聚焦小目标的航拍图像目标检测算法

Focusing on Small Objects Detector in Aerial Images

与通用目标检测不同,无人机(Unmanned Aerial Vehicle,UAV)航拍图像目标检测主要面临两个难题:(1)远距离观察下存在大量小尺寸目标,难以与背景区分;(2)大量区域中目标密集且存在严重遮挡.因此,将通用目标检测器直接应用于航拍图像会导致检测精度下降.本文提出一种聚焦小目标的航拍图像目标检测算法(Focusing on Small objects Detector in aerial images,FocSDet).针对小目标,通过密集高级组合(Dense Higher-Level Composition,DHLC)模式连接双Swin-Transfomer骨干网络,并和特征金字塔(Feature Pyramid Networks,FPN)结合,构建小目标特征聚合网络作为FocSDet的骨干网络,可丰富单层特征表达并提升对图像全局信息的利用,在不损失大目标语义信息的同时得到对小目标更好的特征描述,有效提升了小目标检测能力;针对区域密集遮挡,提出任务平衡样本分配策略,区别于现有样本分配策略只依赖定位位置,本文所提出的策略中样本匹配质量评价分数由定位位置信息和预测分类分数共同构成.基于该新评价分数不断迭代更新样本分配和监督网络优化,取得了更高质量的预测结果.最后,在检测头的分类和回归分支中引入层注意力构成增强检测头,进一步提升了小目标的检测性能.在Visdrone无人机数据集、CARPK航拍数据集上的实验表明,本文提出的FocSDet相较于现有方法ATSS和VFNET,在Visdrone上平均精度(Average Precision,AP)分别提升2%和0.6%,小目标APs分别提升2.6%和1.2%;在CARPK上AP分别提升2.2%和1.7%,小目标APs分别提升5.2%和5.0%.

Different from general object detection in natural images,object detection in unmanned aerial vehicle(UAV)aerial images mainly faces these challenges such as large number of small objects in remote observation,which is difficult to distinguish from the background,and dense objects with serious occlusion in lots of areas.Therefore,the direct application of general object detector to aerial images will lead to the decline of detection performance.In this paper,an aerial image object detection algorithm focusing on small objects(FocsDet)is proposed.For small objects,a small object feature aggregation network is designed,which connects the dual Swin-Transfomer backbone network through dense higher-level composition(DHLC)mode and combines with feature pyramid networks(FPN),so as to improve the utilization of global image information,enrich single-layer feature expression,and obtain better feature description of small objects with-out losing semantic information of large objects.It effectively improves the detection performance of small object.For re-gional dense occlusion,a task-balance label assignment is proposed,in which the label matching quality evaluation score is composed of location cost and classification cost,which is different from the existing evaluation score which only depends on location cost.Based on the evaluation score,label assignment and supervision network optimization are updated iterative-ly,so as to achieve better prediction results.Finally,layer attention is introduced into the classification and regression branch-es of the detection head to form enhanced detection head,which further improves the detection performance of small objects.Experiments on Visdrone dataset and CARPK dataset show that compared with the existing methods such as ATSS and VFNET,the average precision(AP)of FocsDet is improved by 2%and 0.6%,APs is improved by 2.6%and 1.2%on Vsidrone da⁃taset respectively.On CARPK dataset,AP increases by 2.2%and 1.7%,and APs increases by 5.2%and 5.0%respectively.