基于YOLOv7-tiny改进的遥感小目标检测算法
Improved remote sensing tiny object detection algorithm based on YOLOv7-tiny
王子龙 1荣杰2
作者信息
- 1. 江苏科技大学计算机学院 镇江 212100
- 2. 南京邮电大学材料科学与工程学院 南京 210000
- 折叠
摘要
为了解决遥感图像中小目标的误检、漏检难题,提出了一种改进的YOLOv7-tiny算法.首先,引入高效多尺度注意力模块(efficient multi-scale attention,EMA),基于此设计了多尺度特征提取模块ELAN-EMA,这大大增强了骨干网络对于多尺度特征的提取能力;其次,在特征金字塔网络(feature pyramid network,FPN)中引入内容感知特征重组(content-aware reassembly of features,CARAFE)优化最近邻上采样方法,设计了FPN-CARAFE结构,扩大了感受野,从而能够获取小目标更多的细节信息和丰富的语义信息;最后,采用归一化距离损失函数(normalized wasserstein distance,NWD)优化CIoU损失函数,设计了NWD-CIoU损失函数,降低了CIoU对小目标位置偏移的敏感性,能够更好地提升小目标的检测效果.在公开的遥感数据集RSOD和NWPU VHR-10上进行的实验表明,与基准模型相比,在计算量和参数量略增长的情况下,改进的模型在平均精度均值(mAP)mAP@0.5上分别提升了3.6%和1.8%,有效地提高了遥感图像中小目标的检测精度,综合性能优于其他算法,满足部署在遥感检测系统上的要求.
Abstract
Seeking to resolve the issue of missed and incorrect detection of small targets in remote sensing images,this study proposes an optimized YOLOv7-tiny algorithm.Firstly,a multi-scale attention efficient multi-scale attention(EMA)module is introduced,and based on this,the ELAN-EMA,a multi-scale feature extraction module,is incorporated to to greatly enhance the backbone network's proficiency in extracting features across various scales.Secondly,the feature pyramid network(FPN)is introduced with the content-aware reassembly of features(CARAFE)optimization,which expands the receptive field and enables the acquisition of more detailed information and rich semantic information of small targets.Finally,this study adopts the normalized wasserstein distance(NWD)loss function to optimize the complete intersection over union(CIoU)loss function,and designs the NWD-CIoU loss function,which reduces the sensitivity of CIoU to small target position shifts and can better improve the detection performance of small targets.Experiments conducted on the publicly available remote sensing datasets RSOD and NWPU VHR-10 show that compared with the baseline model,the optimized model achieves a 3.6%and 1.8%increase in mAP@0.5,respectively,with slightly increased computational and parameter requirements,markedly enhancing the accuracy with which small targets are detected in remote sensing images.The comprehensive performance meets the requirements for deployment in remote sensing detection systems.
关键词
目标检测/小目标/注意力机制/感受野/损失函数Key words
object detection/tiny object/attention mechanism/receptive field/loss function引用本文复制引用
出版年
2024
NETL
NSTL
万方数据