~图像作为一种信息载体，已经成为人们记录和传递信息的重要途径。伴随着人工智能时代的到来，计算机视觉发挥着越来越重要的作用，无论军事还是民用方面，人们对图像质量的要求越来越高。然而在图像的获取过程中，由于目标物体运动、场景变化等多种因素的影响，导致图像模糊，降低自身的使用价值。因此，研究复原质量高、适用范围广的图像去模糊方法很有必要。
本文针对传统图像去模糊方法估计模糊核的不确定问题，在分析双判别器生成对抗网络的基础上，引入加权的思想重构目标函数，提出基于双判别器加权生成对抗网络的图像去模糊方法，理论证明其在最优判别器下，通过最小化模型与真实数据之间的KL散度和反向KL散度，生成器可以学习到真实数据分布。在人工合成的二维数据、MNIST数据、Cifar-10三个数据集上验证了算法生成多样性样本的能力。分别设计了生成器与判别器的模型结构与损失函数，在GOPRO训练集、Pytorch框架中，对生成对抗网络、双判别器生成对抗网络、DeblurGAN网络、双判别器加权生成对抗网络四种方法进行去模糊效果对比分析，验证双判别器加权生成对抗网络对模糊图像的复原能力。
实验结果表明，双判别器加权生成对抗网络在峰值信噪比、图像结构相似度这两个指标上均有更好的表现，进行去模糊操作后，细节恢复较好，还原图像更加真实，没有棋盘伪影、锐化过度等视觉效果。本文方法为还原复杂场景下的模糊图像提供了理论基础与应用参考。
 

~As a kind of information carrier, image has become an important way for people to record and transmit information. With the advent of the era of artificial intelligence, computer vision is playing an increasingly important role. Whether it is military or civilian, people have higher and higher requirements for image quality. However, in the process of image acquisition, due to the influence of various factors such as the movement of the target object and the scene change, the image is blurred and its use value is reduced. Therefore, it is necessary to develop an image deblurring method with high restoration quality and wide application range.
Aiming at the traditional image deblurring method to estimate the ill-posed problem of the blurring kernel, based on the analysis of the Dual Discriminator Weighted Generation Adversarial Network, the weighted idea is introduced to reconstruct the objective function, and proposed the image deblurring method of the Dual Discriminator Weighted Generative Adversarial Network, theoretically proves that under the optimal discriminator, by minimizing the KL divergence and reverse KL divergence between the model and the real data, the generator can learn the real data distribution. The ability of the algorithm to generate diverse samples is verified on three datasets of artificially synthesized 2D data, MNIST datasets, and Cifar-10 datasets. The model structure and loss function of the generator and the discriminator are designed respectively. In the GOPRO training datasets and the Pytorch framework, the four methods of Generation Adversarial Network, Dual Discriminator Generation Adversarial Network, DeblurGAN, and Dual Discriminator Weighted Generation Adversarial Network are compared and analyzed to verify the ability of the Dual Discriminator Weighted Generation Adversarial Network to recover blurred images.
The experimental results show that, the Dual Discriminator Weighted Generation Adversarial Network has better performance on the two indicators of PSNR and SSIM. After the deblurring operation, the details are restored better, and the restored image is more realistic, without checkerboard artifacts, over-sharpening and other visual effects. In addition, the method in this paper provides a theoretical basis and application reference for restoring blurred images in complex scenes.

[1]甄诚, 杨永胜, 李元祥, 等. 基于多尺度生成对抗网络的大气湍流图像复原[J/OL]. 计算机工程, 2021, 1-8. https://doi.org/10.19678/j.issn.1000-3428.0059695.

[2]Goodfellow I., Pouget-Abadie J., Mirza M., et al. Generative Adversarial Nets[C]//27th International Conference on Neural Information Processing Systems. 2014: 2672-2690.

[3]李明东, 张娟, 伍世虔, 等. 基于RANSAC变换的车牌图像去模糊算法[J]. 传感器与微系统, 2020, 39(2): 153-156+160.

[4]张文义. 基于智能监控系统的图像质量增强算法的研究[D]. 南京: 南京邮电大学, 2015.

[5]马苏欣, 王家希, 戴雅淑, 等. 监控视频下模糊车牌的去模糊与识别探析[J]. 信息系统工程, 2019, 311(11): 111-113.

[6]裴慧坤, 颜源, 林国安, 等. 基于生成对抗网络的无人机图像去模糊方法[J]. 地理空间信息, 2019, 17(12): 4-9+155.

[7]刘晨旭. 基于Wasserstein生成对抗网络的遥感图像去模糊研究[D]. 桂林: 广西师范大学, 2019.

[8]张广明, 高爽, 尹增山, 等. 基于模糊图像和噪声图像的遥感图像运动模糊复原方法[J]. 电子设计工程, 2017, 25(18): 82-86.

[9]谭海鹏, 曾炫杰, 牛四杰, 等. 基于正则化约束的遥感图像多尺度去模糊[J]. 中国图象图形学报, 2015, 20(3): 386-394.

