| Title |
GAN-based Synthetic Generation of Seamless License Plate Images using Pseudo-labeling of Image Registration |
| Authors |
김형래(Hyoungrae Kim) ; 김학일(Hakil Kim) |
| DOI |
https://doi.org/10.5573/ieie.2023.60.4.55 |
| Keywords |
Generative adversarial network; De-identification; License plate dataset; Image registration |
| Abstract |
This paper proposes a vehicle license plate image synthetic generation method using the Generative Adversarial Network(GAN) and an image registration-based pseudo-labeling method for this. The AI-HUB public datasets consist of images with de-identification of personal information such as mosaicking applied to the license plate areas, however it not only contaminates the images, but also reduces usability as training data. This paper introduces a license plate synthetic image generation network that is free from personal information and indistinguishable from real images. For the training of the proposed GAN-based Masked-pix2pix network, a virtual license plate is created, and image pairs are constructed using a pseudo-labeling technique that performs feature matching and image registration with the original image. As a result of training, the PSNR and SSIM are 42.08dB and 97.84%, respectively, which are 0.17dB and 0.82%p higher than Pix2pix, proving the superiority of the network. In addition, the generality of the algorithm is demonstrated by improving its robustness in day and night environmental conditions and its applicability to overseas license plates. |