Research On Key Technologies Of Aircraft Object Detection And Recognition In Remote Sensing Images

With the rapid development of remote sensing imaging technology,the resolution of remote sensing image is higher and higher,the amount of data is more and more large,and the image information is more and more abundant.The real-time and correctness of the algorithm cannot be guaranteed by using traditional image processing algorithms or classical machine learning algorithms to interpret aircraft objects.Therefore,how to quickly and accurately identify and locate aircraft objects in large-scale complex remote sensing images has become an urgent problem in the field of remote sensing aircraft object detection.Based on convolutional neural network(CNN),the paper studies aircraft object detection in remote sensing images.According to the characteristics of complex background and sparse-aggregation of aircraft targets in large-scale remote sensing images,an effective aircraft object detector(EAOD)is proposed.The main research work and contributions are summarized as follows:(1)Aiming at the characteristics of large scale,complex background and sparse target aggregation of remote sensing images,the paper proposes an aircraft area recognition network(A~2RNet).The network uses a lightweight network to extract the overall features of large-scale remote sensing images,and uses a background analysis module to enhance the receptive field of the output features.Then,the potential aircraft areas are screened according to the regional classification results of the image.On the open source DOTA aircraft data set,the A~2RNet proposed in the paper has excellent performance,and the detection accuracy reaches 95.4%.(2)In view of the problem that the image input size of the classical Faster R-CNN algorithm is fixed and it does not have robust rotation detection ability,an improved Faster R-CNN architecture is designed in the paper.Feature pyramid network(FPN)is embedded in the architecture for cross-scale feature fusion and refined rotating classifying module(R~2CM)is embedded to replace RCNN for refining detection.The experimental results show that the AP50 of the improved Faster R-CNN is improved on the self-built aircraft target detection dataset and the open source dataset DOTA.(3)In order to filter out the background region that does not contain the aircraft object and enhance the context feature of the backbone network,the paper proposes a global processing module(GPM).According to the classification results of A~2RNet,the module cuts out the aircraft area in large-scale remote sensing images,and uses the A~2RNet receptive field enhancement feature map corresponding to the aircraft area as the feature mask to supplement the context information of the aircraft area.The experimental results show that after the introduction of GPM,the detection speed of the network on public dataset can be increased by 8.24 times by using A~2RNet,respectively,which greatly improves the efficiency of aircraft object detection.