融合BiFPN和改进Yolov3-tiny网络的航拍图像车辆检测方法  

Vehicle Detection Method in Aerial Images Based on BiFPN and Improved Yolov3-tiny Network

作　　者：鲁博 瞿绍军 LU Bo;QU Shao-jun(College of Information Science and Engineering,Hunan Normal University,Changsha 410081,China;Hunan Xiangjiang Artificial Intelligence Academy,Hunan Normal University,Changsha 410081,China)

机构地区：[1]湖南师范大学信息科学与工程学院,长沙410081 [2]湖南师范大学湖南湘江人工智能学院,长沙410081

出　　处：《小型微型计算机系统》2021年第8期1694-1698,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(12071126)资助;湖南省教育厅科学研究项目(19C1149)资助。

摘　　要：航拍车辆检测能够从航拍图像中检测出车辆的位置和种类信息,可以为与车辆相关的应用问题提供重要信息支撑.为满足实际运用的要求,需要保证目标检测算法的速度与精度.通常精度较高的深度学习目标检测模型速度较慢,而速度优势明显的简单网络结构算法精度难以满足实际需求.针对这一需求与特性,提出了一种航拍图像车辆目标检测算法,在牺牲微量速度的情况下极大提升检测精度.选择Yolov3-tiny的主干网作为主干网,同时引入BiFPN特征金字塔结构与主干网结合以增进多层次特征的融合,提出一种新的上采样结构,替换原有网络中的上采样结构.在vedai数据集上进行实验,与Yolov3相比,每秒检测帧(fps)与检测精度(map)分别提升了73%、6%.与Yolov3-tiny相比,虽然fps下降了5%,帧数减少2帧,但map提升了9.6%,与ssd、mobilenetv3、efficientnet相比,无论是检测精度还是检测速度都具有明显优势,提出方法的map达到了87.6%.本文所提出的Bi-Yolov3-tiny模型与上采样结构综合了轻量级网络和复杂结构网络的优点,使得航拍图像车辆检测的效果更加精准的同时也保持了较快的速度.Aerial vehicle detection can detect the location and category of vehicles from aerial images,which can provide important information support for vehicle related application problems.In order to meet the requirements of practical application,it is necessary to ensure the speed and accuracy of target detection algorithm.Generally,deep learning object detection model with high accuracy is slow,while the simple network structure algorithm with obvious speed advantage is difficult to meet the accuracy demand.In view of this demand and characteristics,a vehicle detection algorithm in aerial images is proposed,which greatly improves the detection accuracy at the expense of a small amount of speed.The backbone network of Yolov3-tiny is selected as the backbone network,and the feature pyramid structure of BiFPN is introduced to combine with the backbone network to enhance the fusion of multi-level features.A new up sampling structure is proposed to replace the upsampling structure in the original network.Experiments were conducted on vedai dataset.Compared with Yolov3,FPS and map are improved by 73%and 6%respectively.Compared with Yolov3-tiny,although FPS decreased by 5%and frame number decreased by 2 frames with no significant change,map increased by 9.6%.Compared with SSD,mobilenetv3 and efficientnet,Bi-Yolov3-tiny have obvious advantages in both detection accuracy and detection speed reaching 87.6%.Bi-Yolov3-tiny model and up sampling structure proposed in this paper combine the advantages of lightweight network and complex structure network,which makes the vehicle detection effect of aerial images more accurate,but also maintains a faster speed.

关 键 词：车辆检测 目标检测 Yolov3 上采样 特征金字塔 小目标 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]