增采样
计算机科学
人工智能
目标检测
计算机视觉
特征(语言学)
特征提取
传感器融合
融合
模式识别(心理学)
对象(语法)
图像融合
雷达探测
图像分割
特征检测(计算机视觉)
遥感
图像处理
杂乱
雷达跟踪器
作者
Zhe Yuan,Wenchao Meng,Baolong Guo,Zhihui Wen,Junbin Zhuang
标识
DOI:10.1109/tgrs.2026.3709139
摘要
Small object detection in uncrewed aerial vehicle (UAV) remote sensing images faces severe challenges, including tiny object size, weak features, and strong background clutter. Single-modality methods depend heavily on lighting and fail in nighttime or complex scenes, while traditional multimodal fusion schemes incur high computational cost and often suffer from the loss of fine-grained features for small object. We propose YOLO-WEH, a multimodal small object detection algorithm with cross-modal feature fusion and lightweight up/downsampling. The dual-branch backbone extracts complementary features from visible (RGB) and infrared (IR) data, and a wavelet-Fourier dual-domain cross self-attention fusion (WFDF) module explicitly preserves high-frequency details and filters thermal noise for semantic alignment. We also introduce a high-frequency enhanced lightweight downsampling (HFE-LDS) module and an enhanced lightweight content-aware reassembly of features (EL-CARAFEs) upsampling module, which recover the geometric and structural boundaries of infinitesimal objects while maintaining computational efficiency. On three public datasets (DroneVehicle, LLVIP, and FLIR), YOLO-WEH achieves mean average precision (mAP) of 74.9%, 97.4%, and 73.1%, using only 5.0 M parameters and 13.3 GFLOPs, outperforming state-of-the-art (SOTA) methods. Inference speeds reach 90.4, 112.4, and 121.5 frames per second (FPS), respectively, demonstrating high accuracy, lightweight design, and real-time capability. Ablation studies and visualizations confirm each core module’s effectiveness and highlight the algorithm’s advantage in reducing missed and false alarms in multisource heterogeneous and extreme scenarios.
科研通智能强力驱动
Strongly Powered by AbleSci AI