CN113066041A - Pavement crack detection method based on stack sparse self-coding deep learning - Google Patents

Pavement crack detection method based on stack sparse self-coding deep learning Download PDF

Info

Publication number
CN113066041A
CN113066041A CN201911383082.1A CN201911383082A CN113066041A CN 113066041 A CN113066041 A CN 113066041A CN 201911383082 A CN201911383082 A CN 201911383082A CN 113066041 A CN113066041 A CN 113066041A
Authority
CN
China
Prior art keywords
crack
image
sub
subblocks
pavement
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN201911383082.1A
Other languages
Chinese (zh)
Inventor
吕建勇
唐振民
黄波
钱彬
徐威
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Science and Technology
Original Assignee
Nanjing University of Science and Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Science and Technology filed Critical Nanjing University of Science and Technology
Priority to CN201911383082.1A priority Critical patent/CN113066041A/en
Publication of CN113066041A publication Critical patent/CN113066041A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/0002—Inspection of images, e.g. flaw detection
    • G06T7/0004—Industrial image inspection
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/187—Segmentation; Edge detection involving region growing; involving region merging; involving connected component labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20021—Dividing image into blocks, subimages or windows
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20081—Training; Learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20084—Artificial neural networks [ANN]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30108—Industrial image inspection
    • G06T2207/30132—Masonry; Concrete

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Geometry (AREA)
  • Quality & Reliability (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a pavement crack detection method based on stack sparse self-coding deep learning, which comprises the steps of firstly, acquiring a pavement image by adopting a high-precision CCD camera and laser lighting equipment when a vehicle runs; then calculating the anisotropic characteristic value of the crack, performing sub-block pre-classification on the image, and extracting a crack sub-block training set; then, parameter training is carried out on the stack sparse self-coding model, and the deep learning model and the classifier are cascaded for fine tuning training again; after the model training is finished, the model is used for identifying the sub-blocks of the crack image; denoising and connecting the whole crack image by using a lattice tower space theory; and finally, carrying out relevant statistics according to the detected crack result, and providing a basis for road maintenance. The invention has higher detection precision and good adaptability to road surfaces with different textures.

