欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 图像与视频处理 • 上一篇    下一篇

基于 Mask R-CNN 与改进 Criminisi 的沥青路面 车道线移除方法

  

  1. (1. 重庆交通大学信息科学与工程学院，重庆 400074；2. 重庆市公路局，重庆 401147)
  • 出版日期:2019-06-30 发布日期:2019-08-02
  • 基金资助:
    重庆市科委民生重点项目(cstc2015shms)

Asphalt Pavement Lane Line Removal Method Based on  Mask R-CNN and Improved Criminisi

  1. (1. School of Information Science and Engineering, Chongqing Jiaotong University, Chongqing 400074, China; 
    2. Highway Bureau of Chongqing, Chongqing 401147, China)
  • Online:2019-06-30 Published:2019-08-02

PDF (PC)

66

摘要： 在对沥青路面病害图像进行自动分类时，含车道线的图像数量较多易造成干扰。 为此，提出一种车道线移除方法以降低其对分类的影响，首先基于 Mask R-CNN 网络训练出复 杂背景下车道线区域的检测模型，通过该模型自动获取车道线区域 mask；然后利用 mask 将车 道线区域全部移除得到破损图像；最后用改进的 Criminisi 图像修复方法对破损图像进行样本块 填充。实验表明，采用 Mask R-CNN 方法对 400 张不同环境下的路面图像进行检测，其漏检率 和误检率分别为 0.50%和 7.87%。在保证图像修复质量的基础上，改进的 Criminisi 方法在修复 速度上比改进前提升约 4~5 倍。同等条件下，采用 VGG 分类模型对比验证，经该算法移除车 道线后的新数据集表现更优。

关键词: 车道线检测, Mask R-CNN, 目标移除, Criminisi, 沥青路面

Abstract: In the automatic classification of the disease images of asphalt pavement, there are a great number of images with lane line, which is subject to interference. A method of lane line removal was proposed to reduce its impact on classification. Firstly, the detection model of the lane line region under complex background was trained based on the Mask R-CNN network, and the mask of the lane line region was automatically obtained through the model. Then the mask was used to completely remove all the lane line areas to get the damaged image. Finally, a modified Criminisi image inpainting method was used to fill the damaged image samples. Experiments show that the missed detection rate and the false detection rate are 0.50% and 7.87% respectively with the application of the Mask R-CNN method to detect the road image in 400 different environments. The improved Criminisi method enhances the repair speed by about 4 to 5 times than before under the premise of ensuring the quality of image restoration. Using VGG classification model for comparison verification, the new data set obtained after removing the lane line by the algorithm performs better under the same conditions.

Key words: lane line detection, Mask R-CNN, target removal, Criminisi, asphalt pavement