CN110120041A - Pavement crack image detecting method - Google Patents
Pavement crack image detecting method Download PDFInfo
- Publication number
- CN110120041A CN110120041A CN201910399544.2A CN201910399544A CN110120041A CN 110120041 A CN110120041 A CN 110120041A CN 201910399544 A CN201910399544 A CN 201910399544A CN 110120041 A CN110120041 A CN 110120041A
- Authority
- CN
- China
- Prior art keywords
- image
- model
- pavement
- crack
- training
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- 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
-
- 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
- G06T2207/30168—Image quality inspection
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of pavement crack image detecting methods, which comprises obtains pavement image to be detected；Obtain training data, the training data crack tag image corresponding with each pavement image that includes multiple pavement images；The depth model of pre-training is obtained, and initial pavement crack detection model is constructed based on the depth model；Based on the training data, the training initial pavement crack detection model；Based on the pavement crack detection model obtained after training, the crack tag image of the pavement image to be detected is obtained, pavement crack image detection result is obtained.The present invention can construct pavement crack detection model according to the depth model of pre-training, improve model training efficiency, also improve the precision of acquired crack tag image.
Description
Technical field
The present invention relates to Crack Detection technical field, especially a kind of pavement crack image detecting method.
Background technique
With the rapid economic development in our country, highway mileage number is continuously increased, the maintenance task of highway pavement becomes increasingly
It is heavy.Crack is the initial symptoms of road surface breakage, and traffic safety is affected in order to prevent, and pavement crack need to find and mend in time
It rescues.But the huge magnitude of traffic flow proposes very high implementation requirement, traditional nourishment's mode to maintenance task on highway
Not there is only security risk and very time-consuming, this pavement crack detection device for allowing for automation has gradually obtained industry
Favor.
Currently, the core content of pavement crack automatic detection device mainly includes crack image detection, crack image detection
Target be target crack, but the variation of image source, the inhomogeneities of crack width, surface texture are detected under a variety of backgrounds
Lack enough background illuminations and the presence of other features makes crack image detection be filled with challenge.
In early stage research, crack image detecting method is mainly based upon the combination of conventional digital image processing technique or changes
Into, it such as threshold process, mathematical morphology and edge detection, but is usually fissured central line by the result that this method obtains, and
Width not comprising crack.
Later, the more information in crack in order to obtain, pre-processes image by using image processing techniques, image
Segmentation and feature extraction develop more quick steady automatic detection and dividing method.For example, Shi Y etc. is in article
“Automatic Road Crack Detection Using Random Structured Forests”(Shi Y,Cui
Li, Qi Z, et al.IEEE Transactions on Intelligent Transportation Systems, 2016) in
There is the problems such as noise of similar grain for the serious inhomogeneities in crack, the topology complexity in crack and crack, disclose one
New road crack detection system (CrackForest) of the kind based on random structure forest；A.Cubero Fernandez etc.
In article " Efficient pavement crack detection and classification " (A.Cubero Fern á
ndez,F.J.Rodriguez-Lozano,R Villatoro,et al.EURASIP Journal on Image and
Video Processing, 2017) it is disclosed after obtaining crack image in, first using one or more algorithms (for example, logarithm
Transformation, two-sided filter, Canny algorithm, morphological filter etc.) fracture image pre-processed to protrude the main spy in crack
Sign reapplies decision tree heuritic approach and carries out classification and Detection to pretreated crack image.But using image procossing
Crack image detecting method is highly prone to the interference of external environment, and stability is poor.
