改进YOLOv5网络在遥感图像目标检测中的应用

Improved YOLOv5 Network in Application of Remote Sensing Image Object Detection

针对遥感图像目标检测存在的尺度多样化、分布密集、小目标检测困难等问题,提出了一种改进YOLOv5网络的遥感图像目标检测的新方法Fca_YOLOv5。该方法引入了频率通道注意力网络,引导模型更加关注信息丰富的特征;将网络输入尺寸优化为1 024,减少了图像缩放带来的影响;采用圆形平滑标签计算角度损失,对船舰目标进行旋转目标检测,进一步提升检测效果。在DOTA遥感图像数据集上进行实验,检测精度最高达到了75.9%,船舰旋转目标检测精度达到了96.1%,并且Fca_YOLOv5s的检测精度比YOLOv5s提高了3.1%。实验结果表明,改进网络对遥感图像中的微小目标具有较好的检测效果,有效提升了遥感图像的检测精度,对实现遥感图像中的微小目标检测具有一定的参考意义。

Aiming at the problems of diverse scales and dense distribution of remote sensing image object detection and difficulties in small object detection,a new method Fca_YOLOv5 for remote sensing image object detection with improved YOLOv5 network is proposed.The method introduces a frequency channel attention network to guide the model to pay more attention to information-rich features;optimizes the network input size to 1 024 to reduce the image scaling brought the impact of image scaling;adopts circular smoothing label to calculate the angular loss and rotating object detection for ship objects to further improve the detection effect.Finally,experiments are conducted on the DOTA remote sensing image dataset,the detection accuracy reaches up to 75.9% and 96.1% for ship rotating objects,and the detection accuracy of Fca_YOLOv5 s improves 3.1% compared with that of YOLOv5 s.The experimental results show that the improved network has better detection effect on the tiny objects in remote sensing images,which effectively improves the detection accuracy of remote sensing images,and has certain reference significance for realizing the detection of tiny objects in remote sensing images.