[10]张建国, 拓洋洋, 蒋瑞娇, 等. Richardson-Lucy算法在模糊图像复原中的改进[J]. 计量学报, 2020, 41(2): 153-158.

[11]Tigga U., Jha J. Image deblurring with impulse noise using alternating direction method of multipliers and Lucy-Richardson method[C]//IEEE International Conference on Computational Intelligence and Communication Networks. 2016: 230-235.

[12]Fergus R., Singh B., Hertzmann A., et al. Removing camera shake from a single photograph[C]//International Conference on Computer Graphics and Interactive Techniques, 2006: 787-794.

[13]Perrone D., Favaro P. Total variation blind deconvolution: The devil is in the details[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2014: 2909-2916.

[14]Xu L., Jia J. Two-phase kernel estimation for robust motion deblurring[C]//11th European Conference on Computer Vision. Heraklion, Crete, Greece, 2010: 157-170.

[15]Shan Q., Jia J., Agarwala A. High-quality motion deblurring from a single image[J]. ACM Transactions on Graphics, 2008, 27(3): 1-10.

[16]Cai F., Ji H., Liu Q., et al. Framelet-based blind motion deblurring from a single image[J]. IEEE Transactions on Image Processing, 2012, 21(2): 562-572.

[17]Xu L., Zheng S., Jia J. Unnatural l0 sparse representation for natural image deblurring[C]//IEEE Conference on Computer Vision and Pattem Recognition. 2013: 1107-1114.

[18]卢晶, 胡钢, 秦新强. 多约束的运动模糊图像盲复原方法[J]. 红外技术, 2017, 39(12): 1098-1106.

[19]Sun J., Cao W., Xu Z., et al. Learning a Convolutional Neural Network for Non-uniform Motion Blur Removal[C]//IEEE Conference on Computer Vision and Pattem Recognition. 2015: 769-777.

[20]Gong D., Yang J., Liu L., et al. From motion blur to motion flow: a deep learning solution for removing heterogeneous motion blur[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2017: 2319-2328.

[21]Nah S., Kim H., Lee M. Deep Multi-scale Convolutional Neural Network for Dynamic Scene Deblurring[C]//IEEE Conference on Computer Vision and Pattem Recognition. 2017: 257-265.

[22]Zhang J., Pan J., Ren J., et al. Dynamic scene deblurring using spatially variant recurrent neural networks[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2018: 2521-2529.

[23]Li J., Li K., Yan B. Scale-aware deep network with hole convolution for blind motion deblurring[C]//IEEE International Conference on Multimedia and Expo. 2019: 658-663.

[24]Liu K., Yeh C., Chung J., et al. A motion deblur method based on multi-scale high frequency residual image learning[J]. IEEE Access, 2020, (8): 66025-66036.

[25]林晨, 尹增山. 基于注意力机制的图像盲去模糊算法[J]. 计算机应用, 2020, 40(2): 151-157.

[26]张嘉晖, 沈文忠. 基于循环卷积神经网络的图像去模糊算法[J]. 科技创新与应用, 2021, (6): 26-27+30.

[27]Kupyn O., Budzan V., Mykhailych M., et al. DeblurGAN: blind motion deblurring using conditional adversarial networks[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2018: 8183-8192.

[28]孙季丰, 朱雅婷, 王恺. 基于DeblurGAN和低秩分解的去运动模糊[J]. 华南理工大学学报, 2020, 48(1): 32-41+50.

[29]陈乔松, 隋晓旭, 官旸珺, 等. 基于多尺度残差生成对抗网络的单图像盲去运动模糊方法[J]. 计算机应用研究, 2021, 38(3): 919-922.

[30]Kupyn O., Martyniuk T., Wu J., et al. DeblurGAN-v2: deblurring (orders-of-magnitude) faster and better[C]//IEEE International Conference on Computer Vision. 2019: 8877-8886.

[31]张柯. 基于双鉴别器的条件生成对抗网络图像去模糊方法研究[D]. 西安: 长安大学, 2019.

[32]Shlens J. Notes on Kullback-Leibler Divergence and Likelihood[J]. Mathematics, 2014, (8): 1-4.

[33]Dai Z., Yang Z., Yang F. Good semi-supervised learning that requires a bad gan[C]//27th International Conference on Neural Information Processing Systems. 2017: 1-13.

[34]Jin Q., Luo X., Shi Y., et al. Image generation method based on improved condition GAN[C]//6th International Conference on Systems and Informatics. 2019: 1290-1294.

[35]Li Y., Zhao K., Ren F., et al. Research on super-resolution image reconstruction based on low-resolution infrared sensor[J]. IEEE Access, 2020, (8): 69186-69199.

[36]Ledig C., Theis L., Huszar F., et al. Photo-realistic single image super-resolution using a Generative Adversarial Network[C]//International Conference on Computer Vision and Pattern Recognition. 2016: 4681-4690.

[37]He Y., Li X., Li R., et al. A Deep-Learning Method for Radar Micro-Doppler Spectrogram Restoration[J]. Sensors, 2020, 20(17): 5007-5021.

