Journal of Frontiers of Computer Science and Technology ›› 2020, Vol. 14 ›› Issue (8): 1368-1379.DOI: 10.3778/j.issn.1673-9418.1909006

Previous Articles     Next Articles

Image Super-resolution Reconstruction of Functional Composite Deep Network

TANG Jiajun, LIU Hui, HU Xueying   

  1. 1. Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming 650000, China
    2. Yunnan Observatory, Chinese Academy of Sciences, Kunming 650000, China
    3. School of Computer Science, Henan Polytechnic University, Jiaozuo, Henan 454150, China
  • Online:2020-08-01 Published:2020-08-07

功能型复合深度网络的图像超分辨率重建

唐家军刘辉胡雪影   

  1. 1. 昆明理工大学 信息工程与自动化学院,昆明 650000
    2. 中国科学院 云南天文台,昆明 650000
    3. 河南理工大学 计算机学院,河南 焦作 454150

Abstract:

Aiming at the problems of weak inter-layer connections, single network concerns and insufficient utilization of layered features in the existing single-image super-resolution reconstruction using simple chain-stacked single network, a composite depth neural network is proposed to improve the performance of image super-resolution reconstruction. Firstly, feature extraction layer is used to extract the initial features of low-resolution images. Then, the initial features are fed into two sub-networks, one is responsible for image detail extraction and calculation, the other is responsible for image noise degradation and elimination. Then the deep abstract features output by the two sub-networks are combined with the initial features, and finally the super-resolution image is obtained through the reconstruction layer. Taking peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) as evaluation indices, the experiment is carried out by using amplification factor 3 in Set14 test set. The experimental results show that PSNR is increased by 2.27 dB, 0.66 dB, 0.54 dB, 0.05 dB and 0.21 dB respectively, while SSIM is improved by 6.08, 1.54, 1.41, 0.36, 0.09 percentage points by comparing the combined network with algorithms Bicubic, SelfEx, SRCNN, VDSR and RED.

Key words: single image super-resolution reconstruction, convolutional neural network (CNN), composite network, sub-network, feature combination

摘要:

针对现有单图像超分辨率重建时主要采用的简单链式堆叠的单一网络存在层间联系弱、网络关注点单一以及分层特征不能充分利用等问题,提出了一种复合的深度神经网络用于提升图像超分辨重建性能。该方法首先使用特征提取层提取低分辨率图像的初始特征;再将初始特征分别送入两个子网络,一个子网络负责图像细节的提取与运算,另一子网络负责图像噪声降解与消除;然后将两个子网络输出的深层次抽象特征与初始特征相结合,最后通过重建层获得超分辨率图像。以峰值信噪比(PSNR)与结构相似性(SSIM)为评价指标,在Set14测试集上使用放大因子3进行实验,将复合网络与算法Bicubic、SelfEx、SRCNN、VDSR和RED等进行对比,实验结果发现,PSNR分别提高了2.27 dB、0.66 dB、0.54 dB、0.05 dB、0.21 dB,而SSIM则分别提高了6.08、1.54、1.41、0.36、0.09个百分点。

关键词: 单图像超分辨率重建, 卷积神经网络(CNN), 复合网络, 子网络, 特征结合