基于改进YOLOv3网络的遥感目标快速检测方法

针对将来卫星在轨实时目标检测需求,且在其内存和算力都受限的条件下,提出一种改进的YOLOv3,利用轻量化网络代替YOLOv3的特征提取网络,实现遥感目标的高效检测。在目标检测精度相近的情况下,改进模型参数相比原先降低了1.5倍,计算量降低了3.3倍。同时提出了一种基于交并比的迭代聚类算法,分别在YOLOv3和改进YOLOv3上实现了7.0%和2.3%的平均精度均值(mAP)提升。实验表明:改进模型的检测速度最快能达到101 frame/s,当其mAP比YOLOv3高6%时,检测速度仍是YOLOv3的1.6倍。本文提出的改进YOLOv3是一种高效遥感目标检测方法,为未来星上应用打下基础。

Efficient Object Detection Method Based on Improved YOLOv3Network for Remote Sensing Images

In the future, satellites need to implement online real-time object detection, but their memory and computing power are limited. An improved YOLOv3 is adopted in this paper,making full use of the lightweight network instead of the feature extraction network of YOLOv3 to achieve high-efficiency object detection for remote sensing images. In the case of similar detection accuracy, parameters and FLOPs (floating point operations) of the improved model are 2.5 times and 4.3 times smaller than YOLOv3, respectively. In addition, an IoU K-medians algorithm is proposed, which improves mAP by 7.0% on YOLOv3 and by 2.3% on the improved YOLOv3. Experiments show that the detection speed of the improved YOLOv3 is 101 frame/s at the fastest,and it is still 1.6 times faster than that of YOLOv3 when its mAP is 6% higher than that of YOLOv3.The efficient remote sensing object detection method proposed in this paper lays the foundation for future satellite on-board applications.