【摘要】 设计了一种新的轻量级多类物体检测网络，该网络可以实现对遥感图像快速准确检测，同时可以满足实时检测的需求.首先，提出了一种更高效的多维注意力网络(FMDA-net),它在只增加少量参数的情况下，就能给检测模型带来明显的性能增益，提高检测器对小物体的灵敏度.然后，基于FMDA-net和更高效的ECA模块对MobileNetV3的Bneck块进行改进，从而获得新的轻量级特征提取网络MobileNetV3+,并将其作为YOLOv4的主干网络，用于遥感图像的目标检测任务.实验证明，网络相对于先进的MobileNetV3,在遥感数据集UCAS-AOD获得了更高的FPS和mAP,相比MobileNetV3检测精度mAP提高6.05%.同时检测速度也高于MobileNetV3达到41.58FPS满足实时检测的要求.更多还原

【Abstract】 In this paper, a new lightweight multi-class object detector is designed.This model can achieve rapid and accurate detection of remote sensing images, which can meet the needs of real-time detection.Specifically, a more efficient multi-dimensional attention network(FMDA-net) is designed, which involves only a few parameters and brings significant performance gains to improve the sensitivity of the detector to small objects.Based on FMDA-net and a more efficient ECA module, the Beck block of MobileNetV3 was improved, and a new lightweight feature extraction network MobileNetV3+ was developed and used as the backbone network of YOLOv4 for target detection tasks in remote sensing images.Experiments have proved that compared with the advanced MobileNetV3,the network in this paper has obtained higher FPS and mAP in the remote sensing data set UCAS-AOD,and the detection accuracy mAP is improved by 6.05% compared with MobileNetV3.At the same time, the detection speed is also higher than that of MobileNetV3,reaching 41.58 FPS to meet the requirements of real-time detection.更多还原