TRANSFORMER-BASED OPTICAL CHARACTER RECOGNITION APPROACH FOR IDENTIFYING MOTOR VEHICLES WITH OVERDUE TAXES

Abstract

The high growth in the number of motorized vehicles in Indonesia has given rise to special attention in managing traffic administration, especially in relation to vehicle taxes. To present innovative solutions in vehicle tax administration, this research was conducted to detect the five-year tax status of motor vehicles in Indonesia using the Transformer Optical Character Recognition (TrOCR) model. The aim of this research is to evaluate the performance of the TrOCR model in recognizing text on motor vehicle number plates in Indonesia and classifying number plates that have and have not paid tax. The data used is primary data in the form of images of motor vehicle number plates taken around the Faculty of Mathematics and Natural Sciences, Universitas Islam Indonesia, using a purposive sampling with constraints on the representation of each class. Although the data collection was limited to this location, Indonesian vehicle plates follow a standardized format, with regional differences primarily in the prefix letters. Additionally, the university attracts students from various regions who often use vehicles registered in their home provinces. Consequently, the collected dataset reflects a diverse range of number plates, making it a reasonable representation of motor vehicle plates across Indonesia. The research results show that the TrOCR model succeeded in achieving a Character Error Rate (CER) value of 2.9% with a data configuration of 90% for training and 10% for testing, and using 8 epochs. Evaluation of model performance indicates that overall text detection is very effective in classifying the five-year tax status of motor vehicles. Although there are some prediction errors, the overall performance of the model can be considered good and is able to provide reliable information regarding the five-yearly vehicle tax status