Number Plate Detection Project
Project Overview
Developed a Number Plate Detection System using deep learning, capable of detecting vehicle license plates from images and videos. The system is built using a custom-trained YOLOv8 model, optimized for fast and accurate object detection in real-world scenarios.
The model is trained on an Indian vehicle dataset sourced from Kaggle , which includes diverse images collected from multiple sources such as OLX listings, Google Images, and video frames. This diversity enables the model to generalize effectively across different lighting conditions, viewing angles, and backgrounds.
The project focuses strictly on object detection, where the model identifies number plates and outputs bounding boxes along with confidence scores.
The system is deployed using a Gradio-based web interface, allowing users to upload images or videos and visualize detection results interactively. The interface also provides structured output in tabular format, displaying frame-wise detection confidence for better analysis.
This project demonstrates a complete end-to-end machine learning pipeline, including data preprocessing, model training, inference, and deployment, with a strong emphasis on usability and performance.
Objectives
The primary objective of this project is to build an efficient and accurate deep learning-based system for detecting vehicle number plates in real-world scenarios. The focus is on developing a robust object detection model capable of handling variations in lighting conditions, viewing angles, and complex backgrounds.
Another key objective is to achieve near real-time performance so that the system can process both images and videos with minimal latency. This is accomplished using an optimized YOLOv8 model that provides a strong balance between detection speed and accuracy.
The project also aims to design a user-friendly deployment interface using Gradio, enabling users to easily upload images or videos and visualize detection results without requiring technical expertise.
Additionally, the system is built to handle multiple detections within a single frame and provide clear visual outputs in the form of bounding boxes along with confidence scores. This makes the solution suitable for practical applications such as traffic monitoring, parking management, and surveillance systems.
Dataset
The model is trained on an Indian vehicle dataset sourced from Kaggle : https://www.kaggle.com/datasets/saisirishan/indian-vehicle-dataset
The dataset contains a total of 3396 images collected from multiple sources, including state-wise OLX listings, Google images, and video frames. This diversity helps the model generalize across different real-world conditions such as varying lighting, backgrounds, vehicle types, and camera angles.
Each image is annotated with bounding boxes around number plates in XML format (Pascal VOC). During preprocessing, only images with valid annotations were used to ensure data quality and consistency.
The annotations were converted from XML to YOLO format (normalized coordinates) to make them compatible with the YOLOv8 training pipeline. The dataset was then split into training and validation sets in an 80:20 ratio.
To further improve model performance, data augmentation techniques such as mosaic, mixup, scaling, translation, and slight rotation were applied during training, allowing the model to perform better on unseen and challenging data.
Technologies Used
The project is developed using Python, chosen for its strong ecosystem in machine learning and computer vision applications. It enables seamless integration of multiple libraries required for building and deploying deep learning systems.
The core object detection model is implemented using the Ultralytics YOLOv8 framework, specifically the YOLOv8m variant, which offers an effective balance between speed and accuracy for real-time detection tasks.
The deep learning backbone is powered by PyTorch, which provides flexibility in model training, supports GPU acceleration, and allows efficient experimentation and optimization.
For image and video processing, OpenCV is used. It handles operations such as frame extraction, drawing bounding boxes, and processing visual inputs for inference.
The deployment interface is built using Gradio, enabling users to interact with the system through a simple web-based application where they can upload images and videos and view detection results.
For data handling and analysis, libraries such as NumPy and Pandas are used, while Matplotlib is utilized for visualizing training performance and evaluation metrics.
Methodology
The development of the Number Plate Detection System follows a structured pipeline consisting of data preparation, model training, evaluation, and deployment.
The process begins with dataset collection and preprocessing. Images are gathered from multiple sources, and only those with valid annotations are selected. The annotations, originally in Pascal VOC (XML) format, are converted into YOLO format with normalized bounding box coordinates to ensure compatibility with the training framework.
