Convolutional neural networks based quantitative crack assessment
Crack detection and quantification is a routine practice and an important component of civil infrastructure inspection and management.To this date,however,visual and manual inspection is still the standard procedure that is costly,time-consuming,and sometimes dangerous to inspectors.With technology advances today,digital imaging has become significantly low-cost and ready for use in practice(e.g.the use of aerial unmanned aerial vehicles with camera payload for remote inspection).With this readiness,the technical challenge becomes how to deal with massive imagery data quickly – not only to detect damage but quantitatively assess damage in a near real-time pace.This paper presents a deep learning-based method for crack damage assessment,including both classification-based detection and quantitative assessment.In the proposed method,the trained and validated Convolutional Neural Networks(CNN)is adopted as a crack damage classifier to detect the crack,which answers the questions of if there is damage and where the approximate location of the damage.With this CNN-based detection step,the traditional threshold-based segmentation method is used to further segment the refined cracks from the imagery background.Based on the extracted cracks skeletons,it is able to further split the cracks and separated crack-zones are obtained.Finally,the trained and validated CNN classifier is utilized again to check each separate crack-zone and to reject the meaningless ones with high level of noise.A novel quantitative assessment step is further developed.For the wider cracks,the crack skeletons are used to calculate the perpendicular orientation of the crack,and then the thickness of crack is obtained by counting thickness pixels.For the thin cracks(thickness pixels less than 5),a crack thickness calculation algorithm based on the Zernike moment method is proposed.A laboratory test is conducted on a concrete beam destroyed in static loading testing.The experimental results obtained from the proposed method agree well with the direct measurements,which demonstrates the effectiveness of proposed method for quantitative cracks detection.We finally envision that the proposed methodology,if implemented to an aerial vehicle(e.g.an automated drone),paves the way for realizing an automated damage assessment system for concrete structures,which would ultimately liberate human-based inspection from the cumbersome and dangerous practice.
Deep learning Convolutional Neural Networks Crack detection Quantitative assessment Thickness extraction
Futao Ni ZhiQiang Chen Jian Zhang
School of Civil Engineering,Southeast University Department of Civil and Mechanical Engineering,University of Missouri-Kansas City
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
2824-2831
2018-07-22（万方平台首次上网日期，不代表论文的发表时间）