通過生成對抗網路模型及起始殘缺密集網路進行動態盲去模糊

去除動態模糊在電腦視覺中是一個被研究已久的課題。在卷積神經網路(CNN)開始被廣泛應用後，也被用來與傳統去模糊的方法結合，多半是用在找出模糊核或顯著的邊緣後再進行去模糊的處理。生成對抗網路(GAN)問世後，在風格轉換的問題上有優秀的表現，於是我們就想將去模糊化的問題也變成一種風格轉換的問題。我們針對去模糊生成對抗網路(DeblurGAN)中的生成器(Generator)架構去做改善，提出了一個結合了起始塊(inception block)、殘缺塊(residual block)以及密集塊(dense block)的新的塊模型來對動態模糊進行去模糊的處理。也因為使用了密集網絡(DenseNet)的概念，可以有效的減少過度擬合(overfitting)的問題。改善後的去模糊生成對抗網路所生成的去模糊圖片在結構相似度以及圖像視覺化中，效果都有顯著的進步。

關鍵字

影像去模糊化；深度學習；生成對抗網絡

並列摘要

Deblurring from a motion blurred image has been studied for some times. After convolution neural network(CNN) be used widely, it can be implemented on finding blur kernel or latent sharp edge from a blurred image. In recent years, the most popular network architecture called generative adversarial network (GAN) which performs well on style transformation. We consider that a deblurring problem as a style transformation problem, so we use GAN to do deblurring. We focus on improving the state-of-the-art deblur method DeblurGAN’s generator, and present a new kind of block which combined inception block, residual block and dense block to do deblurring from motion blur. By using the conception of dense net which can avoid overfitting. The improved DeblurGAN present better in both structural similarity measure and by visual effect.

並列關鍵字

Deblurring ； GAN ； Deep-learning

參考文獻

[1] O. Kupyn, V. Budzan, M. Mykhailych, D. Mishkin, J. Matas. “DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks,” arXiv:1711.07064v4, 2018.

Google Scholar

[2] C. Szegedy, S. Ioffe, V. Vanhoucke, A. Alemi. “Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,” arXiv: 1602.07261v2, 2016.