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
于彦辉 1司占军 1张滢雪 1李雅静 2卢勇拾1
作者信息
- 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
- 折叠
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.
Abstract
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/CARAFEKey words
UAV Object Detection/Improved YOLOv7/Dynamic Sparse Attention/Partial Convolution/CARAFE引用本文复制引用
出版年
2024
NETL
NSTL
万方数据