Evolutionary optimization-based training of convolutional neural networks for OCR applications

This paper presents aspects concerning the implementation of two training algorithms for convolutional neural networks (CNNs) used in optical character recognition (OCR) applications. The two training algorithms involve evolutionary optimization algorithms represented by a Gravitational Search Algorithm (GSA) and a Particle Swarm Optimization (PSO) Algorithm. New CNN training algorithms are offered on the basis of using GSA and PSO algorithms in combination with back-propagation in order to encourage performance improvements by avoiding local minima. A comparison between the new training algorithms is carried out focusing on the analysis of convergence, computational cost and accuracy in the framework of a benchmark problem specific to OCR applications.