多尺度残差深度神经网络的卫星图像超分辨率算法

Satellite Imagery Super-Resolution Algorithm via Multi-Scale Residual Deep Neural Network

卫星图像实现星际对地观测并被广泛的应用到了军事和经济生活领域。受到星载成像设备和星地通讯带宽的限制,卫星图像的地面分辨率常不能完全满足目标识别与分析的需求。卫星图像的成像幅度宽且范围广,地面目标的尺度变化大、纹理信息多样化,给现有图像超分辨率技术带来了新的挑战。针对卫星图像的多尺度特性,提出了一种多尺度残差深度神经网络,首先提取低分辨率卫星图像的多尺度特征,对不同尺度特征建立自适应深度神经网络,然后使用融合网络进行残差融合,融合不同尺度高频信息,最终生成高分辨卫星图像。在Space Net卫星图像数据集中的实验结果证明了本文算法的优越性。

Satellite imagery realizes interstellar-earth observations,which is widely used in military and economic fields. Because the performances of satellite-borne imaging equipment and the band width of satellitecommunications system are limited,the resolution of ground targets in satellite images are often low,thus theycannot fully meet the needs of target identification and analysis. Moreover,satellite images have three features：wide range of imaging,variation of multi-scale of ground targets,and diversification of texture information,which bring new challenges to the existing super-resolution algorithms. Using the multi-scale nature of satelliteimage,a multi-scale residual neural network was proposed in this paper for accurately reconstructing themulti-scale information. Firstly,different scale features of low-resolution satellite images were extracted,thenfor each scale-level, an adaptive deep residual neural network was developed for better reconstructionperformance. Then a fusion network was used to refine different scales of residual information. The proposedfusion network fuses high-frequency information of different scales to output the target high-resolution satelliteimage. Experimental results over the Space Net satellite image database prove the superiority of the proposed algorithm.