Unsupervised Domain Adaptive Corner Detection in Vehicle Plate Images

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• License Plate Recognition (LPR) accuracy is enhanced by image rectification, a geometric transformation requiring plate corner coordinates.
• Detecting these corners accurately is crucial for effective rectification, especially when dealing with diverse image sources.

Purpose of the Study:

• To develop an unsupervised domain adaptation method for robust vehicle license plate corner detection.
• To improve the performance of corner detection models across different countries (domains) without labeled target data.

Main Methods:

• A heatmap-based corner detection model was proposed, outperforming traditional scalar-regression methods.
• An image classifier was used to facilitate domain adaptation between source (Korea) and target (Philippines) plate image datasets.
• The study utilized a dataset of 22,096 Korean and 6,762 Philippine license plate images.

Main Results:

• The proposed heatmap-based approach demonstrated superior performance in unsupervised domain adaptation for corner detection.
• Achieved a 19.1% accuracy improvement compared to baseline discriminator-based domain adaptation methods.
• The model successfully adapted from detecting corners in Korean license plates to detecting them in Philippine license plates.

Conclusions:

• Unsupervised domain adaptation is effective for improving license plate corner detection across different countries.
• The proposed heatmap-based model offers a significant advancement over existing methods for LPR preprocessing.
• This research contributes to more accurate and adaptable LPR systems globally.