A GFP-GAN based method for improving clarity of license plate images

Abstract License plate recognition is a very important component of intelligent transportation system, and its accuracy and speed affect the operation of the entire system. However, it is still difficult to achieve high accuracy in license plate positioning and character recognition for low-quality license plate images. To migrate the GFP-GAN model from blind face reconstruction to license plate image restoration, we made contributions to improve the clarity of license plate images in three aspects: (1). add motion blur present commonly on license plate images rather than facial images; (2). add the license plate recognition loss during training to increase the network's ability to recover correct characters; (3). align the region discriminator to the license plate position to optimize the region recovery effect. As in GFP-GAN, the multi-scale semantic information of low-quality license plate image is extracted by U-Net (extractor), and converted to coding that controls the image content (encoder), then the coding is input to a pre-trained StyleGAN2 network to generate the corresponding high quality license plate image (generator/decoder). The migrated model, datasets requirement and preparation, training process and algorithms are given in detail. As compared with existing methods, the experiments show that our method can significantly improve the accuracy on license plate recognition, with advantages of strong adaptability to various type of blurs, fine and smooth restoration quality, prominent emphasis on license plate area.