Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (15): 243-251.DOI: 10.3778/j.issn.1002-8331.2401-0209

• Graphics and Image Processing • Previous Articles     Next Articles

Detection Model for Small Objects in UAV Image Scene

ZHU Kunhuang, SUN Bo, MAO Guojun   

  1. School of Computer Science and Mathematics, Fujian University of Technology, Fuzhou 350000, China
  • Online:2024-08-01 Published:2024-07-30

面向无人机图像场景的小目标检测模型

朱堃煌,孙博,毛国君   

  1. 福建理工大学 计算机科学与数学学院,福州 350000

Abstract: Due to the flying altitude and shooting angle of UAV, many small targets can be seen in the captured images. The small objects have so few pixels and semantic information that can easily be disturbed by complex background information. Meanwhile, aggregation often takes place. Accurate detection of small targets is the key to improve the performance of the detection model for UAV. In this paper, a detection model called UAIDet for small objects is proposed. An adaptive channel fusion module is developed. In the feature fusion stage, the channel weights are dynamically learned to filter out the information conflicts between the different features levels, to improve the detection ability for small targets. In addition, an offset-sensitive loss function is developed for the location. In the convergence phase of the small object bounding box, the offset-sensitive term solves the sensitivity for geometric error by using root function. The model UAIDet is tested in dataset Visdrone2022, the mAP and AP50 reach 22.0% and 37.1% respectively, which is 3 and 4.7 percentage points higher than that of the benchmark model. The experiment in TinyPerson dataset shows 9.9% mAP and 29.1% AP50, improve 4.29 and 4.2 percentage points separately. The results have verified the robustness and effectiveness of UAIDet model.

Key words: object detection, unmanned aerial vehicles images, small object, feature fusion, loss function

摘要: 无人机因其飞行高度和拍摄角度的独特性,采集的遥感图像中存在大量小目标。小目标物体像素小、语义信息少,容易受背景信息干扰和出现聚集遮挡,是当前检测模型性能不佳的主要原因之一。提出一种面向小目标的无人机图像目标检测模型UAIDet(unmanned aerial vehicles images detector),从解决信息冲突和检测框回归难入手,提升模型的检测性能。其一,构建自适应的通道融合模块,在特征融合阶段动态学习通道权重以过滤不同尺度特征之间的信息冲突,抑制特征融合时的尺度不一致性,提高小目标物体的检测能力;其二,设计误差敏感定位损失函数,在小目标物体检测框的收敛阶段提出偏移敏感损失项以解决小目标对几何误差的敏感性,提高定位损失函数的鲁棒性,优化小目标物体的检测精度。在数据集Visdrone2022上对文章方法进行实验,mAP(means average precision)和AP50(average precision at IOU threshold 50%)分别达到了22.0%和37.1%,相较于基准模型分别提高3和4.7个百分点。TinyPerson数据集上的mAP和AP50为9.9%和29.1%,分别提高了4.29和4.2个百分点,证明UAIDet模型的有效性和鲁棒性。

关键词: 目标检测, 无人机图像, 小目标, 特征融合, 损失函数