Description

Pavement crack detection method based on stack sparse self-coding deep learning
Technical Field
The invention belongs to the field of pavement disease detection, and particularly relates to a pavement crack detection method based on stack sparse self-coding deep learning.
Background
Pavement crack detection belongs to an important component of a pavement maintenance management system, and early discovery and treatment can effectively improve the safety of a highway and greatly reduce the maintenance cost of the highway. The traditional manual detection method has low efficiency, inaccuracy and high risk and is easily influenced by subjective evaluation. Therefore, in recent years, high-speed CCD cameras are generally adopted at home and abroad to acquire road surface images, and then crack images are processed and analyzed by a computer, and the effectiveness of the adopted crack detection algorithm directly influences the acquisition and statistics of crack information.
Current crack detection algorithms are mainly classified into two categories: an image processing-based detection method and a machine learning-based detection method. The traditional image detection-based method mainly adopts algorithms such as a histogram, a morphological operator, Sobel and Canny segmentation to analyze the characteristics of a crack part, mainly obtains the difference between the crack and a background according to the typical characteristic that the crack presents local dark gray, and then identifies the enhanced crack. In order to overcome the above disadvantages, some more complex image processing algorithms have been proposed in succession, which distinguish and identify cracks and noise in the frequency domain by some transform domain methods, such as wavelet transform, NTSC transform, spatial and frequency domain transform of the crack image. Or after the cracks are preliminarily identified, a complex space search strategy is adopted for denoising and connecting, the method has a good processing effect on the cracks, but the algorithm is complex, the crack detection effect is poor under a non-uniform background, and the robustness is low. With the research development of machine learning, some researchers at home and abroad try to perform block detection on cracks by adopting a method based on crack image local feature training and recognition. And classifying the crack sub-blocks by a BP neural network after the crack characteristics are obtained through template operation. Or acquiring the crack image characteristics by adopting a morphological operator and then classifying the crack image characteristics by an SVM (support vector machine). Or by extracting the mean and variance characteristics of the image subblocks, adopting other mode identification methods to detect the crack subblocks. The method can effectively utilize the priori information of the training set to detect the cracks, and can obtain a good detection effect on the road surfaces with different textures, but the methods are insufficient in characteristic excavation of crack sub-blocks, and mostly only adopt the gray scale or variance of the sub-blocks to identify, so that a satisfactory effect cannot be obtained on the accuracy.
The problem that the current pavement crack detection field faces is mainly that the detection precision is low due to the complex pavement environment. The actually shot road surface image comprises various textures and various illumination conditions, and the average gray scale and the roughness of the presented image are inconsistent; white lane lines and other irregular obstacles formed after disintegration; the traditional algorithm cannot meet the variable texture pavement change and interference, crack details are easy to lose during detection, and false detection is easy to occur, so that a new method is necessary to be provided, the crack features are more fully excavated according to the actual application environment, the crack features are separated from the background and the interference, and the detection precision is improved.
Disclosure of Invention
The invention aims to provide a pavement crack detection method based on stack sparse self-coding deep learning, which has good robustness and higher detection precision on pavement cracks with different textures and certain noise.
The technical solution for realizing the purpose of the invention is as follows: a pavement crack detection method based on stack sparse self-coding deep learning comprises the following steps:
step 1, obtaining a pavement gray level image;
step 2, performing presorting on the anisotropic subblocks, namely determining corresponding anisotropic characteristic values according to the gray mean value and the variance of the subblocks by taking the image subblocks as a unit, segmenting the characteristic images of the image subblocks aiming at the cracked image and the crack-free image, extracting the subblocks from the maximum connected region of the cracked region and the crack-free region, and respectively marking the subblocks as a crack set and a crack-free set;
step 3, training the deep learning model, firstly performing normalization processing on the original subblock image data, then performing iterative training, and using the trained model structure parameters for crack subblock identification;
step 4, denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory;
and 5, performing relevant parameter statistics according to the final crack detection image to finish the pavement crack detection.
A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of:
(1) obtaining a road surface grayscale image
Acquiring a pavement gray image through a high-precision CCD camera and a laser lighting device, triggering camera shooting through a GPS-based mileage sensor, and meeting the real-time image acquisition requirement by adopting a multi-thread parallel control technology;
(2) anisotropic subblock pre-sorting
Firstly, taking an image sub-block as a unit, and calculating a corresponding anisotropic characteristic value according to the gray mean value and the variance of the sub-block;
