遥感图像融合是指将不同传感器得到的具有不同观测特性的图像信息有选择、有策略地结合起来, 以得到具有更优观测特性的新图像的方法。提出一种深度学习结合非下采样剪切波变换(NSST)的遥感图像融合算法, 利用改进的超分辨率重建网络对多光谱图像(MS)进行空间分辨率增强, 全色图像(PAN)参考重建后的多光谱图像的每个分量进行直方图匹配。将对应通道的图像进行NSST变换, 分别得到低频子带和若干高频子带。低频子带通过使用基于梯度域的自适应加权平均规则来获得低频融合系数, 高频子带采用局部空间频率最大值规则来获得高频融合系数, 最后经逆NSST变换重构获得融合图像。对不同数据集中的City和Inland多光谱图像采用双三次插值方法进行上采样, 作者提出算法的通用图像质量指数(UIQI)分别为0.988 6和0.932 1, 光谱角映射(SAM)分别为1.872 1和2.143 2。实验结果表明, 图像结构更加清晰, 保存的光谱信息更加完整, 融合图像质量优于对比算法, 融合图像更利于人类视觉观察。

Remote-sensing image fusion refers to the method of selectively and strategically combining image information with different observation characteristics obtained by different sensors to obtain a new image with better observation characteristics. A deep-sensing image fusion algorithm combined with non-subsampled shearlet transform (NSST) was proposed. In this algorithm,the spatial resolution of multi-spectral (MS) image is enhanced by an improved super-resolution reconstruction network. The panchromatic (PAN) image histogram-matched refers to each component of the reconstructed MS image. And the corresponding channel image is subjected to NSST transformation to obtain low-frequency sub-bands and several high-frequency direction sub-bands, respectively. To obtain low-frequency fusion coefficient, the low-frequency region uses an adaptive weighted average rule based on the gradient region, while the high-frequency sub-bands adopt the local spatial frequency maximum rule to obtain the high-frequency fusion coefficient, and finally the fused image can be obtained by inverse NSST transform reconstruction. The MS images City and Inland in different datasets were upsampled by the bicubic interpolation method. With the proposed algorithm, the general image quality index (UIQI) was 0.988 6 and 0.932 1,respectively, and the spectral angle mapping (SAM) was 1.872 1 and 2.143 2,respectively. Experimental results show that the image structure of the fusion algorithm in this paper is more clear, the saved spectral information is more complete, the fusion quality is better than the contrast algorithm, and the fusion image is more conducive to human visual observation.