Recognition Of Concrete Surface Cracks Based On Fully Convolutional Network

The degradation of infrastructure such as bridges,roads and dams is accelerating due to the influence of the environment and loads.As one of the main defects affecting the structural quality of buildings,cracks are essential to monitor and keep buildings in good condition.The traditional manual crack detection results are easily affected,and the detection results are not only unreliable,but also time-consuming.The method of concrete crack detection based on traditional digital images requires high image preprocessing technology,and the detection results are easily affected by factors such as ambient light and noise.In deep learning,the convolutional neural network extracts crack boundary features to accurately identify cracks,and the detection effect has good accuracy and robustness.In this paper,the Seg Net neural network model is used to identify the cracks on the concrete surface,and the quantification after identification is systematically studied,and the quantification method of the crack characteristic parameters is obtained.The main research contents and conclusions of this paper are as follows:（1）This paper identifies concrete surface cracks based on a deep learning model,based on the Seg Net model.Using image enhancement and transfer learning to solve the problem of few image training sets.On this data set,four learning rates of 5×10^-5,1×10^-4,5×10^-4,and1×10^-3 are performed.By comparing the evaluation indicators,it is found that the recognition effect is the best when the learning rate is 1×10^-4.The Seg Net crack identification results are compared with the FCN model and the Otus algorithm,It is concluded that the Seg Net model is more robust and the detection results are more accurate.（2）An accurate fracture characteristic parameter quantification method is proposed to realize the analysis and calculation of fracture area,length,width and other information.First,the binary image after segmentation by the Seg Net model is processed,and morphological denoising is performed on the image with partial noise after segmentation,and the binary image is skeletonized by the Zhang-sun method.For complex cracks after refinement,there may be a large number of burrs around them.The burr is eliminated by setting the threshold length through Matlab,and the crack skeleton with connected segments is disconnected using the separation algorithm.（3）Based on the measurement method of deep learning,this paper conducts experimental tests on single reinforced concrete beams.The entire experiment includes image acquisition,region cropping,crack identification and parameter quantification.By photographing the entire crack area from loading to final failure,on this basis,the static distribution of cracks in the final failure of the beam and the dynamic distribution of cracks during the loading process are quantitatively calculated respectively.The fracture segmentation results were compared with the manual measurement results to verify the accuracy and practicability of the deep learning-based fracture measurement method.