用于全色锐化的相对平均生成对抗网络

A Relativistic Average Generative Adversarial Network for Pan-Sharpening

为解决全色锐化过程中对原图像特征提取不足导致融合结果细节信息易丢失,以及图像融合过程中因忽略不同区域的空间特征差异而导致信息冗余等问题,采用深度学习算法,提出一种用于全色锐化的相对平均生成对抗网络(Pan-RaGAN)。在生成器中利用改进的密集块结构对原图像进行特征提取,充分利用原图像各级特征,获取包含了更多细节信息的融合结果;提出基于空间注意力机制的特征细化模块,用于特征选择,可在保留有效高频信息的同时剔除冗余信息;利用图像重建模块将细化后的特征与上采样的低分辨率多光谱图像进行融合,以保持光谱信息;利用相对平均鉴别器改进网络的损失函数,进一步优化融合效果。在高分2号卫星和快鸟卫星图像数据集上的实验结果表明,与已有用于遥感图像全色锐化的生成对抗网络相比,Pan-RaGAN网络的光谱角映射指标平均降低了0.075,验证了Pan-RaGAN网络的有效性。

A relativistic average generative adversarial network for pan-sharpening(Pan-RaGAN)based on a deep learning algorithm is proposed to solve the problems that details of fusion result are easy to be lost due to insufficient feature extraction from original image in the pan-sharpening process and information redundancy is caused by ignoring spatial feature difference of different regions in image fusion process.Firstly,the improved dense block structure is utilized to extract the features of the original image in the generator,so as to make full use of the features from different layers of the original image and obtain better fusion results with more details.Secondly,a feature refinement module based on spatial attention mechanism is proposed for feature selection,which can make a better trade-off between retaining effective high-frequency information and eliminating redundant information.Furthermore,the image reconstruction module is utilized to fuse the refined features with the up-sampled low resolution multispectral images to preserve the spectral information.Finally,the relativistic average discriminator is utilized to improve the loss function of the network,and further optimize the fusion effect.Experimental results on Gao Fen-2 dataset and Quick Bird dataset and a comparison with the existing generative adversarial network for remote sensing image pan-sharpening show that the spectral angle mapper index of the proposed Pan-RaGAN network is reduced by 0.075 on average,which verifies the effectiveness of the proposed Pan-RaGAN network.