改进的YOLOv5s遥感影像机场场面飞机小目标识别

Improved YOLOv 5 s Based Small Target Recognition of Remote Sensing Image of Aircrafts on Airport

遥感影像具有地物拍摄模糊以及背景环境复杂等特点,导致大面积地面物体识别准确率较低。针对此问题,提出了一种基于YOLOv5s改进网络模型。该模型对YOLOv5s中的骨干提取网络和颈部多尺度特征融合网络进行优化,引入Swin Transformer以获得更多关于目标物体的特征信息;同时对骨干网络中的模块做了修剪;此外,模型中还添加了坐标注意力机制来提升特征提取和融合效果。对于遥感数据集进行了小目标检测识别测试实验,改进后的YOLOv5s网络的mAP值为0.8375,比YOLOv5s网络模型提高了0.0225。实验结果表明,改进后的YOLOv5s网络模型对比YOLO系列网络和EfficientDet模型有效地提高了识别准确率、召回率以及mAP值,并且在训练时间上也比YOLOv5s减少了1/12。

Remote sensing images have the characteristics asblurred terrain capture and complex background environment which leads to a problem of low accuracy in identifying large ground-level objects.To solve this problem an improved network model based on YOLOv5s is proposed.The proposed model adjusts YOLOv5s network model s backbone extraction network and neck multi-scale feature fusion network and introduces Swin Transformer for obtaining more feature information about the target objects.Additionally the model prunes the modules in the main network and adds coordinate attention mechanism to enhance feature extraction and fusion effects.The proposed model is tested on small target recognition by using remote sensing dataset and mAP value of the improved YOLOv5s network is 0.8375 which is 0.0225 higher than that of official YOLOv5s network model.Experimental results show that the proposed model effectively improves the recognition accuracy recall rate and mAP value in comparison with YOLO series network and EfficientDet model and it reduces the training time by 1/12 in comparison with the YOLOv5s model.