Abstract
A feed-forward artificial neural network has been implemented to the problem of removing cosmic-ray hits (CRH) from CCD images. The results of a number of tests demonstrate the effectiveness of this method especially for undersampled stellar profiles. The problem of optimal and low price preparing of training data, which could enable real-time or at least fast post-processing filtering out of CRH is discussed. The training and test ensembles were composed of a number of synthetic stellar profiles involving different S/N ratios and CRH images taken from real data. Certain aspects of the network’s architecture and its training efficiency for different modes of the back-propagation procedure as well as for the pre-process normalization of data have been examined. It is shown that for training set composed of stellar images and CRH at a ratio of 1:2 recognition can reach 99% in the case of stars and 96% for CRH. To determine the extent to which the cognition power of a network trained using an ensemble of circular symmetric stellar profiles of a given radius can be generalised the test data included stellar profiles of different radii, as well as elongated profiles. The goal was to mimic temporal changes in seeing as well as such problems as image defocusing, the lack of isoplanatism and improper sideral tracking of a telescope. The experiments provided us with the conclusion that for S/N > 10 excellent classification property is maintained in cases where the change in the radius of a circular profile is up to 30%, as well as for elongated profiles where the longest dimension is almost double that of the shortest one. Moreover, the generalization capability has been investigated for test images of synthetic pairs of overlapping stars with different distances between components. Almost 99% recognition efficiency was achieved even if the separation was nearly three times the radius of the stellar profile, a case when two stars could be analyzed by appropriate software as separate objects. The example of removal of CRH from real CCD images is presented to give an idea of how an algorithm based on a neural network can work in practice. The result of such an experiment appears fully consistent with the conclusions drawn from the tests made on synthetic data.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.