Research On Robust Classification Of Remote Sensing Images And Few Shot Object Detection Based On Deep Learning

With the development of remote sensing satellite technology,the types of remote sensing images are becoming more and more diverse,such as aerospace remote sensing images and low altitude unmanned aerial vehicle remote sensing images.The acquisition of remote sensing images is becoming more convenient,and the application fields for remote sensing images are becoming wider and wider,such as disaster assessment,military monitorin g,land change detection,industrial patrols,and so on.Due to the impact of equipment failures,human interference,or environmental factors during remote sensing image acquisition,abnormal data maybe introduced,how to extract effective features from complex data containing abnormal values,and improving the model robustness of classification is an important topic.Images with accurate labels is necessary for supervised learning based deep learning algorithms.As remote sensing images are obtained with large size,wide coverage,and complex backgrounds,image annotations are laborcosting and low efficiency,especially for rare targets in remote sensing images.Few shot object detection in remote sensing image is worth researching.In addition,because the shooting angle and height of UAV aerial images are more flexible than remote sensing satellite images,the multi-scale problem of "near large and far small" exists in the images taken by UAV at low altitude and oblique viewing,which has become another challenge of aerial image in object detection.Given the above problems,the main research contents and contribuitons are as follows:(1)Considering the promising performance of Extreme Learning Machine(ELM)in dealing with the robustness of classification on complex data,this study optimizes the robustness of the Stacked ELM(S-ELM),which is a kind of deep learning model based on the least square method.Firstly,replaces the sensitive PCA in S-ELM with the Correntropy Temporal Principle Component Analysis(CTPCA),which is robust to outliers,and CTPCA-based S-ELM is proposed,which improves the robustness of the model to a certain extent;Then,from the perspective of feature extraction,combines ELM Sparse Auto-Encoder(ELM SAE),which is capable of extracting sparse abstract features,with CTPCA-based S-ELM,and CTPCA-based S-ELM with ELM SAE to further improve the robustness of the original model;Finally,the sensitivities of various hyperparameters are analyzed through experiments,which verifies that the methods proposed in this study are robust to outliers,and it is applied to remote sensing image pre-classification tasks,providing a preprocessing method for subsequent remote sensing image object detection tasks.(2)For the problem of zero-shot learning and few-shot learning caused by the scarcity of data annotation of remote sensing satellite images,a synthetic remote sensing image generator SIMPL(Synthetic Implantation)is designed in this study.SIMPL is simple,efficient,reproducible and easy to expand.With experiments on real datasets,it is verified that the remote sensing data synthesized by SIMPL plays a positive role in the problem of zero-shot learning and few-shot learning in remote sensing images.Besides,the importances of key parameters in SIMPL are analyzed,which lays a foundation for subsequent research.(3)To deal with the problem of scarce annotation and multi-scale objects in low altitude UAV aerial images,this study extends the SIMPL,and synthesizes the low altitude UAV aerial images and their corresponding pixel-level labels.Due to the fact that the materials used for synthesizing the UAV aerial images are not from real images,the domain gap between the synthetic images and real images is increased,in order to alleviate the problem of negative domain adaptation during the migration of models from source domain(i.e.,synthetic image datasets)to target domain(i.e.,real image datasets),this study fully utilizes the annotations of synthetic images and designs a soft mask processing method,which is beneficial to extracting domain independent information without significantly increasing the complexity of the model.At the same time,for the multiscale objects problem in low-altitude UAV aerial images,this study proposes a novel framework based on Faster R-CNN with backbone of residual networks and feature pyramid models,which combines soft masks with multi-scale feature maps.Experiments on real datasets have confirmed the detection effect of the model.