基于生成对抗网络的运动模糊车牌图像复原方法

A Method for Restoring Motion-Blurred License Plate Images Based on Generative Adversarial Network

为解决汽车运动过快产生模糊导致车牌识别算法失效的问题,对深度学习的生成对抗网络去模糊方法进行了研究,提出了一种基于生成对抗网络的模糊车牌图像复原方法。主要思路为使用图像复原网络NAFNet中的NAFBlock替换DeblurGAN-v2生成器中的基本卷积块,并在特征提取网络中加入了高效通道注意力机制。对于原模型和修改后的模型,设计了四组不同模型消融实验。实验结果表明,提出方法在复原模糊车辆图像复原任务数据上,峰值信噪比为21.2624,结构相似度为0.6431,较好地解决了模糊车牌复原的问题。

To solve the problem of license plate recognition algorithms failing due to blurriness caused by fast-moving vehicles,this paper studies the Generative Adversarial Network deblurring method of Deep Learning,and proposes a fuzzy license plate image restoration method based on Generative Adversarial Network.The main idea is to use the NAFBlock in the image restoration network NAFNet to replace the basic convolution block in the DeblurGAN-v2 generator,and an Efficient Channel Attention mechanism is added to the feature extraction network.For the original model and the modified model,four groups of different model ablation experiments are designed.The experiment results show that the proposed method has a peak signal-to-noise ratio of 21.2624 and a structural similarity index of 0.6431 on the task data of restoring blurred vehicle image restoration,which better solves the problem of blurred license plate restoration.