基于改进YOLOv5s的遥感目标检测

Remote sensing target detection based on improved YOLOv5s

摘要:

优化遥感目标检测对于军事民生方面都有着重要意义。由于遥感数据中成像模糊、目标较小、检测对象数量多，导致检测精度不高，现提出一种新网络：新网络在YOLOv5s（You Only Look Once v5s）原网络的基础上使用Mish激活函数代替SiLU（Sigmoid Weighted Liner Unit）激活函数；为解决遥感图像中小目标的问题，采用了对小目标、低分辨率友好的SPD-Conv（Space-to-depth-Conv）模块；考虑到使用耦合检测头会存在回归、分类两个任务之间的冲突问题，采用YOLOX (You Only Look Once X)中解耦的检测头，提高了模型查准率。实验结果表明，相比于原始YOLOv5s，新网络在mAP（mean Average Precision）平均精度均值方面提升了7%，查全率（recall）提升了10.9%，检测速度FPS提升了16.95%。改进后的网络模型相对于原始模型具有明显优越性，识别效果提升明显。

Abstract:

Optimizing remote sensing target detection is of great significance to military and people's livelihood. Due to the blurred image, small target and large number of detected objects in the remote sensing data, the detection accuracy is not high. A new network is proposed: the new network uses the Mish activation function to replace the SiLU (Sigmaid Weighted Liner Unit) activation function on the basis of the original YOLOv5s (You Only Look Once v5s) network; In order to solve the problem of small targets in remote sensing images, SPD-Conv (Space-to-depth-Conv) module which is friendly to small targets and low resolution is adopted; Considering the conflict between regression and classification tasks when using coupling detector heads, the decoupled detector heads in YOLOX (You Only Look Once X) are used to improve the model precision. The experimental results show that compared with the original YOLOv5s, the new network has improved the average accuracy of mAP (mean average precision) by 7%, the recall rate by 10.9% and the detection speed by 16.95%. The improved network model has obvious advantages over the original model, and the recognition effect is significantly improved.