北京信息科技大学学报（自然科学版）2025，Vol.40Issue(6)：59-68,10.DOI:10.16508/j.cnki.11-5866/n.2025.06.007
基于小波变换的多模态航拍车辆目标检测网络
Multimodal aerial vehicle object detection network based on wavelet transform
摘要
Abstract
The inherent discrepancies between visible and infrared modalities pose challenges in spatial and semantic alignment for vehicle detection under low-light conditions.Additionally,the limited resolution of infrared images complicates feature extraction for small target vehicles.To address these issues,a wavelet transform-based multimodal aerial vehicle detection network(WAVDNet)for unmanned aerial vehicle(UAV)perspectives was proposed.Firstly,a wavelet transform-based feature enhancement block was designed,leveraging high-frequency information to enhance feature extraction.Based on this,key feature vectors were screened to mitigate interference caused by redundant information across modalities.Furthermore,a deformable attention module was designed to adaptively adjust sampling points using high-frequency information,resolving spatial and semantic misalignment between modalities while enabling multi-level semantic adaptive fusion of visible and infrared modes features.Finally,experiments conducted on two benchmark datasets,DroneVehicle and VEDAI,demonstrate that the proposed method achieves mAP@0.5 scores of 81.7％and 90.7％,respectively,outperforming various state-of-the-art algorithms by margins of 1.4 and 3.5 percentage points over the second-best approaches,thus validating its effectiveness.关键词
可见光-红外/多模态融合/小波变换/特征筛选/可变形注意力Key words
visible-infrared/multimodal fusion/wavelet transform/feature selection/deformable attention分类
信息技术与安全科学引用本文复制引用
李洪玉,韩晶,吕学强..基于小波变换的多模态航拍车辆目标检测网络[J].北京信息科技大学学报（自然科学版）,2025,40(6):59-68,10.基金项目
国家自然科学基金项目(62171043) （62171043）
北京市自然科学基金项目(4254096) （4254096）
北京市教委科研计划科技一般项目(KM202311232003) （KM202311232003）