[38]Zhu J., Park T., Isola P., et al. Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks[C]//IEEE International Conference on Computer Vision. Venice, Italy, 2017: 2242-2251.

[39]Isola P., Zhu J., Zhou T., et al. Image-to-image translation with conditional adversarial networks[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2016: 5967-5976.

[40]Yang T., Chang X., Su H., et al. Raindrop removal with light field image using image inpainting[J]. IEEE Access, 2020, (8): 58416-58426.

[41]Babaee M., Zhu Y., Köpüklü O., et al. Gait energy image restoration using Generative Adversarial Networks[C]//IEEE International Conference on Image Processing. 2019: 2596-2600.

[42]Xiang P., Wang L., Cheng J., et al. A deep network architecture for image inpainting[C]//IEEE International Conference on Computer and Communications. 2017: 1851-1856.

[43]Fuglede B., Topsoe F. Jensen-Shannon divergence and Hilbert space embedding[C]//IEEE International Symposium on Information Theory. 2004: 31-36.

[44]Arjovsky M., Chintala S., Bottou L. Wasserstein Generative Adversarial Networks[C]//International Conference on Machine Learning. 2017: 214-223.

[45]Cai L., Chen Y., Cai N., et al. Utilizing Amari-Alpha Divergence to Stabilize the Training of Generative Adversarial Networks[J]. Entropy, 2020, 22(410): 1-19.

[46]Cao Y., Jia J., Chen Y., et al. Recent advances of Generative Adversarial Networks in computer vision[J]. IEEE Access, 2019, (7): 14985-15006.

[47]Mirza M., Osindero S. Conditional Generative Adversarial Nets[J]. Computer Science, 2014, (6): 2672-2680.

[48]Radford A., Metz L., Chintala S. Unsupervised representation learning with deep convolutional generative adversarial networks[C]//4th International Conference on Learning Representations. 2015: 1-16.

[49]Nguyen T., Le T., Vu H. Dual Discriminator Generative Adversarial Nets[C]//International Conference on Neural Information Processing Systems. 2017: 2667-2677.

[50]Johnson J., Alahi A., Fei L. Perceptual losses for real-time style transfer and super-resolution[C]//European Conference on Computer Vision. 2016: 694-711．

[51]Sun J., Cao W., Xu Z., et al. Learning a Convolutional Neural Network for Non-uniform Motion Blur Removal[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2015: 769-777.

[52]Li C., Wand M. Precomputed Real-time Texture Synthesis with Markovian Generative Adversarial Networks[C]//European Conference on Computer Vision. 2016: 702-716.

[53]于彤. 图像的边缘特征信息提取与质量评价[D]. 天津: 天津理工大学, 2009.

[54]郑圣超, 叶正麟. 基于HVS的若干图像质量度量方法的研究[D]. 西安: 西北工业大学, 2006.

[55]Theis L., Aäron O., Bethge M. A note on the evaluation of generative models[C]//International Conference on Learning Representations. 2016: 1-9.

[56]Yin H., Li Z., Zuo J., et al. Wasserstein Generative Adversarial Network and convolutional neural network (WG-CNN) for bearing fault diagnosis[J]. Mathematical Problems in Engineering, 2020, 2020(6): 1-16.

[57]Mao X., Li, Q., Xie H. On the effectiveness of Least Squares Generative Adversarial Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019, 41(12): 2947-2960.

[58]Somesh D., Jun, S. Mathematical statistics[M]., United Kingdom: Technometrics, 2000.

[59]Kingma D., Ba J. Adam: A method for stochastic optimization[J]. Computer Science, 2014, 41(7): 1-15.

[60]Luke M., Ben P., David P., et al. Unrolled Generative Adversarial Networks[C]//International Conference on Learning Representations. 2017: 1-25.

[61]Lecun Y., Bottou L., Bengio Y. Gradient-based learning applied to document recognition[J]. Proceedings of the IEEE, 1998, 86(11): 2278-2324.

[62]Krizhevsky A., Hinton G. Learning multiple layers of features from tiny images[J]. Handbook of Systemic Autoimmune Diseases, 2009, 1(4): 1-54.

[63]Rubner Y., Tomasi C., Guibas L. The earth mover's distance as a metric for image retrieval[J]. International Journal of Computer Vision, 2000, 40(2): 99-121.

[64]He K., Zhang X., Ren S., et al. Deep residual learning for image recognition[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2016: 770-778.

[65]He K., Zhang X., Ren S., et al. Identity Mappings in Deep Residual Networks[C]//European Conference on Computer Vision. 2016: 630-645.

[66]Odena A., Dumoulin V., Olah C. Deconvolution and Checkerboard Artifacts[J]. Distill, 2016, 1(10): 1-9.

[67]Shi W., Caballero J., Huszar F., et al. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network[C]//IEEE Conference on Computer Vision and Pattern Recognition. 2016: 1874-1883.

[68]Maas L., Hannun Y., Ng Y. Rectifier Nonlinearities Improve Neural Network Acoustic Models[C]//International Conference on Machine Learning. 2013: 1-6.