计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (19): 257-266.DOI: 10.3778/j.issn.1002-8331.2203-0307

• 图形图像处理 • 上一篇    下一篇

跨尺度多分支网络的单帧图像超分辨率重建

刘南艳,马圣祥,魏鸿飞   

  1. 西安科技大学 计算机科学与技术学院,西安 710699
  • 出版日期:2022-10-01 发布日期:2022-10-01

Super-Resolution Reconstruction of Single-Frame Images in Cross-Scale Multi-Branching Networks

LIU Nanyan, MA Shengxiang, Wei Hongfei   

  1. College of Computer Science and Technology, Xi’an University of Science and Technology, Xi’an 710699, China
  • Online:2022-10-01 Published:2022-10-01

摘要: 超分辨率重建算法是计算机视觉领域的重点研究问题之一,目前各个领域对图像要求逐渐提高,现有的超分辨率重建算法通过加深单一网络深度来提升图像质量,忽略了重建时的计算复杂度问题,同时缺少利用图像自身信息来提升图像的重建效果。针对以上问题提出了一种跨尺度多分支的单帧图像超分辨率重建网络,跨尺度模块探索图像内部相似区域的相关性,用于提升重建图像细节信息的能力,并获得更好的视觉效果。多分支结构将图像中重建难度不同的区域,通过不同复杂度的网络分别进行重建,解决了深度网络计算复杂度高的问题。实验结果表明,该模型在Urban100和Manga109测试集上相比其他方法具有更高的峰值信噪比和结构相似度。

关键词: 超分辨率重建, 卷积神经网络, 跨尺度, 多分支

Abstract: Super-resolution reconstruction algorithm is one of the key research problems in the field of computer vision. At present, the image requirements in various fields are gradually increasing, and the existing super-resolution reconstruction algorithms improve the image quality by deepening a single network depth, ignoring the computational complexity problem when reconstructing, and lacking the use of image own information to improve the image reconstruction effect. To address the above problems, a cross-scale multi-branching single-frame image supe-resolution reconstruction network is proposed. The cross-scale module explores the correlation of similar regions within the image and is used to enhance the ability to reconstruct image detail information and obtain better visual effects. The multi-branching structure reconstructs regions in the image with different reconstruction difficulty by networks of different complexity separately, solving the problem of high computational complexity of deep networks. Experimental results show that the model has higher peak signal-to-noise ratio and structural similarity on Urban100 and Manga109 test sets compared with other methods.

Key words: super-resolution reconstruction, convolutional neural network, cross-scale, multi-branching