Research On Remote Sensing Image Object Detection Algorithm Based On Key Points

Remote sensing technology is an important window for mankind to explore and understand the earth in an all-round way,and it macroscopically conveys comprehensive information of various species on the surface and man-made technological products.With the rapid increase in the scale of remote sensing platforms and the rapid advancement of aerospace technology,remote sensing technology has achieved digital,all-round ground monitoring using professional platforms such as radar,aircraft,and satellites.The opening trend,resolution and accessibility of remote sensing images have also increased.Remote sensing image target detection,as a key part of remote sensing information processing,is an important technology related to social people’s livelihood and homeland security.It has been widely used in disaster monitoring,smart cities,traffic management,marine shipping,agricultural monitoring,military technology and other fields.,Designed to identify related targets such as buildings and warships from extremely far apart positions.This paper uses the key point-based detector CenterNet as the detection model.Because the background and features of remote sensing targets are complex,and the scales are diverse,the CenterNet network needs to retain more features,reduce interference from irrelevant backgrounds,and enhance the ability to fuse multi-scale features.In order to retain more remote sensing target features and enhance the CenterNet network’s ability to detect remote sensing targets with multi-scale features,this paper uses a jump interpolation fusion structure and a multi-pooling SPP module.In addition,the CBAM attention module has also been added to enhance the correlation of the characteristics of remote sensing targets in the channel and spatial dimensions.Combine these three modules to get CenterNet based on attention and spatial pyramid pooling.Experiments show that compared with a variety of advanced detection algorithms,the method in this paper has improved accuracy.In remote sensing image target detection,ship targets often have the characteristics of large aspect ratio,close arrangement and arbitrary direction.Therefore,conventional horizontal bounding boxes inspection ships are likely to cause problems such as missed inspections,so remote sensing ships in arbitrary orientation are suitable for detection using rotating bounding boxes.However,the current conventional angle-based rotating bounding box has the problems of being sensitive to angle changes and difficulty in joint training of the parameters of the bounding box.Therefore,this paper adds corner points as key points on the basis of the CenterNet of the jump interpolation fusion structure,and uses the rotation bounding boxes represented by the relative distance between the corner point and the center,and proposes the CenterNet（Center-rd）based on the corner point orientation to enhance the rotation bounding boxes stability.Experiments show that the accuracy of Center-rd has been improved in comparison with a variety of rotating bounding box detectors.