Keypoint-based Distortion Correction and Data Augmentation for High-angle License Plate Recognition

Various techniques are being researched for recognizing distorted license plates. Existing deep learning-based techniques such as segmentation show poor performance in occlusion and quality degradation phenomena, and OCR algorithms trained with conventional augmentation techniques show low performance due to pixel degradation occurring in distortion correction results. To solve these problems, this study introduced keypoint-based distortion correction methods along with new augmentation techniques. The keypoint-based corner detection model reduces dependence on contour information and directly detects edges, enabling accurate corner detection even in occlusion and blur phenomena, and the proposed augmentation technique reproduces the distorted license plate correction process to effectively reproduce the degradation phenomenon that occurs during the correction process. Experimental results show that the OCR algorithm trained with the proposed augmentation technique achieved over 90% recognition rate from front to 30 degrees, and maintained stable performance of over 80% even in high-angle areas of 75-80 degrees where existing methods show weaknesses. Furthermore, it showed a recognition rate of over 50% even at extreme angles of 85 degrees, achieving significant performance improvement compared to existing methods. The results of this study suggest that stable recognition is possible even in license plate shooting situations at various angles that occur in real road environments.