Physics-inspired smoke segmentation via heterogeneous feature modeling with a Kolmogorov–Arnold-inspired activation head
Smoke segmentation systems often use general-purpose backbones with predefined pointwise activations, which may limit channel-specific modeling of nonlinear variations in smoke appearance and apparent density. Differences between black and white smoke, together with interference from clouds and fog, further complicate feature extraction. We propose the smoke KAN-inspired dual-stream SegFormer (SKD-SegFormer), a lightweight framework guided by these smoke characteristics. Its KAN-inspired channel-adaptive activation head applies learnable B-spline mappings to individual feature channels after multi-scale fusion, enabling flexible nonlinear responses across feature-value ranges. Two modules developed in this work address complementary challenges. The dual-stream smoke discrimination module adaptively fuses diffuse patterns and localized details through branches with different receptive fields. The smoke-oriented suppression module combines appearance-dependent confidence with a spatial prior motivated by the typical buoyancy-driven morphology of rising fire smoke to reduce false responses to environmental distractors. On the self-built real smoke dataset, SKD-SegFormer achieves a mean Intersection over Union (IoU) of $$82.85\% \pm 1.14\%$$ over three random seeds (best run 84.03%). Its single-run IoU on the synthetic dataset is $$83.46\%$$ . With 6.45 M parameters and 52 frames per second on an NVIDIA A40 GPU, the model offers a favorable balance between segmentation accuracy and computational efficiency.
Authors
- Yuming Li
- Yongjie Wang
- Changsong Liu
- Anbo Zheng
Institutions
- Tianjin University (CN)
- Jiaxing Nanhu University (CN)
- Luoyang Polytechnic (CN)
- Shijiazhuang Tiedao University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1038/s41598-026-72652-2
- Primary Topic
- Fire Detection and Safety Systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Henan Provincial Science and Technology Research Project