| dc.description.abstract | The license plate image captured by the dashcam in the vehicle or the monitor may blur due to the distance, lack of focus, or high speed of the vehicle. Therefore, the license plate recognition system can’t accurately identify the license plate. Although there is a lot of literature that uses Generative Adversarial Networks (GAN) to achieve and have related research results, the results are different. Some reconstruction success rates are low, and some are only effective for specific blurring methods. After studying different literature, it’s found that different GAN architectures have obvious differences in the results of reconstructed images. Therefore, this thesis searches for and modifies the existing GAN architecture, pairs and combines different generator architectures, discriminator architectures, and loss functions, and compares the effects of reconstructed images under different combinations to find the combination with the best reconstruction effect. In addition, we are also curious about whether increasing the number of image reconstructions will improve the reconstruction effect, and give experiments and analysis.
We divide the reconstruction success criteria into two types, one is ＂the license plate is completely reconstructed correctly＂, and the other is ＂the license plate can be recognized after reconstruction＂. The first type is that the human eye can barely read the license plate number in the blurred image, and the license plate number in the reconstructed image is completely correct. The second type is that the license plate is so blurred that it can not be read by the human eye, and the reconstructed license plate image can be recognized as a reference (the reconstruction may not be completely correct). Both metrics for evaluating the success of reconstruction use SSIM (structural similarity). The final results of this thesis show that no matter which standard is used, the reconstruction effect using DeblurGAN is the best, where the generator uses ResNet with global skip connections, and the discriminator uses multi-scale PatchGAN. Without classification, the overall avgSSIM is 0.8036. Regarding the number of image reconstruction times, if it is the first type of license plate, the avgSSIM of the license plate image reconstructed once has reached 0.8536, which means that the license plate image reconstructed once is clear enough and completely correct, and there is no need for secondary reconstruction. The avgSSIM of reconstruction twice drops to 0.7647 because the background or a few blocks are more different from the original image. Generally, if it is the second type of license plate, the reconstructed license plate is still blurred or not completely correct. If the reconstructed license plate is not clear enough, the license plate can be reconstructed several times until the license plate can be recognized as a reference or can not be changed clearly. Even though the avgSSIM of reconstruction once is higher than that of reconstruction twice, it is more important to be able to recognize the reconstructed license plate than to be correct. Therefore, the second type of license plate has to determine the optimal number of image reconstructions according to the condition of the reconstructed license plate. | en_US |