计算机工程与应用 ›› 2021, Vol. 57 ›› Issue (16): 213-219.DOI: 10.3778/j.issn.1002-8331.2005-0203

• 图形图像处理 • 上一篇    下一篇

图像超分辨率特征浓缩网络

刘星辰,贾俊铖,张莉,胡沁涵   

  1. 苏州大学 计算机科学与技术学院,江苏 苏州 215006
  • 出版日期:2021-08-15 发布日期:2021-08-16

Feature Concentration Network for Image Super-Resolution

LIU Xingchen, JIA Juncheng, ZHANG Li, HU Qinhan   

  1. School of Computer Science and Technology, Soochow University, Suzhou, Jiangsu 215006, China
  • Online:2021-08-15 Published:2021-08-16

摘要:

自从卷积神经网络应用到图像超分辨率领域以来,越来越多的神经网络被提出,并且取得良好的效果,但是当前大多数方法都严重依赖于模型的深度和宽度,而没有充分利用底层信息。针对以上问题,提出了一种新型特征浓缩网络,该网络通过多个特征浓缩块逐步提取有效特征信息。网络包括特征提取模块、特征浓缩模块和重建模块,并添加了双三次插值运算和全局残差学习。通过底层特征处理来提取有用的特征,使用特征浓缩块进一步提取特征,由重建模块恢复高分辨率图像。在实验中,选择4个不同的公开数据集进行不同尺度的测试,通过实验结果可以看出,所提出的网络对比其他方法有更好的客观指标结果。

关键词: 卷积神经网络, 超分辨率, 特征提取, 全局残差学习

Abstract:

Since the application of convolutional neural networks to the field of image super-resolution, more and more neural networks have been proposed and achieved good results. However, most current methods rely heavily on the depth and width of the model without fully utilizing the underlying information. Aiming at the above problems, a new feature enrichment network is proposed, which can gradually extract effective feature information through multiple feature concentration blocks. The network includes a feature extraction module, a feature concentration module and a reconstruction module, and bicubic interpolation operations and global residual learning are added. Firstly, useful features are extracted through the underlying feature processing, then feature concentration blocks are used to further extract features, finally a high-resolution image is reconstructed through a reconstruction module. In the experiment, 4 different public datasets are selected for testing at different scales. It can be seen from the experimental results that the proposed network has better objective index results than other methods.

Key words: convolutional neural network, super-resolution, feature extraction, global residual learning