首页|基于改进YOLOv7的遥感目标检测

基于改进YOLOv7的遥感目标检测

扫码查看
针对遥感目标在检测过程中由于被检测物体尺寸小、数量多而存在误检率高和检测精度较低的问题,提出了一种改进的 YOLOv7 模型.首先,改进特征提取网络,使用 SPD-RConv代替原网络的Repconv,并基于Concat与SE注意力机制提出了一种Concat_SE自适应通道融合全连接层,以增强对高价值信息的关注度.然后,对目标的候选框使用K-means聚类算法进行重新计算.最后,将改进的算法与原算法在增强后的RSOD数据集上进行实验,结果显示改进后算法平均准确率均值提高了 17.1%,查准率和召回率也分别提高了 19.7%和15.5%.证明了改进的YOLOv7 有效增强了模型对于遥感目标的检测能力.
Remote sensing object detection based on improved YOLOv7
To address the issues of high false detection rates and low detection accuracy in remote sens-ing target detection due to small object sizes and large quantitie,an improved YOLOv7 model is proposed by using SPD-RConv to replace Repconv in the original network.And a Concat_SE adaptive channel fusion full connection layer,based on Concat and SE attention mechanism is introduced to enhance the focus on high-val-ue information.Then,the candidate boxes for the target are recalculated using the K-means clustering algo-rithm.The improved algorithm and the original algorithm are tested on the enhanced RSOD data set.The re-sults show that the improved algorithm increased the mAP value by 17.1%,and the precision and recall rates by 19.7%and 15.5%,respectively,which demonstrates that the improved YOLOv7 effectively enhances the model's ability to detect remote sensing targets.

remote sensing targetYOLOv7attention mechanismsK-means clusteringSPD-Conv

梅艺林、崔立堃、耿玺钧、胡雪岩、刘知阳

展开 >

陕西理工大学 机械工程学院,陕西 汉中 723000

遥感目标 YOLOv7 注意力机制 K-means聚类 SPD-Conv

2024

陕西理工大学学报(自然科学版)
陕西理工学院

陕西理工大学学报(自然科学版)

影响因子:0.425
ISSN:2096-3998
年,卷(期):2024.40(4)