GCA:一种结合注意力机制的图像超分辨率重建模型  

GCA:An image super-resolution reconstruction model that combines attention mechanisms

作　　者：彭学桂 PENG Xuegui(School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《智能计算机与应用》2023年第1期12-18,共7页Intelligent Computer and Applications

摘　　要：深度学习在图像超分辨率任务上的重建精度和计算性能均表现优异。通常的方法是将高分辨率图片通过简单的插值法进行下采样得到低分辨率图片,然后将其输入到超分网络中生成高清图片。然而,真实世界的低质量图片往往存在噪声,且其模糊状态按一定分布呈现,直接将通过插值得到的低分辨率图片来训练的模型在真实世界图片上的效果并不好。针对这一问题,本文在数据集的构造上采用了一种“降级”结构,通过退化算法估计各种模糊核以及真实的噪声分布,使得作为网络输入的低分辨率图片与真实图像共享一个公共域;提出一种结合注意力机制的GCA模型,将生成模块得到的超分辨率图片与标签图像共同输入判别模块的胶囊网络中进行二分类,最终达到判别模块无法区分超分辨率图片和对应的高分辨率图片。在City100数据集上的实验表明,本文提出的GCA模型在真实世界图像上取得了更好的效果。Deep learning has performed well in reconstruction accuracy and computational performance on image hyper-resolution tasks. The usual approach is to take a low-resolution picture by sampling it by simple interpolation and then inputting it into a super-divided network to generate a high-definition picture. However, real-world low-quality pictures tend to be noisy, and their fuzzy state is presented in a certain distribution, and models trained directly by interpolation of low-resolution images do not work well in real-world images. In response to this problem, this paper adopts a "degraded" structure in the construction of the data set. The degradation algorithm is used to estimate various fuzzy cores and real noise distribution, so that low-resolution pictures as network inputs share a common domain with real images. At the same time, a GCA model combined with an attention mechanism is proposed, which combines the super-resolution image obtained by the generating module with the label image into the Capsule network of the distinguishing module for two classifications, and finally achieves that the distinguishing module cannot distinguish between the super-resolution picture and the corresponding high-resolution picture. Experiments on the City100 data set show that the GCA model proposed in this paper has achieved better results in real-world images.

关 键 词：图像超分辨率 胶囊网络 注意力机制 模糊核 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]