Infrared small target detection is a critical technology for remote sensing, fire prevention, and surveillance, but existing methods often struggle with tiny, low-contrast targets that lack distinct shape or texture. Now, researchers from the Research Center for Space Optical Engineering at Harbin Institute of Technology have developed a two-stage deep learning network that combines diffusion-based feature enhancement with state-space modeling (SSM) to suppress background clutter while amplifying target signals. The approach achieves state-of-the-art detection rates across three public datasets, significantly reducing both missed detections and false alarms in complex imaging environments. The findings were published on June 30, 2026, in the Journal of Remote Sensing, as detailed in the study (DOI: 10.34133/remotesensing.1046).
The proposed Diffusion-Enhanced Dense Mamba Network (DEDM-Net) addresses a fundamental challenge: infrared targets often occupy fewer than 81 pixels (typically under 9×9), exhibit extremely low energy with signal-to-noise ratios around 3, and lack prominent shape or texture information. These characteristics cause targets to be easily submerged in background clutter. While deep learning methods have improved performance, most focus exclusively on target features while neglecting background information—the vast majority of the image—leading to severe class imbalance between positive and negative samples. DEDM-Net overcomes this by modeling both targets and backgrounds simultaneously.
The team developed a two-stage network that achieves a synergistic effect greater than the sum of its parts. The first stage employs a dual-path diffusion model with a novel blind processing module that predicts each pixel using only surrounding information—never the pixel itself—preventing extremely small targets from being misclassified as background. The second stage introduces a dense nested Mamba architecture based on the state space model (SSM), which captures long-range correlations across global and local features with linear computational complexity—a significant advantage over conventional Transformers. A cross-stage prediction fusion module further integrates features from both stages, improving contour segmentation accuracy. Together, these innovations deliver superior performance across all evaluation metrics compared to 11 state-of-the-art methods.
The DEDM-Net was evaluated on three public datasets: NUAA-SIRST (427 images), NUDT-SIRST (1,327 images at 256×256), and IRSTD-1k (1,000 images at 512×512). On the NUDT-SIRST dataset, the method achieved 93.40% IoU, 93.28% nIoU, 98.37% detection probability (P_d), and a remarkably low false-alarm rate of just 3.75×10⁻⁶—outperforming DNA-Net (92.99% IoU, 93.22% nIoU) and ISTDU-Net (91.69% IoU, 91.84% nIoU). On the IRSTD-1k dataset, DEDM-Net achieved 73.71% IoU and 93.89% P_d with only 11.10×10⁻⁶ false alarms, surpassing all competitors. The blind processing module uses a dual-window structure with outer radius R=4 and inner radius r=2, ensuring that target regions are "blindly processed" while surrounding context is captured. Ablation studies confirmed that each component—the generation path, restoration path, dense nested structure, and residual Mamba blocks—contributes meaningfully to overall performance. The network was trained on an NVIDIA RTX 4080 GPU using the Adam optimizer.
"Infrared small targets are extremely challenging because they lack shape and texture—they're essentially just a few bright pixels in a sea of background," said corresponding author Dr. Shikai Jiang of Harbin Institute of Technology. "By modeling both the target-free background and potential target regions simultaneously, our diffusion-enhanced approach effectively amplifies what matters while suppressing what doesn't. The Mamba architecture then provides the global context needed to distinguish true targets from bright clutter."
The method employs a two-stage training framework. In the diffusion enhancement stage, a U-Net backbone estimates noise across 1,000 diffusion steps, with a dual-path scheme modeling both target masks and background images. The blind processing module generates pixel-wise convolution kernels that exclude the center pixel, effectively removing small targets from background reconstruction. In the detection stage, a dense nested network with feature pyramid connections and residual Mamba blocks extracts multiscale features. The loss function combines binary cross-entropy (BCE) and Dice losses to address class imbalance.
While DEDM-Net achieves superior accuracy, the diffusion-based two-stage design increases inference time compared to single-stage networks. Future work will focus on model distillation, mixed-precision inference, and faster samplers to reduce the required diffusion steps. The approach holds promise for real-time surveillance systems, autonomous drone navigation in low-visibility conditions, and early wildfire detection networks. As the team noted, the framework could also inspire new thinking about how generative models and state-space architectures can be combined for other challenging computer vision tasks where target-background separation is critical.
For Long Beach and surrounding areas, this breakthrough could have significant implications. With increasing wildfire risks in California, early detection systems powered by DEDM-Net could provide faster and more reliable alerts, potentially saving lives and property. The Port of Long Beach, a critical hub for international trade, could benefit from enhanced surveillance capabilities, improving security and operational efficiency. Moreover, autonomous drone navigation in low-visibility conditions—such as fog or smoke—could become more robust, benefiting industries from logistics to emergency response. As the technology matures, it may also stimulate local tech innovation and attract investment in remote sensing and AI, further positioning Long Beach as a forward-thinking community. The research was supported by the National Natural Science Foundation of China and other grants, underscoring the global collaboration driving these advancements.
