第3期 曹锦纲，等：基于生成式对抗网络的道路交通模糊图像增强 ·497· 从图9中可以看到，Sun等的方法在处理 representation for natural image deblurring[C]//Proceed- 图像边缘模糊增强时，效果不是很好，细节不够 ings of 2013 IEEE Conference on Computer Vision and 丰富，文献[11]和文献[14]的方法都会出现失 Pattern Recognition.Portland,USA.2013:1107-1114 真，而经本文模型处理后的图像在主观的视觉感 [7]GOLDSTEIN A.FATTAL R.Blur-kernel estimation from 受方面取得了较好的效果。从表3中可以看到， spectral irregularities[C]//Proceedings of the 12th Sun等四提出的模型峰值信噪比为24.8L,Kpym等 European Conference on Computer Vision.Florence,Italy, 提出的模型峰值信噪比为26.31，而本文提出的模 2012:622-635 型的峰值信噪比达到了27.27，同时本文提出的模 [8]PAN Jinshan,HU Zhe,SU Zhixun,et al.Deblurring text 型的结构相似度达到了0.8991，皆高于另外两个 images via 10-regularized intensity and gradient 模型的结构相似度。因此，本文算法模型无论是 prior[C]//Proceedings of 2014 IEEE Conference on Com- 从峰值信噪比还是从结构相似度上来看都要优于 puter Vision and Pattern Recognition.Columbus,USA, 其他的算法。 2014:2901-2908. [9]PAN Jinshan,SUN Deging,PFISTER H,et al.Blind im- 4结束语 age deblurring using dark channel prior[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern 针对道路交通场景下的模糊图像，基于生成 Recognition.Las Vegas,USA,2016:1628-1636 式对抗网络，本文提出了一个利用多尺度提取图 [10]SCHULER C J.HIRSCH M.HARMELING S,et al. 像特征值多路径学习的模型，通过判别网络和生 Learning to deblur[J].IEEE transactions on pattern ana- 成网络间的对抗训练，端到端生成清晰图像。经 lysis and machine intelligence,2016,38(7):1439-1451. 过实验表明，生成图像的细节更加丰富，无论是 [11]SUN Jian,CAO Wenfei,XU Zongben,et al.Learning a 主观还是客观，本文提出的模型针对道路交通场 convolutional neural network for non-uniform motion 景下的模糊图像增强，都取得了较好的效果。虽 blur removal[C]//Proceedings of 2015 IEEE Conference 然实验证明了该算法的优越性，但是生成图像与 on Computer Vision and Pattern Recognition.Boston, 原始图像相比还存在差距，仍需要进一步优化网 USA,2015:769-777. 络结构和调整参数。 [12]XIAO Lei,WANG Jue,HEIDRICH W,et al.Learning 参考文献 high-order filters for efficient blind deconvolution of doc- ument photographs[C]//Proceedings of the 14th European [1]陈春雷，叶东毅，陈昭炯.多局部模糊核融合的图像盲去 Conference on Computer Vision.Amsterdam,The Neth- 模糊算法U.光子学报，2018,47(10：205-215. erlands.2016:734-749. CHEN Chunlei,YE Dongyi,CHEN Zhaojiong.Blind im- [13]NAH S.KIM T H,LEE K M.Deep multi-scale convolutiona age deblurring via multi-local kernels'fusion[J].Acta I neural network for dynamic scene deblurring[C]//Pro- photonica sinica,2018,47(10):205-215 ceedings of 2017 IEEE Conference on Computer Vision [2]BAHAT Y,EFRAT N.IRANI M.Non-uniform blind and Pattern Recognition.Honolulu,USA,2017:257-265. deblurring by reblurring[Cl/Proceedings of 2017 IEEE In- [14]KUPYN O,BUDZAN V,MYKHAILYCH M,et al. ternational Conference on Computer Vision.Venice,Italy, DeblurGAN:blind motion deblurring using conditional 2017:3306-3314 adversarial networks[Cl//Proceedings of 2018 IEEE/CVF [3]CHAN T F.WONG C K.Total variation blind deconvolu- Conference on Computer Vision and Pattern Recognition tion[J].IEEE transactions on image processing,1998,7(3): Salt Lake City,USA,2018:8183-8192 370-375. [15]MIRZA M.OSINDERO S.Conditional generative ad- [4]CHO S,LEE S.Fast motion deblurring[J].ACM transac- versarial nets[J/OL].[2019-03-29].https://arxiv.org/ tions on graphics,2009,28(5):1-8. abs/1411.1784 [5]XU Li,JIA Jiaya.Two-phase kernel estimation for robust [16]GOODFELLOW I J,POUGET-ABADIE J,MIRZA M. motion deblurring[C]//Proceedings of the 11th European et al.Generative adversarial nets[Cl//Proceedings of the Conference on Computer Vision.Crete,Greece,2010: 27th International Conference on Neural Information Pro- 157-170. cessing Systems.Cambridge,USA,2014:2672-2680. [6]XU Li,ZHENG Shicheng,JIA Jiaya.Unnatural 10 sparse [17]GULRAJANI I.AHMED F,ARJOVSKY M,et al.Improved从图 9 中可以看到，Sun 等 [11] 的方法在处理 图像边缘模糊增强时，效果不是很好，细节不够 丰富，文献 [11] 和文献 [14] 的方法都会出现失 真，而经本文模型处理后的图像在主观的视觉感 受方面取得了较好的效果。