MSADRCN: meta-learning based joint super-resolution fusion of infrared and visible images

Existing infrared and visible image fusion algorithms usually only input images of the same resolution and the obtained fusion image is still of low quality when the source image is of low resolution, which hinders further image analysis. In response to the problem that the source image can only have the same resolution and the up-sampling can only be integer multiples, this paper proposes a joint super-resolution fusion network based on meta-learning at any resolutions. Firstly, this paper proposes a Meta-learning Super-resolution Network (MSN) in the feature extraction section, which can receive source images of any resolutions and change the feature map to the actual required size by dynamically predicting the up-sampling factor through the meta-learning up-sampling module. Secondly, this paper proposes an Autoencoder Dual-discriminator Conditional Generation Fusion Network (A-DCGFN) in the feature fusion section, which strengthens the correlation between global and local regions and obtains preliminary fused images. Finally, this paper proposes a Residual Compensation Dual-backup Fusion Network (RCDFN) to further improve the quality of the fused image by compensating for the lost detailed features and distorted information during the up-sampling process. Through the validation of TNO and MSRS datasets, experimental results show that the proposed method achieves the best performance in fused image quality compared to other existing models.