跨层注意力交互下的多特征交叉无人机图像检测  

Multi-feature cross UAV image detection algorithm under cross-layer attentional interaction

作　　者：张志豪 杜丽霞[1] 侯越[1] 郝紫微 尹杰 ZHANG Zhihao;DU Lixia;HOU Yue;HAO Ziwei;YIN Jie(College of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730000,China)

机构地区：[1]兰州交通大学电子与信息工程学院,甘肃兰州730000

出　　处：《光学精密工程》2024年第24期3616-3631,共16页Optics and Precision Engineering

基　　金：国家自然科学基金(No.62063014,No.62363020);甘肃省自然科学基金(No.22JR5RA365)。

摘　　要：针对无人机交通巡检存在的航拍图像背景复杂、目标密集、目标尺度分布不均匀等问题,提出了一种跨层注意力交互下的多特征交叉无人机目标检测算法(Multi-feature Crossover under cross-layer Attentional Interaction,MCAI)。首先,在主干网络部分设计自适应跨层注意力交互模块(Adaptive Cross-layer Attentional Interaction,ACAI),使模型聚焦于关键特征区域以实现对全局关键特征信息的有效筛选,从而淡化复杂背景的影响。其次,设计了一种基于可变形自注意力的编码器(Deformable Encoder,DeEncoder),该编码器通过扩大特征层感受野,弥补丢失的目标特征。最后,为有效识别区域内不同尺度的微小目标,提出多尺度交叉融合模块(Multi-scale cross fusion module,MSCF),该模块通过小波变换和特征表示结合的方式,融合浅层空间信息和深层语义信息,以有效捕捉不同尺度目标的细粒化特征。通过在Vis⁃Drone2019-DET、BDD-100K数据集和LZTraffic Video数据集上的实验结果表明,MCAI相较于RT-DETR模型,mAP0.5分别提高了3%,2.2%和4.5%,显著提高了无人机巡检的检测精度。此外,在阴雨场景中,MCAI的mAP0.5相较于RT-DETR模型提升了2.1%,极端天气鲁棒性表现更优。Aiming at the problems of complex background of aerial images,dense targets,and uneven tar⁃get scale distribution in UAV traffic inspection,a multi-feature crossover under cross-layer attentional in⁃teraction(Multi-feature crossover under cross-layer attentional interaction,MCAI)UAV target detection algorithm was proposed.Firstly,an Adaptive Cross-layer Attentional Interaction(Adaptive Cross-layer Attentional Interaction,ACAI)module was designed in the backbone network part so that the model fo⁃cused on the key feature regions to achieve effective screening of global key feature information,thus fad⁃ing the influence of the complex background.Secondly,a deformable self-attentive encoder(Deformable Encoder,DeEncoder)was designed,which compensated for the lost target features by expanding the fea⁃ture layer receptive field.Finally,in order to effectively identify tiny targets at different scales in the re⁃gion,the multi-scale cross-fusion module(Multi-scale cross fusion module,MSCF)was proposed,which fused shallow spatial information and deep semantic information by combining the wavelet transform and feature representation in order to efficiently capture the fine-grained features of targets at different scales.The experimental results on the VisDrone 2019-DET,BDD-100K dataset,and LZTraffic Video dataset show that MCAI improves mAP0.5 by 3%,2.2%,and 4.5%,respectively,compared to the RT-DE⁃TR model,which significantly improves the detection accuracy of the UAV inspection.In addition,in the cloudy and rainy scenario,the mAP0.5 of MCAI improves by 2.1%compared to the RT-DETR model,with better extreme weather robustness performance.

关 键 词：无人机巡检 目标检测 注意力交互 编码器 小波变换 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]