| 一种基于改进实例分割模型的路面裂缝检测方法 |
| |
| 引用本文： | 肖力炀,李伟,袁博,崔逸群,高荣,王文庆.一种基于改进实例分割模型的路面裂缝检测方法[J].武汉大学学报(信息科学版),2023,48(5):765-776. |
| |
| 作者姓名： | 肖力炀 李伟 袁博 崔逸群 高荣 王文庆 |
| |
| 作者单位： | 1.西安热工研究院有限公司, 陕西 西安, 710032 |
| |
| 基金项目： | 国家重点研发计划2018YFB1600202国家自然科学基金51978071国家自然科学基金青年科学基金51908059 |
| |
| 摘 要： | 为了解决现有裂缝识别算法准确度不高、检测与分割任务不能同时进行等问题，提出了一种基于改进型Mask R-CNN模型的路面裂缝识别方法。首先，建立裂缝数据集并进行标注，然后使用Mask R-CNN模型对裂缝数据集进行训练和测试，并对模型中锚点的长宽比进行调整，实现在裂缝定位的同时对生成的检测框内的裂缝像素进行分割；其次，针对Mask R-CNN模型生成的裂缝检测框不精准的问题，设计了C-Mask R-CNN多阈值检测方法，通过结合级联不同阈值的检测器来提高候选框质量，实现高阈值检测下的裂缝精准定位；最后，对改进后的模型进行一系列的优化参数和实验对比，并验证所提模型的有效性。实验结果表明，C-Mask R-CNN模型检测部分的平均准确率均值（mean average precision，mAP）达到0.954，与改进前模型相比提升了9.7%，分割部分的mAP达到0.935，与改进前相比提升了13.0%，识别效果较好。综上所述，C-Mask R-CNN模型可以较为完整地对裂缝进行定位及提取，识别精度较高。
|
| 关 键 词： | 路面裂缝识别 深度学习 Mask R-CNN模型 级联阈值检测器 道路工程 |
| 收稿时间： | 2021-12-07 |
A Pavement Crack Identification Method Based on Improved Instance Segmentation Model |
| |
| Affiliation: | 1.Xi'an Thermal Power Research Institute, Xi'an 710032, China2.School of Information Engineering, Chang'an University, Xi'an 710064, China |
| |
| Abstract: | Objectives To solve problems that the existing crack identification algorithms are not accurate and the detection and segmentation tasks cannot be performed simultaneously, we propose a pavement crack identification method via the improved Mask R-CNN model. Methods First, the crack dataset is collected and labeled. And the crack dataset is trained and tested by Mask R-CNN model, and the aspect ratios of the anchor points in the model are adjusted to segment the crack pixels in the generated detection box while the crack is located. Second, to solve the problem that the crack detection boxes generated by Mask R-CNN model are inaccurate, C-Mask R-CNN is designed to improve the quality of crack region proposal by cascading multi-threshold detectors and achieve accurate positioning under high threshold. Finally, a series of optimization parameters and experimental comparison are carried out for the improved model, and the effectiveness of the proposed model is verified. Results The experimental results show that the mean average precision (mAP) of C-Mask R-CNN model in the detection part is 0.954, which is 9.7% higher than that of the conventional model, and its mAP in the segmentation part is 0.935, which is 13.0% higher than that of the conventional model. It confirms that the C-Mask R-CNN model performs well in identifying cracks. Conclusions In summary, the proposed C-Mask R-CNN model can locate and extract cracks with high identification accuracy. |
| |
| Keywords: | |
|
| 点击此处可从《武汉大学学报(信息科学版)》浏览原始摘要信息 |
|
点击此处可从《武汉大学学报(信息科学版)》下载全文 |
|
