StyleFuse: An unsupervised network based on style loss function for infrared and visible fusion

Since infrared images can distinguish targets from backgrounds based on thermal radiation while visible images have considerable details, it is meaningful to fuse two images from these two different types of sensors for subsequent tasks. Although many works have aimed to enhance the content preservation abilities of models, they are still lacking from a principle standpoint: image fusion is the process of aligning the distributions of different image domains. To address this issue, we propose a position-unaware style fusion loss function, which is jointly trained with an unsupervised network to learn the style information of the input images. In our model, we introduce attention-based network connections that can make full use of the information at different scales, and we enhance the transmission of information across these scales. Moreover, we conduct experiments comparing our model with eleven other methods on public datasets. The experimental results demonstrate that our model outperforms the state-of-the-art methods in both qualitative and quantitative evaluations. Finally, additional experiments are performed to demonstrate the generalization ability and complexity of our model.

Automated summary

This paper proposes a position-unaware style fusion loss function, jointly trained with an unsupervised network to learn the style information of the input images. The introduction of attention-based network connections enables full utilization of information at different scales and enhances information transmission across these scales. Experimental results demonstrate that the proposed model outperforms existing methods in both qualitative and quantitative evaluations.