Automated crack detection from large volume of concrete images using deep learning
Most of the industrialized countries are developing economical and effective structural monitoring system using vision-based technique to maintain increasing numbers of aged civil infrastructures.Deep learning-based crack detection methods are emerging as one of the promising methods among many vision-based crack detection techniques.However,the existing deep learning-based crack detection methods are developed for near-ideal condition and have low applicability to field inspection.This paper proposes an enhanced deep learning-based concrete crack detection method using crack images scraped from the Internet,transfer learning and reliability map.A commercial web scraper,ScrapeBox,were used to collect images of diverse concrete crack,intact surface and objects similar to crack from the Internet.A well-known convolutional neural network,AlexNet,was adopted for transfer learning and trained for crack detecting purpose.The crack detection network has four different class consisting of Crack,Joint with Single Line,Joint with Multiple Lines,Intact Surface and Plant,which benefits in the accuracy of crack detection.The reliability map is calculated after 50%overlapping sliding window detection as the average value of softmax layer for Crack class.Due to the diverse training images and the implementation of reliability map,the proposed method shows better performance in detecting crack out of concrete surface having many obstacles for detection.The validation result with 30 crack images shows the high applicability of the proposed method.
Crack Transfer learning Internet-based image set Reliability map
B.Kim S.Cho
Department of Civil Engineering,University of Seoul,Seoul,Republic of Korea
国际会议
The 7th World Conference on Structural Control and Monitoring(7WCSCM)(第七届结构控制与监测世界大会)
青岛
英文
1530-1535
2018-07-22（万方平台首次上网日期，不代表论文的发表时间）