计算机应用 ›› 2020, Vol. 40 ›› Issue (7): 2069-2076.DOI: 10.11772/j.issn.1001-9081.2019122149

• 虚拟现实与多媒体计算 • 上一篇    下一篇

基于混合深度卷积网络的图像超分辨率重建

胡雪影1,2, 郭海儒1, 朱蓉2   

  1. 1. 河南理工大学 计算机科学与技术学院, 河南 焦作 454000;
    2. 嘉兴学院 数理与信息工程学院, 浙江 嘉兴 314000
  • 收稿日期:2019-12-24 修回日期:2020-03-09 出版日期:2020-07-10 发布日期:2020-03-30
  • 通讯作者: 朱蓉
  • 作者简介:胡雪影(1993-),女,河南信阳人,硕士研究生,主要研究方向:智能信息处理、图像处理;郭海儒(1977-),男,山西朔州人,副教授,博士,CCF会员,主要研究方向:智能信息处理、图像处理;朱蓉(1973-),女,浙江嘉兴人,教授,博士,CCF会员,主要研究方向:智能信息处理、机器学习、数据挖掘。
  • 基金资助:
    浙江省重点研发计划项目(2019C03099);浙江省自然科学基金资助项目(LY19F020017)

Image super-resolution reconstruction based on hybrid deep convolutional network

HU Xueying1,2, GUO Hairu1, ZHU Rong2   

  1. 1. College of Computer Science and Technology, Henan Polytechnic University, Jiaozuo Henan 454000, China;
    2. College of Mathematics and Information Engineering, Jiaxing University, Jiaxing Zhejiang 314000, China
  • Received:2019-12-24 Revised:2020-03-09 Online:2020-07-10 Published:2020-03-30
  • Supported by:
    This work is partially supported by the Zhejiang Key Research and Development Program (2019C03099), the Zhejiang Natural Science Foundation (LY19F020017).

摘要: 针对传统图像超分辨率重建方法存在的重建图像模糊、噪声量大、视觉感差等问题,提出了一种基于混合深度卷积网络的图像超分辨率重建方法。首先,在上采样阶段将低分辨率图像放缩至指定大小;然后,在特征提取阶段提取低分辨率图像的初始特征;接着,将提取到的初始特征送入卷积编解码结构进行图像特征去噪;最后,在重建层用空洞卷积进行高维特征提取与运算,重建出高分辨率图像,并且使用残差学习快速优化网络,在降低噪声的同时,使重建图像的清晰度及视觉效果更优。在Set14数据集放大尺度×4的基准下,将所提方法与双三次插值(Bicubic)、锚定邻域回归(A+)、超分辨卷积神经网络(SRCNN)、极深度超分辨网络(VDSR)、编解码网络(REDNet)等超分辨率重建方法进行对比。在超分辨实验中,所提方法与对比方法比较,峰值信噪比(PSNR)分别提升了2.73 dB、1.41 dB、1.24 dB、0.72 dB和1.15 dB,结构相似性(SSIM)分别提高了0.067 3,0.020 9,0.019 7,0.002 6和0.004 6。实验结果表明,混合深度卷积网络能够有效地对图像进行超分辨率重建。

关键词: 图像超分辨率重建, 图像特征去噪, 混合深度卷积网络, 反卷积, 空洞卷积

Abstract: Aiming at the problems of blurred image, large noise, and poor visual perception in the traditional image super-resolution reconstruction methods, a method of image super-resolution reconstruction based on hybrid deep convolutional network was proposed. Firstly, the low-resolution image was scaled down to the specified size in the up-sampling phase. Secondly, the initial features of the low-resolution image were extracted in the feature extraction phase. Thirdly, the extracted initial features were sent to the convolutional coding and decoding structure for image feature denoising. Finally, high-dimensional feature extraction and computation were performed on the reconstruction layer using the dilated convolution in order to reconstruct the high-resolution image, and the residual learning was used to quickly optimize the network in order to reduce the noise and make the reconstructed image have better definition and visual effect. Based on the Set14 dataset and scale of 4x, the proposed method was compared with Bicubic interpolation (Bicubic), Anchored neighborhood regression (A+), Super-Resolution Convolutional Neural Network (SRCNN), Very Deep Super-Resolution network (VDSR), Restoration Encoder-Decoder Network (REDNet). In the super-resolution experiments, compared with the above methods, the proposed method has the Peak Signal-to-Noise Ratio (PSNR) increased by 2.73 dB,1.41 dB,1.24 dB,0.72 dB and 1.15 dB respectively, and the Structural SIMilarity (SSIM) improved by 0.067 3,0.020 9,0.019 7,0.002 6 and 0.004 6 respectively. The experimental results show that the hybrid deep convolutional network can effectively perform super-resolution reconstruction of the image.

Key words: image super-resolution reconstruction, image feature denoising, hybrid deep convolutional network, deconvolution, dilated convolution

中图分类号: