Comparative Transfer Learning Techniques for Plate Number Recognition

Monitoring vehicle activity both on the highway and in certain places such as parking lots needs to be done if there is a specific incident. Unexpected events such as accidents or vehicle theft may occur anytime. Therefore, tracking through number plate recognition has become something important and has become a hot topic with the various methods used. Previous research used machine learning techniques to recognize characters on number plates. The use of this technique has not produced optimal accuracy. Therefore, we propose using transfer learning techniques to achieve better accuracy results. This research evaluated three transfer learning models, namely DenseNet121, MobileNetV2, and NASNetMobile models. The experiment in this research was carried out using the data on number plates in the parking lot. The accuracy calculation counted the number of correctly recognized characters divided by the total characters on the number plate. The experimental results show that the DenseNet121 model produced the best accuracy, 96.42%. Differences in number plate writing style also affected the accuracy results. This research could provide insight into the use of transfer learning techniques in the case of number plate recognition.