改进的PP-YOLO网络遥感图像目标检测

Improved PP-YOLO network for remote sensing image target detection

为了提高遥感图像目标检测精度,提出一种改进的PP-YOLO网络遥感图像目标检测方法。改进的PP-YOLO网络继承了骨干网络结构,改进了PP-YOLO网络的检测颈部分,在保持原检测颈框架基础上,在第4层和第5层加入了由低层网络向高层网络传达的通路,使得网络低层部分可以学习到高层部分的特征信息,加强高层网络学习的特征信息,在保证网络泛化能力的同时比未改进优化同骨干网络的PP-YOLO网络其平均精度均值(mAP)提高了4.4%。同时,优化了PP-YOLO网络训练策略,即基于遥感数据集特点,以更优的CutMix数据增强算法替换掉原有的Mixup数据增强算法,加入GridMask算法增强网络特征的学习,实验取得了最高89.3%的mAP,有效地提高了每一类目标实例的精度。

To improve the accuracy of remote sensing images target detection,an improved PP-YOLO network for remote sensing image target detection is proposed.In the improved PP-YOLO network,the backbone network structure is adopted,the detection neck network is improved,and a path communicated from the low-level network to the highlevel network is added between the fourth layer and fifth layer.In this way,features from the low-level network can be passed to the high-level net,which can improve the accuracy of target detection effectively and ensure the generalization ability of the network.Experiment results show that its mAP is 4.4%higher than that of the unimproved net.At the same time,the training strategy of PP-YOLO network is optimized,which is based on the characteristics of the remote sensing data set.The original Mixup data enhancement algorithm is replaced by CutMix data enhancement algorithm,and the GridMask algorithm is added to enhance the learning of network features.Experiments are done and the mAP is 89.3%,which improves the accuracy of target instances detection effectively.