| Abstract |
Deconvolution reconstructs the distribution of a physical quantity from related quantities. The present work surveys popular deconvolution methods with a comprehensive benchmark targeted on Cherenkov astronomy. Meanwhile, a novel unified view on deconvolution is established from the perspective of machine learning. Within this view, the essential building blocks of deconvolution methods are identified, opening the subject to several directions of future work. Particular attention is turned to DSEA, the Dortmund Spectrum Estimation Algorithm, which employs a classifier to reconstruct the sought-after distribution. An improved version of this algorithm, DSEA+, is proposed here. This version is more accurate and it converges faster than the original DSEA, thus matching the state of the art in deconvolution.
|
