Open AccessNeural network approach to recognition of visible constellations by sky photo image
Author(s) -
В. А. Галкин,
А. В. Макаренко
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1864/1/012014
Subject(s) - observability , constellation , artificial neural network , computer science , metric (unit) , convolutional neural network , a priori and a posteriori , artificial intelligence , convolution (computer science) , stability (learning theory) , channel (broadcasting) , noise (video) , image quality , pattern recognition (psychology) , function (biology) , image (mathematics) , mathematics , machine learning , engineering , telecommunications , philosophy , operations management , physics , epistemology , astronomy , evolutionary biology , biology
The current paper demonstrates the effective capabilities of deep neural networks in solving the problem of identification of constellations from a photo of the sky in conditions of a priori uncertainty, incomplete observability and stochastic disturbances. The quality of solution 0,927 by metric F1 is obtained. In order to achieve the result, the original ResNet-like architecture of the convolution neural network was synthesized; statistical analysis of the dataset was carried out, the function of losses and strategy of neural network training were developed, and an accurate criterion of constellation observability in the image was formed. The observation of noise influence on the quality and stability of solutions was carried out.