Wavelet Interactive Mamba Fusion Network for UAV Small Object Detection with Visible–Thermal Data

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Accurate small object detection in low-altitude unmanned aerial vehicle (UAV) visible–thermal (RGB-T) scenarios remains challenging under complex illumination and background clutter due to thermal noise, weak cross-modal interaction, and attenuation of ultra-small target features. This paper proposes Wavelet Interactive Mamba Fusion Network (WIM-Net), a lightweight RGB-T detector with 10.4 million parameters. The network adopts a dual-branch backbone and introduces two modules for frequency-domain denoising and adaptive feature fusion. The Wavelet Interactive Mamba (WIM) module decomposes visible and thermal feature maps into multi-frequency sub-bands through a discrete wavelet transform and uses bidirectional Mamba interaction to model cross-modal dependencies with linear complexity. Instead of independent single-modal denoising, WIM uses low-frequency semantic cues to constrain high-frequency filtering, suppressing thermal noise while preserving target structures. The Cross-Modal Synergistic Fusion (CMSF) module builds bimodal mutual attention to adaptively combine visible texture and thermal radiation cues, reducing feature attenuation for tiny objects. Ablation studies show that WIM and CMSF contribute to noise suppression and small-object feature retention, respectively. Experiments on the RGBT-Tiny dataset demonstrate that WIM-Net achieves the best mean average precision at an intersection-over-union threshold of 0.5 (mAP50) of 52.0% and Precision of 79.9% among the evaluated single-modal and RGB-T detectors with only 10.4 million parameters, whereas its Recall and F1-score remain below those of QFDet and CALNet, reflecting a precision-oriented operating point. The compact architecture also preserves a favorable balance between detection accuracy and inference efficiency, with measured offline throughput on an NVIDIA Jetson Orin NX, while live airborne performance remains to be tested.