Reading Between the Blurs: A Comparative Analysis of Motion Deblurring Methods for Vehicle License Plate OCR

Optical Character Recognition (OCR) systems have made significant advancements in accurately recognizing text in high-quality images. However, their performance can degrade considerably under challenging conditions, such as motion blur. Meanwhile, motion deblurring methods are often evaluated through well-known metrics such as SSIM and PSNR. Nevertheless, the contribution of motion deblurring in vehicle plate detection and reading has been a widely neglected aspect. The current study investigates and evaluates the impact of motion deblurring on OCR performance along with the use of the PaddleOCR model. To achieve this, our own curated dataset involving real-world vehicle images, captured under varying lighting conditions, camera types, and resolutions, are employed. These images introduce significant variability, allowing us to assess the effectiveness of motion deblurring techniques in enhancing OCR accuracy. In this study, we use the PaliGemma model to detect vehicle license plates. Next, to simulate real world distortions, synthetic motion blur is applied to image samples, mimicking the challenges OCR systems encounter in practical scenarios. Followingly, a series of motion deblurring models—ranging from traditional convolutional neural networks to advanced transformer-based approaches—are employed to restore the clarity of the blurred images. Afterwards, the identified license plate regions are extracted and subsequently analyzed by the PaddleOCR system for text recognition. We evaluate the accuracy of OCR through Levenshtein Distance (LD), which measures the minimum number of edits required to transform the recognized text into its ground truth counterparts. The experimental results demonstrate that the quality of motion deblurring directly influences the accuracy of OCR systems. This highlights the essential role of robust preprocessing, including effective motion deblurring, in improving the performance of vision pipelines for scene text recognition, especially in dynamic, real-world environments where motion blur is common.