基于变分的多尺度遥感图像融合算法

Multiscale Remote Sensing Image Fusion Algorithm Based on Variational Segmentation

全色图像与多光谱图像融合时,忽略了上采样的重要性和通道间细节的差异性.针对前者,利用不同尺度下自相似块,估计出低分辨率图像丢失信息,从而修改了图像上采样的策略,并以此构造目标函数的保真项;针对后者,利用全色图像和光谱图像梯度域结构相似性,提出局部加权动态稀疏约束,构造目标函数的正则项.本文基于变分法理论,构造了新的目标函数,并提出了多尺度迭代融合框架,通过多次迭代逐步提高融合图像的分辨率,每一层的融合结果更加准确,从而提高最终的融合精度.本文算法与Brovey等成分替代算法、P+XS等变分算法、MTF_GLP等多分辨分析算法进行比较.实验结果表明,本算法的融合结果具有良好的视觉效果,且在客观评价指标上比所有对比算法的最优值平均值均有提高.

On the fusion of panchromatic and multispectral images,two important aspects,up-sampling of multispectral images and the difference of channel details,are ignored.For the former,the loss details of low-resolution images are estimated by using self-similar patch at different scales to improve up-sampling.For the latter,the local weighted dynamic sparse constraint is proposed based on the structural similarity between panchromatic images and spectral images in gradient domain.The new objective function based on variational method are proposed,the fidelity term and the regularization term of whose are constructed respectively according to the former and the latter.In addition,a multi-scale iterative fusion framework is presented,where the resolution of the fused image is gradually improved through iterations.The fused results of each iteration are more accurate,so the final fused image is improved.Our algorithm is compared with Brovey and other component substitution algorithms,P+XS and other variational algorithms,MTF_GLP and other multi-resolution analysis algorithms.The experimental results show that the fusion results of this algorithm have good visual effect,and the objective evaluation index is better than the average of the optimal value of all comparison algorithms.