摘要
针对光学遥感图像成像条件复杂,物体角度多变,易漏检、误检的特点,提出一种基于YOLOv8n改进的遥感图像目标检测算法,以提高遥感图像目标检测的性能。首先,在主干网络部分融合了SimAM注意力机制模块,以提高主干网络的特征图质量;其次,结合ECA注意力机制和空洞卷积构建了EC-DCAM注意力机制模块,融入颈部网络的输出阶段,以提高网络对多尺度特征的学习能力;最后,使用WIoU损失函数替换CIoU损失函数进行动态的损失计算。将改进的算法在DIOR数据集和MAR20数据集上进行了实验,在DIOR数据集上的准确率提升了1.3%,达到了74.7%;在MAR20数据集上的准确率提升了3.8%,达到了87.2%,改进算法的参数量和计算复杂度几乎没增加,验证了改进算法的有效性。
In view of the complex imaging conditions of optical remote sensing images,the variability of object angles,and the characteristics of easy missed detection and false detection,based on the YOLOv8n object detection algorithm,an improved remote sensing image object detection algorithm based on YOLOv8n is proposed to improve the performance of remote sensing image object detection.Firstly,the SimAM attention mechanism module is integrated into the backbone network to improve the quality of the feature map of the backbone network.Secondly,the EC-DCAM attention mechanism module is constructed by combining the ECA attention mechanism and the dilated convolution,and is integrated into the output stage of the neck network to improve the network′s ability to learn multi-scale features.Finally,the WIoU loss function is used to replace the CIoU loss function for dynamic loss calculation.The improved algorithm was tested on the DIOR dataset and MAR20 dataset,and the accuracy rate on the DIOR dataset increased by 1.3%to 74.7%;on the MAR20 dataset,the accuracy rate increased by 3.8%to 87.2%.At the same time,the number of parameters and computational complexity of the improved algorithm were almost unchanged.The experimental results verify the effectiveness of the method.
作者
王广川
赵寿为
WANG Guangchuan;ZHAO Shouwei(School of Mathematics,Physics and Statistics,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《智能计算机与应用》
2025年第12期113-116,共4页
Intelligent Computer and Applications
