基于Mask RCNN的裂缝图像识别与渐进式自动标注方法研究
Research on Crack Image Recognition and Progressive Automatic Annotation Method Based on Mask RCNN
邱先志 1李登华 2郭林啸 3徐海涛 1丁勇3
作者信息
- 1. 国家能源集团新疆开都河流域水电开发有限公司,新疆 库尔勒 841009
- 2. 南京水利科学研究院,江苏 南京 210029;水利部土石坝破坏机理与防控技术重点试验室,江苏 南京 210029
- 3. 南京理工大学,江苏 南京 210094
- 折叠
摘要
文章提出了一种渐进式全自动标注算法,分三个阶段标注裂缝样本:首先,在白纸上画线绘制假裂缝,提取裂缝轮廓制作标签,训练生成一级权重文件;利用一级权重文件识别白墙上的裂缝,优化掩膜后制作标签,训练生成二级权重文件;最后,用二级权重文件分批检测混凝土裂缝,优化并提取掩膜轮廓,生成标签并循环训练生成三级权重文件.训练后Mask RCNN模型对三类图像的识别综合评价指标(Evaluation indicator)分别为 95.2%、83.3%、79.2%,识别率较高,可用于裂缝的快速识别.
Abstract
This paper presents a progressive fully automatic annotation algorithm,which annotates crack samples in three stages.Firstly,it draws fake cracks by drawing lines on white paper,and extracts the crack contours to make labels,and then trains to generate the first-level weight file.Secondly,it uses the first-level weight file to identify the cracks on the white wall,and optimizes the mask to make labels,and then trains to generate the second-level weight file.Finally,it utilizes the second-level weight file to conduct batch detection of concrete cracks,optimizes and extracts the mask contours to generate labels,and then trains cyclically to produce the third-level weight file.After training,the comprehensive evaluation indicators of the Mask RCNN model for the recognition of three types of images are 95.2%,83.3%,and 79.2%,respectively.The detection rate is relatively high and this model can be applied to the rapid recognition of cracks.
关键词
裂缝/渐进式自动标注/轮廓提取/深度学习Key words
crack/progressive automatic annotation/contour extraction/Deep Learning引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据