基于改进YOLO算法的遥感图像目标检测

Remote sensing image target detection based on improved Yolo algorithm

针对遥感图像中大尺寸目标检测的平均精度(AP)较低的问题,提出了一种优化的YOLOv3-Tiny算法。首先在YOLOv3-Tiny算法的基础上增加了1×1卷积来实现多个特征图的跨通道交互及信息整合,进而来提取图像的全局信息,使输出的特征图包含更丰富的语义特征;其次将空洞卷积(atrous convolution)引入到YOLOv3-Tiny网络,用来增大特征图的感受野,改善大目标的检测效果;最后在RSOD-Dataset数据集上进行对比实验。结果表明,优化后的YOLOv3-Tiny算法相比原YOLOv3-Tiny算法,其均值平均精度(mAP)提高了1.4%,在较大尺寸的目标Overpass检测中,Overpass检测的平均精度提高了4.05%。

Aiming at the low average precision(AP) problem of large-scale target detection in remote sensing image, an optimized YOLOv3-Tiny algorithm is proposed. Firstly, a 1×1 convolution is added to the YOLOv3-Tiny algorithm to realize the cross channel interaction and information integration of multiple feature maps, and then the global information of the image is extracted to make the output feature map contain more abundant semantic features. Secondly, we introduce the hole convolution into the YOLOv3-Tiny network to increase the receptive field of the feature map and improve the detection effect of large targets. Finally, a comparative experiment is carried out on the RSOD-Dataset. The results show that: compared with the original YOLOv3-Tiny algorithm, the mean average precision mAP of the optimized YOLOv3-Tiny algorithm is improved by 1.4%, and the average accuracy of Overpass detection for large-scale targets is increased by 4.05%.