SEPAL-Fuse: Semantic-preserving alignment for infrared-visible image fusion

In image fusion, higher subjective visual quality does not necessarily translate into stronger support for high-level vision tasks. Existing approaches either (i) impose high-level task losses on the fusion network, (ii) inject semantic features from the task network into the fusion network, or (iii) perform knowledge distillation between the two. However, these approaches lead to either overly weak or excessively strong interactions between the fusion and task networks. To address this issue, we propose SEPAL-Fuse (Semantic-Preserving Alignment for Infrared-Visible Image Fusion), which is the first to employ contrastive learning as a bridge between the fusion and task networks. Specifically, we introduce a semantic-visual disentanglement module that decomposes the fusion representation into semantic and visual subspaces. By confining cross-domain interaction to the semantic subspace, we mitigate the mutual interference between semantic alignment and visual representation learning. We further perform entropy-guided patch-level contrastive learning in the semantic subspace, where a Shannon-entropy-based attention map derived from intermediate task features is used to select task-relevant patches. Patches from the task and fusion encoders at the same spatial location form positive pairs, while patches from different spatial locations serve as negative samples. This contrastive interaction enables effective semantic alignment between fusion-side and task-side semantics while moderately preserving the fusion network's intrinsic semantic distribution, thereby improving semantic-visual compatibility during fusion and ultimately enhancing both perceptual quality and downstream task performance. Extensive experiments on multiple public benchmarks demonstrate the effectiveness and robustness of SEPAL-Fuse.