Automated License Plate Recognition using Object Detection and Optical Character Recognition

In this thesis, we explore the domain of Automated License Plate Recognition (ALPR) systems, seamlessly integrating classic algorithmic programming with the power of Deep Learning. This synthesis results in sophisticated software proficient in accurately recog- nizing and interpreting license plates on automobiles, with implications extending across diverse domains such as modern surveillance, security systems, parking enforcement, traffic management, and law enforcement. Conducted as an internship project at Smart-Interaction, our comprehensive two-part approach centers on license plate detection and optical character recognition (OCR). In the initial phase, we meticulously fine-tune object detection models, including YOLOv8, DETR, and Faster R-CNN, leveraging diverse datasets. YOLOv8 emerges as the pre- eminent performer, showcasing exceptional post-fine-tuning performance. The subsequent phase focuses on OCR fine-tuning, employing advanced models—Easy- OCR, Paddle-OCR, and Parseq-OCR—where Parseq-OCR stands out for its unpar- alleled accuracy. Subsequent fine-tuning further enhances Parseq-OCR’s performance on both individual and combined datasets. This study underscores the efficiency of YOLOv8 in license plate detection and the robustness of Parseq-OCR in character recog- nition, making substantial contributions to the ongoing evolution of automated license plate recognition systems. To enhance result precision, we implement Levenshtein distance in post-processing, fur- ther refining the accuracy of our outcomes. Our research draws from a diverse array of datasets to comprehensively evaluate the proposed methodology.