摘要
针对传统单幅图像超分辨率深度学习方法将不同尺度低分辨率视作独立任务的问题,提出一种以残差通道注意力模块作为特征提取,元上采样模块作为放大模块的超分辨率网络。残差通道注意力机制可以滤除冗余低频信息减少网络深度,使元上采样模块更好地训练不同尺度低分辨率图像特征间的关系,实现任意尺度的超分辨率网络。为了验证该方法有效性,在Set5、Set14、Urban100等公共数据集上实验。实验结果表明,该方法在整数与非整数倍尺度都能很好地恢复高分辨率图像。
Aiming at the problem that the traditional single image super-resolution deep learning methods take low-resolution images of different scales as independent tasks,a super-resolution network based on the meta-upsacle module and residual channel attention module is proposed.The residual channel attention module can filter out the redundant low-frequency information to reduce the network depth and make the meta upscale module train better on the relationship between the features of different scale low resolution image,then a super-resolution network of arbitrary scale can be realized.In order to verify the effectiveness of the method,experiments were made on several public datasets such as Set5,Set14 and Urban100.The results show that this method can recover high-resolution images well in both integer and non-integer sacle.
作者
应凯杰
冯玉田
Ying Kaijie;Feng Yutian(School of Communication and Information Engineering,Shanghai University,Shanghai 200444,China)
出处
《计算机应用与软件》
北大核心
2022年第2期234-238,265,共6页
Computer Applications and Software
关键词
图像超分辨率
深度学习
元上采样
任意尺度
Image super-resolution
Deep learning
Meta upscale
Arbitrary scale
