计算机工程与应用 ›› 2019, Vol. 55 ›› Issue (13): 186-192.DOI: 10.3778/j.issn.1002-8331.1804-0229

• 图形图像处理 • 上一篇    下一篇

基于复合卷积神经网络的图像超分辨率算法

吴嘉昕,胡晓辉,张  明   

  1. 兰州交通大学 电子与信息工程学院,兰州 730070
  • 出版日期:2019-07-01 发布日期:2019-07-01

Image Super-Resolution Algorithm Based on Composite Convolutional Neural Network

WU Jiaxin, HU Xiaohui, ZHANG Ming   

  1. School of Electronics & Information Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China
  • Online:2019-07-01 Published:2019-07-01

摘要: 针对卷积神经网络图像超分辨率算法中的映射函数容易出现过拟合、梯度弥散等问题,提出一种由卷积网络和反卷积网络构成的复合卷积神经网络算法。提出使用RReLUs和Softplus函数结合形式作为激活函数,有效改善了过拟合问题;采用附加修正系数的小批量梯度下降法,避免梯度弥散现象;利用反卷积网络实现高分辨率图像重建。实验证明新的网络模型有效改善了图像的清晰度和边缘锐化,在主观视觉效果和客观评价指标上都获得了显著提升。

关键词: 低分辨率, 超分辨率, 卷积层, 反卷积层

Abstract: An improved image super-resolution algorithm based on composite convolutional neural network which consists of convolution network and deconvolution network is proposed to overcome many problems, such as over fitting of mapping function and problem of gradient dispersion. Firstly, the paper proposes a combination of RReLus and Softplus functions to replace previous activation functions, avoiding over fitting problem. Secondly, the mini-batch gradient descent method with correction coefficients is introduced into the network to avoiding the problem of gradient dispersion. Thirdly, it uses the deconvolution network to reconstruct high resolution image. Finally, experiments prove that the improved network model which obtains effectively improves image sharpness and edge sharpening, and achieves significant improvement in subjective visual effect and objective evaluation index.

Key words: low resolution, super-resolution, convolution layer, deconvolution layer