摘要
针对现有卷积神经网络在超分辨率重建的图像上存在部分细节特征不够突出、边缘模糊等问题,在现有模型三大模块的基础上对映射模块及损失函数进行细致剖析,提出了一种多映射卷积神经网络的超分辨率重建算法.该算法通过构建多映射网络,极大地丰富了图像在聚合高分辨图像时的特征维度.同时在重建模块的卷积层后引入全变分正则项,结合误差反向传播算法,可有效地对解空间约束,从而提取出精确、有效的特征,丰富重建图像的细节信息.在常用数据集上的实验结果表明,该算法生成的网络模型获得了更好的超分辨率结果,主观视觉评价和客观衡量指标有一定的改进,有效地提高了图像的分辨率.
The traditional convolutional neural network for super-resolution will obtain abundant details and edge information with difficulty.By the analysis of the detailed characteristics in three modules of conventional methods,we propose a new multi-mapping convolutional neural network model.By the multimapping module,rich and varied characteristics from each layer can be captured.Combining with the error back propagation algorithm,a novel loss function with total variation regularization is used to train and seek optimal parameters,which reconstruct accurate and effective high-resolution images from the network.Extensive quantitative and qualitative evaluations have shown that the proposed algorithm improves effectively the resolution of the image.
作者
王世平
毕笃彦
刘坤
何林远
WANG Shiping;BI Duyan;LIU Kun;HE Linyuan(Institute of Aeronautics and Astronautics Engineering Air Force Engineering Univ.,Xi＇an 710038,China)
出处
《西安电子科技大学学报》
EI
CAS
CSCD
北大核心
2018年第4期155-160,共6页
Journal of Xidian University
基金
国家自然科学基金资助项目(61701524
61372167)
关键词
图像处理
超分辨率
多映射卷积神经网络
变分约束
image processing
super-resolution
multi-mapping convolutional network
constrainedvariation method
