基于改进Yolov5l的航空小目标检测算法  

Aerial small target detection method based on improved Yolov5l

作　　者：戴得恩 朱瑞飞 陈长征[1] 秦磊 马经宇 DAI De-en;ZHU Rui-fei;CHEN Chang-zheng;QIN Lei;MA Jing-yu(Space Optical DepartmentⅡ,Changchun Institute of Optics,Fine Mechanics and Physics,Chinese Academy of Sciences,Changchun 130033,China;Daheng College,University of Chinese Academy of Sciences,Beijing 100049,China;Data DepartmentⅢ,Changguang Satellite Technology Limited Company,Changchun 130102,China)

机构地区：[1]中国科学院长春光学精密机械与物理研究所空间光学研究二部,吉林长春130033 [2]中国科学院大学大珩学院,北京100049 [3]长光卫星技术有限公司数据三室,吉林长春130102

出　　处：《计算机工程与设计》2023年第9期2610-2618,共9页Computer Engineering and Design

基　　金：国家自然科学基金项目(62105328);国家重点研发计划基金项目(2019YFE0127000);吉林省重大科技专项基金项目(20200503002SF)。

摘　　要：针对航空图像小目标检测存在的检测精度低、误检与漏检严重等问题,提出一种基于改进Yolov5l的航空小目标检测算法(AS-Yolov5)。在Yolov5的主干特征提取网络中引入空洞卷积,使用Transform的Decode模块,在特征融合网络中新增检测头,FPN+PAN特征融合时设置融合权重,输出端采用SE-Net注意力机制,测试时进行多尺寸输入及测试时间增强(TTA)。算法在visdron2021数据集上进行验证,实验结果表明,AS-Yolov5的均值平均精度@0.5(mAP@0.5)为41.0%,较Yolov5l的28.5%提升12.5%,有效提高Yolov5l难以在远距离、暗环境、密集分布和图像模糊的场景下的小目标检测能力。For the detection of small targets in aerial images,there are problems such as low detection accuracy,false detection and serious missed detection,an aviation small target detection algorithm was proposed(AS-Yolov5)based on improved Yolov5l.In the backbone,the atrous spatial pyramid pooling model was inserted,so that the receptive field was expanded without increasing the number of layers.Transformer encoding block was inserted to obtain layer global information.In the neck,a new detection head was added to improve ability of face dense and complex targets.Fusion weights were added to focus on the relationship between different layers and small targets.The detection head used SE-Net attention mechanism,by self-learning the channel information,the learning ability of small targets was improved.Multi-size input and test time enhancement(TTA)were performed during testing to improve model recognition capabilities.The proposed detection method was experimentally verified on the visdron2021 data set.Experimental results show that the mAP@0.5 of AS-Yolov5 is 41.0%,which is 12.5%higher than that of Yolov5l.This method effectively improves the detection ability of aviation small targets.

关 键 词：航空小目标检测 Yolov5l模型 空洞卷积 SE-Net注意力模块 权重融合 深度学习 目标检测 

分 类 号：TP394.1[自动化与计算机技术—计算机应用技术]