Next, the dataset is divided into training and validation sets in an 80:20 ratio. This approach allows the model to learn from one portion of the data while being evaluated on unseen samples, helping to measure its generalization performance.
The model is then trained using the Ultralytics YOLOv8 framework, specifically the YOLOv8m variant. During training, key hyperparameters such as batch size, learning rate, optimizer (AdamW), and number of epochs are configured. Data augmentation techniques including mosaic, mixup, scaling, translation, and slight rotation are applied to improve robustness and reduce overfitting.
After training, the model is evaluated on the validation dataset. Performance metrics such as loss and mean Average Precision (mAP) are analyzed to assess detection accuracy and overall model stability.
Finally, the trained model is deployed using OpenCV for image and video processing, and a Gradio-based interface is developed to allow users to upload images or videos and visualize detection results. This demonstrates the practical usability of the system in real-world scenarios.
Model Details
The core of the system is a custom-trained object detection model based on YOLOv8, specifically the YOLOv8m (medium) variant. This model is chosen for its strong balance between detection accuracy and inference speed, making it well-suited for practical, near real-time applications.
The model is trained for a single class, namely vehicle number plates. During inference, it predicts bounding boxes around detected plates along with corresponding confidence scores that indicate the reliability of each detection.
The training process is carried out using the Ultralytics YOLOv8 implementation, which is built on top of PyTorch. This enables efficient computation, GPU acceleration, and flexibility in optimization during training.
Input images are resized to a fixed dimension (e.g., 800×800) during training to maintain consistency. Various data augmentation techniques are applied to improve the model’s robustness. Through training, the model learns important spatial features of number plates such as shape, edges, and patterns, allowing it to perform reliably under different environmental conditions.
The final trained model is saved as a “best.pt” file, which is used for inference in the Gradio-based application for image and video detection. The model outputs include bounding box coordinates and confidence scores, providing clear and interpretable detection results.
Features
The system provides robust and efficient number plate detection capabilities across multiple input formats, making it suitable for real-world applications.
It supports detection of number plates in images, allowing users to upload images and receive annotated outputs with bounding boxes and confidence scores for each detected plate.
The system also processes video inputs, where each frame is analyzed individually to detect number plates. An annotated video is generated as output, and the system is capable of handling multiple detections within a single frame, ensuring that all visible plates are identified accurately.
The model delivers near real-time performance, enabling efficient processing of both images and videos with minimal latency. This makes the system practical for applications where quick detection is required.
Clear visual outputs are provided by drawing bounding boxes around detected number plates along with confidence scores, making the results easy to interpret and analyze.
Additionally, a web-based interface built using Gradio allows users to interact with the system without requiring technical expertise. The interface supports image uploads, video input, and structured display of detection results.
The system also generates tabular outputs that include frame-wise detection information such as confidence scores, which can be useful for further analysis or reporting.
Deployment
The system is deployed as a web-based application using Gradio, providing an interactive and user-friendly interface. This allows users to upload images or videos and perform number plate detection directly through a browser without requiring any complex setup.
The application processes the input using the trained YOLOv8 model and generates annotated outputs with bounding boxes and confidence scores for each detected number plate. Along with visual results, the system also provides structured detection data in tabular form, enabling better interpretation and analysis.
The deployment ensures smooth handling of both image and video inputs, making the system practical and accessible for real-world use cases. The implementation of this interface is available in app.py .
Results
The trained YOLOv8 model achieved strong performance on the validation dataset, demonstrating high accuracy along with efficient near real-time detection capability.
The evaluation metrics indicate excellent detection performance. The model achieved a precision of approximately 0.997 and a recall of 0.994, showing that it can correctly identify number plates while minimizing false positives and missed detections. The mAP@50 score of 0.995 reflects very high detection accuracy at a standard IoU threshold, while the mAP@50–95 score of 0.845 demonstrates strong performance even under stricter evaluation conditions.
In terms of efficiency, the model performs inference in approximately 12.8 ms per image, making it suitable for real-time applications. Preprocessing and postprocessing times remain minimal, ensuring low overall system latency.
