欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 图像与视频处理 • 上一篇    下一篇

采用双框架生成对抗网络的图像运动模糊盲去除

  

  1. (1. 北京工商大学计算机与信息工程学院，北京 100048；
     2. 北京工商大学食品安全大数据技术北京市重点实验室，北京 100048)
  • 出版日期:2019-12-31 发布日期:2020-01-20
  • 基金资助:
    北京市教委科研计划一般项目(SQKM201610011010)；北京市自然科学基金项目(4162019)；北京市科技计划课题(Z161100001616004)

Blind Motion Image Deblurring Using Two-Frame Generative  Adversarial Network

  1. (1. Department of Computer and Information Engineering, Beijing Technology and Business University, Beijing 100048, China; 
    2. Beijing Key Laboratory of Big Data Technology of Food Safety, Beijing Technology and Business University, Beijing 100048, China)
  • Online:2019-12-31 Published:2020-01-20

PDF (PC)

220

摘要： 传统运动模糊盲去除方法需先预测模糊图像的模糊核，再复原清晰图像。而实际 环境中的复杂的模糊核使此方法不能在视觉上很好地减小实际图像和复原后图像的差异，且直 接将现流行的生成对抗模型应用在图像模糊盲去除任务中会有严重的模式崩塌现象。因此，围 绕去模糊任务的特点提出了一种端到端的生成对抗网络模型——双框架生成对抗网络。该方案 不需要预测模糊核，直接实现图片运动模糊的盲去除。双框架生成对抗网络在原有 CycleGan 基础上将其网络结构和损失函数均作出了改进，提高了运动图像盲去除的精度，并且在样本有 限情况下大幅度增强了网络的稳定性。实验采用最小均方差优化网络训练，最后通过生成网络 和判别网络对抗训练获得清晰图像。在 ILSVRC2015 VID 数据集上的实验结果表明,该方法复原 质量更高，且复原结果在后续目标检测任务中达到了更优的效果。

关键词: 生成对抗网络, 运动模糊盲去除, 循环一致性, 条件模型

Abstract: Traditional methods of motion blur blind removal are required to predict the fuzzy kernel of the blurred images and then restore the clear images. However, the fuzzy kernel in the real environment is complex, causing such method to fail to reduce the difference between the actual image and the restored one. Moreover, the application of the popular generative adversarial model directly to image blurring blind removal will cause serious pattern collapse. Therefore, based on the characteristics of deblurring tasks, we proposed an end-to-end generative adversarial network model—two-frame generative adversarial network. The scheme does not need to predict the fuzzy kernel, and it can directly realize the blind removal of the motion blur of images. Based on the original CycleGan, the two-frame generative adversarial network improved its network structure and loss function to improve the accuracy of blind removal of moving images and greatly improve the stability of the network in the case of limited samples. The minimum mean square error was used to optimize the network training. Finally, a clear image was obtained by the adversarial training between generative network and discriminant network. Experimental results on the ILSVRC2015 VID dataset show that the method has a higher quality of restoration. And the restored results appear to be better in subsequent target detection tasks.

Key words: generative adversarial networks, blind motion blur, cycle consistency, conditional models