Object Detection In Remote Sensing Images Based On Deep Learning And Few Shot Learning

In recent years,the intelligent interpretation of remote sensing images has achieved rapid development driven by deep learning technology,and object detection has therefore become a relatively active research direction in the field of remote sensing.Although the existing object detection methods have achieved remarkable results on natural images,the diversity of object scales,complex background,dense distribution,and small sample characteristics in remote sensing images still restrict the development of object detection in the field of remote sensing.As a result,the task of object detection in remote sensing images is still extremely challenging.Aiming at the problems of object scale diversity,dense distribution and small sample characteristics in remote sensing images,this thesis proposes three remote sensing image object detection methods based on deep learning related theories such as multi-scale features and few shot learning.The specific work content is as follows:(1)Aiming at the problems of object scale diversity in remote sensing images and insufficient multi-scale information representation,a remote sensing image object detection method based on based on multi-level feature adaptive fusion is proposed.This method is based on the RetinaNet detection model,and designs corresponding modules from different angles to enhance the information representation on the feature maps of each level.On the one hand,construct a multi-receptive field feature extraction module to obtain the characteristics of different receptive fields to enhance the ability of the feature map to discriminate the object.On the other hand,build a multi-level feature adaptive fusion module to adaptively aggregate the information of different levels of feature maps,and enrich the information representation of each level of feature maps.In addition,a bottom-up path is designed to enhance the location information on the entire feature level.Experimental verification on the DOTA data set shows that compared with other object detection methods,the performance is better and the accuracy is higher.Compared with the benchmark method RetinaNet,the detection accuracy is increased by 1.8%,which verifies the effectiveness of the proposed method.The results of the ablation experiment further verify the effectiveness of each module.(2)Aiming at the serious problem of missed detection of vehicle objects in dense remote sensing image scenes,a remote sensing image dense vehicle object detection method based on object center point detection and semantic feature enhancement is proposed.This method takes into account the characteristics of vehicle objects under dense distribution,and designs the representation of the object with the center point and the object size(length and width).Furthermore,the constraint of the anchor is abandoned,and the vehicle detection is decoupled into two subtasks: object center point positioning and object size regression,which reduces the impact of object size on object positioning and greatly improves the recall rate of the object.In addition,a semantic self-attention feature enhancement module is constructed to reduce the interference caused by background noise to the classification of the object center point.At the same time,pixel-level classification information is introduced,which reduces the ambiguity between objects,enhances the discrimination between objects,and improves the accuracy of vehicle object detection.Through experimental verification on three data sets DOTA,ITCVD and CARPK,the experimental results show the significant advantages of the method proposed in this chapter in the recall rate.Compared with other vehicle detection methods,it has obtained very good detection accuracy,which verifies the effectiveness of the method in this chapter.(3)Aiming at the problem of object detection under the condition of a few annotated samples in remote sensing images,a remote sensing image few shot object detection method based on rotation matching and adaptive multi-relation is proposed.On the one hand,the rotation matching attention region proposal network is designed to enhance the model’s robustness to the multi-directionality of objects in remote sensing images,and to introduce supporting image information for RPN to improve the quality of proposals.On the other hand,an adaptive multi-relation detector is constructed to model the global,local,and patch relationship between the proposals of the supported image and the query image,and integrate them in an adaptive manner to improve the discrimination ability of the detector.Through experiments on the NWPU VHR-10v2 data set,the experimental results show that under different numbers of small samples,the method proposed in this chapter has shown the superior detection performance.Compared with the benchmark method ARMRNet,the improvements are 1.31%,6.19% and 6.09% under 1-shot,3-shot and 5-shot,respectively,verifying the effectiveness of the proposed method.