Detection and width measurement of concrete apparent cracks based on computer vision
Efficient detection of apparent cracks in reinforced concrete(RC)structures can provide evidence for rapid assessment of earthquake-damaged structures.Such work exhibits large and repetitive characteristics in both earthquake sites and laboratory environments,therefore,it is suitable to adopt the computer vision technology to make up the inefficiency and uncertainty of manual methods.Using images from consumer-grade cameras as data sources,a convolutional neural network(CNN)model suitable for concrete apparent crack detection is constructed by integrating U-Net and VGG-16,and the model training and testing are completed based on a multi-type RC component crack image database.Morphological operations and Otsu threshold segmentation are used to further optimize the crack detection results as input data for width measurement.To reduce the measurement error of crack width caused by the non-perpendicularity of the camera axis to the crack plane,perspective error correction is performed on the original image using specific targets.After verification,the average deviation of the crack width measurement after perspective error correction can be reduced up to 25%.
NETL
NSTL
万方数据