secondly, segmenting feature maps of the crack image and the crack-free image, extracting sub-blocks from the maximum communication areas of the crack area and the crack-free area to serve as model training databases, and respectively marking the sub-blocks as a crack set and a crack-free set;
(3) deep learning model training
Firstly, normalizing the original sub-block image data, and if the sub-block is too large, performing data dimension reduction by adopting PCA (principal component analysis).
And step two, adopting a stack sparse self-coding model, cascading the output of the stack sparse self-coding model with a softmax classifier, setting the number of layers and the number of nodes, and performing iterative training.
And thirdly, using the trained model structure parameters for crack sub-block identification.
(4) Space denoising and linking based on lattice tower theory
And denoising and connecting the identified crack sub-blocks by adopting the characteristics of common circularity and continuity according to the visual principle of the lattice tower.
(5) Calculating fracture related parameters
And carrying out relevant parameter statistics according to the final crack detection image, wherein the parameters comprise crack types, lengths, areas and the like.
Compared with the prior art, the invention has the beneficial effects that: according to the method, various crack characteristics can be effectively described through an anisotropic characteristic value algorithm; various cracks can be effectively detected through a stack sparse self-coding deep learning model; through the lattice tower theoretical model, the denoising and the connection can be effectively realized. The invention has higher detection precision and good adaptability to road surfaces with different textures.
Drawings
FIG. 1 is a flow chart of a pavement crack detection method based on stack sparse self-coding deep learning of the present invention
FIG. 2 is a 4-direction neighborhood anisotropic model diagram of the pavement crack detection method based on the stack sparse self-coding deep learning
FIG. 3 is a deep learning model diagram of the pavement crack detection method based on the stack sparse self-coding deep learning of the present invention
Fig. 4 is a transverse crack image detection effect diagram of the pavement crack detection method based on stack sparse self-coding deep learning, wherein the diagram (a) is a transverse crack, and the diagram (b) is a detection effect diagram of the method.
Fig. 5 is a longitudinal crack image detection effect diagram of the pavement crack detection method based on stack sparse self-coding deep learning, wherein the diagram (a) is a longitudinal crack, and the diagram (b) is a detection effect diagram of the method.
Fig. 6 is a graph of the detection effect of the network crack image of the pavement crack detection method based on the stack sparse self-coding deep learning, wherein the graph (a) is the network crack, and the graph (b) is the detection effect graph of the method.
Detailed Description
With the combination of the attached drawings, the invention discloses a pavement crack detection method based on stack sparse self-coding deep learning, which comprises the following steps:
step 1, obtaining a pavement gray level image;
step 2, performing presorting on the anisotropic subblocks, namely determining corresponding anisotropic characteristic values according to the gray mean value and the variance of the subblocks by taking the image subblocks as a unit, segmenting the characteristic images of the image subblocks aiming at the cracked image and the crack-free image, extracting the subblocks from the maximum connected region of the cracked region and the crack-free region, and respectively marking the subblocks as a crack set and a crack-free set;
the specific steps for determining the characteristic value of the anisotropy are as follows:
step 2-1, partitioning the collected crack image into blocks with the size of NxN, and then determining the gray mean and variance of each sub-block; n is a natural number; said N is preferably 32.
Step 2-2, fusing the mean value and the variance of the gray scale of the subblocks as the characteristic values of the subblocks, namely
B＝-Bmean+λBstd
Wherein B represents the calculated syndrome of the crack sub-block, BmeanRepresenting the mean of the grey levels of the sub-blocks, BstdRepresents the variance of the sub-block, and λ represents the relative influence coefficient of the mean and variance;
step 2-3, determining the anisotropic characteristic value of the subblock, wherein the formula is as follows:
Bθ＝median{B-d,θ,B-d+1,θ,…,B-1,θ,B0,θ,B-1,θ,…,Bd-1,θ,Bd,θ},
in the formula, theta belongs to {0 degrees, 45 degrees, 90 degrees, 135 degrees }, d is larger than or equal to 1, theta represents the neighborhood direction, d represents the neighborhood size, mean represents the median operation, the difference between the maximum direction characteristic value and the minimum direction characteristic value is taken as the final anisotropy characteristic value T of the subblock, namely the difference between the maximum direction characteristic value and the minimum direction characteristic value is taken as the final anisotropy characteristic value T of the subblock
T＝max(Bθ)-min(Bθ)。
Classifying the whole image through an anisotropic algorithm, judging whether the image contains cracks or not, segmenting a characteristic image of the image with the cracks according to the image with the cracks, extracting sub-block images from the maximum connected domain of a crack area, marking a label as 1, and taking the sub-block images as a crack sub-block training set; and (3) segmenting the characteristic graph of the crack-free image, extracting the sub-block image from the maximum connected domain of the crack-free area, marking the label as 0, and taking the sub-block image as a crack-free sub-block training set.
