MRF-BASED DISCONTINUITY ADAPTIVE ROBUST MOTION SUPER-RESOLUTION

Advisor: Dr. A.N. Rajagopalan

Professor, Department of EE

IIT Madras, Chennai

Super-resolution is a cost effective method to obtain a high-resolution image from a set of low-resolution observations by performing dealiasing, deblurring, and denoising operations. The focus here was on developing a robust and discontinuity adaptive method based on the Markov random field model for super-resolution of license plates of vehicles in real traffic video. A graduated non-convexity algorithm was used to arrive at a high quality license plate image. Further, super-resolution of images captured with real-aperture cameras was also considered. A computationally efficient algorithm based on iterated conditional modes which is also robust to any modeling and estimation errors of motion and space-variant blur was developed. Also, a new approach to super-resolution that fuses both motion and defocus cues was proposed. This is particularly useful when the sampling is non-uniform or when the number of observations is inadequate.