Search

CN-118587183-B - Concrete crack identification method based on synthetic data set and semi-supervised learning

CN118587183BCN 118587183 BCN118587183 BCN 118587183BCN-118587183-B

Abstract

The invention provides a concrete crack identification method based on a synthetic data set and semi-supervised learning, which comprises the steps of 1) collecting an original data set, selecting a certain mainstream neural network model for supervised learning, 2) updating a reference identification model based on the semi-supervised learning method, 3) classifying label-free data in personal data sets based on the updated reference identification model, 4) selecting an image with higher confidence level, marking a pixel-level mask for the image to determine a crack segmentation model, 5) marking the image with higher confidence level by using the crack segmentation model, 6) obtaining skeleton lines of cracks based on the mask marking result of the cracks, calculating to obtain maximum width and average width, and carrying out error analysis of crack width, 7) verifying the concrete crack identification model, and 8) importing the shooting result into the concrete crack identification model to obtain the skeleton lines, the maximum width and the average width result of each crack.

Inventors

  • LI YONGGANG
  • FANG CHENG
  • JIA WANGLONG
  • ZHANG FAN
  • Chen Runke

Assignees

  • 中国二十二冶集团有限公司
  • 同济大学

Dates

Publication Date
20260512
Application Date
20240605

Claims (6)

  1. 1. A concrete crack identification method based on a synthetic data set and semi-supervised learning is characterized by comprising the following steps: step 1, collecting an original data set, respectively performing supervised learning by using a plurality of main stream neural network models, and selecting a reference recognition model from the main stream neural network models; step2, updating a reference recognition model based on a semi-supervised learning method; Step 3, classifying the unlabeled data in the personal data set based on the updated reference recognition model; step 4, selecting an image with higher confidence, labeling the image with a pixel level mask, comparing a plurality of semantic segmentation models based on the image and the mask labeling result, and selecting MaskIoU with the highest index as a crack segmentation model; step 5, carrying out pixel-level labeling on the picture with higher confidence coefficient by using a crack segmentation model; Step 6, obtaining skeleton lines of the cracks based on the mask data set and edge pixels of the cracks, which are obtained by dividing in the step 5, calculating to obtain the maximum width and the average width of the cracks, and carrying out error analysis on the width of the cracks to obtain a concrete crack identification model; step 7, verifying the concrete crack identification model; and 8, based on the shooting result of the camera on the surface of the structure, importing the shooting result into a concrete crack identification model, and dividing each crack to obtain the skeleton line, the maximum width and the average width result of each crack.
  2. 2. The concrete crack identification method based on the synthetic data set and the semi-supervised learning, which is characterized in that in the step 1, specifically, six public data sets are based, 500 images with and without cracks are screened out from each data set to form a total of 3000 comprehensive data sets, the supervised learning is carried out by adopting VGGNet, resNet, inception-ResNet and SE-ResNet-18 neural network models respectively, parameters, accuracy, precision, recall and F β score indexes of confusion matrixes and models are introduced to test the neural network models, and SE-ResNet-18 is selected as a reference identification model according to test results.
  3. 3. The concrete crack identification method based on the synthetic data set and the semi-supervised learning of claim 1, wherein in the step 2, the reference identification model is updated by adopting an active learning method of a stacked convolution self-encoder based on the unlabeled data in the personal data set.
  4. 4. The method for identifying the concrete cracks based on the synthetic data set and the semi-supervised learning according to claim 1, wherein in the step 4, the crack segmentation model is a Mask RCNN model with MaskIoU indexes highest, and the MaskIoU comparison index formula is as follows: inter=∑ (x,y)∈P (M c (x,y)+M G (x,y)==2) union=∑ (x,y)∈P (M C (x,y)＝＝1)+∑ (x,y)∈P (M G (x,y)==1)-inter Mask_IoU=(TP)/(TP+FP+FN)。
  5. 5. the method for identifying concrete cracks based on the synthetic data set and semi-supervised learning according to claim 1, wherein in step 6, The crack width is identified as that a picture is shot and the oblique photo is processed into a positive photo, the converted positive photo crack image is input into a Mask RCNN model which is trained, a binary mask with a foreground as a crack pixel and a background as a concrete surface is obtained, two crack edges of a crack are defined as L 1 and L 2 respectively, and then the width, the maximum width and the average width are defined as follows: d i ＝min[(x i ,y i ),(x j ,y j )],(x i ,y i )∈L 1 ,(x j ,y j )∈L 2 d max ＝max(d i ) d Ave ＝sum(Mask)/L Ave L Ave in the formula 7 is the edge, and the length of the crack skeleton line is determined by L 1 and L 2 ; The maximum width and average width of the crack are calculated as follows, according to the definition of the width of the crack, based on the extracted mask edge, any pixel on the edge L 1 is replaced by the outer edge corner point of the pixel, euclidean distance from the corner point to the outer edge corner point of the edge L 2 is calculated, and after any pixel on the L 1 traverses all the corner points on the L 2 , the minimum value of the calculation result is the width of the crack generated by the point: w max ＝min[(x i ,y i ),(x j ,y j )],(x i ,y i )∈L 1 ,(x j ,y j )∈L 2 The calculation of the crack length comprises the steps of adopting a connecting line of width midpoints corresponding to each point on an envelope line of the outer edge, extending the connecting line to an external rectangular frame of all the crack pixels, calculating the midpoints of the crack width connecting line according to L 1 and L 2 respectively, obtaining midpoints of the corresponding widths of each point of L 1 and L 2 , connecting all the calculated midpoints with nearest points to obtain a skeleton line of the crack, calculating the sum of lengths of folding lines of the crack skeleton, and obtaining the total length of the crack, wherein the average width of the crack can be obtained by the following formula: W Ave ＝γ×sum(Mask)/L Ave ; The maximum width error analysis of the crack is that the error of the maximum width mainly originates from the approximation of the pixel in the width direction of the crack, the width direction of the crack is theta, the actual width of the crack is D, and the theoretical horizontal coordinate difference value Deltax and the vertical coordinate difference value Deltay and the actual calculated horizontal coordinate difference value x and the vertical coordinate difference value y are as follows: Δx=Dcosθ,Δy=Dsinθ,0°≤θ≤90° Δx-0.5λ<x<Δx+0.5λ,Δy-0.5λ<y<Δy+0.5λ the theoretical error is calculated as follows: further simplifying and obtaining: 。
  6. 6. The concrete crack identification method based on the synthetic data set and the semi-supervised learning of claim 1 is characterized in that in step 7, a camera is calibrated based on Camera Calibrator tool boxes in MATLAB, and the concrete identification model result is verified based on a simple scene, a complex scene and an actual engineering scene.

