基于注意力与通道重排的无人机对地目标检测算法
Detection algorithm of ground target based on attention and channel rearrangement for UAV
李佳一 1闫振纲 2闫克丁 3赵英然 1檀蕊莲 4梁超5
作者信息
- 1. 西安工业大学 电子信息工程学院,西安 710021;北京航宇创通技术股份有限公司,西安 710038
- 2. 西安工业大学 电子信息工程学院,西安 710021;西安现代控制技术研究院,西安 710065
- 3. 西安工业大学 电子信息工程学院,西安 710021
- 4. 北京航宇创通技术股份有限公司,西安 710038
- 5. 西安现代控制技术研究院,西安 710065
- 折叠
摘要
无人机自主察打对地攻击场景中,针对无人机作战时效性强,地面目标识别场景复杂,存在模型训练、推理速度慢,小目标检测漏检、误检的问题,提出一种基于注意力机制与通道重排思想的无人机对地目标检测算法.该算法引入 CA(coordinate attention)注意力机制,可提高网络对关注部分的特征提取能力;且对主干网络进行通道重排(channel shuffle)轻量化处理,可有效减少多次卷积造成的特征损失;最后,为提升战时训练及推理速度,替换部分激活函数为 H-Swish,优化其损失函数为CIoU(complete intersection over union).实验证明:采用改进的新算法,提升了28.4%训练速度,目标识别的平均精度均值(mean average precision,mAP)达99.1%,可实现最小目标检测为19∗25像素,经TensorRT加速后检测速率达72.99 FPS,满足实时检测需求,针对复杂地形下的坦克小目标检测性能较好.
Abstract
In the context of autonomous unmanned aerial vehicle(UAV)ground attacks,there are challenges such as time-sensitive UAV combat,complex ground target recognition scenarios,slow model training and inference speeds,and missed and false detections of small targets.To address these challenges,a novel UAV-to-ground target detection algorithm based on attention mechanism and channel shuffle ideas is proposed.The algorithm incorporates the Coordinate Attention(CA)mechanism to enhance the network's feature extraction capability for specific regions of interest.Additionally,the main network is subject to lightweight processing using channel shuffle to mitigate feature loss caused by multiple convolutions.Lastly,the algorithm replaces some activation functions with H-Swish and optimizes the loss function as Complete Intersection over Union(CIoU)to improve the training and inference speed during wartime.Experimental results demonstrate that the proposed algorithm achieves a 28.4%improvement in training speed,a mean Average Precision(mAP)of 99.1%for target recognition,can detect targets as small as 19×25 pixels,and a detection rate of 72.99 FPS after acceleration using TensorRT,meeting the requirements for real-time detection.Moreover,the algorithm shows strong performance in detecting small targets such as tanks in complex terrain.
关键词
小目标检测/深度学习/注意力机制/通道重排/轻量化模型Key words
small object detection/deep learning/attention mechanism/channel shuffle/lightweight引用本文复制引用
出版年
2024
NETL
NSTL
万方数据