Step 3, training the deep learning model, firstly performing normalization processing on the original subblock image data, then performing iterative training, and using the trained model structure parameters for crack subblock identification;
step 3-1, carrying out normalization processing on the pixel data of the original subblocks, and carrying out stack sparse self-coding optimization on a crack subblock training set, wherein a single sample loss function is as follows:
Figure BDA0002342759850000051
wherein { (x)(1),y(1)),(x(2),y(2)),…,(x(m),y(m)) The method comprises the following steps of (1) representing m training samples, wherein x is a single sample characteristic, y is a sample corresponding output, and a loss function of a sample training set is as follows:
Figure BDA0002342759850000052
by adding sparse constraint to the hidden layer of the sample, more robust detection characteristics are obtained, namely
Figure BDA0002342759850000053
Wherein beta controls the weight of the sparse penalty factor, s is the number of hidden nodes,
Figure BDA0002342759850000054
representing the average value of the output of a single node of the hidden layer, wherein rho is a sparse parameter, and KL is a relative entropy measurement criterion:
Figure BDA0002342759850000055
and 3-2, extracting the features of all the training samples by using a feature detector, putting the training samples into a softmax classifier for supervised learning to obtain classification parameters of the softmax, cascading stack sparse self-coding and the softmax classifier, and carrying out the supervised learning on the whole network again to complete the optimization of the parameters of the whole network.
Step 4, denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory; the method specifically comprises the following steps:
step 4-1, performing second-order tensor coding on the whole image by taking the pixel as a unit so as to enable the image gray information and the direction information to be fused into a unified tensor field;
4-2, carrying out tensor voting on each point to surrounding neighborhoods, and counting the difference of primary and secondary direction vote values of each point after the voting is finished to be used as a spatial significance value of the point;
and 4-3, performing threshold segmentation on the finally formed tensor field, reserving pixel points with larger tensor values, and realizing binary representation of the cracks so as to finish extraction of the cracks. The larger tensor is the tensor that exceeds the threshold. The threshold value is set according to specific conditions.
And 5, performing relevant parameter statistics according to the final crack detection image to finish the pavement crack detection. The parameters include crack type, length and area.
A pavement crack detection system based on stack sparse self-coding deep learning comprises:
the acquisition module is used for acquiring a road surface gray level image;
the presorting module is used for presorting the anisotropic subblocks, firstly, the image subblocks are taken as a unit, corresponding anisotropic characteristic values are determined according to the gray mean value and the variance of the subblocks, the characteristic images of the image with cracks and the image without cracks are segmented, the subblocks are extracted from the maximum connected region of the crack region and the non-crack region and respectively marked as a crack set and a crack set;
the training module is used for training the deep learning model, firstly carrying out normalization processing on the original subblock image data, then carrying out iterative training, and using the trained model structure parameters for crack subblock identification;
the denoising and connecting module is used for denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory;
and the counting module is used for counting related parameters according to the final crack detection image to complete pavement crack detection.
A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the following steps when executing the computer program:
step 1, obtaining a pavement gray level image;
step 2, performing presorting on the anisotropic subblocks, namely determining corresponding anisotropic characteristic values according to the gray mean value and the variance of the subblocks by taking the image subblocks as a unit, segmenting the characteristic images of the image subblocks aiming at the cracked image and the crack-free image, extracting the subblocks from the maximum connected region of the cracked region and the crack-free region, and respectively marking the subblocks as a crack set and a crack-free set;
step 3, training the deep learning model, firstly performing normalization processing on the original subblock image data, then performing iterative training, and using the trained model structure parameters for crack subblock identification;
step 4, denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory;
and 5, performing relevant parameter statistics according to the final crack detection image to finish the pavement crack detection.
A computer-readable storage medium, on which a computer program is stored which, when executed by a processor, carries out the steps of:
step 1, obtaining a pavement gray level image;
step 2, performing presorting on the anisotropic subblocks, namely determining corresponding anisotropic characteristic values according to the gray mean value and the variance of the subblocks by taking the image subblocks as a unit, segmenting the characteristic images of the image subblocks aiming at the cracked image and the crack-free image, extracting the subblocks from the maximum connected region of the cracked region and the crack-free region, and respectively marking the subblocks as a crack set and a crack-free set;
step 3, training the deep learning model, firstly performing normalization processing on the original subblock image data, then performing iterative training, and using the trained model structure parameters for crack subblock identification;
step 4, denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory;
