Research On Detection And Classification Algorithm For Aircraft Target In Remote Sensing Images

As important military equipment,aircraft is of great strategic significance.It has important research and application value for detecting and classifying aircraft targets in remote sensing images.Due to the different imaging periods and environmental conditions of the remote sensing images,the resolution of the visible remote sensing platform,the camera F number,the flying height,the viewing angle and other parameters are different.Detection of aircraft target in remote sensing images often requires to consider complex and variable background information and unstable target characteristics.After the aircraft targets are detected,further automatic identification of the aircraft type and even automatic identification of the aircraft model is an inevitable option to solve the problem of delay in the manual identification and real-time recognition of the aircraft targets in remote sensing images.In view of above problems,this thesis focuses on the aircraft detection and classification techniques in remote sensing images.In this thesis,the data in the "DOTA aircraft target detection data set in remote sensing images" were processed and classified firstly.To address the lack of public data set on the aircraft target classification task in remote sensing images,this thesis proposed a classification standard based on shape of targets to classified the aircrafts in the remote sensing images into four categories roughly: civil aircraft 1,civil aircraft 2,old-fashioned aircraft,and fighter aircraft.Then,the "DOTA-T4 aircraft target classification data set in remote sensing images" was constructed based on above classification standard.Later,the "DOTA-F9 fighter target fine classification data set in remote sensing images" was constructed by further classifying the DOTA-T4 fighter aircraft samples into 9 categories according to the differences in models,colors and assembly status of aircraft targets.To address the instability of aircraft target shape,scale and radiation characteristics in large-format remote sensing images,this thesis proposed an aircraft target detection algorithm in remote sensing images(G-DYOLT)based on ground sampling distance.The method considers the distribution of different aircraft target sizes in different ground sampling distances.As the target size is relatively limited in the detection of the aircraft targets in the remote sensing images in the general deep learning network framework,the ground sampling distance information was introduced to realize the adaptive selection of the detection network.Next,large-scale experiments verified the great detection performance for aircraft targets with widely distributed features.To classify the aircraft targets after detection,this thesis proposed an aircraft target classification algorithm(SPFC)in remote sensing images based on the sum pooling feature.This method constructed the sum pooling features and retained the channel construction of original features with the feature map extracted by the convolutional neural network,which handles the position information of the original feature map in a more reasonable manner,and improves the rough aircraft target classification and the fighter target classification results.At the same time,the structure can generate a score contribution rate map to observe the classification basis in the classification process,which means the features in the network presents strong interpretability,thereby contributing to improving the classification algorithm in a more targeted manner.