Subgrade distresses recognition based on convolutional neural network(PDF)
长安大学学报（自然科学版）[ISSN:1006-6977/CN:61-1281/TN]
- Issue:
- 2019年02期
- Page:
- 1-9
- Research Field:
- 道路工程
- Publishing date:
Info
- Title:
- Subgrade distresses recognition based on convolutional neural network
- Author(s):
- SHA Aimin; CAI Ruonan; GAO Jie; TONG Zheng; LI Shuai
- (1. School of Highway, Changan University, Xian 710064, Shaanxi, China;2. School of Materials Science and Engineering, Changan University, Xian 710061, Shaanxi, China)
- Keywords:
- road engineering; convolutional neural network; GPR; subgrade distresses detection; image processing
- PACS:
- -
- DOI:
- -
- Abstract:
- Aimed at the data analysis of ground penetrating radar (GPR) technology for detecting subgrade distresses relies on manual identification，which classifying distresses was inefficient and inaccurate. The application of convolutional neural networks (CNN) was put forward to classify subgrade distresses automatically. The cascaded CNN consists of two convolutional neural networks, which were separately utilized to recognize distresses using lowresolution images and highresolution images. The processes for developing convolutional neural networks mainly include training, validating and testing. After developing the cascaded CNN, training and testing results were used to verify the stability of the cascaded CNN. The cascaded CNN was then compared with Sobel edge detection and Kvalue clustering analysis to demonstrate its superiority. The results show that the accuracies of the cascaded CNN in the training and validating processes were 97.46％ and 95.80%. The cascaded CNN has a high accuracy in identifying subgrade distress. When the cascaded CNN was operated at frequencies of 300, 500 and 900 MHz, the accuracies of image classification were 94.20%, 93.89%, and 94.57%, respectively. When dealing with different highway structures, the accuracies of the images were 94.80%, 94.78%, 94.28%, and 94.21%. The cascaded CNN has great stability with respect to both emission frequencies and highway structures. When the image resolution is low, Sobel edge detection and Kvalue clustering analysis cannot extract the distress information accurately. However, CNN can extract the distress information accurately with classifier 2. When the image resolution is high, Sobel edge detection and Kvalue clustering analysis can extract only some of the subgrade distress. The remaining distress information needs to be extracted manually. The cascaded CNN can identify subgrade cracks accurately and efficiently，compared with other algorithms. 3 tabs, 10 figs, 28 refs.
Last Update: 2019-04-01