Vietnam Vehicle Number Recognition Based on an Improved CRNN with Attention Mechanism

Applications for automatic license plate recognition (ALPR) are evaluated as an important technology for effective traffic control. In particular, the most used application today for license plate recognition is at highways, toll stations, agencies, parking, and schools, with the support of camera equipment that provides high accuracy. In this study, Convolutional Recurrent Neural Networks (CRNNs) are utilized for the license plate recognition task. We are developing an automatic license plate recognition system designed for Vietnamese license plates, intended for use in indoor parking facilities with fixed cameras. In this context, the problem presents three main steps: (1) using the YOLO model to detect vehicles in the given images, (2) using the WPOD-NET model to extract license plates, and (3) introducing a new method based on an improved Convolution Recurrent Neural Network (CRNN) with combination between connectionist temporal classification (CTC) and attention mechanism to recognize characters on license plates. Our CRNN model is jointly trained with both CTC and attention objective functions. Experimental results on a license plate database collected from an indoor parking area achieved a Word Error Rate (WER) of 0.014 in the optical character recognition (OCR) task. The experimental results demonstrate that the proposed model performs well for vehicle number plate recognition and can be applied to real-time applications.