基于YOLOv8L遥感图像旋转目标检测

Rotating object detection of remote sensing image based on YOLOv8L

针对遥感图像复杂背景下的目标(如船舶、飞机等)具有朝向任意、尺度变化较大、数量多、目标排列密集等特点,提出一种基于改进YOLOv8L的旋转目标检测算法,用带有角度的旋转框能够更加精确定位目标。首先,在网络Head部分增加解耦角度预测头,预测目标的角度信息;其次,融合坐标注意力机制模块提高模型抑制噪声的能力;最后,在Neck部分引入自适应空间特征融合模块,抑制不同尺度特征图之间融合特征信息时的不一致性,保留有效的信息并进行融合。实验结果表明,所提算法在DOTA数据集上的检测精度达到了73.85%,较原有YOLOv8L模型提升了3.53%。

The proposed algorithm utilizes an improved YOLOv8L model to detect rotating objects(such as ships and aircraft)in complex remote sensing images with arbitrary orientation,large scale variation,and dense array of objects.By incorporating a rotating frame with angle,the algorithm achieves more accurate target localization.Firstly,the decoupling angle prediction head is incorporated into the network′s head section to accurately forecast the angular information of the target object.Secondly,by integrating a coordinate attention mechanism module,the model′s capability to suppress noise is significantly enhanced.Lastly,an adaptive spatial feature fusion module is introduced in the neck section to effectively address inconsistencies in feature information fusion across different scales and retain valuable information for optimal fusion.The experimental results demonstrate that the proposed algorithm achieves a detection accuracy of 73.85%on the DOTA dataset,surpassing the original YOLOv8L model by 3.53%.