MDRDA-GAN: A Multi-discriminator Generative Adversarial Network Based on Residual Dense Attention for Remote Sensing Image Fusion

The aim of high spatial resolution panchromatic image and high spectral resolution multispectral image fusion is to obtain both high spatial and spectral resolution multispectral images for further remote sensing applications. Therefore, it is crucial to preserve the spatial information of the panchromatic image and the spectral information of the multispectral in the fusion image. To solve this problem, a new multi-discriminator and residual dense attention for the generative adversarial network (MDRDA-GAN) is proposed for remote sensing image fusion. Compared with the interpolation or deconvolution amplification operations of other methods, the super-resolution module is designed in the generator of MDRDA-GAN, and a specific super-resolution loss function is designed to generate super-resolution multispectral images so that the multispectral images contain more spatial and spectral information for subsequent fusion. The multi-scale feature extraction module and the residual dense module are introduced in the generator, which can make full use of the features of different perceptual fields and different convolution layers. With an embedding attention mechanism in these two modules, the network can adapt feature refinement to retain more important features and ignore redundant information. And a spatial discriminator and a spectral discriminator are designed to learn against the generator to generate high-quality, high-resolution fused images. Quantitative comparison of experimental results and qualitative analysis show that our method outperforms existing methods.