无人机视角下的小目标检测方法研究  

Study on Small Object Detection Methods from Unmanned Aerial Vehicle Perspective

作　　者：于彦辉 司占军[1] 张滢雪 李雅静 卢勇拾 YU Yan-hui;SI Zhan-jun;ZHANG Ying-xue;LI Ya-jing;LU Yong-shi(College of Artificial Intelligence,Tianjin University of Science and Technology,Tianjin 300457,China;Engineering Research Center of Integration and Application of Digital Learning Technology,Ministry of Education,The Open University of China,Beijing 100039,China)

机构地区：[1]天津科技大学人工智能学院,天津300457 [2]国家开放大学数字化学习技术集成与应用教育部工程研究中心,北京100039

出　　处：《印刷与数字媒体技术研究》2024年第1期60-69,共10页Printing and Digital Media Technology Study

摘　　要：针对传统卷积网络对无人机图像中小目标检测精度低和误检问题,本研究提出一种改进的无人机图像小目标检测算法,提高航拍检测精度。本算法采用YOLOv7作为基本框架,并在空间金字塔池化中融入动态稀疏注意力,形成SPPCSPC-B模块,增强了对小目标的检测能力。同时,本算法使用局部卷积替代了高效聚合网络中的部分群卷积,形成ELAN-P模块,提高了检测速度。最后,使用轻量级上采样算子CARAFE对特征进行重组,进一步提高了检测精度。在Aerial-airport数据集上的实验结果表明,本算法在参数量减少9%、模型缩小8%的情况下,检测精度达94.7%,召回率达到90.8%,比基准算法提高了3.9个百分点,且有效改善了小目标误检、漏检现象。In response to the challenges posed by the low accuracy and false positives in small object detection in unmanned aerial vehicle(UAV)images using traditional convolutional networks,an enhanced algorithm for small object detection in UAV images was presented in this study,aimed at improving the precision of aerial surveillance.YOLOv7 was employed as the foundational framework for this algorithm,incorporating Dynamic Sparse Attention into Spatial Pyramid Pooling to form SPPCSPC-B module to enhance the detection capabilities for small objects.Simultaneously,the algorithm replaced certain group convolutions in the efficient aggregation network with local convolutions,forming the Enhanced Lightweight Aggregation Network(ELAN-P)module,thereby boosting detection speed.Finally,the feature recombination was accomplished using the lightweight upsampling operator CARAFE,further augmenting detection accuracy.Experimental results conducted on the Aerial-airport dataset demonstrated that,with a 9%reduction in model parameters and an 8%reduction in model size,the proposed algorithm achieves a detection accuracy of 94.7%and a recall rate of 90.8%.This represents an improvement of 3.9 percentage points over the baseline algorithm and effectively mitigates issues related to false positives and false negatives in small object detection.

关 键 词：无人机目标检测 YOLOv7 动态稀疏注意力 部分卷积 CARAFE 

分 类 号：TN911.73[电子电信—通信与信息系统] TP753[电子电信—信息与通信工程]