作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2022, Vol. 48 ›› Issue (6): 228-234. doi: 10.19678/j.issn.1000-3428.0061892

• 图形图像处理 • 上一篇    下一篇

基于深度可分离卷积的轻量级图像超分辨率重建

柳聪, 屈丹, 司念文, 魏紫薇   

  1. 中国人民解放军战略支援部队信息工程大学 信息系统工程学院, 郑州 450000
  • 收稿日期:2021-06-09 修回日期:2021-07-14 发布日期:2022-06-11
  • 作者简介:柳聪(1997—),男,硕士研究生,主研方向为智能信息处理、超分辨率重建;屈丹(通信作者),教授、博士;司念文,博士;魏紫薇,硕士研究生。
  • 基金资助:
    国家自然科学基金(62171470,61673395);郑州市重大科技攻关项目(188PCXZX773)。

Lightweight Image Super-Resolution Reconstruction Based on Depthwise Separable Convolution

LIU Cong, QU Dan, SI Nianwen, WEI Ziwei   

  1. College of Information Systems Engineering, PLA Strategic Support Force Information Engineering University, Zhengzhou 450000, China
  • Received:2021-06-09 Revised:2021-07-14 Published:2022-06-11

摘要: 图像超分辨率重建旨在依据低分辨率图像重建出接近真实的高分辨率图像,现有基于卷积神经网络的图像超分辨率重建方法存在网络参数量大、重建速度慢等问题,从而限制其在内存资源小的终端设备上的应用。提出一种基于深度可分离卷积的轻量级图像超分辨率重建网络,利用深度可分离卷积提取图像的特征信息,减少网络的参数量,采用对比度感知通道注意力机制获取图像的对比度信息,并将其作为全局信息,同时对提取特征的不同通道权重进行重新分配,增强重建图像的细节纹理信息。在此基础上,采用亚像素卷积对图像特征进行上采样操作,提高整体重建图像质量。实验结果表明,当放大倍数为2、3和4时,该网络的参数量分别为140 000、147 000和152 000,重建时间为0.020 s、0.014 s和0.011 s,相比VDSR、RFDN、IDN等网络,在保证重建效果的前提下能够有效减少网络参数量。

关键词: 卷积神经网络, 图像超分辨率重建, 轻量级网络, 深度可分离卷积, 注意力机制

Abstract: Image super-resolution reconstruction aims to reconstruct a high-resolution image close to the real image according to the low-resolution image.The existing image super-resolution reconstruction methods based on the Convolutional Neural Network(CNN) usually have problems such as large network parameters and slow reconstruction speeds, which have limited their application in terminal devices with small memory resources.This paper proposes a lightweight image super-resolution reconstruction network based on depthwise separable convolution.The depthwise separable convolution is used to extract the feature information of the image and reduce network parameters.The contrast information of the image is obtained using the contrast perception channel attention mechanism, which is used as the global information.At the same time, the weights of different channels of the extracted features are reassigned to enhance the detailed texture information of the reconstructed image.On this basis, sub-pixel convolution is used to upsample the image features and improve the overall reconstructed image quality.The experimental results show that when the magnification is 2, 3, and 4, the network parameters are 140 000, 147 000, and 152 000, respectively, and the reconstruction time is 0.020 s, 0.014 s, and 0.011 s.Compared with VDSR, RFDN, IDN, and other networks, it can effectively reduce the amount of network parameters on the premise of ensuring the reconstruction effect.

Key words: Convolutional Neural Network(CNN), image super-resolution reconstruction, lightweight network, depthwise separable convolution, attention mechanism

中图分类号: