面向图像复原的残差密集生成对抗网络新方法

New Method of Residual Dense Generative Adversarial Networks for Image Restoration

针对成像设备抖动和场景目标物体移动等因素导致的图像运动模糊问题,提出了一种基于残差密集生成对抗网络的图像去运动模糊新方法,采用端到端的方式对模糊图像进行直接复原,避免了对模糊核进行估计.使用残差密集网络结构作为生成式模型的核心组件,并在对抗损失、感知损失、L1损失及梯度L1正则化的共同约束下,进行生成器模型与判别器模型之间的对抗训练,重建出细节丰富、真实锐利的去模糊图像.实验结果表明,本文方法对图像边缘特征和细节信息的提取和复原具有良好的作用,与已有的代表性图像复原算法相比,在主观视觉和图像质量定量评价上均有明显提高.

Aiming at image motion blurring caused by shaking of imaging equipment and moving of scene target,the paper proposes a new method of image motion blurring removal based on residual dense generative adversarial networks.The motion blurred image is directly restored in an end-to-end manner without estimating blur kernels.Residual dense network structure is adopted as the core component of the generative model.Adversarial training is performed between the generator model and the discriminator model under the joint constrains of adversarial loss,perceptual loss,L1 loss function as well as gradient L1 regularization so that blurred images are reconstructed with rich details and real sharpness.The experimental results show that the proposed method has a better effect on extracting and restoring image edge features and detail information.Compared with existing representative image restoration algorithms,the proposed method gains significant improvement in both the subjective visual effect and the quantitative image quality metrics.