Research on object detection in remote sensing images based on improved horizontal target detection algorithm

This paper aims to enhance target detection in remote sensing images, crucial for military defense, urban planning, disaster response, and intelligent transportation. It proposes refining the You Only Look Once Version 7 (YOLOv7) target detection technology to address challenges like complex backgrounds, arbitrary space target directions, and unbalanced sample categories. The method integrates the YOLOv7 algorithm with the intersection ratio loss function of the Kalman filter, transforming horizontal frame detection into rotating frame object detection. This adaptation accurately represents the real shape of the target. Additionally, Self-Correcting Convolution (SCC) is introduced to further improve detection capability, addressing the low accuracy issue. Experimental results demonstrate enhanced accuracy, with an overall average improvement from 63.52% to 65.24%. Notably, roundabout detection accuracy improves significantly from 33.94% to 45.00%. The refined algorithm facilitates more precise target identification and localization in high-resolution remote sensing images. In conclusion, the proposed algorithm provides a valuable contribution to fine recognition of space targets in remote sensing imagery, benefiting both artificial intelligence research and practical engineering applications.