DEDM-Net: Diffusion-Enhanced Dense Mamba Network for Infrared Small Target Detection

Infrared small target detection plays a pivotal role in remote sensing situational awareness, forest fire prevention, and surveillance and early warning systems. Current mainstream deep learning methods, however, primarily focus on target-specific features while underutilizing the nontarget background regions. Furthermore, extremely small-scale targets are often undetectable due to the difficulty of extracting their feature information. To address these challenges, we propose a Diffusion-Enhanced Dense Mamba Network for small target detection. The network comprises 2 main stages: a feature enhancement stage based on the dual-path diffusion model and a target detection stage based on the dense link state space model. In the feature enhancement stage, the diffusion model, combined with low-frequency image features and a blind processing module, performs dual-path modeling of target-free backgrounds and candidate target masks, effectively enhancing target features while mitigating background clutter. In the target detection stage, the state space model captures long-range correlations across both global and local features. Utilizing a dense link nested structure along with a cross-stage prediction fusion module, features of different scales, semantics, and stages are thoroughly integrated and enhanced, enabling effective detection of weak and small infrared targets. Experimental results on 3 publicly available high-quality datasets demonstrate that the proposed method achieves state-of-the-art performance in detection probability, intersection over union, and other evaluation metrics compared with other advanced methods.