this is not a real-time operation). The image can be

submitted to the server for processing tasks, the server replying with the high-resolution image.

8/6/2019 Kolte Arora Image Super Resolution

2/2

An example of super-resolution using the method in [5] is shown above. If the resolution of the original image is

increased via bicubic interpolation we can see that the high frequency components are not recovered. The result of

super-resolution using the method in [5] recovers high frequency details.

References

[1] S.C. Park, M. K. Park, and M. G. Kang, Super-resolution image reconstruction: a technical overview,IEEE

Signal Processing Magazine, vol. 20, no. 3, pp. 21-36, 2003.

[2] Elad, M.; Feuer, A.; , Restoration of a single superresolution image from several blurred, noisy, and

undersampled measured images,Image Processing, IEEE Transactions on , vol.6, no.12, pp.1646-1658, Dec 1997.

[3] Jian Sun; Zongben Xu; Heung-Yeung Shum; , Image super-resolution using gradient profile prior, Computer

Vision and Pattern Recognition, 2008. CVPR 2008. IEEE Conference on , vol., no., pp.1-8, 23-28 June 2008.

[4] J. Yang, J. Wright, T. Huang and Y. Ma Image Super-Resolution via Sparse Representation in IEEE

Transactions on Image Processing, 2010.

[5] D. Glasner, S. Bagon, and M. Irani. Super-resolution from a single image, In Proc. of ICCV '09. IEEE

Computer Society Press, 2009.

[6] William T. Freeman, Thouis R. Jones, Egon C Pasztor, Example-Based Super-Resolution, IEEE Computer

Graphics and Applications, v.22 n.2, p.56-65, March 2002.

[7] Sunil Arya and David M. Mount, Approximate nearest neighbor queries in fixed dimensions, InProceedings

of the fourth annual ACM-SIAM Symposium on Discrete algorithms(SODA '93).