基于多光谱交互注意力融合的多尺度无人机小目标检测  

Multiscale UAV small target detection based on multispectral interactive attention fusion

作　　者：吴长柯 陈虎[1] 潘涛 黄菊 刘洪[1,3] 张萍 吴志红[3] 苏强 WU Chang-Ke;CHEN Hu;PAN Tao;HUANG Ju;LIU Hong;ZHANG Ping;WU Zhi-Hong;SU Qiang(National Key Laboratory of Fundamental Science on Synthetic Vision,Sichuan University,Chengdu 610065,China;61287 Troops,Chengdu 610036,China;College of Computer Science,Sichuan University,Chengdu 610065,China;Dongfang Electric Corporation,Chengdu 611731,China;School of Optoelectronic Science and Engineering,University of Electronic Science and Technology of China,Chengdu 611731,China)

机构地区：[1]四川大学视觉合成图形图像技术重点学科实验室,成都610065 [2]61287部队,成都610036 [3]四川大学计算机学院,成都610065 [4]中国东方电气集团有限公司,成都611731 [5]电子科技大学光电科学与工程学院,成都611731

出　　处：《四川大学学报（自然科学版）》2024年第3期136-143,共8页Journal of Sichuan University(Natural Science Edition)

基　　金：国家自然科学基金重点项目(U20A20161)。

摘　　要：针对无人机检测中存在的目标较小、受背景环境影响大、以及多光谱特征难以深度融合等问题,本文提出了针对无人机小目标检测的多尺度多光谱交互注意力融合目标检测模型.首先,将骨干网络设计为双流网络,分别提取不同尺度红外和可见光特征,并增加小目标检测层和BiFPN级联操作,提升对无人机小目标特征的提取能力.其次,创新性的设计了多光谱交互注意力融合模块,在该融合模块的指导下,网络可以在不同尺度融合红外和可见光模态的信息,使红外和可见光的特征进行深度聚合,发挥各自模态的优势,指导开展无人机小目标检测.实验结果表明,与最先进的多光谱目标检测模型相比,本文提出的模型在FLIR、LLVIP两个公开的多光谱目标检测数据集上都达到了优越的性能,在构建的多光谱无人机数据集上,本文提出的模型有效提升了无人机的检测精度和鲁棒性.Aiming at the challenges in UAV detection,such as small target size,significant influence by background environment,and difficulties in deep fusion of multispectral features,this paper proposes a multiscale multispectral interactive attention fusion target detection model for small UAV target identification.Firstly,the backbone network is designed as a dual-flow network to extract the infrared and visible features at different scales,with the addition of small target detection layer and BiFPN cascade operation to enhance the ex‐traction capability of UAV small target features.Secondly,an innovative multispectral interactive attention fusion module is designed.Under the guidance of this fusion module,the network can fuse the infrared and visible modalities’information at different scales,allowing for deep aggregation of their repective features,leveraging the strengths of each modality,and guiding the UAV small target detection.The experimental results demonstrate that compared to state-of-the-art multispectral target detection models,the model proposed in this paper achieves superior performance on the two public multispectral target detection datasets,FLIR and LLVIP.Furthmore,on the constructed multispectral UAV dataset,the proposed model effectively improves UAV detection accuracy and robustness of UAV.

关 键 词：无人机检测 小目标检测 多光谱交互注意力融合 多尺度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]