ABSTRACT
This chapter examines multiframe image super-resolution in a probabilistic framework. It deals with the generative model and simple Maximum Likelihood and Maximum a Posteriori (MAP) solutions for the super-resolution image. The chapter illustrates the consequences of inaccurate point estimates of the latent parameters on the simple MAP algorithm. It introduces the simultaneous MAP algorithm, which estimates the super-resolution image together with parameters such as the image registration, allowing the high-resolution information to influence and improve the estimates of the latent parameters. The chapter shows how Bayesian marginalization can integrate the latent parameters out of the problem, leading to a cost function in terms of the low-resolution images that can be optimized with respect to the high-resolution pixels directly. It highlights the importance of considering latent quantities such as image registration or point-spread function size as part of the super-resolution problem instead of estimating and fixing them in advance.
