Improving License Plate Recognition Accuracy in Motion and Out-of-Focus Blur Environments

This report aims to enhance the recognition accuracy of vehicle license plate characters by restoring motion blur and out-of-focus blur caused by high-speed motion or lens focus errors using deep learning techniques. The primary factors contributing to recognition accuracy degradation in deep learning-based vehicle license plate recognition are image blurring caused by rapid vehicle movement or camera shake (e.g., due to wind or vibrations). This issue is particularly prominent in rolling shutter cameras, where motion blur becomes more severe depending on the speed of motion and exposure time. Therefore, preprocessing techniques capable of addressing motion blur from high-speed vehicles and focus blur caused by lens setting changes are essential. To solve this issue, a deblurring process was added to the conventional license plate recognition workflow, generating deblurred license plate images for recognition. A comparative analysis was then conducted to evaluate the difference in recognition accuracy between the conventional method and the deblurring-enhanced approach. A total of 219,756 license plate images from 8 classes were collected, and two types of artificial blurring, Gaussian blur and motion blur, were applied to simulate degraded images. The effectiveness of deblurring preprocessing on recognition performance was analyzed using four deep learning models. The evaluation results confirmed that applying preprocessing models capable of effectively removing motion and focus blur reduced false negatives and false positives, leading to a significant overall improvement in recognition accuracy.