recognition, Vehicle detection.

License plate detection is critical in intelligent transportation systems and is used in traffic monitoring, toll collection, and access control applications. Automated License Plate Recognition (ALPR) systems typically involve two main components: License plate detection and character recognition. This project focuses on license plate detection, aiming to build a robust and efficient model capable of identifying license plates in diverse conditions.

The You Only Look Once (YOLO) model, known for its real-time processing capability, has been chosen as the foundation for this project. Specifically, YOLOv8 a state of the art object detection model offers improved accuracy and speed compared to its predecessors, making it well suited for dynamic, real world applications. This paper presents the design, training, and evaluation of a YOLOv8 based model optimized for detecting license plates with high accuracy and minimal latency.

License plate detection has undergone significant evolution over the years. Initial methods primarily relied on heuristic techniques such as edge detection and morphological operations, which, while effective under controlled conditions, often failed in complex scenarios involving variable lighting, occlusions, and cluttered backgrounds. The advent of deep learning, particularly Convolutional Neural Networks (CNNs) [1], has revolutionized the field, enabling robust object detection even in challenging environments. Among these approaches, the YOLO family of models has emerged as a key solution due to its real-time processing capabilities.

This section presents the framework for developing the license plate detection model using YOLOv8.

I. YOLOv8 architecture: YOLOv8 integrates enhancements such as a refined feature pyramid and attention mechanisms that improve small object detection, which is crucial for license plates [13]. The model is trained with a custom data set specific to license plates, enhancing its performance for this task.

II. Preprocessing and data augmentation: The data set is preprocessed with transformations to improve model generalization, including rotation, scaling, and brightness adjustments. Data augmentation helps the model learn to detect license plates under varied conditions.

III. Training: YOLOv8 is trained using labeled images of vehicles with annotated license plates. The training process involves: Setting a batch size and learning rate optimal for license plate detection.

IV. Regular monitoring of validation loss to prevent overfitting.

V. Hyperparameter tuning, focusing on confidence threshold and input image size to improve detection rates. VI. Fine-tuning: Parameters are adjusted based on initial performance to achieve optimal results. Techniques such as adjusting the learning rate, using a larger data set, and increasing training epochs are applied iteratively to fine-tune the model.

This section outlines the experimental setup used to evaluate the performance of multiple YOLO models specifically YOLOv4, YOLOv5, YOLOv7, and YOLOv8 on the task of license plate detection. The goal is to determine the most effective model regarding accuracy, speed, and robustness under various environmental conditions.

The performance of four YOLO models YOLOv4, YOLOv5, YOLOv7, and YOLOv8 was evaluated and compared on the task of license plate detection. The analysis was based on quantitative and qualitative metrics to assess the models' ability to detect license plates across different conditions. This section presents the results obtained from testing these models and a comparative analysis with other state-of-the-art detection models like Faster R-CNN and SSD.

This study demonstrates the effectiveness of YOLOv8 for real-time Automatic License Number Plate Detection (ALNPD). The YOLOv8 model achieved exceptional performance, providing both high accuracy and fast inference times, making it highly suitable for applications such as traffic management, law enforcement, and automated toll collection systems. YOLOv8's robust architecture, which integrates advanced features like transformer-based backbones and self-supervised learning, allowed it to outperform other YOLO models (YOLOv4, YOLOv5, YOLOv7) in terms of precision, recall, and F1-score while maintaining efficient processing speeds.

Although YOLOv8 showed impressive results, specific challenges remain, particularly in detecting partially occluded or highly angled license plates. Though less frequent, these scenarios can still cause detection errors, which may impact the overall robustness of the system in real-world settings. Future research can focus on improving the model's capability to handle these edge cases by incorporating more diverse datasets, advanced data augmentation techniques, and hybrid models that combine the strengths of YOLO with other detection frameworks.

In conclusion, YOLOv8 is a powerful tool for ALNPD, offering a balanced trade-off between speed and accuracy. It is well-suited for real-time deployment in automated systems. Further work on enhancing its ability to detect partially occluded plates and refining the model's adaptability in challenging conditions will ensure its broader applicability and robustness in intelligent transportation systems.

We are profoundly grateful to Prof. Shubadip Pramanik for his expert guidance and continuous encouragement throughout the process of seeing that this project reached its target from its commencement to completion.

Parthiv Patnaik led the critical data collection, annotation, and preprocessing tasks. He gathered a large and diverse dataset of license plate images from multiple sources and meticulously annotated them with bounding boxes to ensure accuracy in detecting plates. Parthiv also handled essential preprocessing tasks such as image re-sizing, augmentation, and normalization, optimizing the dataset to enhance model performance. He then prepared the dataset for training, ensuring it was in the correct format for the YOLOv8 model.

Soumyajit Maity was responsible for model development, training, and optimization. He implemented and configured the YOLOv8 model specifically for the license plate detection task and fine-tuned it by adjusting hyperparameters to improve performance. Additionally, he developed code to augment and preprocess data dynamically during training, such as real-time augmentation techniques, to enrich the model's learning N oughout the training process, making necessary adjustments to the architecture and techniques to achieve higher detection accuracy.

Saswata Dey took on the role of project manager, overseeing planning and research paper writing. He defined the project's scope, set milestones, and created an overall timeline, ensuring the project stayed on track. Saswata also coordinated various tasks, ensuring team members effectively met their objectives. In addition, he conducted a comprehensive literature review and contributed to the research paper by writing sections such as the introduction, background, and methodology, ensuring they met academic standards.

Ayush Kumar Dubey focused on evaluation, testing, deployment, and final reporting. He rigorously tested the trained YOLOv8 model on new, unseen data to assess its accuracy and robustness under different N s deployment for real-time detection, ensuring it could operate efficiently with low latency in a practical application. In the final stages, llenges, and conclusions. Additionally, Ayush documented the deployment process and any necessary adjustments to the system architecture, providing a comprehensive overview of the deployment strategy.

The authors declare that they have no conflicts of interest. This means the work was conducted independently, without any sponsorship, financial support, or influence from commercial entities, institutions, or organizations that might benefit from the research results. The authors confirm that they did not receive any grants or funding that could influence or bias the outcomes or interpretations of this study. None of the authors have personal or professional relationships with companies, industry partners, or other stakeholders with a vested interest in the findings. This helps to ensure that the research was conducted impartially and that the study's results and interpretations remain objective.