A Deep Neural Network for Automatic License Plate Recognition with Hyper parameters Study and Regularization

Abstract

 License plate recognition has been a hot research topic in academia and transportation industry. With automatic license plate recognition, real-time vehicle identification application can be implemented and helps authorities to enable multi-lane to allow high volume of traffic passing through toll collection plaza in metropolis cities. However, to avoid any error in charging the driver, a matching mechanism between registered license plate in the RFID and passing vehicle license plate required a robust optical character recognition (OCR) system. Recent deep learning research contributed significant breakthrough in computer vision and scene text recognition. In this study, we improve the deep learning basedsegmentation-less scene text recognition framework, CRNN,by tuning the hyper parameter that could fit well to Malaysia license plate format for automatic OCR. In CRNN model training, raw license plate images retrieved from deployed highway CCTV appeared to be severe character class imbalance and synthetic license plate is added to the real license plate images pool to create a 70k images final training data set. Malaysia license plates exist in different length and can cause spatial distortion, as the nature of convolution neural network only accept predefined input image size. Image pre-processing is included to pad the license plate images to improve the OCR training and recognition accuracy. With proposed image pre-processing and neural network regularization, our proposed license plate recognition model achieved 93% of accuracy compared to vanilla CRNN’s 74%.