Abstract: Aerial remote sensing imaging has a wide range of military and civilian applications. The detection of small targets in aerial remote sensing images is a challenging problem in the field of remote sensing images. General target detection methods are insensitive to small targets, and have low detection accuracy for aerial remote sensing images with complex backgrounds and few target feature information. To solve the above problems, this paper proposes a small target detection network ARST-YOLOv7 for aerial remote sensing images for the first time. Firstly, the paper proposes a new feature enhancement module DSPPCFF to improve the feature expression ability of model due to the complicated background of remote sensing images and weak target features. Secondly, improved the feature pyramid structure, named DC-FPN, which combines multi-scale convolution (DC-MSC) modules to solve the semantic difference problem between cascade feature maps effectively. Furthermore, to retain more spatial information in the network, the DCA module is proposed to improve the network’s emphasis on critical targets, hence enhancing detection robustness. The quantitative and qualitative comparison results on the challenging NWPU VHR-10 dataset, RSOD dataset, and HRRSD dataset show that the proposed method is more effective for aerial remote sensing small target detection compared to current advanced target detection methods.

Key words: aerial remote images, small target, YOLOv7, dilated convolution, multi-scale convolution

摘要： 航空遥感成像具有广泛的军事和民用应用。航空遥感图像中的微小目标检测是遥感图像领域的一个具有挑战性的问题。通用的目标检测方法对小目标不敏感，对于背景复杂、目标特征信息量少的航空遥感图像检测精度较低。为解决上述问题，首次提出了一种用于航空遥感图像的小目标检测网络ARST-YOLOv7。针对遥感图像背景复杂，且目标特征较弱，提出一种新的特征增强模块DSPPCFF（dilated spatial pyramid pooling convolution feature fusion），增强模型的特征表达能力。提出一种新的特征金字塔结构（DC-FPN），来解决级联特征图之间的语义差异问题。此外，为了让网络保留更多的空间信息，提出了DCA（dilated convolution attention）模块，增强网络对重要目标的关注，从而提高检测的鲁棒性。在NWPU VHR-10数据集、RSOD数据集和HRRSD数据集上与当前先进的检测方法比较结果表明，该方法对于航空遥感小目标检测更有效。

关键词: 航空遥感图像, 小目标, YOLOv7, 空洞卷积, 多尺度卷积