航拍图像小目标检测算法设计  

Design of Small Target Detection Algorithm for Aerial Images

作　　者：于立君 孙超[1] 王辉[1] 徐博[1] 李广东 YU Lijun;SUN Chao;WANG Hui;XU Bo;LI Guangdong(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin 150001,China)

机构地区：[1]哈尔滨工程大学智能科学与工程学院,哈尔滨150001

出　　处：《实验室研究与探索》2024年第10期21-25,共5页Research and Exploration In Laboratory

基　　金：黑龙江省自然科学基金项目(LH2022F014);黑龙江省高等教育教学改革工程项目(SJGY20220082);黑龙江省高等教育学会高等教育科学研究规划课题项目(23GJZD002)。

摘　　要：针对传统检测算法在航拍图像小目标检测上准确率低,并存在误检、漏检等问题,提出一种基于YOLOv5的改进算法RBN-YOLOv5。设计基于RepVGG模块的C3RepBlock特征提取模块,增加小目标检测层更具判别性的浅层特征,通过局部和全局信息的联合表征获得更大的感受野;引入BiFormer注意力机制,提升模型检测精度,并基于归一化Wasserstein距离改进损失函数,增强小目标定位能力。在VisDrone2019数据集上的训练结果表明,RBN-YOLOv5相较于YOLOv5在检测精度上提高了9.8%,而且模型参数量大幅降低。In view of the low accuracy of the traditional detection algorithm in aerial image target detection,and the problems of false detection and missed detection,an improved algorithm RBN-YOLOv5 based on YOLOv5 was proposed.A C3RepBlock feature extraction module based on RepVGG was designed to add more discriminative shallow features of the small target detection layer,and a larger receptive field was obtained through the joint representation of local and global information.The BiFormer attention mechanism was introduced to improve the detection accuracy of the model,and the loss function was improved based on the normalized Wasserstein distance to enhance the localization ability of small targets.The training results on the VisDrone2019 dataset show that the detection accuracy of RBN-YOLOv5 is improved by 9.8%compared with the original YOLOv5 model,and the number of model parameters is greatly reduced.

关 键 词：目标检测 航拍图像 小目标 特征提取 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]