计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (21): 232-242.DOI: 10.3778/j.issn.1002-8331.2103-0530

• 图形图像处理 • 上一篇    下一篇

基于模糊核校正的未知退化方式图像超分辨率

汪澜,孔祥屹,张海涛   

  1. 辽宁工程技术大学 软件学院,辽宁 葫芦岛 125100
  • 出版日期:2022-11-01 发布日期:2022-11-01

Image Super-Resolution Based on Blur Kernel Correction with Unknown Degradation Method

WANG Lan, KONG Xiangyi, ZHANG Haitao   

  1. College of Software, Liaoning Technical University, Huludao, Liaoning 125100, China
  • Online:2022-11-01 Published:2022-11-01

摘要: 真实世界图像包含未知模糊和噪声信息,现有以双三次下采样为退化方式的数据集无法训练出适用于真实世界图像的超分辨率网络。为此设计预测器初步估计模糊核并设计校正器得到准确的图像模糊核信息。高分辨率图像通过模糊化处理并注入噪声构建低分辨率真实世界图像。设计新型超分辨率网络结构,每个卷积层后根据模糊核对特征图进行空间特征变换,提高网络处理不同模糊图像的能力。将非线性映射部分以残差密集块结构相连,并整合入生成对抗网络框架增强纹理细节恢复能力。在Flickr2K和DIV2K两个数据集上的测试结果表明新方法的峰值信噪比、结构相似度和感知指数高于EDSR和ESRGAN等经典方法。

关键词: 真实世界图像, 模糊核, 残差密集块, 空间特征变换, 感知指数

Abstract: Real-world images contain unknown blur and noise information, and the existing datasets using bicubic downsampling as the degradation method cannot train super-resolution networks suitable for real-world images. Therefore, a predictor is designed to estimate the blur kernel and a corrector is designed to obtain precise image blur kernel information. High-resolution images are processed by blurring and injecting noise to construct low-resolution real-world images, which are utilized to train a new super-resolution network. After each convolutional layer, spatial feature transformation is performed on the feature map according to the blur kernel, which improved the network’s ability to process different blurred images. The non-linear mapping part is connected with a residual dense block structure and integrated into the generative adversarial network framework to enhance the ability of texture detail recovery. The experiment results on the datasets of Flickr2K and DIV2K show that the peak signal-to-noise ratio, structural similarity and perception index of the new method are higher than those of classic methods such as EDSR and ESRGAN.

Key words: real-world image,  , kernel,  , residual dense block,  , spatial feature transformation,  , perception index