Using Generative Adversarial Network Technology for Repairing Dynamically Blurred License Plates

Abstract

In recent years, due to the rapid development of artificial intelligence, many related technologies have been widely used in various fields, including the deep learning-based license plate recognition technology. However, there are still some problems with the deep learning-based license plate recognition technology, such as the inability to process license plate images with low light and dynamic blur. In addition, in real life, due to factors such as the speed of vehicle movement and camera exposure time, license plates often appear blurred, causing difficulties in license plate recognition. Therefore, this study proposes a method for restoring dynamic blur license plates based on Generative Adversarial Network (GAN) technology. Using a dataset of 16,900 original license plates and 25,000 iterations of training, a high-fidelity license plate model was trained and a dataset of 3,000 high-fidelity license plates was randomly generated, with dynamic blur effects added to the high-fidelity license plate dataset. Then, using the structure of the cGAN network in the pix2pix technology, the clear license plate was restored from the dynamic blur license plate. Our model was able to effectively restore 2,873 dynamic blur license plates out of 3,000 license plates with a blur level of 85 or more in the preliminary experiment on the dataset, with a restoration rate of 95.7%. The proposed method is more excellent and adaptable to most physical environments than traditional image processing methods. In the future, we will further improve and optimize the model, and introduce effects such as pollution, exposure, darkness, and obstruction to train a license plate restoration model with multiple functions to meet the increasingly wide-ranging needs of license plate recognition applications.