基于改进YOLOv8的路面病害检测系统应用

Application of Pavement Disease Detection System Based on Improved YOLOv8

摘要:

针对道路病害检测中人工标注效率低、主观性强且难以规模化推广的问题，提出了一种基于改进YOLOv8的路面病害检测系统。该系统以CiCSI-C多功能路况快速检测车采集的线阵相机图像为数据来源，构建了包含68 310张图像、涵盖7类路面病害的分割数据集；在模型改进方面，引入多头自注意力机制（MHSA）与蛇形动态卷积（DySnakeConv）的组合对YOLOv8基础模型进行增强，并提出面向工程实际的“网格化掩膜IoU”评价指标。结果表明，改进模型精确率达80.4%，较基础模型提升0.9%，增强了对复杂路面病害的识别能力和整体检测性能。在此基础上，设计并实现了集算法推理、网格化处理和标准化输出于一体的交互式识别系统，形成了“算法初筛+人工复核”的作业模式，使检测效率较纯人工方式提高2倍以上，同时能保证病害标注质量，满足工程交付要求。

Abstract:

To address the issues of low efficiency, strong subjectivity and limited scalability of manual marking in road disease detection, a pavement disease detection system based on an improved YOLOv8 is proposed. Taking the line array camera images collected by the CiCSI-C multi-functional road condition rapid detection vehicle as the data source, a segmented data set containing 68310 images covering 7 types of pavement diseases is constructed. For the model improvement, a combination of multi-head self-attention (MHSA) mechanism and dynamic snake convolution (DySnakeConv) is introduced to enhance the YOLOv8 baseline model, and an evaluation index of “grid-based mask IoU” oriented to engineering practice is proposed. The results show that the improved model achieves a precision of 80.4%, an increase of 0.9 % over the baseline model, thereby enhancing the recognition ability for complex pavement diseases and the overall detection performance. On this basis, an interactive recognition system integrating algorithm inference, grid-based processing and standardized output is designed and implemented, forming an “algorithm-based preliminary screening + manual review” operational mode, which can increase the detection efficiency by more than twice compared with the pure manual method, while ensuring the quality of disease marking and meeting the requirements for project delivery.