A hybrid method for recognizing vehicle registration plates based on a pipeline architecture
TL;DR
Experimental validation of the proposed hybrid method for recognizing vehicle license plates confirmed its high efficiency and ability to ensure robust recognition under real-world conditions, demonstrating the method’s ability to ensure robust recognition under real-world conditions.
Abstract
The purpose of the research is improving the accuracy and speed of real-time recognition of vehicle registration plates through the use of a hybrid method adapted to Russian standards and resistant to environmental conditions. Methods . This paper proposes and tests a hybrid method for recognizing vehicle registration plates based on the YOLOv12 – EasyOCR pipeline architecture. A unique training dataset was synthesized to ensure maximum variability. A comparative analysis of detectors (YOLOv10, YOLOv12, RT-DETR) was conducted, confirming the feasibility of using the YOLOv12 model, which demonstrated the best performance on the most important metrics (Precision = = 0,99, Recall = 0,98, and mAP 0,5:0,95 = 0,91). A multivariate image preprocessing pipeline for OCR was developed, performing sequential transformations of the cropped image, including adaptive scaling, contrast enhancement, noise reduction, and morphological operations. To minimize typical OCR errors, a limited set of acceptable characters was used, and an algorithm for checking compliance with the Russian license plate format was implemented. Results. The OCR performance was assessed by comparing the recognized text with the reference text using quantitative metrics such as Character Recognition Rate (CRR) and Plate Recognition Rate (PRR), which is the industry equivalent of Word Recognition Rate (WRR). A series of experiments were conducted on a test dataset. Experimental validation of the proposed hybrid method for recognizing vehicle license plates confirmed its high efficiency (quantitative metric values of mAP 0,5:0,95 = 0,91, CRR = 99,0 %, PRR = 98,5 %), demonstrating the method’s ability to ensure robust recognition under real-world conditions. Conclusion . The obtained results correspond to the upper limit of the performance range of modern automatic recognition systems, including commercial solutions. This demonstrates the high competitiveness of the proposed method.