基于改进Yolov3的飞机目标检测算法  

Aircraft Object Detection Algorithm Based on Improved Yolov3

作　　者：张驰 王晓峰[1] 刘树光[2] 白宁宁 赵泽坤 胡幸福 ZHANG Chi;WANG Xiaofeng;LIU Shuguang;BAI Ningning;ZHAO Zekun;HU Xingfu(College of Science,Xi'an University of Technology,Xi'an 710048,China;School of Equipment Management and Unmanned Aerial Vehicle Engineering,Air Force Engineering University,Xi*an 710051,China)

机构地区：[1]西安理工大学理学院,西安710048 [2]空军工程大学装备管理与无人机工程学院,西安710051

出　　处：《无人系统技术》2023年第2期81-94,共14页Unmanned Systems Technology

基　　金：装备预研国防科技重点实验室基金(6142219200301);陕西省重点研发计划(2022GY-087)。

摘　　要：为了解决军用飞机目标检测过程中难以兼顾检测精度与检测速度的问题,提出了一种新的飞机目标检测算法。该算法建立在Yolov3的基础上,其特点是在保证检测速度的情况下,大幅提升检测精度。在提出的方法中,首先使用K-Means++聚类算法,解决了由随机初始化聚类中心带来的误差问题;其次对通过聚类得到的先验框(Anchors)进行线性拉伸,使其在贴合数据集目标大小的同时具有不同的尺度;再次,用CARAFE上采样算子构建上采样过程,使得网络能够捕捉到丰富的语义信息;最后,在网络中加入改进的通道显著性注意力机制CS-SE,使得网络能够有效关注图像前景内容,从而提高检测精度。实验表明,相比于Yolov3,所提方法 mAP@0.5增加了5.3%,mAP@0.5:0.95增加了8.0%,提高了飞机的目标检测准确率和可靠性,减少误判和漏判,使其在不同的气象条件、光线条件和目标形态下实现准确的目标检测。In order to address the issue of achieving a balance between detection accuracy and speed in military aircraft object detection,a new aircraft object detection algorithm is proposed.The algorithm is based on Yolov3,and its advantage is that the detection accuracy can be greatly improved while maintaining the detection speed.In the proposed method,the K-Means++clustering algorithm is used to solve the error problem caused by randomizing the initial cluster centers.Next,linearly stretch the anchors obtained by clustering algorithms so that it has a richer scale while fitting the size of the target data set.Then,the up-sampling process is constructed by CARAFE,which enables the network to capture rich semantic information.Finally,the network is equipped with an improved channel saliency attention mechanism(CS-SE)to improve detection accuracy by directing the network's attention to the foreground content of the image.The results of experiments show that compared to Yolov3 the mAP@0.5 of the proposed method increased by 5.3%,the mAP@0.5:0.95 increased by 8.0%.The proposed method enhances the accuracy and reliability of aircraft target detection,reduces false positives and false negatives,and enables accurate target detection under different weather conditions,lighting conditions,and target forms.

关 键 词：飞机目标检测 线性拉伸 先验框 CARAFE上采样算子 通道显著性注意力机制 边界框聚类 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]