Description

Concrete crack identification method based on synthetic data set and semi-supervised learning Technical Field The invention relates to the field of structural health monitoring, in particular to concrete crack identification and detection based on computer vision. Background At present, the bridge detection work still mainly depends on manual detection, generally, a telescope, a crack observer and other instruments are used for determining the position of a crack in a short distance, and then the width of the crack is measured. Although the method can finish the detection of most middle and small bridges, the method consumes very much manpower, material resources and time, which is a difficult problem to be solved in bridge detection work. Under the current environment of the rapid increase of traffic and maintenance workload, the traditional method for detecting cracks cannot meet engineering requirements more and more, and especially, the management of manual drawing files easily causes the loss of project information, so that continuous tracking is difficult to carry out. Therefore, a new intelligent crack identification method is needed in engineering to realize the automation and the intellectualization of the detection process. Along with the development of computer science, the digitization of the image can convert the actual image into a digital matrix which can be processed in the computer, which lays a foundation for the image recognition of the crack. In recent decades, software and hardware rapidly progress, so that the speed and the precision of digital image processing are improved, and the application of digital images to the detection field is possible. As the computing power is improved, the processing method of the digital image is also continuously developed to be complicated and refined, and the effect is gradually diversified. At present, a deep convolutional neural network model is mostly used for crack detection based on image recognition, positioning and pixel level classification tasks can be completed in engineering projects, but the problems still exist that (1) the deep convolutional neural network trained from scratch, the final result is related to various factors such as a neural network structure, initialization of network layer parameters, value taking of custom super parameters, an optimization algorithm and the like, so that the test result after each training has certain fluctuation. (2) When training a model, hundreds of images are often marked to meet the precision requirement, but the marked images can only be marked manually, which consumes a great deal of time and effort. Meanwhile, if all the images need to be marked manually, the method is equivalent to the fact that engineering has completed manual detection, and the meaning of training the neural network to conduct identification is not great. (3) The neural network model obtained by training can only reflect the characteristics of the training set, and when the model after training is applied to a new project, the model effect is also reduced because the characteristic distribution of the new project and the original training set is not completely the same. Disclosure of Invention The invention aims to solve the problems, and provides a concrete crack identification method which has high detection speed and can obtain stable and accurate result data. The invention solves the problems, and adopts the following technical scheme: a concrete crack identification method based on a synthetic data set and semi-supervised learning, Step 1, collecting an original data set, respectively performing supervised learning by using a plurality of main stream neural network models, and selecting a reference recognition model from the main stream neural network models; step2, updating a reference recognition model based on a semi-supervised learning method; Step 3, classifying the unlabeled data in the personal data set based on the updated reference recognition model; step 4, selecting an image with higher confidence in the personal data set, labeling the image with a pixel level mask, comparing a plurality of semantic segmentation models based on the image and the mask labeling result, and selecting MaskIoU with the highest index as a crack segmentation model; step 5, carrying out pixel-level labeling on the picture with higher confidence coefficient by using a crack segmentation model; Step 6, obtaining skeleton lines of the cracks based on the mask data set and edge pixels of the cracks, which are obtained by dividing in the step 5, calculating to obtain the maximum width and the average width of the cracks, and carrying out error analysis on the width of the cracks to obtain a concrete crack identification model; step 7, verifying the concrete crack identification model; and 8, based on the shooting result of the camera on the surface of the structure, importing the shooting result into a concrete crack identification model,