面向航拍路面裂缝检测的AC-YOLO

AC-YOLO for Aerial Pavement Crack Detection

针对当前道路巡检智能化程度不足以及效率低等现状,为提高利用无人机道路巡检检测效率与检测精度,在YOLOv8s的基础上针对无人机航拍场景提出改进模型AC-YOLO,在主干网络引入动态大卷积核注意机制LSK-attention来扩展模型的感受野,提高对路面裂缝范围检测的准确性。在颈部结构设计多尺度特征融合策略,融入BiFPN网络,改善对细小裂缝的检测。替换损失函数为WIoUv3,优化梯度分配策略,使模型更加关注裂缝主体。在数据集UAV-PDD2023上进行实验验证,改进后AC-YOLO精准度达到0.895,较原模型提高0.128,mAP50达到0.791,提高0.071,F1得分提高0.051,模型大小减小8.5%,FPS达到了129,提高4%。同时采用不同任务验证了模型泛化性能,实验结果证明AC-YOLO具有更强的检测性能,能有效应用于无人机视角下的路面裂缝检测。

In response to the current issues of insufficient automation and low efficiency in road inspection,this study aims to enhance the efficiency and accuracy of UAV-based road inspection.Based on YOLOv8s,an improved model,AC-YOLO,is proposed specifically for UAV aerial scenarios.First,the model integrates a dynamic large-kernel convolutional attention mechanism,LSK-attention,into the backbone network to expand the receptive field and improve the model’s accuracy in detecting road crack areas.Second,a multi-scale feature fusion strategy is introduced in the neck structure by incorporating the BiFPN network,enhancing the model’s ability to detect fine cracks.Finally,the loss function is replaced with WIoUv3,optimizing the gradient allocation strategy to enable the model to focus more precisely on crack regions.Experimental validation on the UAV-PDD2023 dataset demonstrates that the improved AC-YOLO achieves an accuracy of 0.895,representing a 0.128 increase compared to the original model.The mAP50 reaches 0.791,reflecting an improvement of 0.071,while the F1 score increases by 0.051.Moreover,the model’s size is reduced by 8.5%,and the FPS reaches 129,showing a 4%improvement.The model’s generalization performance is verified across multiple tasks,and the experimental results confirm that AC-YOLO offers superior detection capabilities,making it highly effective for UAV-based road crack detection.