IMFNet: Infrared multispectral fusion for camouflaged personnel detection in jungle environments
Haoyu Wang, Zhixiang Xue, Junlong Zheng, Feng Huang, Ying Shen · Optics & Laser Technology · 2026
Detecting camouflaged personnel in military operations is severely hampered by complex jungle illumination, where shadows distort visible-light reflectance and cause a substantial decline in detector performance. This paper presents a multimodal acquisition and preprocessing system combining near-infrared multispectral (NIRMS) with long-wave infrared (LWIR) imaging, and proposes a feature-level fusion framework—IMFNet (Infrared Multispectral Fusion Network). Specifically, the system employs multispectral complementary feature extraction to distill discriminative spatial and spectral information. IMFNet incorporates an efficient fusion mechanism based on the State Space Model, utilizing multi-directional scanning to enhance global contour modeling. Furthermore, IMFNet incorporates a feature extraction architecture with 3D attention and introduces a dynamic loss function to improve the robustness of feature capture and localization for strongly camouflaged personnel. To support this research, we construct a real-world jungle-shadow infrared multispectral camouflaged personnel dataset (JSIMCPD). Experimental results demonstrate that by combining the three-band NIRMS (836/868/938 nm) with LWIR and employing spatial-information compensation, IMFNet achieves a mAP 0.5 of 94.7% and a Recall of 88.3%, outperforming the runner-up method MSTF (MultiSpectral Transformer Fusion) by 1.7% and 3.4%, respectively, while operating at 35 FPS to meet real-time requirements. This performance provides effective technical support for near-ground-perspective battlefield situational awareness and personnel rescue tasks.