AUTOMATIC NUMBER PLATE RECOGNITION USING YOLO

• Main Author: Fawwaz Zuhdi, Ahmad
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66343
• Institution: Institut Teknologi Bandung

Automatic number plate recognition (ANPR) is a technology that uses optical character recognition to extract information and read vehicle license plates from images or sequences of images. ANPR technology must take into account variations in vehicle number plates from one place to another according to the standards of each country, which generally have different formats, sizes, and number plate characters. Motorized Vehicle Number (TNKB) is an aluminium plate that marks the number of motorized vehicles in Indonesia which has been registered at the Joint Office for One-Stop Administration System (SAMSAT). The automatic number plate recognition system basically has 3 main components, namely number plate detection, character segmentation, and character recognition. Automatic number plate recognition systems can be used by law enforcement to identify unregistered or stolen vehicles, assist traffic law enforcement, and identify wanted criminal vehicles. The ANPR system can be integrated into the parking system to increase transaction speed without the need for a parking operator. The use of a number plate recognition system in a traffic monitoring system can serve to assist in making decisions regarding traffic conditions such as congestion. The development of the number plate recognition system in this final project aims to collect datasets of Indonesian vehicle number plates and implement the YOLO algorithm on datasets of Indonesian vehicle number plates. vii You Only Look Once (YOLO) is a deep neural network architecture for object detection. The main advantage of YOLO compared to other object detection methods is the speed of detection and classification. YOLO only performs image processing once, and immediately produces output in the form of a bounding box and an image prediction class in the bounding box. YOLOv4 on a darknet architecture has the following sections, the input section is a training image, the backbone section for feature or pattern extraction using CSPDarknet53, the neck section for aggregation using SPP Block and PANet, and the head section for detection, prediction, localization, or classification using YOLOv3. The performance of the YOLO algorithm is measured in recognizing Indonesian vehicle number plates through an evaluation process during the training and testing stages through several scenarios. The results of the evaluation of the number plate detection model have an average precision and recall of 80.31% and 97.82% and the results of the evaluation of the character recognition model have an average precision and recall of 98.45% and 99.22%, respectively. The test results in the scenario of rotating the image which is 15% and the scenario of blurring the image which is 40% raises the suspicion that the dataset used in the training phase has not been able to meet the needs of the test. In addition, based on the test results, it takes time to generate darknet architecture for YOLO on initial calls with durations ranging from 319 – 381 milliseconds for license plate detection and 324 – 373 milliseconds for character recognition. However, once the initial call has been made, reads for license plate detection can be made with a duration of 2.2 milliseconds per image. For further research, the addition of the amount and variation of data can be done to improve system performance. The data needed to meet the needs of testing as a representation of the real world include data with variations in lighting, license plate conditions, number plate size in the image, and angle of capture.