Feed forward network for vehicle license character recognition

Abstract

This paper describes the construction of a system that recognizes vehicle license numbers using feed forward neural networks, once they have been extracted using classical methods. The system has been trained and tested on real-world data. In order to reduce the total amount of required memory and increase the process speed, an additional step has been added to the learning algorithm, that produces low precision weights {+1,0,-1}. The network obtained after this training process has a similar behaviour to those networks using a floating point representation for weights. A special hardware accelerator has been developed to achieve high speed recognition.