Scientific Journal of KubSAU

Car license plates recognition problem is one of the typical tasks of computer vision. Video surveillance software usually provides license plates recognition function. Meanwhile, there are many approaches to solve this problem, where template-based methods are the most common. Such methods providing predictable and short enough execution time, and little percent of mistakes. However, such methods are far less effective in case there is a need to recognize car’s license plate, which may be located in unpredictable place, typed in undefined font and on non-standard background, or without strict formatting. For example, USA car license plates. One of the methods to increase effectiveness and quality of such license plates recognition is to use neural networks. It is assumed, that neural networks usage can significantly increase recognition quality. Nevertheless, neural networks usage entails difficulties of it’s training, and often becomes less efficient as template-based methods usage. This article discusses probability of usage of convolutional neural network, which was trained using MNIST (Mixed National Institute of Standards and Technology) database. This article is a review of usage of templates and neural networks for car’s license plate recognition in terms of quality, performance and complexity of the usage