Original research

ENHANCING LICENSE PLATE RECOGNITION USING YOLO-NAS, YOLOV8, AND SORT ALGORITHMS

Pages 19-26

10.61552/JAI.2025.01.003

Charvi Khanna, Bhavini Bisht, Dhruv Kamshetty, Harshit Bhardwaj, Monica Bhutani

• View Article
• PDF 985 K

Abstract Targeting the challenge posed by the traditional license plate recognition approach's deficiencies in precision and speed, a novel end-to-end deep learning model has been introduced. This model employs YOLO-NAS for the accurate detection and recognition of license plates in real-world scenarios. Employing the YOLO-NAS model, our approach to license plate identification involves comprehensive training on diverse datasets, spanning small, medium, and large scales to achieve optimal accuracy. YOLO-NAS introduces an innovative quantization-friendly basic block, mitigating a key limitation in earlier YOLO models. Performance is further heightened through the incorporation of advanced training methodologies and post-training quantization techniques. In conjunction, YOLOv8 serves to categorize vehicles into specific types, such as cars or bikes. The SORT algorithm assigns distinct identity numbers to vehicles, facilitating seamless linkage with their corresponding detected license plates. This associational data is systematically stored in a CSV file for reference. For visualization, EasyOCR is deployed to recognize alphanumeric characters on license plates. This recognition output is visually represented as a box above the identified vehicles. Leveraging YOLO-NAS for license plate detection not only ensures superior accuracy but also optimizes performance through quantization support and strategic accuracy-latency trade-offs, contributing to a more refined and efficient recognition system. The accuracy that we obtained for our YOLO-NAS (small) model was 90.2%. Using YOLO-NAS for license plate detection we are able to develop a model which combines high speed with accuracy.

Keywords: License Plate Recognition (LPR) Technology, Computer Vision, YOLO (You Only Look Once) Algorithm, Deep Learning Neural Network Models, Convolutional Neural Networks (CNN), YOLO-NAS (You Only Look Once - Neural Architecture Search), Quantization Support, Deformable Convolution, Multi-scale Detection Layers, YOLOv5, Recurrent Neural Networks (RNNs), Model Optimization, Character Segmentation.

Recieved: 09.06.2024. Revised: 15.07.2024. Accepted: 25.08.2024.