多尺度特征聚合扩散和边缘信息增强的小目标检测算法  

Multi-Scale Feature Aggregation Diffusion and Edge Information Enhancement Small Object Detection Algorithm

作　　者：江旺玉 王乐 姚叶鹏 毛国君 JIANG Wangyu;WANG Le;YAO Yepeng;MAO Guojun(College of Computer Science and Mathematics,Fujian University of Technology,Fuzhou 350118,China;Fujian Provincial Key Laboratory of Big Data Mining and Applications,Fujian University of Technology,Fuzhou 350118,China;Technology Innovation Center of Factored Transaction Data in Tourist Attractions,Ministry of Culture and Tourism,Fuzhou 350000,China;Institute of Information Engineering,Chinese Academy of Sciences,Beijing 100084,China)

机构地区：[1]福建理工大学计算机科学与数学学院,福州350118 [2]福建理工大学福建省大数据挖掘与应用技术重点实验室,福州350118 [3]景区交易数据要素化文化和旅游部技术创新中心,福州350000 [4]中国科学院信息工程研究所,北京100084

出　　处：《计算机工程与应用》2025年第7期105-116,共12页Computer Engineering and Applications

基　　金：国家重点研发项目(2019YFD0900800/05);国家自然科学基金(61773415);福建省自然科学基金(2023J01954);中国科学院青年创新促进会(2022159);中央引导地方科技发展专项(2023L3030)。

摘　　要：无人机航拍图像中,目标尺寸变化剧烈、背景复杂且小目标比例较高等特点为目标检测任务带来巨大挑战。尽管现有的基于卷积的目标检测算法能有效获取空间信息,但在实现不同尺度特征的全局交互及边缘细节信息的有效利用上仍存在不足。因此,提出了一种结合多尺度特征聚合扩散和边缘信息增强的小目标检测算法ADEYOLO。构建了多尺度特征聚合扩散金字塔网络(MFADPN),通过在中间层聚合不同层级特征,并将其直接扩散至相邻层以缩短传播路径,有效减少了信息在传递过程中的损失,增强了模型的多尺度表达能力,显著提升了对不同尺度目标的检测能力。设计了自适应上下文融合模块(ACFM),利用通道注意力机制自适应地调整不同特征图的贡献,进一步强化多尺度特征的融合效果,使得重要特征在信息融合过程中更加突出。提出的C2f-Sobel模块通过额外分支结合Sobel算子来提取图像的边缘信息,从而为模型提供了更丰富的细节信息,提升了其在复杂场景下目标定位能力。实验结果表明,ADE-YOLO相较于基线YOLOv10s,在VisDrone2019和TinyPerson数据集上分别提高了8.6个百分点和4.0个百分点(mAP0.5),并且在与其他先进模型的对比中也展示了显著的优势。In drone aerial images,characteristics such as drastic changes in object sizes,complex backgrounds,and a high proportion of small objects pose significant challenges for object detection tasks.Although existing convolution-based object detection algorithms can effectively capture spatial information,there are still shortcomings in achieving global interaction of features at different scales and effectively utilizing edge detail information.Therefore,this study proposes a small object detection algorithm,ADE-YOLO,which combines multi-scale feature aggregation diffusion and edge information enhancement.First,a multi-scale feature aggregation diffusion pyramid network(MFADPN)is constructed.This network aggregates features of different levels in intermediate layers and directly diffuses them to adjacent layers to shorten the propagation path,effectively reducing information loss during the transmission process,enhancing the model’s multiscale expressive ability,and significantly improving the detection capability for objects of varying scales.Next,an adaptive context fusion module(ACFM)is designed,which uses a channel attention mechanism to adaptively adjust the contribution of different feature maps,further strengthening the fusion effect of multi-scale features and making important features more prominent during the information fusion process.Finally,the proposed C2f-Sobel module extracts edge information from images by combining an additional branch with the Sobel operator,thereby providing the model with richer detail information and enhancing its object localization ability in complex scenes.Experimental results show that ADEYOLO improves mAP0.5 by 8.6 percentage points and 4.0 percentage points compared to the baseline YOLOv10s on the VisDrone2019 and TinyPerson datasets,respectively,and demonstrates significant advantages when compared with other advanced models.

关 键 词：小目标检测 航拍图像 特征金字塔 自适应特征融合 边缘信息 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]