中文核心期刊
CSCD来源期刊
中国科技核心期刊
RCCSE中国核心学术期刊

重庆交通大学学报(自然科学版) ›› 2017, Vol. 36 ›› Issue (7): 39-45.DOI: 10.3969/j.issn.1674-0696.2017.07.07

• 道路与铁道工程 • 上一篇    下一篇

基于容差PR曲线的路面裂缝识别算法性能评价机制

彭博1, 姬然2   

  1. (1. 重庆交通大学 交通运输学院, 重庆 400074; 2. 俄克拉荷马州立大学,土木和环境工程学院,俄克拉荷马 74078, 美国)
  • 收稿日期:2016-01-05 修回日期:2016-02-24 出版日期:2017-07-31 发布日期:2017-07-31
  • 作者简介:彭博(1986—),男,四川南充人,博士,主要从事交通领域的图像与视频检测研究工作。
  • 基金资助:
    重庆市社会事业与民生保障科技创新专项项目(cstc2015shms-ztzx30002;cstc2015shms-ztzx0127);重庆市科委基础科学与前沿技术研究项目(cstc2017jcyjAX0473);重庆市教委科学研究项目(KJ1600513);重庆交通大学科研启动项目(15JDKJC-A002)

Performance Evaluation Mechanism of Pavement Crack Recognition Algorithms Based on PR Curve with Tolerance

PENG Bo1, JI Ran2   

  1. (1. School of Traffic and Transportation, Chongqing Jiaotong University, Chongqing 400074, P. R. China; 2. School of Civil & Environmental Engineering, Oklahoma State University, OK, 74078, USA )
  • Received:2016-01-05 Revised:2016-02-24 Online:2017-07-31 Published:2017-07-31

摘要: 为了客观合理地评价路面裂缝图识别算法的性能,提出了基于容差PR曲线的算法性能评价机制。首先,建立路面裂缝图像基准库,为裂缝识别算法性能评价提供统一规范的测试图像;其次,提出考虑容差的准确率-召回率和F1值计算方法,作为算法性能的衡量指标;然后,提出裂缝识别算法的PR曲线评价方法;最后,以两套裂缝识别算法为例进行PR曲线性能评价分析。结果表明:PR曲线评价机制确立了统一的基准测试图像、衡量指标及评价方法,能评价无参数、具有1个或多个参数的裂缝识别算法的性能(包括准确率-召回率、F1值及算法稳定性),有助于进行算法性能评价与改进。

关键词: 道路工程, 评价机制, 图像处理, 裂缝识别算法, 容差, PR曲线

Abstract: In order to assess the performance of pavement crack detection algorithm objectively and reasonably, an algorithm performance evaluation mechanism based on PR curve with tolerance was proposed. Firstly, a benchmark library of pavement crack images was built to provide the unified and standard test images for evaluating crack detection algorithm performance. Secondly, calculation methods of precision ratio, recall ratio and F1 value considering tolerance were put forward as measurement indicators of algorithm performance. Then, a PR curve based on evaluation method for pavement crack detection algorithms was proposed. At last, two crack detection algorithms were evaluated and analyzed as examples by using PR curve. Results show that, the proposed PR curve evaluation mechanism establishes the unified benchmark test images, measurement indicators and evaluation method, which can assess performance (including precision ratio, recall ratio, F1 values and algorithm stability) of crack detection algorithms without or with 1 or more parameters. It is helpful for algorithm performance evaluation and improvement.

Key words: highway engineering, evaluation mechanism, image processing, crack recognition algorithm, tolerance, PR curve

中图分类号: