In optical remote sensing images, the aircraft to be detected is very small; external environmental factors such as cloud occlusion, aircraft, and the site background are easily fused; and the interference of objects to aircraft has a great impact on the aircraft characteristics in remote sensing images. In response to the above problems, we designed a remote sensing aircraft detection method based on deep learning. First, to ensure the feature extraction capability and limit the number of calculations of the network, the LightNet v2 network unit is designed, and it constitutes an efficient backbone network. In addition, spatial pyramid pooling of residual ideas (Res-SPP) is performed on the output results of the backbone network. Res-SPP is used to separate more important contextual features while using almost no computing space. A multi-scale fusion prediction network (MFPN) is proposed to perform feature fusion from multiple angles to achieve a rich combination of gradients. The MFPN enhances the network’s ability to detect extremely small objects and can improve accuracy while ensuring that the method is lightweight. Finally, according to the judgment of the threshold, the dark channel defogging method, which enhances the ability to detect aircraft in cloudy and foggy scenes, is used for remote sensing images full of clouds and fog. The experimental results show that the proposed method can detect airplanes, especially very small airplanes, in various scenarios. The amount of calculation of the method is 23.56 BN, the model volume is 56 MB, the speed on a GTX 1080 platform reaches 157 frames per second (FPS), and the F1 % on the remote sensing aircraft data set reaches 99.2. In particular, this method can be embedded in an ordinary field programmable gate array platform due to its lightweight characteristics, and the calculation speed can reach 32 FPS.

中文翻译：

---

基于深度学习的光学遥感图像飞机快速检测方法

在光学遥感图像中，要检测的飞机很小。外部环境因素（例如云遮挡，飞机和站点背景）很容易融合；物体对飞机的干扰对遥感图像中飞机的特性影响很大。针对上述问题，我们设计了一种基于深度学习的遥感飞机检测方法。首先，为了确保特征提取能力并限制网络的计算数量，设计了LightNet v2网络单元，它构成了有效的骨干网络。此外，对主干网络的输出结果执行残差想法（Res-SPP）的空间金字塔池。Res-SPP用于分离更重要的上下文功能，而几乎不使用任何计算空间。提出了一种多尺度融合预测网络（MFPN），用于从多个角度执行特征融合，以实现梯度的丰富组合。MFPN增强了网络检测极小的物体的能力，并在确保方法轻巧的同时提高了准确性。最终，根据阈值的判断，使用暗通道除雾方法来增强在多云和有雾场景中检测飞机的能力，该方法用于遥感充满云雾的图像。实验结果表明，所提出的方法可以在各种情况下检测飞机，尤其是非常小的飞机。该方法的计算量为23.56 BN，模型容量为56 MB，在GTX 1080平台上的速度达到了每秒157帧（FPS），遥感飞机数据集上的F1％达到99.2。特别地，该方法由于其轻巧的特性可以被嵌入到普通的现场可编程门阵列平台中，并且计算速度可以达到32 FPS。