摘要
针对遥感图像细节信息丰富、内容复杂、部分特征信息重建丢失以及现有遥感图像重建模型对感受野尺度关注不足带来的问题,本文提出一种融合多路径与混合注意力机制的算法用于遥感图像的超分辨率重建.该算法首先利用浅层特征提取模块提取浅层特征图;然后通过多层级特征提取模块进行深层特征的自适应学习;接着使用全局特征融合模块进行多层级特征融合和全局残差学习;最后经上采样模块和重建模块得到高分辨率图像.算法重点关注各层级特征提取和全局特征的融合,并通过横向扩展网络以便于利用不同的卷积核来获得不同感受野,同时结合混合注意力机制来关注高频信息,最终取得不错的重建效果.实验表明,该算法在峰值信噪比和结构相似度两个评价指标上相对于主流算法均有提升,视觉效果上也能够恢复更多的细节信息.
Aiming at the problems caused by rich detail information,complex content,loss of partial feature information and insufficient attention to receptive field scale in existing remote sensing image reconstruction models,an algorithm integrating multi-path and mixed attention mechanism is proposed for super-resolution reconstruction of remote sensing images.Firstly,the shallow feature extraction module is used to extract the shallow feature map;Then,the multi-level feature extraction module is used for adaptive learning of deep features;Then,the global feature fusion module is used for multi-level feature fusion and global residual learning;Finally,the high-resolution image is obtained by up sampling module and reconstruction module.The algorithm focuses on the feature extraction at all levels and the fusion of global features,and expands the network horizontally to obtain different receptive fields by using different convolution kernels.At the same time,it focuses on high-frequency information by combining the mixed attention mechanism,and finally achieves good reconstruction results.Experiments show that the algorithm improves the two evaluation indexes of peak signal-to-noise ratio and structural similarity compared with the mainstream algorithms,and can recover more detailed information in visual effect.
作者
李千
赵逢禹
LI Qian;ZHAO Feng-yu(School of Optical-Electrical&Computer Engineering,University of Shanghai for Science&Technology,Shanghai 200093,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2023年第7期1508-1513,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61602305)资助
上海市自然科学基金项目(19ZR1477600)资助。
关键词
遥感图像
超分辨率重建
特征融合
注意力机制
卷积神经网络
remote sensing image
super-resolution reconstruction
feature fusion
attention
convolutional neural network
