Two-bit weights are enough to solve vehicle license number recognition problem

The construction of a system that recognizes vehicle license numbers using feedforward neural networks, after the numbers have been extracted using classical methods, is described. The system is trained and tested on real-world data. In order to reduce the total amount of memory required and increase the process speed, an additional step is added to the learning algorithm that produces low precision weights (+1, 0, -1). The network obtained after this training process has a behavior similar to those networks using floating point representation for weights. A special hardware accelerator is developed to achieve high-speed recognition.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">></ETX>