改进YOLOv5的无人机影像小目标物体识别算法  

Improved Algorithm for Small Target Object Recognition in Drone Images Based on YOLOv5

作　　者：窦同旭 曾勇[1] 杨冲 李昂 

机构地区：[1]盐城工学院机械学院,江苏 盐城

出　　处：《建模与仿真》2023年第6期5395-5407,共13页Modeling and Simulation

摘　　要：无人机图像目标检测是流行的图像识别任务之一。然而,由于无人机航拍影像具有目标尺度变化大、背景复杂等特点,现有的检测器难以准确检测小目标。为了解决这个问题,本文提出了改进的算法模型YOLO-ADOP,用于提高在无人机影像中小目标的检测效果。首先,通过调整检测分支以提高模型在小目标下的检测能力。其次,将分类任务与回归任务解耦,设计了解耦头结合辅助训练头的预测头。最后,使用EIOU优化损失函数,使用OTA改进标签分配策略。通过VisDrone数据集的实验结果表明,改进后的YOLO-ADOP模型较YOLOv5模型在640分辨率和1536分辨率上的AP50指标分别提升了7.18%和4.65%。对比其他主流模型对于小目标拥有更好的检测效果,能够有效完成无人机航拍影像的小目标检测任务。Unmanned aerial vehicle (UAV) image object detection is a popular task in image recognition. How-ever, existing detectors face difficulties in accurately detecting small objects in UAV aerial images due to their large scale variations and complex backgrounds. To address this issue, this paper pro-poses an improved algorithm model called YOLO-ADOP, aiming to enhance the detection perfor-mance of small objects in UAV images. Firstly, the detection branch is adjusted to enhance the mod-el’s detection capability for small targets. Secondly, the classification task and regression task are decoupled, and a decoupled head with an auxiliary head is designed for prediction. Finally, the EIOU optimization loss function and OTA improved label assignment strategy are employed. Experi-mental results on the VisDrone dataset demonstrate that the improved YOLO-ADOP model achieves a 7.18% and 4.65% improvement in AP50 metric at resolutions of 640 and 1536, respectively, compared to the YOLOv5 model. It outperforms other mainstream models in detecting small targets and can effectively accomplish the task of small target detection in drone aerial images.

关 键 词：无人机影像 YOLOv5 小目标 辅助训练头 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]