摘要
目前在单帧图像超分辨率(SISR)研究领域中,一些深度网络在重构阶段通过简单级联、通道注意、空间注意等方式,利用中间特征来提高图像重构效果,但是它们通常只注意到其中一个方向。为此,文章研究了一种新的注意力,即基于空间特征变换(SFT)的空间通道注意力,并提出了基于SFT的空间通道注意力机制重构的渐进式网络算法。该算法多角度地利用中间特征进行图像重构,首先基于SFT提供更多的相似性特征,然后在重构时利用SFT空间通道注意力模块(SFTCA模块)提供通道贡献力度和空间依赖关系。实验结果表明,相对于大多数超分辨重建算法,该算法在图像超分辨重构时各评价指标均有较大提升,所重构的图像纹理信息更加清晰。
Presently,in the field of single-frame image super-resolution(SISR),some deep networks are used to improve the image reconstruction effect through some intermediate features,such as simple cascading,channel attention,spatial attention,etc.in the reconstruction stage.However,people usually only pay attention to one of the directions.In this paper,the spatial channel attention based on SFT,and a progressive network are proposed,which based on the reconstruction of the spatial channel attention mechanism of spatial feature transform(SFT).The network uses intermediate features for image reconstruction from multiple angles.Firstly,it provides more similarity features based on SFT during feature extraction,and then it uses SFT spatial channel attention module(SFTCA module)to provide channel contribution strength and spatial dependence for image reconstruction.The experimental results show that,compared with most super-resolution algorithms,the proposed method has greatly improved the evaluation indexes during image super-resolution reconstruction,and the texture information of the reconstructed image is also clearer.
作者
秦玉
谢超宇
王晓明
QIN Yu;XIE Chaoyu;WANG Xiaoming(School of Computer and Software Engineering,Xihua University,Chengdu 610039 China)
出处
《西华大学学报（自然科学版）》
CAS
2022年第2期39-50,共12页
Journal of Xihua University:Natural Science Edition
基金
国家自然科学基金资助项目(61602390)。
关键词
注意力机制
空间特征变换
渐进式上采样
超分辨率
深度学习
attention mechanism
spatial feature transform
progressive upsampling
super-resolution
deep learning
