Research On Optical Remote Sensing Image Target Detection Technology

Optical remote sensing image target detection can be used for military reconnaissance,weapon guidance,ocean monitoring,urban planning,etc.,and has important applications in both military and civilian fields.The traditional target detection method spends a lot of time on artificial feature extraction and candidate region extraction,which has a large amount of calculation and high time complexity,and cannot meet the practical application requirements.With the development of deep learning,the deep features extracted by neural networks have stronger semantic representation ability and discrimination,and higher efficiency,which has gradually become the mainstream detection method.Because remote sensing images have the characteristics of dense and diverse targets,uncertain target direction,wide field of view and complex background,the target detection method in natural scenes is difficult to be directly applied to remote sensing image target detection.In this paper,two detection methods are proposed for the difficulty of deep learning optical remote sensing image target detection.The specific research contents are as follows :Aiming at the problem of target scale diversity and dense target in remote sensing images,this paper proposes a multi-scale remote sensing image target detection method based on Faster RCNN.Firstly,Res Net50 is used to replace the original VGG16 as the backbone network for feature extraction.The Group Norm method is used to solve the numerical stability problem when Batchsize is 1,and the deformable convolution module is added to make the extracted features focus on the target.Secondly,FPN is used to improve the receptive field of the network to better collect the context information of the image.Finally,Soft NMS is used to improve the accuracy of the generated detection box.Experiments are carried out on the UCAS-AOD remote sensing image dataset.The experimental results show that the average accuracy AP of the improved detection model is 87.4 %,which is 7 % higher than the original algorithm,and the target can be better detected.Aiming at the directional problem of remote sensing image and the problem of large wide field of view and complex background,this paper proposes a Transformer target detection method based on rotating frame.Combining with the advantages of Transformer in establishing global image connection,a rotating region proposal network is designed.This method can learn the characteristics of the rotating region and accurately locate the rotating target.Visualization is realized by Grad-CAM,which intuitively displays the information concerned by the feature extraction network,reflects the interpretability of the model,and helps the model learn more distinguishing features.Finally,experiments are designed to verify and analyze the detection performance of the algorithm.The average accuracy m AP of the DOTA dataset is 0.8496,which verifies the reliability and rationality of the algorithm.