摘要
由于现有的基于卷积神经网络的超分辨率重建模型无法从多个尺度上利用特征进行加权,导致对于高频细节的恢复能力较差;同时随着网络深度的不断增加导致浅层信息在传输过程中容易弱化或丢失。为此,提出一种基于特征融合和混合注意力的超分辨方法。利用不同尺度间特征的相似性这一特点设计一种多尺度注意力机制,捕获特征图之间和特征图内部的依赖关系;与空间注意力相结合,捕获空间域内隐含的高频信息。设计一种层次特征融合结构对提取出的特征进行充分保留。实验结果表明,该算法在恢复高频细节方面表现良好。
The existing super-resolution reconstruction model based on convolutional neural networks cannot be weighted with features from multiple scales,resulting in poor recovery of high-frequency details.At the same time,with the continuous increase of network depth,shallow information is easily weakened or lost during transmission.To this end,a super-resolution method based on feature fusion and hybrid attention was proposed.Taking advantage of the similarity of features between diffe-rent scales,a multi-scale attention mechanism was designed to capture the dependencies between feature maps and that within feature maps.It is combined with spatial attention to capture high-frequency information implicit within the spatial domain.A hierarchical feature fusion structure was designed to fully preserve the extracted features.Experimental results show that the proposed algorithm performs well in recovering high-frequency details.
作者
左云瑞
陈东方
王晓峰
ZUO Yun-rui;CHEN Dong-fang;WANG Xiao-feng(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China;Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System,Wuhan University of Science and Technology,Wuhan 430065,China)
出处
《计算机工程与设计》
北大核心
2023年第11期3387-3394,共8页
Computer Engineering and Design
基金
国家自然科学基金项目(61572381、61273225)。
关键词
卷积神经网络
特征相似性
多尺度注意力
空间注意力
特征融合
跳跃链接
图像超分辨率
convolutional neural networks
feature similarity
multi-scale attention
spatial attention
feature fusion
skip connection
image super resolution
