Rotating Object Detection In High Resolution Remote Sensing Images Based On Deep Learning

With the continuous upgrading of drone technology and Beidou satellites,the methods for acquiring remote sensing images have become more straightforward.In the field of optical remote sensing image processing,high-resolution remote sensing image object detection is a crucial task.Its purpose is to accurately locate and identify highvalue ground targets,which have significant military and civilian value.Many excellent object detection algorithms have achieved superior performance in natural image processing,but there are still many constraints that limit their detection accuracy in highresolution remote sensing image object detection.Theis is attributed to the complexity of remote sensing image scenes,the arbitrariness of target distribution directions,and the dense arrangement of targets with large aspect ratios and scales,which pose significant difficulties for detection.Theis article addresses the shortcomings of existing methods and focuses on the research of remote sensing image object detection from the perspectives of rotation detection,angle optimization,feature extraction,and loss function.The main research contents are as follows:The YOLOv5 algorithm can achieve detection in sparsely distributed target scenes.However,in complex scenes where targets are densely distributed and oriented in arbitrary directions,there is a phenomenon of missed detection.Additionally,using rotation detection causes problems with periodicity and boundary issues,leading to a sudden increase in loss values,which is not conducive to loss learning.To address these issues,theis article improves the data augmentation technique and adds angle information to enable rotation rectangle learning based on the original baseline algorithm.Furthermore,circular smooth label technology is added to the network to transform continuous angle information into discrete classification,which solves the problems of periodicity and boundary issues and increases the fault tolerance between adjacent angles.Experimental results show that the constructed R-YOLOv5 algorithm is more suitable for high-resolution remote sensing image object detection in complex scenes,avoiding the accuracy loss caused by non-maximum suppression in horizontal object detection and improving detection accuracy.A single remote sensing image may contain hundreds of targets,but extracting features from the entire image leads to a large amount of redundant information,and optimizing the loss function independently for detection is not conducive to holistic learning.To address these issues,theis article integrates convolutional block attention module into the network,which adaptively adjusts the feature response of different spatial positions to improve model performance.In terms of the loss function,the KLD loss function is used to convert the independently optimized five parameters into a twodimensional Gaussian distribution,coupling the various parameters together and dynamically adjusting the gradient based on the target features.Finally,the improved algorithm is analyzed and verified through experiments,and multiple ablation experiments are designed to validate the effectiveness of the convolutional block attention module and the KLD loss function.