Abstract: Aiming at the requirement of digital traffic marking damage diagnosis data, this paper proposes FE-Mask R-CNN, a Mask R-CNN-based model. Through target detection and instance segmentation, the damaged traffic marking and its corresponding repair image are obtained. Firstly, the traffic marking data set in a complex environment is collected. According to the size of the data set and the environmental characteristics of the road traffic marking data, the Mask R-CNN trunk is replaced by the improved VGG16. Secondly, a cross-linking FPN network is built to enable the model to adaptively realize the weight fusion of different feature layers and extract  important information between different feature layers. Finally, efficient channel attention (ECA) is added to each feature extraction network to capture the relationship between different channels, which enhances the feature extraction ability of the model in complex environments. The experimental results demonstrate that the improved VGG16 model achieves a 21.18% reduction in parameter count compared to the original model. By employing the cross-linking FPN network and the FE-Mask R-CNN model, the following performance metrics are obtained: under low threshold conditions (IoU=0.5), the object detection and semantic segmentation precision reach 97.93%, 98.7%, 97.74% and 97.8%; under high threshold conditions (IoU=0.75), the corresponding precision rates are 95.56%, 97.90%, 81.48% and 92.0%. The detection speed decreases from 33?FPS to 24?FPS. Notably, while the segmentation precision improvement is observed across thresholds, the substantial enhancement in high-threshold segmentation precision provides a robust data foundation for image restoration tasks.

Key words: complex environment, object detection, instance segmentation, damaged traffic marking, improved VGG16, cross-linking FPN, efficient channel attention

摘要： 针对数字化交通标线损伤诊断数据需求，提出了基于Mask R-CNN的FE-Mask R-CNN模型，通过目标检测与实例分割，获得损伤交通标线及其对应的修复图像。采集了复杂环境下的交通标线数据集，根据数据集的大小及分析数据中路面交通标线的环境特征，将Mask R-CNN主干替换为改进的VGG16。构建了跨层连接的特征金字塔（cross-linking FPN）网络，使模型能自适应实现不同特征层权重融合，挖掘不同特征层之间的重要性信息。在每一特征层提取网络中添加了efficient channel attention（ECA）注意力机制，以捕捉不同通道之间的关系，增强了模型在复杂环境下的特征提取能力。实验结果表明，采用改进的VGG16，参数量较原模型降低了21.18%；采用cross-linking FPN网络及改进后的Mask R-CNN模型，在低阈值（IoU=0.5）下，目标检测和语义分割精度分别为97.93%、98.7%、97.74%、97.8%；在高阈值（IoU=0.75）下，目标检测和语义分割精度分别为95.56%、97.90%、81.48%、92.0%；检测速度由33?FPS降低至24?FPS。相对于低阈值分割精度提升效果，高阈值下的分割精度的大幅提升，为图像的高精度修复提供了数据基础。

关键词: 复杂环境, 目标检测, 实例分割, 损伤交通标线, 改进的VGG16, 跨层连接特征金字塔, 注意力机制