CMAFusion: Cross modal attention based end-to-end infrared and visible image fusion network

In this paper, a novel cross modal attention based end-to-end infrared and visible image fusion network (CMAFusion) is proposed to generate a fused image with rich texture and salient object information. On the one hand, we design a cross modal feature fusion module to fully fuse the deep complementary features of infrared and visible images. On the other hand, we design a joint loss function that constrains the network from structure, texture and similarity perspectives. Comprehensive comparison and generalization experiments demonstrate that CMAFusion over state-of-the-art alternatives on both subjective qualitative and objective quantitative metrics. More importantly, the performance evaluation of various fusion algorithms on high-level visual perception tasks reveals that our network is beneficial for autonomous driving perception.