改进的密集连接网络遥感图像超分辨重建  

An improved remote sensing images super-resolution method based on densely connected network

作　　者：柏宇阳 朱福珍[1] 巫红[1] Bai Yuyang;Zhu Fuzhen;Wu Hong(College of Electrical Engineering,Heilongjiang University,Harbin 150080)

机构地区：[1]黑龙江大学电子工程学院,哈尔滨150080

出　　处：《高技术通讯》2021年第10期1037-1043,共7页Chinese High Technology Letters

基　　金：国家自然科学基金(61601174);黑龙江省自然科学基金(F2018026);黑龙江省博士后科研启动金(LBH-Q17150);黑龙江省普通高等学校电子工程重点实验室开放课题及省高校科技创新团队(2012TD007)资助项目。

摘　　要：遥感图像超分辨增加了遥感图像的细节信息,在遥感图像处理中有重要的地位。为了进一步提高遥感图像超分辨的重建效果,本文提出一种改进的密集连接网络遥感图像超分辨重建算法。首先对基于残差网络的深度超分辨算法(VDSR)进行改进,结合密集连接网络(DenseNet),将残差网络中的残差块替换成密集块,并且添加一组密集层与瓶颈层,实现DenseNet网络结构的改进,同时,修改网络激活函数为PReLU函数,网络训练采用L1损失函数。为了使网络在遥感图像上具有更好的效果,训练网络时,数据集全部采用遥感图像作为训练样本。当训练的epoch达到了大约35次时网络已经收敛。实验结果表明,与VDSR算法相比,本文改进的算法对遥感图像的效果更优,峰值信噪比(PSNR)平均增加了1.05 dB,结构相似度(SSIM)平均增加了0.042。Remote sensing image super-resolution improves the detail information of remote sensing image,which is of great significance in remote sensing image processing.In order to further improve the super-resolution reconstructed effect of remote sensing image,an improved remote sensing image super-resolution algorithm based on densely con-nected network is proposed.First,the very deep super-resolution algorithm(VDSR)based on residual network is improved.Combined with densely connected network(DenseNet),the residual blocks in the residual network are replaced by densely blocks,and a group of densely layers and bottleneck layers are added to improve the DenseNet network structure.Second,the activation function is modified to PReLU function,and the L1 loss function is adopted in training.In order to make the super-resolution network have better effect on remote sensing images,re-mote sensing images are used as the training data set in training.The network converges when the epoch reaches about 35 times.Finally,the experimental results show that the effect of the improved algorithm is better than the effect of VDSR algorithm on remote sensing images.The PSNR increases by 1.05 dB and the SSIM increases by 0.042 on average.

关 键 词：遥感图像 超分辨率 密集连接网络(DenseNet) 深度学习 

分 类 号：TP751[自动化与计算机技术—检测技术与自动化装置]