基于Tansformer的生成对抗网络图像去模糊

Image Deblurring Using Transformer-Based Generative Adversarial Network

摘要: 现实场景中的图像模糊通常由相机抖动运动等诸多因素造成，由于模糊核未知，图像去模糊问题成为图像处理逆问题。目前去模糊方法虽取得较好效果，但在处理复杂真实厂家的模糊时，会存在纹理细节恢复不足和去模糊结果伪影的问题。为此，本文提出了一种新颖的transformer方法的生成对抗网络。该网络融合局部–全局双元Transformer模块。我们在不同基准数据集(如GoPro和HIDE)上进行了实验，主观客观结果均表明，本方法在恢复图像边缘信息和真实纹理特征方面取得较好的结果。

Abstract: Blur in real-world images is often caused by various factors such as camera shake. Due to the unknown blur kernel, image deblurring becomes an inverse problem in image processing. Although current deblurring methods have achieved promising results, they still suffer from insufficient texture detail recovery and artifacts when dealing with complex real-world blur. To address this issue, this paper proposes a novel transformer-based generative adversarial network. The network integrates a local-global dual Transformer module. We conducted experiments on different benchmark datasets, such as GoPro and HIDE. Both subjective and objective results demonstrate that the proposed method achieves favorable outcomes in restoring image edge information and realistic texture features.