摘要
针对无人机航拍时物体尺度变化大,检测目标大多较小且物体较密集的问题,提出一种混合特征增强结构(mix feature enhancement, MFE)方法。通过在超分辨率方法中加入注意力机制以增强小目标信息提取,利用一种新的特征层融合计算方法,加强不同特征层间的融合效率,提高了中小型目标的检测精度;设计了尾端感受野扩大层以扩大尾端特征层感受野,使检测头可接收丰富的物体信息来定位并区分密集物体。实验在数据集VisDrone2021的测试集上进行测试,MFE-YOLOX网络的AP50结果为47.78%,在参数量、计算量与原网络相近的情况下精度提高了9.43个百分点。
A mixed feature enhancement(MFE)method is proposed for the problem that the object scale varies greatly during UAV aerial photography,and most of the detection targets are small and dense objects.First,an attention mechanism is added to the super-resolution method to enhance small target information extraction;then,a new fusion calculation method of feature layers is proposed to enhance the fusion efficiency between different feature layers and improve the detection accuracy of small and medium-sized targets.Finally,a tail-end perceptual field expansion layer is designed to expand the tail-end feature layer perceptual field so that the detection head can receive rich object information to locate and distinguish dense objects.The experiments are tested on the test set of dataset VisDrone2021,and the results show that the AP50 result using the MFE-YOLOX network is 47.78%,and the accuracy is improved by 9.43%with similar number of parameters and computational load compared to the original network.
作者
马俊燕
常亚楠
MA Junyan;CHANG Ya’nan(College of Mechanical Engineering,Guangxi University,Nanning 530004,P.R.China;Key Laboratory of Guangxi Manufacturing System and Advanced Manufacturing,Guangxi University,Nanning 530004,P.R.China)
出处
《重庆邮电大学学报（自然科学版）》
CSCD
北大核心
2024年第1期128-135,共8页
Journal of Chongqing University of Posts and Telecommunications(Natural Science Edition)
基金
国家自然科学基金项目(52165062)
广西自然科学基金重点项目(2020JJD160004)~~。
关键词
小目标检测
无人机
注意力机制
特征融合
YOLOX
small object detection
unmanned aerial vehicle
attention mechanism
feature fusion
YOLOX
