Automatic Malaysia license plate recognition using deep learning

• Auteur principal: Tay, Choon Kiat
• Format: Thèse
• Langue: anglais; anglais
• Publié: 2021
• Sujets: T Technology (General); TK Electrical engineering. Electronics Nuclear engineering
• Accès en ligne: http://eprints.utem.edu.my/id/eprint/25990/1/Automatic%20Malaysia%20license%20plate%20recognition%20using%20deep%20learning.pdf; http://eprints.utem.edu.my/id/eprint/25990/2/Automatic%20Malaysia%20license%20plate%20recognition%20using%20deep%20learning.pdf

Since 2008, Malaysia government has initiated the multi-lane free flow (MLFF) highway plans to improve traffic quality. To support the initiative, a robust license plate recognition system is required to track the highway vehicle passing the toll as MLFF do not have barrier gate to control the vehicle if the driver payment card has insufficient credit. Conventional license plate recognition system uses template matching, a simple image processing technique with manually defined image template to identify target character in the captured license plate images. Template matching based license plate recognition system proven to be useful in many countries where the license plate is issued by relevant authority but not in Malaysia. Correctly recognizing all the characters with various readable font types and spacing on the captured Malaysia license plate image will required deep learning type of technique to reduce the handcrafting of the matching templates/features. The first problem faced in this study was the collected 80,000 Malaysia license plate images suffering from character imbalance (skewed class). To form a good dataset for both training and testing, 297,840 synthetic images were generated, together with 73,000 original images to form the training and validation dataset of 370,840 images (remaining 7,000 images were used for the testing). A deep learning-based end to end segmentation-free character recognition model, Convolutional Recurrent Neural Network (CRNN), was used to train on the 370,840 training images and achieved only 55% of preliminary recognition accuracy. With the proposed image pre-processing of input license plate image, hyper parameters tuning (regularization, LSTM time step optimization) and better decoder, the recognition accuracy of the CRNN model increased to 95%.