基于Faster R-CNN的轻量化遥感图像军用飞机检测模型

A lightweight remote sensing image military aircraft detection model based on Faster R-CNN

遥感图像军用飞机目标检测对侦察预警和情报分析等领域具有重要意义。针对该任务中图像背景复杂、目标尺度变化大和分布密集等挑战,提出了一种基于Faster R-CNN的轻量化检测模型。该模型使用残差拆分注意力网络来捕获目标区域特征的全局上下文信息以提升模型的表征能力;利用可变形卷积来动态学习目标区域的形变特征,适应不同尺度和形状的目标;采用对比实验的方法精简骨干网络,降低过深的骨干网络与过低的采样率对于小目标检测的影响,提高模型的识别速度。在目标候选框筛选阶段,引入Soft NMS算法,根据置信度降序排名去除重叠度高的候选框,降低密集分布目标的漏检率。实验结果表明,提出的Faster R-CNN模型在参数量为23.844 MB的情况下,mAP0.5-0.95达到了77.1%,检测速度达到了43.7帧/秒,相比于多个主流模型具有较好的综合性能。

Military aircraft target detection in remote sensing images is of great significance for reconnaissance and early warning and intelligence analysis.In view of the challenges of complex image background,large target scale variation and dense distribution in this task,a lightweight detection model based on Faster R-CNN is proposed.The model uses residual split attention network to capture the global context information of target region features to improve the representation ability of the model;it uses deformable convolution to dynamically learn the deformation features of target region,adapt to targets of different scales and shapes;it uses the method of comparative experiment to streamline the backbone network,reduce the impact of too deep backbone network and too low sampling rate on small target detection,and improve the recognition speed of the model.In the target candidate box screening stage,Soft NMS algorithm is introduced to remove candidate boxes with high overlap according to the descending order of confidence,and reduce the miss detection rate of densely distributed targets.The experimental results show that the Faster R-CNN model proposed in this paper has a mAP0.5-0.95 of 77.1%when the number of parameters is 23.844 MB,and the detection speed reaches 43.7 frames per second.Compared with multiple mainstream models,it has better comprehensive performance.