A super-resolution framework for license plate recognition based on deep learning

Accurately recognizing low-resolution license plates from real-world scenes remains a challenging task. In recent years, deep learning-based image super-resolution techniques have been increasingly applied to license plate datasets. However, most of these datasets contain low-resolution plates that are artificially synthesized with added noise. Furthermore, due to limitations in image quality assessment metrics, the quality scores of the super-resolved images generated by the model do not show a direct positive correlation with the improvement in their text recognition accuracy. This study generated a novel paired low-resolution and high-resolution license plate image dataset extracted from a dashcam video-based dataset and proposed a super-resolution framework specifically for license plate recognition. We analyze existing loss functions in the framework and introduce two novel perceptual loss functions based on different feature extractor backbones. By combining various types of loss functions, our approach effectively enhances license plate recognition accuracy while demonstrating visually pleasing results.