[author_cn_name].[cn_title][J].空军工程大学学报:自然科学版,[year_id],[volume]([issue]):[start_page]-[end_page] 基于并行残差卷积网络的图像超分辨重建-Image Super Resolution in Combination with Convolution Neural Network
文章摘要
杨伟铭,张钰.基于并行残差卷积网络的图像超分辨重建[J].空军工程大学学报:自然科学版,2019,20(4):84-89
基于并行残差卷积网络的图像超分辨重建
Image Super Resolution in Combination with Convolution Neural Network
  
DOI:
中文关键词: 卷积神经网络  图像超分辨率  扩张神经网络  跳跃链接  深度学习
英文关键词: convolution neural network  image super resolution  dilated convolutional network  skip connection  deep learning
基金项目:国家自然科学基金(41671409)
作者单位
杨伟铭,张钰 1.军事科学院系统工程研究院后勤科学与技术研究所,北京,100800
2.陕西师范大学计算机科学学院西安710119 
摘要点击次数: 9
全文下载次数: 10
中文摘要:
      针对VDSR模型卷积核单一和DRRN模型不能全局利用的问题,提出了基于并行残差卷积神经网络的联合卷积图像超分辨重建模型。模型首先利用原始卷积层和扩张卷积层融合,建立联合卷积层,然后利用跳跃链接,将多种抽象层次的特征进行融合,最后完成整个超分辨网络的模型构建。提出的模型具有以下优点:①扩张卷积神经网络与原始卷积神经网络融合,在计算机复杂度不变的情况下,可以获取更多尺度的信息,因此具有更强的表达能力;②跳跃链接方式,将抽象层度较低与较高抽象层次的信息融合,获取更多的信息,使得模型具有更强的学习能力。通过在多个数据集上进行实验,模型在大多数任务中与VDSR、DRRN和SRCNN等先进模型相比,IFC值取得了大于0.1的提升。
英文摘要:
      Aimed at the problems that the VDSR model convolution kernel is single and the DRRN model fails to take advantages of global features, a combined convolution image super resolution model is proposed based on parallel residual convolution neural networks. Firstly, the combined convolution neural layer is structured by the original convolution layer and dilated convolution layer, and the skip connection approach is employed to connect the different layers to take advantage of different level features, completing super resolution network. There are two advantages of this model:①Combination of dilated convolution neural layers and original convolution layers can capture multi scale features without computation consuming. Based on this approach, the network can get more presentation capacity. ②Skip connection approach fuse low level information and high level information. From this approach, different level features can be learned. This means that stronger learning ability can be obtained. Based on the experiment results on multiple data sets, more than 0.1 IFC improvement is achieved, compared with the state of the art models VDSR, DRRN, SRCNN in most tasks.
查看全文   查看/发表评论  下载PDF阅读器
关闭