YOLO-SWLFP: Rotating Object Detection Method in Remote Sensing Images with Embedded Displacement Window and Leap-Forward Feature Pyramid Structure

The existing remote sensing image detection technology often has problems such as difficult to accurately locate the target frame, overlapping and missing detection results, which is because remote sensing images often have problems such as large information, large size span of the target, and random direction arrangement. To solve these problems, based on the latest one-stage high-precision YOLOv8x-OBB rotating target detection model, a rotating target detection method with embedded displacement window and leap-forward feature pyramid structure (YOLO-SWLFP) is proposed in this paper. Firstly, a displacement window is added to the Backbone part of YOLOv8x-OBB network to solve the problems of small-scale target stacking and missing detection, and the original attention unit module is replaced by an improved DCWF module to optimize the computational load. Secondly, a leapforward feature fusion structure is proposed to characterize multi-scale target features to solve the problem that the scale difference of the target to be detected is large, which affects the detection accuracy. For the problems such as poor generalization performance of the original model and incompatibility of classification Loss function, image clipping and Mosaic data enhancement are used to improve the generalization performance. VariFocal Loss, which is more suitable for rotation detection tasks, is used to improve the robustness of the network. The proposed method is validated on the public DOTAv1.0 dataset, and the validation indicators include precison, recall, and mAP. The results show that the proposed method achieves a mAP yield of 5% compared with the baseline, and the improved model structure proposed in this paper can significantly improve the precision of rotating target detection in satellite remote sensing images.