Bayesian-based image/video super-resolution techniques
In many imaging systems, the resolution of the detector array of the camera is not sufficiently high for a particular application. Furthermore, the capturing process introduces additive noise and the point spread function (PSF) of the lens and the effects of the finite size of the photo-detectors further degrade the acquired video frames. The goal of super-resolution (resolution enhancement) is to estimate a high-resolution (HR) image from a sequence of low-resolution (LR) images while also compensating for the above-mentioned degradations. Super-resolution (resolution enhancement) using multiple frames is possible when there exists subpixel motion between the captured frames. The objective of super-resolution is to reconstruct a high-resolution image from a sequence of low-resolution images, under the assumption that there exists subpixel motion between the low-resolution frames. Each of the frames provides a unique look into the scene. Super-resolution can reduce the expensive cost of the optical devices and overcome the resolution limitations of the optical sensor size. The success of super-resolution techniques mainly depends on three coupled aspects: registration, deblurring and noise filtering, corresponding to the three degradations: sub-pixel motion, blur and additive noise. In this work, I first explore the existing techniques in image/video super-resolution. Then, I focus on a Bayesian framework and study the following topics: (1) Joint MAP registration and super-resolution with Gaussian-Markov random field (GMRF) as the image prior. (2) Huber-Markov random field (HMRF) as the image prior and related topics. (3) Choice of regularization parameters. (4) Joint PSF blur estimation and super-resolution. (5) Super-resolution with inaccurate PSF and motion estimates. (6) EM-based super-resolution. Mathematical derivations and analysis are provided on these topics, with conclusions drawn based on the experimental results using both synthetic images and real data. Finally, the direction of future work is suggested.