As deep learning is in the development of speech recognition and computer vision field, people gradually by deep learning (for example,
Convolutional neural networks) it is applied to overcome the shortcoming in traditional images processing method in pavement crack image detection.For example,
Zhang A etc. is in article " Automated Pixel-Level Pavement Crack Detection on 3D Asphalt
Surfaces Using a Deep-Learning Network”(Zhang A,Wang K C P,Li B,et
Al.Computer-aided Civil&Infrastructure Engineering, 2017) it is disclosed in a kind of based on convolution
The efficient framework (CrackNet) of neural network (Convolutional Neural Network, CNN), but CrackNet due to
It requires further improvement to detect compared with minute crack, and network structure is related to input picture size, so that the extensive energy of this method
Power is poor；Young-Jin Cha etc. is done brought by the variation (for example, variation of illumination and shade) of true environment to overcome
It disturbs, in article " Deep Learning-Based Crack Damage Detection Using Convolutional
Neural Networks”(Cha Y J,Choi W,O.Computer-Aided Civil and
Infrastructure Engineering, 2017) a kind of method based on computer vision is disclosed in, this method uses volume
The deep layer framework of neural network (CNN) is accumulated to detect distress in concrete, i.e., first divides the image into block using sliding window, then pass through
Whether include crack in CNN forecast image block, is split since this method can only find bulk in the case where not considering pixel scale
Seam, leading to the output result of the network is not complete crack image, but image block；Pauly L etc. is in article " Deeper
Networks for Pavement Crack Detection”(Pauly L,Peel H,Luo S,et al.34th
International Symposium in Automation and Robotics in Construction, 2017) public in
The method that deeper network is used in pavement crack based on computer vision detection is opened, but network is deeper also implies that ginseng
Number is more, and the required training time is longer；Zou Q etc. is in article " DeepCrack:Learning Hierarchical
Convolutional Features for Crack Detection”(Zou Q,Zhang Z,Li Q,et al.IEEE
Transactions on Image Processing, 2018) disclose in a kind of can train depth convolutional Neural end to end
Network (DeepCrack) carries out the automatic detection in crack by learning the advanced features in crack.
During proposing the disclosure, inventors have found that almost each pavement crack image detection side in the prior art
Method has the specific scope of application, so that existing pavement crack image detecting method is single in crack image clearly, background
Under the conditions of, performance is good, it can be difficult to meeting the engineering demand of actual multiplicity.For example, being illuminated by the light, the noises such as shade
It influences, crack local environment is usually relatively complex, although the detection method based on CNN has the effect of certain, segmentation result
It is image block, has over-evaluated the width in crack；All existing compared in minute crack image in illumination shade, crack is not divided completely
It cuts out, there are less divided phenomenons；In order to make FCN network convergence, needs successively to train each network stage by stage, cause to train
Journey is complicated and takes a long time, and is unfavorable for the realization of real-time process.
Summary of the invention
In order to solve above-mentioned problems of the prior art, the present invention provides a kind of pavement crack image detecting method.
The pavement crack image detecting method includes:
Step S1 obtains pavement image to be detected；
Step S2 obtains training data, and the training data includes multiple pavement images and each pavement image pair
The crack tag image answered；
Step S3 obtains the depth model of pre-training, and constructs initial pavement crack based on the depth model and detect mould
Type；
Step S4 is based on the training data, the training initial pavement crack detection model；
Step S5 obtains the crack of the pavement image to be detected based on the pavement crack detection model obtained after training
Tag image obtains pavement crack image detection result.
Optionally, the depth model is taken as VGG16 model, and the pavement crack detection model is VGG-U-net model.
Optionally, the VGG16 model includes input layer, m coding convolutional coding structure, p full connection structures and output
Layer, wherein m and p is positive integer, and each coding convolutional coding structure includes at least two convolutional layers and a pond layer, described complete
Connection structure includes multiple full articulamentums.
Optionally, the VGG-U-net model includes:
Coded portion, the coded portion include the input layer and preceding n coding convolutional coding structure of the VGG16 model,
In, n is the positive integer less than or equal to m；
Decoded portion, the decoded portion include n decoding convolutional coding structure and export structure, each decoding convolution knot
Structure includes at least two convolutional layers and a up-sampling layer, and the export structure includes at least two convolutional layers and an output
Layer.
Optionally, the up-sampling layer carries out up-sampling calculating to the image that previous convolutional layer exports.
Optionally, the up-sampling calculate the following steps are included:
Step S21, up-sampling are based on deconvolution core, determine the image of the previous convolutional layer output of the up-sampling layer
Fisrt feature image, the fisrt feature image include p channel, wherein p is the integer greater than 2；
Step S22, image duplication replicate the second feature image that port number in the coded portion is p；
The second feature image cropping is and the fisrt feature image pixel quantity phase by step S23, image cropping
Same third feature image；
Step S24, image merge, the third feature image and the fisrt feature image are merged, and obtain the
Four characteristic images.
Optionally, the pond layer uses maximum Chi Huafa, and sampling window is 2 × 2；The up-sampling layer uses warp
Area method, and the matrix that deconvolution core is 2 × 2.
Optionally, the coding convolutional coding structure further includes normalization layer, active coating and random deactivating layer；
The decoding convolutional coding structure further includes active coating and random deactivating layer.
Optionally, the VGG16 model for obtaining pre-training includes:
Based on the training data, the training VGG16 model；Or
The VGG16 model trained in load ImageNet data set.