and 5, performing relevant parameter statistics according to the final crack detection image to finish the pavement crack detection.
The invention has higher detection precision and good adaptability to road surfaces with different textures.
The specific steps are the same as those described above, and are not described herein again.
The following is a more detailed description with reference to examples.
Examples
A pavement crack detection method based on stack sparse self-coding deep learning comprises the following steps:
step 1, obtaining a pavement gray level image;
step 2, performing presorting on the anisotropic subblocks, namely determining corresponding anisotropic characteristic values according to the gray mean value and the variance of the subblocks by taking the image subblocks as a unit, segmenting the characteristic images of the image subblocks aiming at the cracked image and the crack-free image, extracting the subblocks from the maximum connected region of the cracked region and the crack-free region, and respectively marking the subblocks as a crack set and a crack-free set; the method specifically comprises the following steps:
in the first step, the acquired crack image is divided into blocks with the size of N × N (such as 32 × 32) pixels, and the mean and variance of the gray scale of each sub-block are calculated.
Second, the mean and variance of the sub-block gray scale are fused as the eigenvalues of the sub-block, i.e. the eigenvalues of the sub-block
B＝-Bmean+λBstd
B denotes the calculated syndrome of the crack sub-block, BmeanRepresenting the mean of the grey levels of the sub-blocks, BstdRepresents the variance of the sub-block, and λ represents the relative influence coefficient of the mean and variance, and λ is determined experimentally and is usually taken to be 5.
Thirdly, calculating the anisotropic characteristic value of the subblock according to the 4 neighborhood directions
Bθ＝median{B-d,θ,B-d+1,θ,…,B-1,θ,B0,θ,B-1,θ,…,Bd-1,θ,Bd,θ},
Wherein theta belongs to {0 degrees, 45 degrees, 90 degrees, 135 degrees }, d is more than or equal to 1, theta represents a neighborhood direction, d represents a neighborhood size, and median represents a median operation. The difference between the maximum and minimum directional eigenvalues is taken as the final eigenvalue of the sub-block, i.e.
T＝max(Bθ)-min(Bθ)
Step 3, training the deep learning model, firstly performing normalization processing on the original subblock image data, then performing iterative training, and using the trained model structure parameters for crack subblock identification; the specific method comprises the following steps:
firstly, classifying the whole image through an anisotropic algorithm, and judging whether the image contains cracks. And (3) segmenting the characteristic graph of the image with the crack, extracting a sub-block image from the maximum connected domain of the crack area, and marking a label as 1 (serving as a crack sub-block training set). Segmenting the characteristic graph of the non-crack image according to the non-crack image, extracting sub-block images from the maximum connected domain of the non-crack area, and marking a label as 0 (serving as a non-crack sub-block training set)
Step two, normalizing the pixel data of the original subblocks, and performing stack sparse self-coding optimization on a crack subblock training set, wherein a single sample loss function is
Figure BDA0002342759850000081
Wherein { (x)(1),y(1)),(x(2),y(2)),…,(x(m),y(m)) Denotes m training samples, x is a single sample feature, y is a sample corresponding output, and the loss function of the sample training set is
Figure BDA0002342759850000082
By adding sparse constraint to the hidden layer of the sample, more robust detection characteristics can be obtained, namely
Figure BDA0002342759850000083
Wherein beta controls the weight of the sparse penalty factor, s is the number of hidden nodes,
Figure BDA0002342759850000084
representing the average value of the output of a single node of the hidden layer, wherein rho is a sparse parameter and can be set to be 0.2 according to experimental experience, and KL is a relative entropy measurement criterion:
Figure BDA0002342759850000085
and thirdly, extracting the characteristics of all the training samples (the crack sub-blocks and the non-crack sub-blocks) through a characteristic detector, and putting the training samples into a softmax classifier to perform supervised learning to obtain classification parameters of softmax. And (4) cascading the stack sparse self-coding and the softmax classifier, and performing supervised learning on the whole network again to optimize the parameters of the whole network.
Step 4, denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory; the specific method comprises the following steps:
firstly, carrying out second-order tensor coding on the whole image by taking a pixel as a unit so as to enable image gray information and direction information to be fused into a uniform tensor field.
And secondly, carrying out tensor voting on each point to a surrounding neighborhood, wherein the voting mode is a mode of stacking the votes considering directionality and distance, and counting the difference of the primary and secondary direction vote values of each point after the voting is finished to be used as a spatial significance value of the point.
And thirdly, performing threshold segmentation on the finally formed tensor field, reserving pixel points with larger tensor values, and realizing binary representation of cracks.
And 5, performing relevant parameter statistics according to the final crack detection image to finish the pavement crack detection.
According to the method, various crack characteristics can be effectively described through an anisotropic characteristic value algorithm; various cracks can be effectively detected through a stack sparse self-coding deep learning model; through the lattice tower theoretical model, the denoising and the connection can be effectively realized.

