Diffusion-Mamba network boosts infrared small-target detection

Researchers at Harbin Institute of Technology have unveiled a two-stage deep learning model that improves infrared small-target detection by combining diffusion-based feature enhancement with state-space modeling. The system outperformed 11 leading methods on three public datasets, cutting missed detections and false alarms in applications such as wildfire warning, surveillance and remote sensing.

Why it matters: - Infrared small-target detection is used in remote sensing, fire prevention and surveillance, where missed detections and false alarms can have serious consequences. - The new Diffusion-Enhanced Dense Mamba Network, or DEDM-Net, is designed to find tiny, low-contrast targets that are easily lost in background clutter. - The method could improve early wildfire detection, threat assessment and other systems that need reliable target-background separation.

What happened: - Researchers from the Research Center for Space Optical Engineering at Harbin Institute of Technology published the work in the Journal of Remote Sensing on June 30, 2026. - The team developed DEDM-Net, a two-stage network that combines diffusion-based enhancement with state-space modeling. - The paper reports state-of-the-art results across three public datasets and compares performance against 11 other methods. - The original source is available in the full paper.

The details: - Infrared small targets often occupy fewer than 81 pixels, usually under 9×9, and typically have signal-to-noise ratios around 3. - These targets usually lack clear shape or texture, which makes them hard to separate from background clutter. - The first stage of DEDM-Net uses a dual-path diffusion model and a blind processing module that predicts each pixel from surrounding context rather than the pixel itself. - The second stage uses a dense nested Mamba architecture built on a state space model to capture long-range dependencies with linear computational complexity. - A cross-stage prediction fusion module combines outputs from both stages to improve contour segmentation. - The network was evaluated on NUAA-SIRST, with 427 images; NUDT-SIRST, with 1,327 images at 256×256; and IRSTD-1k, with 1,000 images at 512×512. - On NUDT-SIRST, DEDM-Net reached 93.40% IoU, 93.28% nIoU, 98.37% detection probability and a false-alarm rate of 3.75×10⁻⁶. - On the same dataset, DEDM-Net outperformed DNA-Net, which posted 92.99% IoU and 93.22% nIoU, and ISTDU-Net, which posted 91.69% IoU and 91.84% nIoU. - On IRSTD-1k, DEDM-Net achieved 73.71% IoU, 93.89% detection probability and 11.10×10⁻⁶ false alarms. - The blind processing module uses a dual-window structure with outer radius R=4 and inner radius r=2. - Ablation studies found that the generation path, restoration path, dense nested structure and residual Mamba blocks each contributed to performance. - The model was trained on an NVIDIA RTX 4080 GPU using the Adam optimizer. - The loss function combines binary cross-entropy and Dice loss to address class imbalance.

Between the lines: - The central technical move is to model both the target and the target-free background, instead of focusing only on foreground features. - That approach is aimed at a common weakness in infrared detection, where the background contains most of the image and overwhelms tiny targets. - The Mamba-based stage adds global context without the heavier computational cost of conventional Transformers. - The two-stage design improves accuracy, but it also increases inference time versus single-stage networks. - The quoted perspective from corresponding author Dr. Shikai Jiang underscores the paper’s claim that combining generative modeling with state-space architectures helps separate real targets from bright clutter.

What's next: - The researchers plan to focus on model distillation, mixed-precision inference and faster samplers to reduce the number of diffusion steps. - The team says the approach could be adapted for real-time surveillance, autonomous drone navigation in low-visibility conditions and early wildfire detection networks. - The authors also suggest the framework may inform other computer vision tasks that require strong target-background separation.

The bottom line: - DEDM-Net brings diffusion modeling and Mamba-based state-space processing together to set a new benchmark for infrared small-target detection, with the biggest gains coming where clutter and faint signals make traditional methods fail.