基于改进YOLOv8s的无人机目标检测算法

UAV target detection algorithm based on improved YOLOv8s

针对目前无人机航拍图像目标尺寸较小,图像背景复杂,导致现有的无人机目标检测算法检测精度较低的问题,提出一种改进YOLOv8s的无人机目标检测算法。首先使用可变形卷积替换标准卷积,以增强网络对不规则形状目标的特征提取能力;然后使用可分离大核注意力机制(LSKA)改进快速空间金字塔池化(SPPF)模块,改善因目标尺度差异较大导致检测精度较低的问题。在网络颈部结合双向特征金字塔网络(Bi-FPN)实现多尺度特征融合,改善网络对小目标的漏检和错检问题。在网络头部,使用自注意力机制动态检测头(DyHead)替换原检测头,增强对遮挡物体和小目标的检测能力。最后,针对数据集中存在大量低质量样本对训练过程产生负面影响的问题,使用Wise-IOU损失函数,提升模型收敛速度和检测精度。实验结果表明,改进后的方法在VisDrone2019数据集上获得了41.7%的平均精度均值(mAP),与原YOLOv8s算法相比,mAP@0.5提升了3.0%,mAP@0.5∶0.95提升了1.9%,参数量下降了17.5%,计算量下降了12.63%。实现了模型轻量化和检测精度双重提升。

Aiming at the problem that the target size of the current UAV aerial images is small and the image background is complex,which leads to low detection accuracy of existing UAV target detection algorithms,this article proposes an improved YOLOv8s UAV target detection algorithm.First,deformable convolution is used to replace standard convolution to enhance the network's feature extraction ability for irregularly shaped targets.Then the separable largekernel attention mechanism(LSKA)is used to improve the SPPF module to improve the problem of low detection accuracy due to large differences in target scales.The weighted bi-directional feature pyramid network(Bi-FPN)is combined at the neck of the network to achieve multi-scale feature fusion and improve the network's missed detection and false detection of small targets.At the head of the network,the dynamic detection head(DyHead)is used to replace the original detection head to enhance the detection ability of occluded objects and small targets.Finally,in order to solve the problem that a large number of low-quality samples in the dataset have a negative impact on the training process,the Wise-IOU loss function was used to improve the model convergence speed and detection accuracy.Experimental results show that the improved method achieved 41.7%mAP on the VisDrone2019 dataset.Compared with the original YOLOv8s algorithm,mAP@0.5 increased by 3.0%,mAP@0.5∶0.95 increased by 1.9%,the number of parameters decreased by 17.5%,and the amount of calculation dropped by 12.63%.It achieves both model lightweight and detection accuracy.