Research map / Detection and segmentation / CNN features
TCI-Former: Thermal Conduction-Inspired Transformer for Infrared Small Target Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TCI-Former, a U-Net-like transformer for infrared small target detection whose attention and boundary modules are derived from a pixel-movement differential equation analogous to thermal conduction.
Problem
Infrared small target detection (ISTD) segments tiny, low-contrast targets from cluttered backgrounds, and existing networks lack an explainable view of how feature maps evolve toward targets during ISTD.
Approach
Models feature-map pixel movement toward targets as heat conduction, deriving a pixel movement differential equation (PMDE); a finite-difference-based attention (TCIA) captures target body features via neighbor-shift conduction, and a Laplacian-based boundary module (TCBM) refines boundaries.
Key results
- On IRSTD-1k, TCI-Former reaches 70.14 IoU, 67.69 nIoU, 96.31 Pd and 14.81 Fa, the best on all four metrics versus DNANet and ISNet.
- On NUAA-SIRST, TCI-Former achieves 80.79 IoU and 79.85 nIoU, the highest among compared methods; its Fa (4.189) is second to RKformer (1.580).
- Ablation on IRSTD-1k shows the baseline ViT at 62.82 IoU rising to 70.14 with TCIA and TCBM combined.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use TCI-Former instead of ISNet | ISTD methods · NUAA-SIRST | IoU: 80.02 -> 80.79 | Helped |
| use TCI-Former instead of DNANet | ISTD methods · IRSTD-1k | IoU: 69.01 -> 70.14 | Helped |
| add TCIA to the basic pyramid ViT baseline (ablation) | TCI-Former encoder · IRSTD-1k | IoU: 62.82 -> 67.26 | Helped |
| add TCBM to the TCIA-equipped baseline (ablation) | TCI-Former encoder · IRSTD-1k | IoU: 67.26 -> 70.14 | Helped |
| replace MHSA with TCIA (ablation) | TCI-Former attention block · IRSTD-1k | IoU: 66.73 -> 70.14 | Helped |
| replace CSWSA with TCIA (ablation) | TCI-Former attention block · IRSTD-1k | IoU: 68.23 -> 70.14 | Helped |
| replace MHRA with TCIA (ablation) | TCI-Former attention block · IRSTD-1k | IoU: 68.86 -> 70.14 | Helped |
| replace ResBlock with TCBM (Laplace+Resblock) (ablation) | TCI-Former boundary module · IRSTD-1k | IoU: 68.93 -> 70.14 | Helped |
| replace Roberts+ResBlock with TCBM (Laplace+Resblock) (ablation) | TCI-Former boundary module · IRSTD-1k | IoU: 69.51 -> 70.14 | Helped |
Limitations
- The thermal-conduction analogy is heuristic; the derivation assumes p_s=0 and constant coefficients, and the finite difference approximation introduces boundary errors the paper acknowledges.
- Evaluated only on two ISTD datasets (IRSTD-1k and NUAA-SIRST).
Experimental setup
Datasets: NUAA-SIRST IRSTD-1k
Benchmarks: NUAA-SIRST IRSTD-1k
Metrics: IoU nIoU Pd Fa
Baselines: PSTNN MSLSTIPT MDvsFA ACM AlcNet DNANet Dim2Clear FC3-Net IAANet RKformer ISNet
Builds on: ViT ISNet Taylor finite difference
Closest papers by meaning
- Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection (AAAI 2025 · similarity 0.84)
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.83)
- IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.83)
- Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (AAAI 2025 · similarity 0.83)
- Pseudo Visible Feature Fine-Grained Fusion for Thermal Object Detection (CVPR 2025 · similarity 0.82)
- MOCID: Motion Context and Displacement Information Learning for Moving Infrared Small Target Detection (AAAI 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.