改进YOLOv8的无人机航拍道路损伤检测方法

Road damage detection method based on improved YOLOv8 in UAV aerial images

针对无人机道路损伤检测任务存在精度低,小尺寸目标易漏检等问题,提出了一种无人机图像道路损伤检测模型。该模型基于YOLOv8,采用数据增强策略应对道路损伤样本分布不均的问题,提升模型训练效果的充分性;在骨干网络集成了全维度动态卷积,加强模型对道路损伤特征的提取;特征融合网络引入一种“聚集-分发”机制,增强融合不同特征图;模型中采用ADown下采样模块降低计算复杂度;最后利用Powerful-IoU损失函数,合理引导边界框的回归,加快模型收敛的同时提升对道路损伤目标的定位精度。经过真实数据集测试,改进模型的平均精度均值相较其他7种经典目标检测模型分别提高11%、14.1%、8%、5.3%、1.6%、3.8%和4.2%,证实了该模型在无人机图像道路损伤检测中的有效性。

Aiming to address the issues of low accuracy and missed detection of small-size targets in UAV road damage detection tasks,this paper proposes a road damage detection model for UAV images.The model is based on YOLOv8 and incorporates a data augmentation strategy to tackle the uneven distribution of road damage samples and enhance the effectiveness of model training.It integrates omni-dimensional dynamic convolution into the backbone network to improve feature extraction for road damage.Additionally,the feature fusion network employs a“gather-and-distribute”mechanism to better combine different feature maps.To reduce computational complexity,the model utilizes the ADown downsampling module.Finally,the Powerful-IoU loss function is employed to guide bounding box regression,speed up model convergence,and enhance localization accuracy of road damage targets.Testing on real datasets demonstrates that the improved model achieves increases in mean average precision of 11%,14.1%,8%,5.3%,1.6%,3.8%,and 4.2%,respectively,compared to seven other classical target detection models,confirming its effectiveness in UAV-based road damage detection.