基于改进YOLOv4算法的遥感图像飞机目标检测

Aircraft Object Detection Based on Improved YOLOv4 Algorithm for Remote Sensing Images

针对在遥感图像上对飞机目标检测的精度低问题,论文通过对PANet特征融合网络结构的加深使得YOLOv4算法对小目标的检测更加敏感,进而提高算法的平均检测精度;另外,利用K-means++算法产生了能够自适应与数据集的检测先验框以减少YOLOv4检测算法对边界框回归损失计算过程中的冗余。在RSOD(Remote Sensing Object Detection)数据集上的对比实验表明,综合改进后的YOLOv4算法AP值达到了80.25%。特别地,改进后的YOLOv4算法对小目标检测的置信度得分较高。

Aiming at the problem of low accuracy of aircraft target detection on remote sensing images,this paper deepens the PANet feature fusion network structure to make the YOLOv4 algorithm more sensitive to the detection of small objects,thereby im-proving the average detection precision of the algorithm.In addition,the K-means++algorithm is used to generate adaptive data sets.In order to reduce the redundancy of the YOLOv4 detection algorithm in the calculation of the bounding box regression loss.Comparative experiments on the RSOD data set show that the AP value of the improved algorithm reaches 80.25%.In particular,the improved YOLOv4 algorithm has a higher confidence score for small object detection respectively.