A genetic programming algorithm for synthesis of object detection systems is proposed and applied to the task of license plate recognition in uncontrolled lighting conditions. The method evolves solutions represented as data flows of high-level parametric image operators. In an extended variant, the algorithm employs implicit fitness sharing, which allows identifying the particularly difficult training examples and focusing the training process on them. The experiment, involving heterogeneous video sequences acquired in diverse conditions, demonstrates that implicit fitness sharing substantially improves the predictive performance of evolved detection systems, providing maximum recognition accuracy achievable for the considered setup and training data.
@INPROCEEDINGS { Krawiec:2013:EvoIASP-short,
ABSTRACT = { A genetic programming algorithm for synthesis of object detection systems is proposed and applied to the task of license plate recognition in uncontrolled lighting conditions. The method evolves solutions represented as data flows of high-level parametric image operators. In an extended variant, the algorithm employs implicit fitness sharing, which allows identifying the particularly difficult training examples and focusing the training process on them. The experiment, involving heterogeneous video sequences acquired in diverse conditions, demonstrates that implicit fitness sharing substantially improves the predictive performance of evolved detection systems, providing maximum recognition accuracy achievable for the considered setup and training data. },
ADDRESS = { Vienna, Austria },
AUTHOR = { Krzysztof Krawiec and Mateusz Nawrocki },
BOOKTITLE = { Applications of Evolutionary Computing, EvoApplications 2012 },
DOI = { doi:10.1007/978-3-642-37192-9_38 },
EDITOR = { Anna I. Esparcia-Alcazar et al. },
ISBN13 = { 978-3-642-37191-2 },
KEYWORDS = { genetic algorithms, genetic programming, pattern recognition, image analysis, implicit fitness sharing, car number plate recognition },
MONTH = { {3-5 } # apr },
NOTES = { http://www.kevinsim.co.uk/evostar2013/cfpEvoApplications.html EvoApplications2013 held in conjunction with EuroGP2013, EvoCOP2013, EvoBio'2013 and EvoMusArt2013 },
ORGANISATION = { EvoStar },
PAGES = { 376--386 },
PUBLISHER = { Springer },
PUBLISHER_ADDRESS = { Berlin },
SERIES = { Lecture Notes in Computer Science },
SIZE = { 11 pages },
TITLE = { Implicit Fitness Sharing for Evolutionary Synthesis of License Plate Detectors },
VOLUME = { 7835 },
YEAR = { 2013 },
1 = { https://doi.org/10.1007/978-3-642-37192-9_38 },
}