Categorizing license plates using convolutional neural network with residual learning

• Main Authors: Jose, John Anthony C., Maningo, Jose Martin Z., Rogelio, Jayson P., Bandala, Argel A., Vicerra, Ryan Rhay P., Sybingco, Edwin, Ching, Phoebe Mae L., Dadios, Elmer Jose P.
• Format: text
• Published: Animo Repository 2019
• Subjects: Automobile license plates > Philippines—Classification; Transfer learning (Machine learning); Neural networks (Computer science); Stochastic systems; Stochastic models
• Online Access: https://animorepository.dlsu.edu.ph/faculty_research/3019
• Institution: De La Salle University

© 2019 IEEE. ike other countries, the Philippines uses various license plate standards wherein some purely text while some are hybrid graphic-text plates. And to harness its generalizability, this study developed a classification algorithm utilized as a pre-processing scheme for the multi-standard license plate. With an input image captured at a different perspective, it was feed into the neural network and classify as Rizal monument series (2001 base and 2003 base), 2014 series and conduction sticker for new vehicles. In total, there are 303 different images captured for this study. Around 100 conduction sticker images, 103 Rizal Monument images, 100 black and white images. Furthermore, this study focused on using transfer learning technique, wherein a trained network utilized, then only the last layer was reset and retrained on the new dataset. To measure the performance of the classification model and optimized it cross-entropy and stochastic gradient descent was employed respectively at a learning rate of 0.001 and reduced by 10 for every seven (7) epochs. The progression of accuracy results in increasing the epochs, and for the 25 epochs, the training completed in 4 minutes and 7 seconds with the best validation accuracy of 82.61%.