Open AccessSubpixel precision in registration of multimodal datasets
Author(s) -
Matěj Lébl,
J. Blažek,
Jana Striová,
Raffaella Fontana,
Barbara Zitová
Publication year - 2020
Publication title -
iop conference series. materials science and engineering
Language(s) - English
Resource type - Journals
eISSN - 1757-899X
pISSN - 1757-8981
DOI - 10.1088/1757-899x/949/1/012007
Subject(s) - subpixel rendering , computer science , modalities , task (project management) , artificial intelligence , mutual information , focus (optics) , image registration , rotation (mathematics) , machine learning , data mining , computer vision , pattern recognition (psychology) , pixel , image (mathematics) , social science , sociology , physics , management , optics , economics
The motivation for our research is the huge demand for registration of multimodal datasets in restorers practice. With an increasing number of various screening modalities, each analysis built on the acquired dataset starts with the registration of images acquired from different scanners and with varying levels of mutual correspondence. There is currently no well-suited state of the art method for this task. There are many existing approaches, i.e. based on control points or mutual information, but they do not provide satisfying (subpixel) precision, thus the registration is very often realized manually in Adobe Photoshop or any similar tool. Another popular option is to use scanners able to produce registered datasets by design. During the last 10 years, datasets from these devices have extended available analytical techniques the most. In our research, we focus on solving the mentioned registration task. In [1] we concluded that the work with misregistered modalities is possible but limited. Now we present results of our experiments challenging these limits and conditions under which we can precisely register data from different modalities. The achieved results are promising and allow usage of more complex artificial neural networks (ANN) for dataset analysis e.g. [2]. We describe the construction of registration layers for estimation of shift, rotation and scale and a useful strategy and parametrization for ANN optimizer.