BSPFusion: A Bottom-Up Semantically Progressive Network for Infrared and Visible Image Fusion

• A bottom-up semantically progressive network is proposed for image fusion. • An entity anchoring pipeline provides fine-grained entity-region semantic priors. • A spectrum synergy module preserves visual details via frequency-domain processing. • An entity-anchored semantic alignment module enables text-free semantic fusion. Infrared and visible image fusion aims to integrate complementary spectral information to enhance scene perception. Current text-guided methods mainly rely on textual priors to guide a fusion process. However, these approaches lack fine-grained entity-region correspondences, and their dominant semantic modulation often suppresses text-unrelated visual contents, thereby compromising fusion quality. Moreover, the reliance on textual inputs during an inference process fundamentally limits the deployment flexibility of these methods. To overcome these limitations, we propose BSPFusion, a novel Bottom-up Semantically Progressive network for infrared and visible image Fusion. Our network introduces a custom entity anchoring pipeline to provide fine-grained entity-region priors, establishing a progressive fusion flow from low-level visual contents to high-level semantic concepts through two core modules: (i) the spectrum synergy module performs frequency-aware cross-modal interaction by decomposing visual features into multi-band components and regulating their integration via intra- and inter-modal attention, thereby enhancing the completeness and balance of low-level visual representations before semantic modulation, which helps mitigate the attenuation of text-unrelated visual contents during the fusion process; (ii) the entity-anchored semantic alignment module enables semantic-aware fusion without textual input by learning entity-level semantic alignment into the fusion network through entity-aware attention during training, allowing the network to activate semantically relevant regions using only visual features at inference stage. Extensive experiments on three public benchmarks ( i.e. , MSRS, M 3 FD, and RoadScene) demonstrate that BSPFusion achieves state-of-the-art performance by effectively balancing low-level visual preservation with high-level semantic awareness. Our code and annotated benchmarks will be released at https://github.com/wyshao-01/BSPFusion .