基于深度学习的车牌图像去运动模糊技术  

License Plate Motion Deblurring Based on Deep Learning

作　　者：毛勇 陈华华[1] MAO Yong;CHEN Huahua(School of Communication Engineering,Hangzhou Dianzi University,Hangzhou Zhejiang 310018,China)

机构地区：[1]杭州电子科技大学通信工程学院,浙江杭州310018

出　　处：《杭州电子科技大学学报（自然科学版）》2018年第5期29-33,共5页Journal of Hangzhou Dianzi University：Natural Sciences

摘　　要：针对目前车牌去运动模糊算法的估计模糊核过程复杂、复原质量低等问题,提出一种基于生成对抗网络的车牌去运动模糊算法。运用深度学习的方法对运动模糊图像直接进行复原,省去了估计模糊核的过程,并且增加梯度图像l1正则化,保护复原图像的强边缘特征;最后以分割后的车牌字符作为网络输入,随机抽取等量数据进行训练并测试,以增强网络泛化能力。实验结果表明,提出的模型能够有效去除合成运动模糊图像和真实场景下运动模糊图像中存在的运动模糊,对比测试阶段的峰值信噪比指标,所提模型比当前最新的去运动模糊模型提升了0.40dB。In terms of the problem of the complex fuzzy kernel estimation and the low quality of restoration of the current license plate motion deblurring algorithm,agenerative adversarial network（GAN）model for license plate motion deblurring was proposed.Apply deep learning to restore the motion blurred image directly,eliminating the process of estimating the fuzzy kernel.And,adopt the l_1-regularization of the gradient image to protect the strong edge feature of the restored image.Finally,segment license plate characters as input of the network and randomly select equal amount of data for training and testing to enhance the network generalized capabilities.The experimental results show that the proposed model can effectively remove the motion blur in the synthetic motion blurred image and the motion blurred image in the real scene.Compared with the latest motion deblurring algorithm in the test phase,the proposed model can improve the peak signal to noise ratio（PSNR）by 0.40 dB.

关 键 词：深度学习 去运动模糊 生成对抗网络 L1正则化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]