从表 3 中可以看到， Sun 等 [11] 提出的模型峰值信噪比为 24.81，Kupyn 等 [14] 提出的模型峰值信噪比为 26.31，而本文提出的模 型的峰值信噪比达到了 27.27，同时本文提出的模 型的结构相似度达到了 0.899 1，皆高于另外两个 模型的结构相似度。因此，本文算法模型无论是 从峰值信噪比还是从结构相似度上来看都要优于 其他的算法。 4 结束语 针对道路交通场景下的模糊图像，基于生成 式对抗网络，本文提出了一个利用多尺度提取图 像特征值多路径学习的模型，通过判别网络和生 成网络间的对抗训练，端到端生成清晰图像。经 过实验表明，生成图像的细节更加丰富，无论是 主观还是客观，本文提出的模型针对道路交通场 景下的模糊图像增强，都取得了较好的效果。虽 然实验证明了该算法的优越性，但是生成图像与 原始图像相比还存在差距，仍需要进一步优化网 络结构和调整参数。 参考文献： 陈春雷, 叶东毅, 陈昭炯. 多局部模糊核融合的图像盲去 模糊算法 [J]. 光子学报, 2018, 47(10): 205–215. CHEN Chunlei, YE Dongyi, CHEN Zhaojiong. Blind im￾age deblurring via multi-local kernels’ fusion[J]. Acta photonica sinica, 2018, 47(10): 205–215. [1] BAHAT Y, EFRAT N, IRANI M. Non-uniform blind deblurring by reblurring[C]//Proceedings of 2017 IEEE In￾ternational Conference on Computer Vision. Venice, Italy, 2017: 3306−3314. [2] CHAN T F, WONG C K. Total variation blind deconvolu￾tion[J]. IEEE transactions on image processing, 1998, 7(3): 370–375. [3] CHO S, LEE S. Fast motion deblurring[J]. ACM transac￾tions on graphics, 2009, 28(5): 1–8. [4] XU Li, JIA Jiaya. Two-phase kernel estimation for robust motion deblurring[C]//Proceedings of the 11th European Conference on Computer Vision. Crete, Greece, 2010: 157−170. [5] [6] XU Li, ZHENG Shicheng, JIA Jiaya. Unnatural l0 sparse representation for natural image deblurring[C]//Proceed￾ings of 2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland, USA, 2013: 1107−1114. GOLDSTEIN A, FATTAL R. Blur-kernel estimation from spectral irregularities[C]//Proceedings of the 12th European Conference on Computer Vision. Florence, Italy, 2012: 622−635. [7] PAN Jinshan, HU Zhe, SU Zhixun, et al. Deblurring text images via l0-regularized intensity and gradient prior[C]//Proceedings of 2014 IEEE Conference on Com￾puter Vision and Pattern Recognition. Columbus, USA, 2014: 2901−2908. [8] PAN Jinshan, SUN Deqing, PFISTER H, et al. Blind im￾age deblurring using dark channel prior[C]//Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, USA, 2016: 1628−1636. [9] SCHULER C J, HIRSCH M, HARMELING S, et al. Learning to deblur[J]. IEEE transactions on pattern ana￾lysis and machine intelligence, 2016, 38(7): 1439–1451. [10] SUN Jian, CAO Wenfei, XU Zongben, et al. Learning a convolutional neural network for non-uniform motion blur removal[C]//Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition. Boston, USA, 2015: 769−777. [11] XIAO Lei, WANG Jue, HEIDRICH W, et al. Learning high-order filters for efficient blind deconvolution of doc￾ument photographs[C]//Proceedings of the 14th European Conference on Computer Vision. Amsterdam, The Neth￾erlands, 2016: 734−749. [12] NAH S, KIM T H, LEE K M. Deep multi-scale convolutiona l neural network for dynamic scene deblurring[C]//Pro￾ceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, USA, 2017: 257−265. [13] KUPYN O, BUDZAN V, MYKHAILYCH M, et al. DeblurGAN: blind motion deblurring using conditional adversarial networks[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA, 2018: 8183−8192. [14] MIRZA M, OSINDERO S. Conditional generative ad￾versarial nets[J/OL]. [2019 –03 –29].https: //arxiv.org/ abs/1411.1784. [15] GOODFELLOW I J, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[C]//Proceedings of the 27th International Conference on Neural Information Pro￾cessing Systems. Cambridge, USA, 2014: 2672−2680. [16] [17] GULRAJANI I, AHMED F, ARJOVSKY M, et al. Improved 第 3 期 曹锦纲，等：基于生成式对抗网络的道路交通模糊图像增强 ·497·