杨爱萍,张 兵,王 前,李晓晓,何宇清
AuthorsHTML:杨爱萍，张 兵，王 前，李晓晓，何宇清
AuthorsListE:Yang Aiping,Zhang Bing,Wang Qian,Li Xiaoxiao,He Yuqing
AuthorsHTMLE:Yang Aiping，Zhang Bing，Wang Qian，Li Xiaoxiao，He Yuqing
Unit:天津大学电气自动化与信息工程学院，天津 300072
Unit_EngLish:School of Electrical and Information Engineering，Tianjin University，Tianjin 300072，China
Abstract_Chinese:基于深度学习的方法在去模糊领域已经取得了很大的进展，但目前的 CNN 结构通常会对特征图进行下采 样，导致图像特征在网络学习过程丢失．针对这些问题，本文设计了一种基于高分辨率特征保持的图像去模糊网 络，与现有的串行连接由高-低分辨率图像特征图不同，所设计网络并行连接由高至低各分辨率特征子网络，而无需 经过由低分辨率到高分辨率的重建过程来恢复高分辨率信息，可以在整个去模糊过程中较好地保持图像的高分辨率 信息不丢失．另外，本文通过交叉连接各并行子网络来融合不同分辨率特征图，从而使得网络可以自适应选择模糊 特征来复原图像．最后，使用动态卷积核网络以非线性方法聚合多个卷积核信息提取图像特征，自适应调节感受野 大小，提高网络的泛化能力．与已有方法相比，本文使用的高分辨率特征保持网络去模糊较为彻底，能更好地保持 图像的特征信息，复原图像细节更为清晰．
Abstract_English:Recent researches have revealed the potential of using deep learning algorithm to performing single image deblurring，but existing learning-based CNN deblurring network has the down-sampling feature extraction module， which results in the detail looseness of image feature. In this paper，we propose an image deblurring architecture with high-resolution preservation to overcome these problems. Different from the previous networks in connecting high-to\u0002low resolution feature map in serial manner，our approach connects the high-to-low resolution subnet work in parallel manner without the low-to-high resolution reconstruction operation. So，our deblurring architecture can preserve the high-resolution information more efficiently. Further，we perform cross-connection among the paralleled subnets to fuse variant resolution feature maps so that they are chosen adaptively for image deblurring. Furthermore，the dy\u0002namic convolution kernel convolutional network is performed to extract the image feature non-linearly from several convolution kernels for adjusting the receptive field adaptively. When compared to the existing methods，the pro\u0002posed deblurring framework can maintain feature maps and recover a high quality deblurring image with vivid details.
Keyword_Chinese:图像去模糊；高分辨率保持；自适应；动态卷积核网络
Keywords_English:image deblurring；high-resolution preservation；adaptively；dynamic convolution kernel convolutional network

PDF全文下载地址:http://xbzrb.tju.edu.cn/#/digest?ArticleID=6609