TagT: A versatile ANPR solution for diverse applications

Tan, Jo Fang (2025) TagT: A versatile ANPR solution for diverse applications. Final Year Project, UTAR.

PDF Download (13Mb)

Abstract

Traditional Automatic Number Plate Recognition (ANPR) systems, which focus solely on license plate numbers detection and recognition are vulnerable to fraud. This project presents the design and implementation of TagT, an advanced ANPR framework that enhances security through multi-attribute car recognition. TagT integrates three key components: a YOLO11n model for high-speed car detection, a ResNet18 model with cosine similarity for intelligent frame optimization and the Gemini model for robust recognition of a car's license plate number, brand and colour. An extensive preliminary investigation justifies the selection of these models over numerous alternatives. The final, implemented system features a native iOS application and a Python back-end. A comprehensive evaluation was conducted to validate the prototype's performance, focusing on two key areas: Efficiency and Accuracy. The evaluation of the architecture's efficiency demonstrated a 92.4% reduction in frames sent for analysis, which resulted in a 91.9% decrease in API costs and an 87.6% decrease in API latency compared to a baseline approach. Furthermore, the system's real-world accuracy was validated across 160 demanding tests in varied conditions, achieving an average Overall Success Rate of 83.75% and a near-perfect Car Brand Accuracy of 98.13%. Overall, TagT provides a versatile, costeffective and scalable solution that successfully addresses the limitations of traditional ANPR, enhancing public safety and car management.

Actions (login required)