Super Resolution

This chapter is concerned with super-resolution methods for image enhancement. Development of super-resolution methods is dated back to early 1980s, and the effort intensifies greatly in the past two decades. The essential objective of super-resolution is to enhance the quality of a low-resolution image, using a high-resolution counterpart. The super-resolution research started with combining information content from multiple low-resolution images and fusing them into an output image that can show finer details unavailable in any of the original low-resolution images. This branch of super-resolution research is referred to as multi-frame super-resolution, to be discussed in Sect. 11.1. Then the focus was shifted to the example-based approach or learning-based approach, in which external high-resolution and low-resolution patch pairs are created and a relationship between the high-resolution and low-resolution patch pairs are learned through a training dataset. When given a single new low-resolution image, which may not necessarily be related to any of the images in the training set, the learned algorithm is supposedly able to boost its resolution to a higher level. This second branch of super-resolution research is referred to as single-image super-resolution, to be discussed in Sect. 11.2. In the single-image super-resolution, the low-resolution image in the training set is typically created as a blurred and downsampled version of a high-resolution physical image. In the material science image acquisition, researchers often acquire a pair of images at different physical resolutions by using the same electron microscope over the same material sample. This process produces a pair of physical high-resolution and low-resolution images, rather than a physical image with its downsampled counterpart. If the resolution ratio is two-to-one, then the high-resolution image covers about 25% of the view field of the low-resolution image. The objective is then to enhance the quality of the low-resolution image area where there is no high-resolution counterpart. This last branch of research is referred to as paired images super-resolution, to be discussed in Sect. 11.3.