Detection And Fine-grained Recognition Of Aircrafts In Optical Remote Sensing Images

Detection of aircraft targets in optical remote sensing images is of great importance in military reconnaissance and aerial monitoring.The background environment of remote sensing images is generally complex,and the types of aircraft targets are diverse and the scales are various,which increase the difficulty of target detection.The fine-grained recognition is embodied in model identification,which plays an important role in intelligence acquisition and precision strike in the military field.However,many models of aircraft have similarities in basic features such as appearance and texture,so to achieve accurate fine-grained identification is a great challenge.Methods based on deep learning on target detection and recognition have excellent performance in suppressing complex backgrounds and extracting target features,but currently there are relatively few related studies on aircraft detection and fine-grained recognition in remote sensing images.This thesis deeply studies the problems and difficulties in this field based on deep learning methods and carries out experiments.The main work is as follows:The background and significance of the thesis are systematically explained,and the current research progress of aircraft detection and fine-grained recognition in remote sensing images is summarized.Then the target detection methods based on deep learning and the theories and technologies involved in this thesis are outlined.The first remote sensing dataset that can be used for aircraft target detection and finegrained recognition is created and a 3D model simulation method for data augmentation is proposed.After statistical analysis of aircraft sizes,a high-precision improved YOLT method for target detection is proposed.YOLT is the first light algorithm focusing on target detection in remote sensing images.The thesis improves its network structure,designs a new feature method based on receptive field improvement and adopts an optimized nonmaximum suppression strategy.The results show that our method has better performance than other main target detection algorithms,especially for small targets and cross-scale samples.Aiming at the problems of small differences between large classes within aircraft target classes and difficulty in accurate classification,this thesis uses the structure relationship of aircraft category labels and adopts a hierarchical network structure design to propose an aircraft fine-grained recognition method based on attention mechanism.First,two subnetworks called category attention and densely-connection learn the attention of different channels about 7 categories,transfer features and expand channels,for helping the recognition of 21 fine categories.Then,we design a spatial attention module to locate discriminative regions.Finally,we introduce non-local operations to better combine context information and enhance discrimination ability.The comparative experiments and ablation experiments both prove the effectiveness of the method.In view of the practical application of target detection and recognition tasks in remote sensing field,an end-to-end aircraft target detection and fine-grained recognition framework is proposed,which can accurately and quickly implement detection and recognition in an end-to-end manner.The main network of the framework adopts the design ideas of target detection and fine-grained recognition methods proposed in the previous parts of this thesis,making sure the accuracy of detection and recognition.Then,to solve the problem of high false detection rate and missed detection rate of densely arranged targets,we propose redetection mechanism.In order to reduce the large amounts of calculations of deep network and improve real-time performance,we introduce depthwise separable convolution to optimize networks.Finally,a weight mapping method based on transfer learning is adopted to solve the problem of data labeling and also helps the detection and fine-grained recognition.The results show that the proposed end-to-end framework has good robustness,versatility and efficiency in aircraft detection and fine-grained recognition tasks.