During training, both training and validation losses showed smooth and consistent convergence, indicating stable learning and good generalization capability. The model performs reliably on both images and videos under normal conditions and is capable of detecting multiple number plates within a single frame.
However, performance shows slight degradation in challenging scenarios such as low-light conditions, motion blur, occlusion, and very small or distant number plates. Despite these limitations, the overall results demonstrate that the system is accurate, fast, and reliable for practical number plate detection applications.
Challenges Faced
One of the major challenges in this project was the variation in number plate appearances. Number plates differ significantly in size, font style, color, and layout across different vehicles and regions, making it difficult for the model to generalize effectively across all scenarios.
Another challenge was handling low-quality data, including blurred images, low-resolution frames, and partially visible plates. These conditions made accurate detection more difficult, particularly in video inputs where motion and compression artifacts are common.
Lighting conditions also posed a significant difficulty. Images captured under low light, shadows, or overexposed environments negatively impacted detection performance and reduced confidence scores in certain cases.
Dataset imbalance was another issue, as well-lit and clearly visible number plates were more prevalent than challenging samples. This imbalance can cause the model to perform better in ideal conditions while struggling with edge cases.
Detecting very small or distant number plates required the model to learn fine-grained features, which is inherently challenging in object detection tasks and can affect detection consistency.
Finally, achieving near real-time performance while maintaining high accuracy required careful model selection and optimization. The use of YOLOv8 (YOLOv8m variant) involved tuning hyperparameters to balance speed and precision effectively.
Future Improvements
The current system focuses on accurate number plate detection; however, several enhancements can be made to extend its functionality and performance.
One major improvement is the integration of OCR into the pipeline to enable complete number plate recognition (detection along with text extraction) within a unified system. This would significantly increase the practical applicability of the solution in areas such as automated toll collection and traffic enforcement.
Model performance can be further improved in challenging conditions such as low-light environments, motion blur, and occlusions by incorporating more diverse training data and applying advanced data augmentation techniques. This would enhance the robustness and generalization capability of the system.
Another important enhancement is optimizing the model for edge devices such as embedded systems or mobile platforms. This would allow deployment in real-time monitoring environments without requiring high-end computational resources.
Inference speed can also be improved using model optimization techniques such as quantization and pruning, making the system faster and more resource-efficient while maintaining acceptable accuracy.
Expanding and balancing the dataset with a wider variety of real-world scenarios can further improve detection performance and reduce bias toward ideal conditions.
Finally, integrating additional features such as object tracking, automated alert systems, and database storage can evolve the project into a complete intelligent traffic monitoring and management solution.
Applications
The Number Plate Detection System has a wide range of real-world applications, particularly in areas related to traffic management, security, and automation.
It can be used in traffic monitoring systems to automatically detect and analyze vehicle movement on roads, helping authorities manage congestion and enforce traffic regulations more effectively. The system can assist in identifying patterns such as high-traffic zones and peak hours.
The system is also useful in automated parking management, where it can detect vehicles entering and exiting parking areas without manual intervention. This improves efficiency and reduces dependency on human operators.
In toll collection systems, number plate detection can be integrated to enable automatic vehicle identification, reducing waiting time and improving throughput. This makes the process faster and more seamless.
For law enforcement, the system can help in identifying vehicles involved in traffic violations or suspicious activities. It can support authorities in monitoring and tracking vehicles when required.
Additionally, it can be deployed in surveillance systems for monitoring restricted areas, residential societies, and commercial spaces, enhancing overall security.
Overall, the project demonstrates strong potential in contributing to intelligent transportation systems and smart city solutions by automating vehicle detection and identification processes.
Project Links
Live Demo : https://huggingface.co/spaces/Aryan2301/Number_Plate_Detection_Project
Source Code (GitHub) : https://github.com/Aryanupadhyay23/Number_Plate_Detection_Project