Investigation of Different Image Super Resolution Methods on Paired Electron Microscopic Images
Loading...
Date
2020-11-11
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis is concerned with investigating super-resolution algorithms and solutions for handling electron microscopic images. Please note two main aspects differentiating the problem discussed here from those considered in the literature. The first difference is that in the electron imaging setting, a pair of physical high-resolution and low-resolution images is used, rather than a physical image with its downsampled counterpart. The high-resolution image covers about 25\% of the view field of the low-resolution image, and the objective is to enhance the area of the low-resolution image where there is no high-resolution counterpart. The second difference is that the physics behind electron imaging is different from that of optical (visible light) photos. The implication is that super-resolution models trained by optical photos are not effective when applied to electron images. Focusing on the unique properties, a global and local registration method is devised to match the high- and low-resolution image patches and different training strategies are discussed for applying deep learning and non deep learning based super-resolution methods to the paired electron images.
This thesis investigates the uniqueness of the super-resolution problem on paired electron microscopic images. After extensive experimentation and comparison on 22 pairs of electron images, it is now believed that the self-training strategy, in which the training images come from the same image pair of the test set, leads to better super-resolution outcomes, despite the relatively small training data size. Deep learning-based super-resolution methods show the best performances, whereas a revised paired library-based non-local mean method shows advantage in training time and interpretability.
Paired images super-resolution has important implications in many research areas. Paired electron images are rather common in scientific experiments, especially in material and medical research. Due to the destructive imaging process while using electron sources, researchers tend to use low-energy beams or subject the samples to a short duration of exposure to protect the sample. As a consequence, low-resolution images are generated. Super-resolution methods, which can subsequently boost these low-resolution images to a higher resolution, are much desired in scientific researches using electron imaging.
Description
Keywords
Paired-image super resolution, electron microscopic images, deep learning, library-based non-local mean.