基于改进YOLOX的无人机航拍图像小目标检测算法  

A small target detection algorithm of UAV aerial photography images based on improved YOLOX

作　　者：潘翔 陈前斌[1,2] 黄昂 罗佳[1,2] 唐伦 PAN Xiang;CHEN Qianbin;HUANG Ang;LUO Jia;TANG Lun(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China;Key Laboratory of Mobile Communication Technology,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065 [2]重庆邮电大学移动通信技术重点实验室,重庆400065

出　　处：《南京邮电大学学报（自然科学版）》2024年第1期90-100,共11页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(62071078);重庆市教委科学技术研究项目(KJZD-M201800601);川渝联合实施重点研发(2021YFQ0053)资助项目。

摘　　要：在无人机上运用目标检测技术具有广泛的应用前景,但和自然拍摄的图像不同,无人机航拍的图像更加复杂,且大多数为小目标。而现有的检测算法缺乏对小目标的特征提取能力,从而导致严重的误检和漏检问题。针对上述问题,提出一种基于YOLOX框架的高效的无人机小目标检测算法。首先,在特征融合网络中增加一层检测微小目标的特征融合结构,通过利用浅层特征图中丰富的位置信息和轮廓信息来加强网络对小目标的识别能力;同时,为了防止额外参数的增加,将减少头网络中的一层卷积层并缩减其通道数。其次,提出一种通道-空间注意力机制模块(Channel Spatial Attention Module,CSAM),利用最优的权重分配使网络聚焦于特征图中小目标密集的区域。最后,提出一种带位置引导的标签分配策略(LB-SimOTA),根据每个预测框和真实框的交并比(IOU)的大小,分别配以不同的权重,以改善网络中整体预测框的质量。在小目标居多的数据集VisDrone2019上的实验结果表明,文中提出的算法和YOLOX-S相比,针对车和人的检测精度提升了8.63%,检测速度FPS也可达到86,因此更适合在无人机对地面小目标检测的场景下部署。Target detection for unmanned aerial vehicles(UAVs)has great potential.However,different from natural images,aerial images taken by UAVs are more complex,containing considerable small targets.This has imposed challenges for the existing detection algorithms,since they lack the feature extraction ability for small targets,leading to serious problems of false detection and missing detection.Therefore,this paper proposes an efficient small target detection algorithm based on YOLOX framework.Firstly,a layer of feature fusion structure is added to the feature fusion network to detect small targets,and the recognition ability of the network to small targets is enhanced by using the abundant position information and contour information in the shallow feature map.,A layer of convolution layer in the head network is reduced and the number of channels is reduced,thus the increase of additional parameters is prevented.Secondly,a channel spatial attention module(CSAM)is proposed.It uses the optimal weight allocation to drive the network to focus on the small and target-dense regions in the feature graph.Finally,a position-guided label allocation strategy(LB-SimOTA)is proposed.According to the intersection ratio(IOU)between each prediction box and the real box,different weights are assigned to improve the quality of the overall prediction box in the network.Experimental results on VisDrone2019,a data set with most small targets,show that compared with those of YOLOX-S,the detection accuracy of the proposed algorithm for vehicles and people is improved by 8.63%,and the detection speed is up to 86 FPS.The proposed algorithm is more suitable for the scenario of UAV detection of small targets on the ground.

关 键 词：无人机 小目标检测 多尺度检测 注意力机制 标签分配策略 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]