DCFusion: Dual-Headed Fusion Strategy and Contextual Information Awareness for Infrared and Visible Remote Sensing Image

Abstract

In remote sensing, the fusion of infrared and visible images is one of the common means of data processing. Its aim is to synthesize one fused image with abundant common and differential information from the source images. At present, the fusion methods based on deep learning are widely employed in this work. However, the existing fusion network with deep learning fails to effectively integrate common and differential information for source images. To alleviate the problem, we propose a dual-head fusion strategy and contextual information awareness fusion network (DCFusion) to preserve more meaningful information from source images. Firstly, we extract multi-scale features for the source images with multiple convolution and pooling layers. Then, we propose a dual-headed fusion strategy (DHFS) to fuse different modal features from the encoder. The DHFS can effectively preserve common and differential information for different modal features. Finally, we propose a contextual information awareness module (CIAM) to reconstruct the fused image. The CIAM can adequately exchange information from different scale features and improve fusion performance. Furthermore, the whole network was tested on MSRS and TNO datasets. The results of extensive experiments prove that our proposed network achieves good performance in target maintenance and texture preservation for fusion images.