Research On Object Detection Algorithm Of Aerial Remote Sensing Images Based On Full Convolutional Network

With the continuous progress of remote sensing technology and UAV technology,the quality of remote sensing images and aerial images has been gradually improved,and the related target detection tasks are also gradually studied in depth.In this paper,we take aerial remote sensing images as the research object and select the current advanced rotating target detection model FCOSR for improvement,and conduct relevant experiments on the rotating target detection dataset constructed based on aerial remote sensing images to prove the effectiveness of the method in this paper.The research in this paper is as follows:1.To address the problem that the current human-vehicle target detection dataset taken by UAVs uses horizontal frames to label targets,which leads to inaccurate target localization,this paper uses UAVs to collect a large number of street scenes containing people and vehicles to make a dataset specifically for human-vehicle target detection,refines the category of small vehicles in the dataset to enhance the detection of small vehicles,and expands the human-vehicle target detection task to a rotating target detection task,while The research on the task of rotating target detection of remote sensing images is carried out.2.To address the problem of unbalanced detection speed and accuracy of the current rotating target detection model,this paper improves the single-stage unanchored frame rotating target detection model FCOSR based on the feature-aligned pyramid structure,and improves the detection accuracy of the model while ensuring the high detection speed of the model.3.To address the problem that the current rotating target detection model has insufficient application to multi-scale features and extraction,this paper combines the fine-grained multi-scale feature extraction method and attention mechanism to propose Focused Multi-Scale Feature Module(FMSM),and the feature map containing fine-grained multi-scale features extracted by this module is used as The top-down feature fusion is performed in the PAFPN structure,and then the bottom-up feature fusion is performed to transfer the extracted multi-scale features to the whole network to improve the responsiveness of the model to the extraction of multi-scale information,and the module is called Multi-scale Feature Fusion Module(MSFM)here.Module(MSFM),and incorporates this module into the single-stage anchorless frame rotation target detection model FCOSR to improve the detection performance of the model.In this paper,the improved model is trained and evaluated on the challenging DOTA dataset and the human-vehicle rotating target detection dataset produced in this paper,and the experimental results are compared and analyzed with those of some advanced rotating target detection methods at this stage,and it can be seen from the experimental results that the proposed method in this paper can effectively improve the overall performance of the model.