Image Restoration of Degraded Images (Super Resolution of Color Image from Multiple Low Resolution Images)

Abstract

The use of digital imaging systems is widespread in several applications. These applications include taking images for astronomical and meteorological uses, consumer photography and images taken aerially. It is important to restore distorted images taken for such applications. Theoretically and practically this process of restoration of images poses an interesting problem of image processing. There are many types of distortion: 1) Blurring - due to incorrect focus and movement 2) Zooming in – Due to zooming of images etc. In zooming problems we have to find the missing image points and that is a challenging problem. Imaging systems often provides us low resolution images of the same scene. Every LR image got some information. Super resolution approach gathers this information from multiple LR images and combines it to get an HR image. Super resolution approach gathers this information from multiple LR images and combines it to get an HR image. Basic assumption is the availability of multiple images of the same scene to obtain different information in each LR image, such that some relative motion must exist between LR images or video sequence. The main objective of Super-Resolution (SR) is to recover a high resolution (HR) image from multiple Low resolution (LR) images such that zooming problem is solved. Super-Resolution (SR) image reconstruction is the procedure of joining several spatially misaligned low resolution (LR) images into a single high resolution (HR) image. In this route accurate image registration is the key step and homography serves as basis for image registration. Homography estimation between LR images was done using Harris corner detector and Random Sampling Consensus (RANSAC). Harris corner detector finds the corner points of LR images and RANSAC finds the transformation matrix by filtering out mismatched points. A new SR approach is proposed in which LR images are then projected to HR grid using transformation matrix. New SR approach avoids iterative process for convergence making computation faster for better results. Remaining unfilled spaces in HR image are filled with interpolation.