Research map / Detection and segmentation / CNN features
IRMamba: Pixel Difference Mamba with Layer Restoration 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 IRMamba, an encoder-decoder for infrared small target detection combining a Pixel Difference Mamba (center pixel difference inside the SSM state equation) with a Layer Restoration Module based on unrolled double-DIP gradient descent.
Problem
Infrared small target detection struggles with small, dim targets amid noise and clutter; CNNs miss global context, ViT hybrids are costly, and flattening 2D images into 1D Mamba tokens breaks local 2D dependencies needed for small targets.
Approach
Injects horizontal/vertical pixel-difference terms between scan positions and their neighborhoods into the Mamba state equation, gated by a local-standard-deviation spatial distribution function, and adds an unfolded iterative layer-separation module (BITGD + cross-proximal attention) that restores masks by alternating high- and low-level gradient paths.
Key results
- On NUDT-SIRST, IRMamba reaches IoU 95.18, Pd 99.26, Fa 1.309e-6, above SCTransNet (IoU 94.09) and other baselines.
- On IRSTD-1k, IRMamba reaches IoU 70.04 and Pd 95.81, versus 69.96 IoU for UIU-Net; on NUAA-SIRST it reaches IoU 80.2 and Fa 0.887e-6.
- Ablation on NUDT-SIRST: removing both PDMamba and LRM drops IoU from 95.18 to 91.09; with 10.51M parameters IRMamba is smaller than RKformer and IAANet.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LRM (ablation) | IRMamba · NUDT-SIRST | IoU: 92.96 -> 95.18 | Helped |
| add LRM (ablation) | IRMamba · NUDT-SIRST | Pd: 97.87 -> 99.26 | Helped |
| add LRM (ablation) | IRMamba · NUDT-SIRST | Fa: 2.65 -> 1.309 | Helped |
| add PDMamba (ablation) | IRMamba · NUDT-SIRST | IoU: 92.28 -> 95.18 | Helped |
| add PDMamba (ablation) | IRMamba · NUDT-SIRST | Pd: 97.01 -> 99.26 | Helped |
| add PDMamba (ablation) | IRMamba · NUDT-SIRST | Fa: 1.98 -> 1.309 | Helped |
Limitations
- Evaluated only on three IRSTD datasets at 256x256 resolution.
- Layer restoration and PDMamba add architectural complexity; the spatial distribution function uses hand-set thresholds (gamma=1, window n=2).
Experimental setup
Datasets: NUAA-SIRST IRSTD-1k NUDT-SIRST
Benchmarks: NUDT-SIRST IRSTD-1k NUAA-SIRST
Metrics: IoU Pd Fa Params inference time
Baselines: IPI Top-Hat MSLSTIPT ACM ALCNet FC3-Net ISNet DNA-Net UIU-Net Dim2Clear GCI-Net RKformer
Builds on: Mamba Deep Image Prior Double DIP Deep unfolding networks VMamba
Closest papers by meaning
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.88)
- IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.86)
- Pinwheel-shaped Convolution and Scale-based Dynamic Loss for Infrared Small Target Detection (AAAI 2025 · similarity 0.84)
- SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · similarity 0.84)
- Infrared Small Target Detection with Scale and Location Sensitivity (CVPR 2024 · similarity 0.84)
- Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (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.