Research On Satellite Image Object Detection Based On Key Points And Rotating Bounding Box

Satellite image object detection is divided into horizontal bounding box detection and rotating bounding detection according to different detection tasks.Horizontal bounding box detection in the object detection task of satellite remote sensing images,due to the dense arrangement of object and the irregular shape,the detection result may contain a large proportion of background and the target overlaps.The detection method of rotating bounding box improves the proportion of foreground in the detection result through the regression method of rotating the quadrilateral around the target.However,for targets with important military value such as airplanes,ships,the corners of the detection quadrilateral can not reflect the semantic information of the target feature points.Because the foreground and background features of the label are not obvious during the network learning process,it is also more difficult to make the network converge normally and learn features effectively.Firstly,the horizontal bounding box detection and rotation bounding box detection in satellite image object detection task are studied,and the comparison datum of the research is established.Taking DOTA and i SAID satellite image remote sensing data sets as the research objectives,the benchmarks for horizontal bounding box and rotating bounding box object detection frames are designed.Aiming at the changeable target scale in the data set,the characteristics of small targets are mostly,the detection framework of FPN+Faster-rcnn is used as a benchmark,and the distribution of Anchor is as close as possible to the distribution of target size through the K-means clustering Anchor method.By increasing the output of the rotation corner point regression layer,a rotating bounding box object detection frame is designed on the frame of the horizontal bounding box detection,improve the detection result of the network on rotating bounding box and reduce the proportion of background in the detection results.Aiming at the problems existing in horizontal bounding box detection and rotation bounding box detection,the detection method based on key points is used,directly detect the four key points of the top right,bottom right,top left and bottom left of the satellite image target,which can reduce the proportion of the background in the detection result and increase the ability of the background semantic information in front of the network learning.A key point prediction network based on regression method is designed,and feature extraction is performed through deep neural network to complete the regression and category output after feature point coding.Realizing the model fusion with the binary classification horizontal frame detection network to achieve the end-to-end output of the model during inference,output the key point coordinates of the target from the original picture.In order to compare and analyze the performance effect of the key points of the regression method,design the migration application of key point detection network based on Heat-map,directly detect the four feature points of the top,bottom,left,and right of the target.The detection results are approximated to the segmentation results of the target by interpolation.At last,the performance test and time evaluation of each network on the uniform minimum enclosing rectangle validation set are carried out,and the results under different IOU thresholds are tested.The experimental results show that the key point detection network improves the target effect of most categories significantly on the minimum bounding rectangle verification set compared with the horizontal bounding box and the rotation bounding box,and the key point detection speed of the regression method is improved significantly compared with the Heat-map method.