Research And Implementation Of Dense Small Object Detection Algorithm In Remote Sensing Images

Remote sensing technology is currently widely applied in a variety of military and civilian fields.Accurate detection of objects in remote sensing images so as to analyze and utilize these images can be of significant application value in security,transportation and rescue.With the rapid development of satellite remote sensing and aerial photography technology,people can capture a greater amount of information about ground objects since the resolution of remote sensing images gets higher.Yet,the remote sensing image processing has turned out to be more difficult with the increase of image resolution.Deep learning methods have achieved remarkable results in the field of computer vision with the emergence of many excellent network structures along with the advancement of computing power.The emergence of deep learning has brought unprecedented opportunities to the field of object detection in remote sensing images.Researchers have proposed various solutions based on deep learning techniques to enable the network to learn finer features.However,a large quantity of densely arranged,small scaled rotated objects still pose a challenge to remote sensing image object detection.This thesis detects the dense small objects in remote sensing images effectively based on the S~2A-NET algorithm.The specific work of this thesis is as follows.1)Aiming at the problem that small objects with insufficient feature information in remote sensing images are poorly detected,a small object feature enhancement module S~2ANET-SR based on Super-Resolution is proposed.S~2ANET-SR enhances the feature extraction for small objects by utilizing the information of large objects with the same features but different scales in remote sensing images.The introduced Super-Resolution loss of S~2ANET-SR is computed in a mask manner to focus on the regions with foreground objects in the image,as well as in a patch manner to ensure the consistency of texture information of the feature map in local areas.The experimental results demonstrate the superiority of the proposed S~2ANET-SR for detecting small objects in remote sensing images.2)In view of the problem that some objects of the same category in dense scenes within remote sensing images are missed when detected,a module S~2ANET-REP is proposed to improve the classification confidence of hard positive samples based on representative features.Firstly,S~2ANET-REP selects the representative features to represent the semantic information of each category in the image,and improves the reliability of representative features by setting a suitable confidence threshold.Then,the similarity between the missed hard positive samples and the representative features is calculated.Finally,the classification confidence of the hard positive samples is added adaptively according to the similarity to further infer the existence of hard positive samples.The parameters of the classification branches in both stages are shared in the model to ensure the consistency of the similarity calculation process and also to reduce the complexity and number of parameters of the model.Experiments show that S~2ANET-REP can effectively improve the classification ability of the model for hard positive samples.3)The proposed S~2ANET-SR module and S~2ANET-REP module are integrated to obtain the final S~2ANET-SREP model,and an object detection system for remote sensing images is implemented.Extensional experiments are conducted to evaluate the performance on the general remote sensing dataset DOTA,and the results show that the proposed S~2ANET-SREP model can achieve 75.16%m AP,which is 1.72%better than the accuracy of S~2A-NET baseline model.