Multi-Scale Object Detection in Satellite Images Based on Improved YOLOv7
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
维普
万方数据
针对卫星遥感影像目标检测中,小目标检测精度低、漏检率高,以及实际应用场景中检测效率低等问题,文章提出一种基于改进YOLOv7(You Only Look Once)的卫星遥感影像多尺度目标检测方法.在检测网络中,聚焦提升小目标检测能力,添加类注意力机制的卷积模块(ConvNeXt Block,CNeB),提升对小目标细粒度特征的提取及利用能力;同时,提出后处理机制,通过建立小目标与大目标的相互关系,实现使用单个模型对多种尺度目标进行检测.实验结果表明,在TGRS-HRRSD数据集 4个小目标上,改进后的检测模型相较原始YOLOv7 在平均精确率均值指标上提升了 16.6 个百分点.在检测特定大目标任务中,通过后处理机制,在保持精度的条件下,相较YOLT(You Only Look Twice)时间减少了 70%.相较于主流的面向遥感影像的检测方法,该方法在检测多尺度目标上,检测精度更高、速度更快.
To address the problems of low detection accuracy,high missed detection rate in small target detection,and low detection efficiency in practical application scenarios in satellite remote sensing image target detection,a multi-scale target detection method based on improved YOLOv7 for satellite remote sensing images is proposed.In the detection network,the focus is on improving the detection capability of small targets by adding a ConvNeXt Block(CNeB)with class attention,which enhances the ability of extraction and utilization of fine-grained features of small targets.At the same time,a post-processing mechanism is proposed to establish the mutual relationship between small and large targets,enabling the detection of multiple-scale targets using a single model.Experimental results show that on four small targets in the TGRS-HRRSD dataset,the improved detection model achieved an average improvement of 16.6%in mean average precision compared to the original YOLOv7.In specific large target detection tasks,the post-processing mechanism reduced the time by 70%compared to YOLT while maintaining accuracy.Compared to mainstream remote sensing image detection methods,this method is more accurate and faster in detecting multi-scale targets.
NETL
NSTL
维普
万方数据
