面向超分辨率重建的层次间局部特征增强网络
Hierarchical local feature enhancement network for super-resolution reconstruction
王晓峰 1黄煜婷 1张文尉 1张轩 1陈东方1
作者信息
- 1. 武汉科技大学计算机科学与技术学院,湖北武汉 430070;武汉科技大学智能信息处理与实时工业系统湖北省重点实验室,湖北武汉 430070
- 折叠
摘要
基于卷积神经网络的超分辨率重建模型以单项传播为主,层次越靠后感知信息的能力越微弱,导致层次间局部特征部分丢失,难以实质提升网络的特征表达能力.针对此问题,提出层次间局部特征增强网络.该方法由级联残差模块、层次间特征增强块和特征感知注意力机制组成.级联残差模块通过有效残差连接增加对残差分支信息的利用;层次间特征增强块提取不同深度特征的依赖关系,自适应调整中间层特征权值增强捕获关键信息的能力;特征感知注意力机制采用方向感知和位置判断的方式准确定位和识别感兴趣对象.多项标准数据集的实验结果表明,该方法能改善超分辨率的视觉重建效果,整体性能优于现有方法.
Abstract
The super-resolution reconstruction model based on convolutional neural network is mainly based on single propaga-tion.In addition,the ability to perceive information is weaker than the further back in the level,which leads to the partial loss of local features between levels,and it is difficult to substantially improve the feature expression ability of the network.To solve this problem,the hierarchical local feature enhancement network for single image super resolution was proposed.The method consisted of a cascading residual module,an inter-level feature enhancement block,and a feature-aware attention mechanism.The cascaded residuals module increased the utilization of residual branch information through effective residual connections.The inter-level feature enhancement block extracted the dependencies of features of different depths.Besides,the middle layer feature weights were adaptively adjusted to enhance the ability to capture key information.The feature-aware attention mechanism accu-rately located and identified objects of interest by means of direction perception and position judgment.Experimental results of a number of standard datasets show that this method can improve the visual reconstruction effect of super-resolution,and the overall performance is better than that of the existing methods.
关键词
卷积神经网络/超分辨率/局部特征增强/级联残差模块/注意力机制/方向感知/位置判断Key words
convolutional neural network/super resolution/local feature enhancement block/cascade residual module/attention mechanism/direction perception/position judgement引用本文复制引用
基金项目
智能信息处理与实时工业系统湖北省重点实验室开放研究课题基金项目(ZNXX2022008)
出版年
2024
NETL
NSTL
万方数据