License Plate Recognition via Convolutional Neural Networks

The problem of license plate recognition has traditionally been tackled by methods such as optical character recognition and multi-layer perceptron + SVM, and these methods have been widely deployed in practice. Despite their success, they can only be used in idea situations. Recently, several attempts have been made to exploit the representation power of deep convolutional neural networks (CNN). In this paper, we also try to exploit the power of CNN in license plate recognition. In the proposed system, the SSD detector [1] is firstly used to detect the license plate. Then after obtaining the rough location of the license plate, we use the vertical projection method for character segmentation. Finally, characters are classified with a small but powerful CNN network. The proposed license plate recognition system achieves a very high accuracy with a high speed in a test dataset collected by ourselves.