Classic versus deep learning approaches to address computer vision challenges : a study of faint edge detection and multispectral image registration
Year: 2021
Type: dissertation
Abstract: Computer Vision involves many challenging problems. \nWhile early work utilized classic methods, in recent years \nsolutions have often relied on deep neural networks. In this \nstudy, we explore those two classes of methods through two applications that are at the limit of the ability of current computer \nvision algorithms, i.e., faint edge detection and multispectral \nimage registration. We show that the detection of edges at a \nlow signal-to-noise ratio is a demanding task with proven lower \nbounds. The introduced method processes straight and curved \nedges in nearly linear complexity. Moreover, performance is of \nhigh quality on noisy simulations, boundary datasets, and real \nimages. However, in order to improve accuracy and runtime, a \ndeep solution was also explored. It utilizes a multiscale neural \nnetwork for the detection of edges in binary images using edge \npreservation loss. The second group of work that is considered \nin this study addresses multispectral image alignment. Since \nmultispectral fusion is particularly informative, robust image \nalignment algorithms are required. However, as this cannot be \ncarried out by single-channel registration methods, we propose \na traditional approach that relies on a novel edge descriptor using a feature-based registration scheme. Experiments demonstrate that, although it is able to align a wide field of spectral channels, it lacks robustness to deal with every geometric \ntransformation. To that end, we developed a deep approach for \nsuch alignment. Contrarily to the previously suggested edge \ndescriptor, our deep approach uses an invariant representation \nfor spectral patches via metric learning that can be seen as a \nteacher-student method. All those pieces of work are reported \nin five published papers with state-of-the-art experimental results and proven theory. As a whole, this research reveals that, \nwhile traditional methods are rooted in theoretical principles \nand are robust to a wide field of images, deep approaches are \nfaster to run and achieve better performance if, not only sufficient training data are available, but also they are of the same \nimage type as the data on which they are applied.
Author: Yehonatan Nati Ofir
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3191159868