Cascaded YOLO-based approach for Bangladeshi license plate identification and number recognition.

Automatic vehicle license plate detection and recognition has become an important component in intelligent transportation systems, traffic monitoring, and law enforcement applications. However, existing approaches for Bangladeshi license plates often rely on optical character recognition (OCR) or recognition models with a limited number of character classes, which restricts their ability to handle the complex structure and large variation of Bangla license plate components. To address these challenges, this study proposes a vehicle license plate detection and recognition (VLPDR) framework based on a cascading architecture of three YOLOv12 models. The first model detects vehicles, the second model localizes the license plate region, and the third model performs character-level recognition without relying on conventional OCR engines. The proposed recognition model incorporates 104 classes, covering a wide range of Bangladeshi license plate components, including city codes, vehicle types, and numeric digits. Two datasets consisting of 9,357 images have been used for training, validation, and testing, while an additional 150 images have been used for external validation. The license plate detection (LPD) model has achieved testing precision, recall, and mAP@50 of 0.986, 0.965, and 0.963, respectively. The license plate number recognition (LPNR) model has obtained testing precision, recall, and mAP@50 values of 0.95, 0.898, and 0.949. During external validation, the cascaded framework achieved 100% license plate detection accuracy, 95.33% full-number recognition accuracy, 98.82% character-wise mean accuracy, and 99.57% Levenshtein-based similarity accuracy. The system has also been evaluated on multiple hardware platforms, demonstrating inference times of 2311.53 ms on a CPU-based laptop and 284.31 ms on a GPU-enabled computer. These results indicate that the proposed cascaded YOLO-based framework provides an effective and practical solution for Bangladeshi license plate detection and recognition with improved class coverage and strong generalization capability.