|Table of Contents|

Defect recognition for steel bridge based on convolutional neuralnetwork and transfer learning(PDF)

长安大学学报（自然科学版）[ISSN:1006-6977/CN:61-1281/TN]

Issue:
2021年3期
Page:
52-63
Research Field:
桥梁与隧道工程
Publishing date:

Info

Title:
Defect recognition for steel bridge based on convolutional neuralnetwork and transfer learning
Author(s):
ZHU Jinsong1; 2;  LI Huan2;  WANG Shifang3
(1. School of Civil Engineering, Tianjin University, Tianjin 300072, China; 2. Key Laboratory of Coast CivilStructure Safety, Ministry of Education, Tianjin University, Tianjin 300072, China；3. School of Electrical Engineering, Anhui Polytechnic University, Wuhu 241000, Anhui， China)
Keywords:
bridge engineering;  steel bridge;  defect recognition;  convolutional neural network;  transfer learning;  Inceptionv4;  image processing
PACS:
-
DOI:
-
Abstract:
In view of the low efficiency and accuracy of steel bridge defect recognition, a steel structure apparent defect identification method based on deep learning was proposed. In this method, convolutional neural network Inceptionv4 and transfer learning were combined, and two training methods of feature extraction and finetuning in transfer learning were used to obtain two models, and an original Inceptionv4 model was trained for comparison. Firstly, 656 images of steel bridge diseases were collected, including 176 images of coating deterioration, 173 images of corrosion, 151 images of weld cracking and 156 images of intact. The data set was expanded to 3 742 images by rotating, flipping and adjusting contrast. According to the ratio of 8∶1∶1, the training set, verification set and test set were divided. Secondly, 3 365 steel bridge defect images were used to train and verify the feature extraction model, finetuning model and new training model respectively. The effects of batchsize and learning rate on the training effect of the model were compared, and the two parameters were optimized. Thirdly, 377 defect images were used to test these models. The results show that the epoch training time of feature extraction model, finetuning model and new training model was 47.2, 119.2 and 121.8 s, and the test accuracy was 89.39%, 97.88% and 91.25%, respectively. Compared with the new training model, the two models of transfer learning reduce the defect of data, improve the operation efficiency and the accuracy of defect recognition, but the finetuning model can achieve higher test accuracy with less epoch, which is more suitable for the practical application of steel bridge defect recognition. 5 tabs, 10 figs, 28 refs.

References:

-

Memo

Memo:
-
Last Update: 2021-06-04