熊承义,李世宇,高志荣,金鑫.级联模型展开与残差学习的压缩感知重构[J].中南民族大学学报自然科学版,2019,(2):265-272
级联模型展开与残差学习的压缩感知重构
Compressive sensing reconstruction via stacked unfolding model and residual learning
  
DOI:10.12130/znmdzk.20190221
中文关键词: 压缩感知  深度学习  模型展开  残差学习
英文关键词: compressive sensing  deep learning  model unfolding  residual learning
基金项目:国家自然科学基金资助项目(61471400),中央高校基本科研业务经费专项资金项目(CZY19016)
作者单位
熊承义1,李世宇1,高志荣2,金鑫1 1 中南民族大学 电子信息工程学院武汉430074 2中南民族大学 计算机科学学院武汉 430074 
摘要点击次数: 261
全文下载次数: 251
中文摘要:
      基于传统优化模型展开的深度网络由于集成了深度学习与传统优化方法的优点,具有良好的可解释性,在当前图像处理与计算机视觉领域得到广泛关注。提出了一种级联模型展开与残差学习的图像压缩感知重构深度网络框架,以实现重构图像质量的进一步改善。第一级的基于模型展开的深度网络根据输入的压缩测量值得到初始的重构图像,第二级的深度残差网络对初始重构图像进行去噪处理,最终得到高质量的重构结果。该两级级联网络的训练分别独立完成,训练过程简单易实现,将ADMM-Net与ResNet级联实现对磁共振图像重构,将ISTA-Net+与ResNet级联实现对自然图像重构。大量实验结果比较验证了所提出方法的有效性。
英文摘要:
      Deep networks based on unfolding conventional optimization model have been paid widely attention in many fields including image processing, computer vision and so on, because they not only combine the advantages of current deep leaning and conventional optimization-based approach, but also character well interpretability. A novel deep network architecture for compressive sensing image reconstruction is proposed by cascading model unfolding and residual learning, which aims to further improving the reconstructed image quality. The first stage of deep network is designed based on model unfolding to transform the compressed measurements of input into the initial reconstruction, and the second stage is a deep residual network to remove the noise in the initial reconstruction, consequently producing higher quality of reconstruction image. The training of the two-stage network is completed independently, which is simple and easy to conduct. Specifically, stacking the ADMM-Net and ResNet to reconstruct magnetic resonance imaging, and stacking the ISTA-Net+ and ResNet to reconstruct natural images. Extensive experimental results comparison demonstrates the effectiveness of the proposed method.
查看全文   查看/发表评论  下载PDF阅读器
关闭