Claims (10)

1. A pavement crack detection method based on stack sparse self-coding deep learning is characterized by comprising the following steps:
step 1, obtaining a pavement gray level image;
step 2, performing presorting on the anisotropic subblocks, namely determining corresponding anisotropic characteristic values according to the gray mean value and the variance of the subblocks by taking the image subblocks as a unit, segmenting the characteristic images of the image subblocks aiming at the cracked image and the crack-free image, extracting the subblocks from the maximum connected region of the cracked region and the crack-free region, and respectively marking the subblocks as a crack set and a crack-free set;
step 3, training the deep learning model, firstly performing normalization processing on the original subblock image data, then performing iterative training, and using the trained model structure parameters for crack subblock identification;
step 4, denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory;
and 5, performing relevant parameter statistics according to the final crack detection image to finish the pavement crack detection.
2. The pavement crack detection method based on the stack sparse self-coding deep learning of claim 1, wherein the specific step of determining the anisotropy eigenvalue in the step 2 is as follows:
step 2-1, partitioning the collected crack image into blocks with the size of NxN, and then determining the gray mean and variance of each sub-block; n is a natural number;
step 2-2, fusing the mean value and the variance of the gray scale of the subblocks as the characteristic values of the subblocks, namely
B＝-Bmean+λBstd
In the formula, B tableIndicating the calculated syndrome of the crack sub-block, BmeanRepresenting the mean of the grey levels of the sub-blocks, BstdRepresents the variance of the sub-block, and λ represents the relative influence coefficient of the mean and variance;
step 2-3, determining the anisotropic characteristic value of the subblock, wherein the formula is as follows:
Bθ＝median{B-d,θ,B-d+1,θ,…,B-1,θ,B0,θ,B-1,θ,…,Bd-1,θ,Bd,θ},
in the formula, theta belongs to {0 degrees, 45 degrees, 90 degrees, 135 degrees }, d is larger than or equal to 1, theta represents the neighborhood direction, d represents the neighborhood size, mean represents the median operation, the difference between the maximum direction characteristic value and the minimum direction characteristic value is taken as the final anisotropy characteristic value T of the subblock, namely the difference between the maximum direction characteristic value and the minimum direction characteristic value is taken as the final anisotropy characteristic value T of the subblock
T＝max(Bθ)-min(Bθ)。
3. The pavement crack detection method based on stack sparse self-coding deep learning of claim 1, wherein the step 2 is to segment the characteristic graph of the image with cracks and the image without cracks, extract sub-blocks from the maximum connected domain of the crack region and the non-crack region, and respectively mark the sub-blocks as a crack set and a crack-free set, and specifically comprises the following steps:
classifying the whole image through an anisotropic algorithm, judging whether the image contains cracks or not, segmenting a characteristic image of the image with the cracks according to the image with the cracks, extracting sub-block images from the maximum connected domain of a crack area, marking a label as 1, and taking the sub-block images as a crack sub-block training set; and (3) segmenting the characteristic graph of the crack-free image, extracting the sub-block image from the maximum connected domain of the crack-free area, marking the label as 0, and taking the sub-block image as a crack-free sub-block training set.
4. The method for detecting the pavement cracks based on the stack sparse self-coding deep learning of claim 1, wherein the step 3 is to normalize the image data of the original sub-blocks, and then to iteratively train the image data, and the method comprises the following specific steps:
step 3-1, carrying out normalization processing on the pixel data of the original subblocks, and carrying out stack sparse self-coding optimization on a crack subblock training set, wherein a single sample loss function is as follows:
Figure FDA0002342759840000021
