基于密集卷积神经网络的遥感影像超分辨率重建

Super Resolution Reconstruction of Remote Sensing Images Based on Dense Convolution Neural Network

针对传统遥感影像超分辨率重建方法依赖同一场景多时相图像序列且需预先配准等缺点,本文提出了一种基于密集卷积神经网络的遥感影像超分辨率重建的方法。该网络直接将低分辨率遥感影像作为网络的初始输入,通过密集卷积神经网络学习影像的高阶表示,获得更具有表达能力的深层特征;同时,在网络中采用并行的1×1卷积滤波器结构,通过该结构减少模型参数;在重建网络中使用亚像素卷积可以更快地实现特征图的重建。在UCMerced_LandUse公共数据集上的实验表明:本文的网络模型提升了传统深度网络的影像重建性能,增强了重建图像的纹理细节并改善影像边缘失真,提升了重建影像的性能。

Regarding the disadvantages of traditional remote sensing image super-resolution reconstruction methods,such as relying on multi-temporal image sequences of the same scene and requiring pre-registration,this paper proposes a method of remote sensing image super-resolution reconstruction based on DenseNet of dense convolution neural network.In this network,low-resolution remote sensing images are directly taken as the initial input of the network,and DenseNet is used to learn the higher-order representation of images to obtain deeper features with more expressive ability.Meanwhile,a parallel 1×1 convolutional filter structure is adopted in the network,which reduces model parameters and avoids over-fitting.Using sub-pixel convolution in the reconstruction network can realize the reconstruction of feature graph more quickly.Experiments on the public data sets of UCMerced_LandUse indicate that the network model of this paper improves the image reconstruction performance of the traditional deep network,enhances the texture details of the reconstructed image,reduces the image edge distortion,and improves the performance of the reconstructed image.