Implicit fitness sharing for evolutionary synthesis of license plate detectors

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. © Springer-Verlag Berlin Heidelberg 2013.