wherein { (x)(1),y(1)),(x(2),y(2)),…,(x(m),y(m)) The method comprises the following steps of (1) representing m training samples, wherein x is a single sample characteristic, y is a sample corresponding output, and a loss function of a sample training set is as follows:
Figure FDA0002342759840000022
by adding sparse constraint to the hidden layer of the sample, more robust detection characteristics are obtained, namely
Figure FDA0002342759840000023
Wherein beta controls the weight of the sparse penalty factor, s is the number of hidden nodes,
Figure FDA0002342759840000024
representing the average value of the output of a single node of the hidden layer, wherein rho is a sparse parameter, and KL is a relative entropy measurement criterion:
Figure FDA0002342759840000025
and 3-2, extracting the features of all the training samples by using a feature detector, putting the training samples into a softmax classifier for supervised learning to obtain classification parameters of the softmax, cascading stack sparse self-coding and the softmax classifier, and carrying out the supervised learning on the whole network again to complete the optimization of the parameters of the whole network.
5. The pavement crack detection method based on stack sparse self-coding deep learning of claim 1, wherein the concrete method for denoising and connecting the recognized crack sub-blocks in the space based on the lattice tower theory in the step 4 is as follows:
step 4-1, performing second-order tensor coding on the whole image by taking the pixel as a unit so as to enable the image gray information and the direction information to be fused into a unified tensor field;
4-2, carrying out tensor voting on each point to surrounding neighborhoods, and counting the difference of primary and secondary direction vote values of each point after the voting is finished to be used as a spatial significance value of the point;
and 4-3, performing threshold segmentation on the finally formed tensor field, reserving pixel points with larger tensor values, and realizing binary representation of the cracks so as to finish extraction of the cracks.
6. The method for detecting the pavement cracks based on the stack sparse self-coding deep learning of claim 1, wherein the parameters in the step 5 comprise crack types, lengths and areas.
7. The method for detecting the pavement cracks based on the stack sparse self-coding deep learning of claim 4, wherein the larger tensor in the step 4-3 is a tensor exceeding a threshold value.
8. The utility model provides a road surface crack detecting system based on stack sparse self-encoding degree of depth study which characterized in that includes:
the acquisition module is used for acquiring a road surface gray level image;
the presorting module is used for presorting the anisotropic subblocks, firstly, the image subblocks are taken as a unit, corresponding anisotropic characteristic values are determined according to the gray mean value and the variance of the subblocks, the characteristic images of the image with cracks and the image without cracks are segmented, the subblocks are extracted from the maximum connected region of the crack region and the non-crack region and respectively marked as a crack set and a crack set;
the training module is used for training the deep learning model, firstly carrying out normalization processing on the original subblock image data, then carrying out iterative training, and using the trained model structure parameters for crack subblock identification;
the denoising and connecting module is used for denoising and connecting the identified crack sub-blocks based on the space of the lattice tower theory;
and the counting module is used for counting related parameters according to the final crack detection image to complete pavement crack detection.
9. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the steps of the method of any of claims 1 to 7 are implemented when the computer program is executed by the processor.
10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 7.
CN201911383082.1A 2019-12-27 2019-12-27 Pavement crack detection method based on stack sparse self-coding deep learning Pending CN113066041A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911383082.1A CN113066041A (en) 2019-12-27 2019-12-27 Pavement crack detection method based on stack sparse self-coding deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911383082.1A CN113066041A (en) 2019-12-27 2019-12-27 Pavement crack detection method based on stack sparse self-coding deep learning

