Research map / Detection and segmentation / CNN features
Infrared Small Target Detection with Scale and Location Sensitivity
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 a Scale and Location Sensitive (SLS) loss and a simple Multi-Scale Head on plain U-Net (MSHNet) for infrared small target detection, achieving SOTA IoU/Pd on IRSTD-1k and NUDT-SIRST.
Problem
Infrared small target detection (IRSTD) models are trained with IoU and Dice losses that are insensitive to target scale and location, limiting detection performance.
Approach
A loss that weights the IoU loss by a scale ratio (min/max of predicted and ground-truth pixel counts, with variance smoothing) plus a polar-coordinate center-point location penalty, applied at every scale of a plain U-Net with a lightweight multi-scale prediction head.
Key results
- MSHNet reaches 67.16% IoU and 93.88% Pd on IRSTD-1k, versus 65.71% IoU for DNANet.
- On NUDT-SIRST, MSHNet reaches 80.55% IoU and 97.99% Pd, versus 79.98% IoU and 96.93% Pd for DNANet.
- Applying SLS loss improves IoU for DNANet, ISNet and MSHNet, but raises false alarm rate (e.g., DNANet Fa rises from 17.61 to 35.15 x10^-6).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the scale weight w to the IoU-based scale loss (ablation) | MSHNet · IRSTD-1k | IoU (%): 64.83 -> 65.82 | Helped |
| add the location-sensitive loss LL to LS (ablation) | MSHNet · IRSTD-1k | IoU (%): 65.82 -> 67.16 | Helped |
| replace LL with L2 center-point distance (ablation) | MSHNet · IRSTD-1k | IoU (%): 67.16 -> 64.89 | Hurt |
| replace LL with L1 center-point distance (ablation) | MSHNet · IRSTD-1k | IoU (%): 67.16 -> 64.75 | Hurt |
| increase the number of supervised prediction scales from 1 to 4 (ablation) | MSHNet · IRSTD-1k | IoU (%): 63.10 -> 67.16 | Helped |
| replace IoU loss with SLS loss | DNANet · IRSTD-1k | IoU (%): 65.71 -> 67.09 | Helped |
| replace IoU loss with SLS loss | ISNet · IRSTD-1k | IoU (%): 62.88 -> 64.42 | Helped |
| replace IoU loss with SLS loss (ablation) | MSHNet · IRSTD-1k | IoU (%): 64.83 -> 67.16 | Helped |
| replace SLS loss with IoU loss (ablation) | MSHNet · IRSTD-1k | Fa (10−6): 15.03 -> 5.28 | Helped |
| replace SLS loss with Dice loss (ablation) | MSHNet · IRSTD-1k | IoU (%): 67.16 -> 65.36 | Hurt |
| use SLS loss instead of IoU or Dice loss (ablation) | MSHNet · IRSTD-1k | Fa (10−6): 5.28; 14.12; 15.03 | Mixed |
Surprising findings
- The SLS loss consistently increases false alarms (Fa), which the authors attribute to the location sensitive term treating noise as targets.
- Replacing the polar-coordinate location loss with simple L1 or L2 distances gives higher Fa, showing a trade-off between discriminability and false alarms.
Limitations
- The location sensitive loss increases false alarm rate, reducing precision.
- Evaluation limited to two IRSTD datasets with fixed 256x256 input size.
Experimental setup
Datasets: IRSTD-1k NUDT-SIRST
Benchmarks: IRSTD-1k NUDT-SIRST
Metrics: IoU Pd Fa FLOPs inference speed
Baselines: Top-Hat Max-Median WSLCM TLLCM IPI NRAM RIPT PSTNN MSLSTIPT MDvsFA ALCNet ISNet
Builds on: U-Net IoU loss Dice loss DNANet ISNet
Closest papers by meaning
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.90)
- Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection (AAAI 2025 · similarity 0.87)
- IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.86)
- Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (AAAI 2025 · similarity 0.86)
- SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · similarity 0.85)
- IRMamba: Pixel Difference Mamba with Layer Restoration for Infrared Small Target Detection (AAAI 2025 · similarity 0.84)
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.