Automatic Cambodian License Plate Recognition Using Deep Learning With Robust Skew Handling

Read full text
Authors: Oudom Bol, Bunheang Kem, Sidavid Sin
Cited by: 0

Abstract

This study addresses the challenges of Automatic License Plate Recognition (ALPR) in Cambodia, with a specific focus on accurately detecting and recognizing license plates at varying skew angles. While recent advances in deep learning have improved ALPR systems, the unique characteristics of Cambodian license plates including dual script formats and distinct layout types combined with skewed orientations present significant technical challenges that current solutions do not adequately address. The research employs YOLOv8 instance segmentation for license plate detection, segmentation and type classification, followed by perspective correction using Canny Edge Detection, Douglas-Peucker algorithm, and Homography Transformation to rectify plate skew. Text zone detection utilizes separate YOLOv8 models for Type 1 and Type 2 plates to localize Latin script and plate number, with EasyOCR performing final character recognition enhanced through preprocessing and post-processing optimization. The study utilizes a purpose-built dataset of 1,200 real-world Cambodian license plate images systematically distributed across four skew angle categories ranging from 0° to 60° and two plate types. Performance evaluation demonstrates exceptional results with YOLOv8m-seg achieving 99.5% mAP@70 for detection, a 26.9% improvement over previous benchmarks, while maintaining 100% recall across all skew categories. The perspective correction and EasyOCR integration delivered substantial recognition improvements, with character recall exceeding the previous study by margins ranging from 1.2% to 32.6%, particularly achieving 30% exact match rate improvements for severely skewed plates (45-60°). The findings establish the first systematic skew-handling methodology for Cambodian license plates and demonstrate robust performance across varying skew conditions, providing a methodological foundation for advancing Cambodian ALPR research beyond previous limitations in handling skewed plates.

Fields of Study & Subjects

Vehicle License Plate Recognition Advanced Neural Network Applications Handwritten Text Recognition Techniques

Publication details

Journal / source
AIJR Proceedings
Volume
8 (1)
Pages
22–30
Publication date
20 January 2026
Language
EN