Inverse Asymptotic Fusion Framework for Fusion of Infrared and Visible Images of Fires

In situations where smoke or other light sources obstruct the view of a fire scene, using visible light images for real-time detection and response can be challenging. In contrast, infrared images, which are more sensitive to heat sources, have been proven to be more effective. This sensitivity enables the detection and tracking of high-temperature areas at fire scenes, even in conditions of limited visibility. Combining infrared with visible light images provides a more comprehensive understanding of fire situations, aiding in the assessment of fire intensity and the prediction of fire spread. In this paper, we propose the IAFusion network, a fusion network for fire infrared and visible light images constructed using a progressive inverse feature pyramid. The network uses the IAFPN module to initially retain high semantic information and progressively fuses layers of multi-texture information upwards, ensuring that the fused image incorporates more detailed features on the basis of highlighting significant targets, thereby creating a binary mask focused on the subject. Subsequently, in the newly proposed fusion strategy, the fusion with infrared and visible light images is followed by selective feature reconstruction with the RFN module under the guidance of attention, serving as the input to the decoder, thus completing the image reconstruction. Quantitative and qualitative experiments have validated the advancement of our algorithm. Compared with nine competitive algorithms on the general dataset TNO, our network can generate images with clearer and more comprehensive information. In comparison with RFN-Nest on the Corsican fire dataset, our network can add more texture information while highlighting the main subject of flames in images.