A Novel License Plate Image Reconstruction System using Generative Adversarial Network

This paper deals with the problem of license plate reconstruction, which is a method used for enhancing the quality of images of vehicle license plates in parking lot management systems. More specifically, poorly capture images of vehicle license plates which are unrecognizable by both human eyes and computer will be reconstructed so that they can be perceptible. This paper proposes a two-stage deep learning based algorithm for this problem. In the first stage, the position of the license plate in the image is detected using a YOLOv4 based transfer learning model. In the second stage, the image area of the license plate detected in the previous stage is fed to Pix2Pix, which is a type of Generative Adversarial Networks for the reconstruction. The experiment results show that by applying the proposed algorithm, license plate images with blur and flare can be transformed in to clear images which can be read by human eyes or can be used as inputs for computer vision applications such as license plate recognition.