@article{M39F7ED49,
title = "Lightweight Adversarial Domain Adaptive Model for License Plate Detection",
journal = "The Journal of Korean Institute of Communications and Information Sciences",
year = "2024",
issn = "1226-4717",
doi = "10.7840/kics.2024.49.1.41",
author = "Hangjae Cho, Jeonghyeon Kim, Kyungkoo Jun",
keywords = "Deep learning, license plate detection, adversarial domain adaptation",
abstract = "The technology of car license plate location recognition using deep learning is a crucial prerequisite in solving the car license plate recognition problem. However, due to the inherent characteristics of deep learning models, there exists a challenge of performance degradation when making inferences in environments different from the ones they were trained on. To address this issue, performance can be enhanced by leveraging adversarial domain adaptation techniques. However, the experiments conducted using the DA (Domain Adaptation) RetinaNet model, which applied adversarial domain adaptation techniques to RetinaNet, revealed no performance improvement. This lack of improvement was attributed to the fact that the size of the car license plate data used was smaller compared to the model size. To overcome these challenges, this study proposes a method to enhance the performance of the car license plate location recognition problem by customizing DA RetinaNet to match the dataset size and utilizing pre-training. Applying the proposed method resulted in up to a 45% improvement based on the F1 Score and up to a 91% improvement based on mAP@0.5 compared to the performance achieved when only source domain data was used for training, and inference was conducted in other domains."
}