Design And Implementation Of Rotating Object Detection Method In Remote Sensing Images

The continuous development of UAVs and Beidou satellites has brought people a new way of life,and it has also brought people more access to remote sensing images.Now people can use drones to take images from different perspectives almost anywhere.Existing object detection methods can perform the object detection task relatively well,but there will be a certain loss of accuracy when performing detection in remote sensing images.In this scenario,the non-maximum suppression algorithm will have a large error.Therefore,this thesis proposes a rotating object detection method to perform the object detection task in remote sensing images with a more accurate representation.Combined with cutting edge technology,this thesis improves the existing horizontal object detection method to make it suitable for rotating object detection tasks.The improved method can predict the angle of the target and represent the target with a more precise oriented box,which solves the problem.Accuracy loss caused by non-maximum suppression algorithms.Since the proportion of the target in the remote sensing image is generally small,thesis introduces the attention mechanism of the spatial domain and the channel domain in the feature extraction network,and improves the attention module according to the characteristics of the target in the image,so that the backbone network can pay more attention to the features of the targets in remote sensing image in the feature extraction and feature fusion part.In the loss function calculation part,thesis proposes a short shift loss suitable for the rotating object detection task,which is used to correct the error caused by the angle to the box loss calculation in the rotating object detection task,and let the model initially learn the relationship between the angle deviation and the box loss.In order to verify the effectiveness of this method,several sets of experiments are designed in thesis,and the effectiveness of the rotating object detection method in thesis is verified in the experiments.And the ablation experiments of the modules used in thesis are carried out to verify the effectiveness of each module in the proposed or improved methods.The verification results show that the method proposed in thesis has significantly improved compared with the benchmark on remote sensing image datasets,and has also achieved certain improvements compared with other methods.Finally,based on the rotating object detection method proposed in this thesis and the relevant theoretical knowledge of distributed architecture,an object detection prototype system is designed and implemented to provide support for the implementation of the algorithm project.