使用對抗生成網路去除影像動態模糊

Motion Deblur Using Generative Adversarial Network

徐旨暘(Jr-Yang Shiu)

指導教授 ： 顏淑惠

共同指導教授 ： 林慧珍

淡江大學/工學院/資訊工程學系資訊網路與多媒體碩士班/碩士(2020年)

https://doi.org/10.6846/TKU.2020.00659

摘要

本篇論文以對抗生成網路作為基本架構，訓練目標為取得模糊影像和清晰影像間的差異，利用結果將模糊影像重建為清晰影像來達到去模糊的效果。所提出的方法架構中使用了ResNet[4]中的residual block堆疊將訓練的重點放在模糊和清晰影像差異上，並使用Atrous Spatial Pyramid Pooling (ASPP) [13]的方法來加強高頻特徵的傳遞。而在對抗生成的部分使用兩個鑑別器針對不同目標做判斷，加強影像細部還原同時還要保持影像整體的完整性，最後在計算loss時加入計算從模糊影像所取出的邊界影像和清晰影像的邊界影像的差異，來針對影像邊界作為主要目標。

關鍵字

對抗生成網路 ； 影像去模糊

並列摘要

In this paper, we use a generative adversarial network (GAN) as the basic structure to solve the motion blurring problem. Short cuts and residual blocks are adopted so the learning focuses on the difference between the blurred image and the clear image. Atrous Spatial Pyramid Pooling (ASPP) [13] method is used to enlarge receptive fields while preserving the high-frequency features. In the part of the discriminator, global and local discriminators are used to strengthen the restoration of image details as well as to maintain the integrity of the image. Finally, the edge loss is used to insure the restored image can preserve the structure details as the clear one. Test on two public data sets, Kohler and Gopro test sets, our method shows a good result on Kohler but not as good on Gopro.

並列關鍵字

Deblurring ； Generative Adversarial Network

參考文獻

[1] S. Nah, T. Hyun, K. Kyoung, and M. Lee. Deep Multi-scale

Convolutional Neural Network for Dynamic Scene Deblurring. 2016.

[2] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych,

Dmytro Mishkin, and Jiˇr´ı Matas. Deblurgan: Blind motion

deblurring using conditional adversarial networks. 2017.