Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation
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 S3D, a semi-supervised pipeline for single-frame infrared small target detection that uses thermodynamic-inspired chromatic and spatial uneven perturbations, a weighted strong-to-strong consistency loss, and a Truncated Squeeze loss for class imbalance.
Problem
Single-frame infrared small target (SIRST) detection suffers from scarce labeled data, foreground-background class imbalance, and label noise.
Approach
Augments IR images with thermodynamics-inspired chromatically uneven (strong) and spatially uneven (weak) perturbations, weights unlabeled consistency by prediction agreement with a weak-view anchor, and uses a truncated squeeze loss that zeroes loss on well-classified pixels with asymmetric thresholds.
Key results
- On NUDT-SIRST with 1/8 labels, S3D (ResUNet-based) reaches 90.12 IoU with Fa 6.85e-6, about 98% of fully-supervised DNANet (IoU 91.89 with DNANet... as reported in Table 2 cited comparison).
- On NUAA-SIRST with 1/8 labels, S3D reaches IoU 73.12 and Fa 10.28e-6, about 92% of fully-supervised DNANet IoU.
- Applying the proposed techniques to ResUNet on NUDT-SIRST with 1/32 labels boosts IoU from 46.35 to 77.61 and cuts Fa from 160.25 to 21.04.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add chromatic and spatial uneven perturbations (ablation) | Semi-supervised ACMNet · NUDT-SIRST 1/8 labeled | IoU (%): 79.61 -> 85.07 | Helped |
| Add spatial uneven perturbation to chromatic perturbation (ablation) | Semi-supervised ACMNet · NUDT-SIRST 1/8 labeled | IoU (%): 83.25 -> 85.07 | Helped |
| Add chromatic uneven perturbation to spatial perturbation (ablation) | Semi-supervised ACMNet · NUDT-SIRST 1/8 labeled | IoU (%): 83.00 -> 85.07 | Helped |
| Replace IoU loss with TS loss (ablation) | ACMNet · NUDT-SIRST 1/32 labeled | IoU (%): 54.29 -> 58.42 | Helped |
| Replace IoU loss with TS loss (ablation) | DNANet · NUDT-SIRST 1/32 labeled | IoU (%): 63.57 -> 69.39 | Helped |
| Use TS loss instead of BCE loss (ablation) | Semi-supervised ResUNet · NUDT-SIRST 1/32 labeled | IoU (%): 36.01 -> 77.61 | Helped |
| Remove truncation from TS loss (ηc∗=1) (ablation) | Semi-supervised ResUNet · NUDT-SIRST 1/32 labeled | Fa (10−6): 27.75 vs 21.94 | Hurt |
| Replace confidence-adaptive consistency with UniMatch consistency (ablation) | Semi-supervised ResUNet · NUDT-SIRST 1/32 labeled | IoU (%): 77.61 vs 77.13 | Hurt |
| Add the proposed semi-supervised techniques (ablation) | ACMNet · NUDT-SIRST 1/32 labeled | IoU (%): 58.42 -> 61.04 | Helped |
| Add the proposed semi-supervised techniques (ablation) | ResUNet · NUDT-SIRST 1/8 labeled | IoU (%): 76.15 -> 85.07 | Helped |
| Add the proposed semi-supervised techniques (ablation) | DNANet · NUDT-SIRST 1/8 labeled | IoU (%): 84.30 -> 90.12 | Helped |
Surprising findings
- Table 2 reports S3D (1/8) IoU 90.12 on NUDT-SIRST, below the fully-supervised DNANet+LTS IoU of 91.89, yet the text claims near-full performance; the bold/variant mapping is inconsistent across tables.
Limitations
- Relies on a fixed 256x256 resize and specific hyperparameters (thresholds 0.99/0.9, lambda 0.5) with limited sensitivity analysis.
- Evaluated only on two single-frame IR benchmarks with pixel-level metrics.
- Paper notes label noise and boundary annotation issues but does not quantify their effect beyond ablations.
Experimental setup
Datasets: NUDT-SIRST NUAA-SIRST
Benchmarks: NUDT-SIRST NUAA-SIRST
Metrics: IoU Fa (false alarm rate)
Baselines: ST++ CPS MDvsFA ALCNet ISNet GCI-Net ACMNet DNANet ResUNet UniMatch loss
Builds on: FixMatch UniMatch Focal Loss DNANet
Closest papers by meaning
- Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm (NeurIPS 2026 · similarity 0.91)
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.88)
- Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection (AAAI 2025 · similarity 0.88)
- IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.87)
- SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · similarity 0.86)
- Infrared Small Target Detection with Scale and Location Sensitivity (CVPR 2024 · similarity 0.86)
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.