New Diffusion-Mamba Hybrid Model Significantly Enhances Infrared Small Target Detection Performance

Researchers from the Harbin Institute of Technology have introduced the Diffusion-Enhanced Dense Mamba Network (DEDM-Net), a breakthrough in infrared small target detection. By integrating diffusion-based feature enhancement with advanced state-space modeling, this two-stage deep learning framework effectively suppresses background clutter while amplifying weak, low-contrast signals. This innovation addresses the persistent challenge of identifying targets that occupy fewer than 81 pixels, providing a robust solution for environments where traditional methods frequently fail.

The DEDM-Net architecture utilizes a unique blind processing module to ensure target regions are analyzed without center-pixel bias, followed by a dense nested Mamba structure for superior global context capture. Evaluated across three major public datasets, including NUAA-SIRST and IRSTD-1k, the model consistently outperformed eleven state-of-the-art methods. On the NUDT-SIRST dataset, it achieved a remarkable 98.37% detection probability, setting a new benchmark for accuracy in complex imaging scenarios.

This technology holds transformative potential for critical sectors, including forest fire early warning systems, autonomous drone navigation, and military threat assessment. While the current diffusion-based design prioritizes precision over inference speed, future research will focus on model distillation and mixed-precision optimization. This advancement marks a significant leap in computer vision, proving that combining generative models with state-space architectures can solve complex target-background separation problems.