摘要
结合深度学习,提出一种多尺度全卷积神经网络驱动的遥感影像修复方法。构建高斯金字塔实现多尺度变换,影像通过FNEA（fractal net evolution approach）算法进行分割,利用卷积层之间的激活函数值来寻找最佳匹配像素,最后通过全卷积神经网络反向传播函数对修补区域进行像素填充。多组实验验证了多尺度全卷积神经网络方法可以良好的修复遥感影像,有自动化高、效率高、目视修复效果佳的优势。
Combined with deep learning,a multi-scale and full-convolution neural network-driven remote sensing image restoration method is proposed.The Gaussian pyramid was constructed for multi-scale transformation.The image was segmented by FNEA（fractal net evolution approach）algorithm.The optimal matching pixels are found by using the activation function values between the convolution layers,and finally the pixels are filled in the repaired regions by the full convolution neural network back propagation function.Several groups of experiments have verified that the multiscale full-convolution neural network method can well repair the remote sensing images,which has the advantages of high automation,high efficiency and good visual restoration.
作者
石斌斌
何海清
游琦
SHI Binbin;HE Haiqing;YOU Qi(Faculty of Geomatics ,East China University of Technolog, Nangchang 330013, China;Fuzhou Investieation and Surveying Institute,Fuzhou 350000, China)
出处
《测绘地理信息》
2018年第3期124-126,共3页
Journal of Geomatics
基金
国家自然科学基金资助项目(41401526)
测绘遥感信息工程国家重点实验室开放研究基金资助项目((13)重04)
关键词
多尺度
全卷积神经网络
高斯金字塔
FNEA分割
反向传播函数
multi-scale
full convolution neural network
gaussian pyramid
FNEA segmentation
reverse broadcast function
