The Comparison of the Blurred License Plate Reconstruction Effects in Four Modified GANs

This article considers four kinds of Generative Adversarial Networks (GANs), which are Pix2pix, Pix2pixHD, LSGAN, and DeblurGAN, and modify their architectures on the generators, discriminators, and/or loss functions. Then we used these modified GANs to reconstruct the blurred license plate images and evaluated and compared their reconstruction effects. The reconstruction index is the structural similarity (SSIM) between the original and blurred images. The experimental results show that the modified DeblurGAN, which is with the multi-scale PatchGAN discriminator and ResNet generator, has the highest average SSIM among the four modified GANs. It was found that adding SSIM into the content loss of DeblurGAN does not obviously improve the reconstruction effect. In addition, the image reconstruction once is almost better than that twice on the SSIM index. However, if the blurred image has not the corresponding original image, the multiple reconstructions may make the license plate numbers and letters much clearer.