卷积神经网络遥感影像飞机目标检测方法研究

Research on Aircraft Target Detection Method in Remote Sensing Images Based on Convolution Neural Network

针对如何利用高分辨率遥感影像对飞机目标实施快速精准检测的问题,提出并构建了基于深度学习方法的单阶段遥感影像飞机目标检测网络。首先使用小尺寸卷积核通过密集连接的方式搭建了深层特征提取骨干网络,让网络在特征提取过程中能够充分保留不同层的特征图信息,然后根据目标尺寸较小的特点设计了四层的多尺度特征金字塔进行特征图语义强化,并根据训练数据集中样本框特点,采用聚类获得了合适的锚点框数据集,并选择了合适的锚点框个数。然后对网络进行了训练,采用精度均值AP和FPS对模型的检测精度和速度进行评价,并将结果与Faster-Rcnn、YOLOv3和SSD的测试结果进行对比。实验结果表明,本文所提出的网络架构与Faster-Rcnn和YOLOv3、SSD相比在检测精度分别提高了4.5%、9.3%和16.6%。在速度方面检测速度可以达到29FPS,相比Faster-Rcnn提高了27.5%。从测试结果来看,所提出网络对于遥感影像中不同尺度的飞机目标都具有良好的检测效果。

Aiming at the problem of how to use high-resolution remote sensing images to detect aircraft targets quickly and accurately,a single-stage remote sensing image aircraft target detection network based on deep learning method is proposed and constructed.Firstly,a deep feature extraction backbone network is built by using small-size convolution kernel through dense connection,so that the network can fully retain the feature map information of different layers in the feature extraction process.Then,according to the characteristics of small target size,a four layer multi-scale feature pyramid is designed to enhance the semantic meaning of feature map,and clustering is adopted according to the characteristics of sample box in training data set.The appropriate anchor box data set is obtained,and the appropriate number of anchor box is selected.Then,the network is trained,and the accuracy mean AP and FPs are used to evaluate the detection accuracy and speed of the model,and the results are compared with those of Faster-Rcnn,YOLOv3 and SSD.Experimental results show that the proposed network architecture improves the detection accuracy by 4.5%,9.3%and 16.6%compared with Faster-Rcnn,YOLOv3 and SSD,respectively.In terms of speed,the detection speed can reach 29fps,which is 27.5%higher than Faster-Rcnn.From the test results,the proposed network has a good detection effect for different scales of aircraft targets in remote sensing images.