Zhang Shizheng, Liu Zhihao, Wang Kunpeng, Huang Wanwei, Li Pu
Software Engineering College, Zhengzhou University of Light Industry, Zhengzhou, Henan, China.
PeerJ Comput Sci. 2025 Jan 7;11:e2593. doi: 10.7717/peerj-cs.2593. eCollection 2025.
Effective and efficient detection of pavement distress is very important for the normal use and maintenance of roads. To achieve this goal, a new road damage detection method based on YOLOv8 is proposed in this article. Firstly, omni-dimensional dynamic convolution (ODConv) block is employed to better grasp the complex and diverse features of damage objects by making dynamic adjustment according to the features of input images. Secondly, to extract the global and local feature information simultaneously to better improve the feature extraction ability of the model, BoTNet is added to the end of the backbone, which can combine the advantages of convolutional neural network (CNN) and Transformer. Finally, the coordinate attention mechanism (CA) is incorporated into the Neck section to make more accurate speculations and enhance detection accuracy further which can effectively mitigate irrelevant feature interference. The new proposed model is named OBC-YOLOv8 and the experimental results on the RDD2022-China dataset demonstrate its superiority compared with baselines, with 1.8% and 1.6% increases in mean average precision 50 (mAP@0.5) and F1-score, respectively.
有效且高效地检测路面病害对于道路的正常使用和维护非常重要。为实现这一目标，本文提出了一种基于YOLOv8的新型道路损伤检测方法。首先，采用全维动态卷积（ODConv）模块，根据输入图像的特征进行动态调整，从而更好地把握损伤对象复杂多样的特征。其次，为同时提取全局和局部特征信息以更好地提高模型的特征提取能力，在主干网络末尾添加了BoTNet，它可以结合卷积神经网络（CNN）和Transformer的优势。最后，将坐标注意力机制（CA）融入Neck部分，以进行更准确的推断并进一步提高检测精度，这可以有效减轻无关特征干扰。新提出的模型名为OBC-YOLOv8，在RDD2022-China数据集上的实验结果表明，与基线相比它具有优越性，平均精度均值50（mAP@0.5）和F1分数分别提高了1.8%和1.6%。
