摘要
压缩感知图像重构要求从远低于奈奎斯特采样率的数据中恢复高分辨率图像,在光学和雷达系统、医学成像、射电天文学等成像领域具有广泛需求。深度学习网络能够自动提取数据的高层次特征,近年来在图像重构领域得到广泛应用。为进一步提高图像重构性能,提出了一种基于高倍特征残差网络结构的重构模型。该模型中采样部分应用非重叠卷积核卷积采样,重构部分在保证网络参数量合理的同时,设计了特征图倍增的残差网络结构。训练模型时采用l1范数和结构相似度指标组合的损失函数。实验结果表明,基于本模型的图像重构质量优于其它重构模型。
Compressed sensing image reconstruction needs to recover high resolution images from far fewer samples than is possible using Nyquist sampling rate,which is essential for optical and radar systems,medical imaging,radio astronomy and other imaging fields.Deep learning networks automatically extract high-level features from data and performs well in the field of image reconstruction.In order to further improve image reconstruction performance,a reconstruction model based on wider feature residual network is proposed.The sampling part of the model uses non-overlapping convolution kernel for convolution sampling,and the reconstruction part uses the residual network units with the feature map multiplication while ensuring the reasonable number of network parameters.When training the model,the loss function combined by the l1 norm and the structural similarity index.Experimental results show that the image reconstruction quality based on this model is better than other reconstruction models.
作者
涂云轩
冯玉田
应凯杰
高萌
Tu Yunxuan;Feng Yutian;Ying Kaijie;Gao Meng(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)
出处
《电子测量技术》
2020年第7期113-118,共6页
Electronic Measurement Technology
关键词
图像重构
残差网络
卷积核
压缩采样
image reconstruction
residual network
convolution kernel
compression sampling