End to End Afghan License Plate Detection and Recognition Using Deep Learning Models
Description
This dataset contains approximately 8,500 annotated images of vehicles and license plates from Afghanistan and was prepared for the development and evaluation of an automatic license plate detection system. The dataset focuses on the detection stage of an Automatic License Plate Recognition (ALPR) pipeline, where the objective is to identify and localize the license plate region within a full vehicle or scene image. Each image contains a vehicle or scene in which a license plate is visible. The license plate regions are annotated using bounding boxes that define the location of the plate within the image. These annotations enable object detection models to learn how to locate Afghan license plates within complete images before the recognition stage is performed. The dataset includes variations in vehicle type, license plate appearance, viewing angle, distance, background, illumination, and image quality. These variations provide realistic conditions for developing and evaluating license plate detection models. The dataset is organized into training, validation, and testing subsets to support model training and performance evaluation. The annotations are provided in COCO JSON format, which stores image information, object categories, and bounding-box coordinates in a structured format. This makes the dataset suitable for different object detection frameworks and allows the annotations to be converted into other formats when necessary. In our research, this dataset was used for the detection component of an end-to-end Afghan License Plate Recognition pipeline. A YOLO-based object detection model was trained to locate license plates within full images. Once a license plate is detected, its region can be cropped from the original image and passed to a separate Optical Character Recognition (OCR) model for character recognition. Therefore, this published dataset specifically represents the license plate detection component of the system and does not include the separate augmented dataset used for training the recognition model. The dataset can support research in automatic license plate detection, Automatic License Plate Recognition (ALPR), computer vision, intelligent transportation systems, traffic monitoring, and object detection. By making the dataset publicly available, it can support the reproducibility of the detection experiments and provide a resource for researchers interested in developing and evaluating computer vision approaches for Afghan license plates.
Files
Steps to reproduce
1. Download the dataset ZIP file and extract it to a local directory. 2. The dataset contains approximately 8,500 images with license plate regions annotated using bounding boxes. 3. The dataset is divided into training, validation, and testing subsets. Keep the provided folder structure and data splits unchanged when reproducing the experiments. 4. Bounding-box annotations are provided in COCO JSON format. The annotation files contain the information required to locate the license plate regions within the corresponding images. 5. The COCO annotations can be used directly with object detection frameworks that support COCO format or converted to another annotation format, such as YOLO, when required. 6. For the detection experiment, use the training subset to train the license plate detection model, the validation subset for model validation and parameter selection, and the testing subset for final evaluation. 7. In our study, the dataset was used to train and evaluate a YOLO-based license plate detector. The trained detector identifies the license plate region within a full image, which can subsequently be cropped and provided to a separate recognition model as part of an end-to-end Automatic License Plate Recognition (ALPR) system.
Institutions
- Asian University for WomenChittagong, Chittagong