Research On Remote Sensing Image Target Detection Algorithm Based On Deep Learning

With the rapid development of remote sensing technology,remote sensing image target detection algorithms have developed rapidly.However,many remote sensing image target detection algorithms still have problems such as insufficient detection accuracy.This article deeply analyzes some pain points of current remote sensing image target detection algorithms,and conducts research on remote sensing images with characteristics such as large image size,tilted and densely arranged targets,large target scale changes,and complex target background.In view of the characteristics that remote sensing images have inclined and densely arranged targets,a detection algorithm using rotation frame-YOLOv5s-rd is proposed,which is improved based on YOLOv5 s algorithm.Although YOLOv5 s algorithm has excellent performance,it is a horizontal target frame detection algorithm.In the face of the tilted and densely arranged objects in the remote sensing image,multiple prediction frames will overlap and cause visual confusion.Therefore,to solve this problem,YOLOv5 s is improved to use the rotation frame detection algorithm.Because the conventional detection algorithm using rotation frame adopts regression method for angle prediction,and the regression method is easy to produce boundary problems,this paper proposes a detection method using angle classification to avoid boundary problems.Gaussian function is used as window function in angle classification,and the influence of different Gaussian function radii on algorithm performance is compared.The experiment shows that the improved detection algorithm of rotation frame is suitable for the scene where the objects are inclined and densely arranged.In order to improve the detection accuracy of YOLOv5s-rd,further improvements have been made to the algorithm based on some characteristics of remote sensing images.Due to the huge size of the image,the original image is first cut,and then the cut image is subjected to Mosaic-9 data enhancement;An improved feature fusion network structure is proposed to fully integrate the information of low-level feature maps and high-level feature maps for the characteristics of large changes in target scales;Aiming at the complex characteristics of the target background,the Shuffle Attention mechanism is introduced to weaken the background and enhance the target information features.After analyzing the shortcomings of the Adam optimization algorithm,an improved Adarad optimization algorithm is proposed.Finally,the improved algorithm combined with the above improved points is compared with YOLOv5s-rd and popular algorithms for target detection in remote sensing images.Experiments show that compared to YOLOv5s-rd,S2A-Net,Re Det,Oriented R-CNN,Oriented Repoints,and SA-S&SA-M,the improved YOLOv5s-rd algorithm in this paper performs better in target detection on remote sensing images.