Research On Accurate Detection Technology Of Optical Remote Sensing Target Based On Rotating Bounding Box

With the development of remote sensing technology,the spatial resolution of optical remote sensing images has gradually increased,providing clearer outline and texture information for ground objects in a wide area.The application of remote sensing images is no longer limited to the previous rough remote sensing image classification.Remote sensing target detection has gradually become an important part of remote sensing application technology.Optical remote sensing target detection takes the image captured by the remote sensing platform sensor working in the visible light band as the research object,and uses image processing technology to classify and locate the target of interest in the image.In recent years,remote sensing image target detection has been widely used in the fields of land resource survey,ecological environment monitoring,disaster prevention and disaster assessment,etc.Research on remote sensing image target detection has important theoretical significance and application value.Because remote sensing images have the characteristics of complex background,large changes in target scale and aspect ratio,arbitrary orientation of targets,and uneven spatial distribution,traditional target detection methods are difficult to robustly detect targets in remote sensing images.The advancement of deep convolutional neural network technology has rejuvenated remote sensing image target detection,and the detection direction has also moved from horizontal bounding box target detection to more refined rotating bounding box detection.Based on remote sensing image target detection technology and general target detection technology,this paper analyzes and summarizes the shortcomings of existing methods,combined with the current multi-scale,multi-angle and other new challenges in remote sensing image detection.Corresponding improvements are made in the detection accuracy and detection speed of remote sensing targets in complex backgrounds,and a multiscale and multi-stage directional remote sensing target detection algorithm based on spacenon-local attention and a directional remote sensing target detection algorithm based on rotation center point estimation are proposed.The main tasks include:1.In terms of image preprocessing,aiming at the problems of large overlapping area of sub-images,large number of repeated samples,and single cropping size in the uniform sliding window method,a dynamic cropping method based on grid clustering is proposed to crop the high-resolution large remote sensing image into suitable subgraph input into the model.Explore a variety of data augmentation methods on remote sensing images,such as geometric transformation,color transformation,mixed sample enhancement,and multi-scale samples.The experimental results show that a reasonable data enhancement method can greatly improve the detection effect of remote sensing targets.The dynamic cropping method based on grid clustering improves the quality of the sample compared with the uniform sliding cropping,and can get better training effect with a smaller sample size.2.Aiming at the problems of low accuracy of remote sensing target positioning,unstable detection of dense small targets and multi-scale samples,and serious background noise interference,this paper first analyzes the advantages of multi-stage networks,and on this basis,proposes a multi-scale multi-stage network based on spatial-non-local attention.The network mainly includes three stages: the first stage is to predict HBB by the enhanced multi-scale feature extraction network and RPN subnet,the second stage is to predict the OBB initially by the Ro I transformation network,and the refined OBB is predicted through the aligned rotation invariant features to achieve high-precision rotation target positioning in the third stage.In order to suppress background noise and enhance the expression of target features,a binary mask label is designed as a spatial-non-local attention structure combining the spatial attention of the supervisory signal and the NL module.The experimental results on the DOTA dataset show that the detection effect of the multi-scale and multi-stage algorithm based on spatialnon-local attention is better than that of commonly used directional target detection algorithms such as Ro I Transformer and SCRNet.3.In order to speed up the forward inference speed of the network and enable remote sensing target detection tasks to run on terminal devices with limited computing resources for real-time processing,this paper proposes a directional remote sensing target detection algorithm RCNet based on the estimation of the rotation center point based on the framework of anchor-free horizontal target detection.The algorithm predicts the center point of the target by generating a key point heat map,converts the rotation frame detection into a key point estimation,and adds a predicted rotation angle and width and height to each center point.The RCNet network has fewer hyperparameters and a flat structure.It consists only of the backbone network and different lighter detection branches,which solves the efficiency problem of remote sensing target detection.By quantifying the trained model,the purpose of real-time detection is achieved.The experimental results verify that RCNet has a faster detection speed with good detection accuracy,and has strong practicability.