Publications (1)

Publication Number Publication Date
CN113066041A true CN113066041A (en) 2021-07-02

Family

ID=76559307

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911383082.1A Pending CN113066041A (en) 2019-12-27 2019-12-27 Pavement crack detection method based on stack sparse self-coding deep learning

Country Status (1)

Country Link
CN (1) CN113066041A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113706472A (en) * 2021-07-30 2021-11-26 中国公路工程咨询集团有限公司 Method, device and equipment for detecting road surface diseases and storage medium
CN116630813A (en) * 2023-07-24 2023-08-22 青岛奥维特智能科技有限公司 Highway road surface construction quality intelligent detection system

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113706472A (en) * 2021-07-30 2021-11-26 中国公路工程咨询集团有限公司 Method, device and equipment for detecting road surface diseases and storage medium
CN116630813A (en) * 2023-07-24 2023-08-22 青岛奥维特智能科技有限公司 Highway road surface construction quality intelligent detection system
CN116630813B (en) * 2023-07-24 2023-09-26 青岛奥维特智能科技有限公司 Highway road surface construction quality intelligent detection system

Similar Documents

Publication Publication Date Title
Akagic et al. Pavement crack detection using Otsu thresholding for image segmentation
Ai et al. Computer vision framework for crack detection of civil infrastructure—A review
Akarsu et al. A fast and adaptive road defect detection approach using computer vision with real time implementation
Zhou et al. Seed-based approach for automated crack detection from pavement images
CN103048329A (en) Pavement crack detecting method based on active contour model
CN107798293A (en) A kind of crack on road detection means
Yun et al. Crack recognition and segmentation using morphological image-processing techniques for flexible pavements
CN110660065A (en) Infrared fault detection and identification algorithm
CN115797813B (en) Water environment pollution detection method based on aerial image
CN111667655A (en) Infrared image-based high-speed railway safety area intrusion alarm device and method
Zhao et al. Image-based comprehensive maintenance and inspection method for bridges using deep learning
CN113066041A (en) Pavement crack detection method based on stack sparse self-coding deep learning
CN113313107A (en) Intelligent detection and identification method for multiple types of diseases on cable surface of cable-stayed bridge
CN107516315B (en) Tunneling machine slag tapping monitoring method based on machine vision
Gui et al. Object-based crack detection and attribute extraction from laser-scanning 3D profile data
CN116452613B (en) Crack contour extraction method in geological survey
Tsai et al. Automatic detection of deficient video log images using a histogram equity index and an adaptive Gaussian mixture model
CN103065296B (en) High-resolution remote sensing image residential area extraction method based on edge feature
KR102585146B1 (en) Pothole detection device and method based on deep learning
Peng et al. Automatic Crack Detection by Multi-Seeding Fusion on 1mm Resolution 3D Pavement Images
CN115512310A (en) Vehicle type recognition method and system based on face features under video monitoring
CN104504702A (en) Cement notch pavement crack recognition method based on square lattice searching method
Kembhavi et al. Low-cost image processing system for evaluating pavement surface distress
CN113920055A (en) Defect detection method
CN110929757A (en) Method for quickly classifying crack types of asphalt pavement

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination