西安电子科技大学学报 ›› 2019, Vol. 46 ›› Issue (5): 15-23.doi: 10.19665/j.issn1001-2400.2019.05.003

• • 上一篇    下一篇

一种对称残差CNN的图像超分辨率重建方法

刘树东,王晓敏,张艳()   

  1. 天津城建大学 计算机与信息工程学院, 天津 300384
  • 收稿日期:2019-01-16 出版日期:2019-10-20 发布日期:2019-10-30
  • 通讯作者: 张艳
  • 作者简介:刘树东(1965—),男,教授,E-mail:liushudong@tcu.edu.cn.
  • 基金资助:
    天津市企业科技特派员项目(18JCTPJC60000)

Symmetric residual convolution neural networks for the image super-resolution reconstruction

LIU Shudong,WANG Xiaomin,ZHANG Yan()   

  1. School of Computer and Information Engineering,Tianjin Chengjian University, Tianjin 300384, China
  • Received:2019-01-16 Online:2019-10-20 Published:2019-10-30
  • Contact: Yan ZHANG

摘要:

基于卷积神经网络的图像超分辨率重建方法具有很高的重建性能。但该类方法存在网络参数多、训练难度大,梯度消失和网络退化等问题。针对这些问题,提出一种基于对称残差卷积神经网络的图像超分辨率重建方法。通过将对称融入到残差块中,采用对称连接实现局部特征融合,提取尽可能多的有价值特征;残差块外采用跳跃连接实现全局特征融合,以提高图像的重建质量。该方法使用峰值信噪比和结构相似度作为评价指标,在Set5、Set14和BSD100标准数据集上进行2倍、3倍和4倍因子重建后的结果大部分优于比较方法,平均峰值信噪比和结构相似度值较比较方法均有提高。实验结果表明,该方法重建的图像纹理更清晰,细节更丰富,具有较好的主观视觉效果。

关键词: 超分辨率重建, 卷积神经网络, 深度学习, 对称残差网络

Abstract:

The image super-resolution reconstruction methods based on the convolutional neural network have high reconstruction performance, but they have many network parameters and are difficult to train. Also, they are prone to problems such as gradient disappearance and network degradation. To solve these problems, a super-resolution image reconstruction method based on the symmetric residual convolution neural network is proposed. This method integrates symmetry into residual blocks. It realizes local feature fusion by adopting the symmetric connection and extracts as many valuable features as possible. To improve the quality of image reconstruction, the global feature fusion is realized by skip connection. In this method, the peak signal to noise ratio (PSNR) and structural similarity (SSIM) are used as evaluation indexes. The results by the proposed reconstruction method on Set5, Set14, and BSD100 are superior to those by most of other methods in comparison. The average PSNR and SSIM values are improved compared with those methods. Experimental results show that the image reconstructed by this method has clearer textures, richer details and a better subjective visual effect.

Key words: super-resolution reconstruction, convolution neural network, deep learning, symmetric residual network

中图分类号: 

  • TP391