Optionally, described to be based on the training data, the training initial VGG-U-net model includes:
Step S31 is based on the VGG16 model, carries out to the parameter of the coded portion of the VGG-U-net model initial
Change；
Step S32 determines the output characteristic pattern of pavement image in the training data by the VGG-U-net model；
The output characteristic pattern is converted to probability distribution graph by step S33；
Step S34 calculates the deviation between the probability distribution graph and the crack tag image；
Step S35 updates the parameter of decoded portion in the VGG-U-net model according to the deviation, then based on more
VGG-U-net model after new parameter repeats the step S32, step S33 and step S34, until the deviation is small
In preset threshold, the VGG-U-net model of training completion is obtained.
Main advantages of the present invention are as follows: the method for the present invention is a kind of pavement crack image detection based on pre-training model
On the one hand method increases the depth of model using depth model, improve precision of prediction, on the other hand utilizes pre-training model
Training process is simplified, training effectiveness is improved.
Detailed description of the invention
Fig. 1 is the flow chart of pavement crack image detecting method according to an embodiment of the present invention；
Fig. 2 is the structural schematic diagram of VGG16 model according to an embodiment of the present invention；
Fig. 3 is the structural schematic diagram of VGG-U-net model according to an embodiment of the present invention；
Fig. 4 is the flow chart that up-sampling according to an embodiment of the present invention calculates；
Fig. 5 is the training flow chart of VGG-U-net model according to an embodiment of the present invention；
Fig. 6 is VGG-U-net model training schematic diagram according to an embodiment of the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference
Attached drawing, the present invention is described in more detail.
Fig. 1 is the flow chart of pavement crack image detecting method according to an embodiment of the present invention.
As shown in Figure 1, in an embodiment of the present invention, the pavement crack image detecting method includes:
Step S1 obtains pavement image to be detected；
Step S2 obtains training data, and the training data includes multiple pavement images and each pavement image pair
The crack tag image answered；
Wherein, the crack tag image refers to carrying out pavement crack inspection for the pavement image in the training data
After survey, for the image that Crack Detection result is marked, is marked in the tag image of the crack and stated institute in pavement image
The crack information for including, for example, the position in crack, width, length etc..
Step S3 obtains the depth model of pre-training, and constructs initial pavement crack based on the depth model and detect mould
Type；
In an embodiment of the present invention, the depth model is taken as VGG16 model, and the pavement crack detection model is
VGG-U-net model, next with the depth model for VGG16 model, the pavement crack detection model is VGG-U-net
For model, the present invention is further expalined and is illustrated.
In an embodiment of the present invention, the VGG16 model M 1 connects entirely including input layer, m coding convolutional coding structure, p
Binding structure and output layer, wherein m and p is positive integer, and each coding convolutional coding structure includes at least two convolutional layers and a pond
Change layer, each full connection structure includes multiple full articulamentums.Fig. 2 is the knot of VGG16 model according to an embodiment of the present invention
Structure schematic diagram, as shown in Fig. 2, the VGG16 model M 1 includes input layer, 5 coding convolution in an embodiment of the present invention
Structure, 1 full connection structure and output layer, wherein each coding convolutional coding structure includes 2~3 convolutional layers and 1 pond layer,
The full connection structure includes multiple full articulamentums.For example, the first two coding convolutional coding structure of the VGG16 model M 1 includes 2
A convolutional layer and 1 pond layer, rear three coding convolutional coding structures include 3 convolutional layers and 1 pond layer, the full connection structure
Including 3 full articulamentums, i.e., in this embodiment, the VGG16 model be include 13 layers of convolutional layer and 3 layers of full articulamentum
16 layers of convolutional neural networks.
In an embodiment of the present invention, in same coding convolutional coding structure, the convolution nuclear volume of each convolutional layer is identical.
In a convolutional layer, the corresponding channel for generating characteristic image of a convolution kernel, therefore, in one and same coding convolutional coding structure
In, the port number for the characteristic image that each convolutional layer determines is identical.
In an embodiment of the present invention, convolution kernel in the convolutional layer of the VGG16 model is 3 × 3 matrixes, for from
Feature is extracted in the pavement image.The pond layer of the VGG16 model use pond window size for 2 × 2 maximum pond
Method, i.e., the maximum value in window that each record size is 2 × 2, for carrying out dimensionality reduction to the matrix after convolution, to avoid tieing up
Number explosion.
In this embodiment, the VGG16 model is a kind of artificial neural network based on convolutional neural networks structure,
And the parameter configuration of the VGG16 model be it is publicly available, be used as reference characteristic extractor for many other applications
In program.Therefore, in an embodiment of the present invention, the VGG16 model for obtaining pre-training can be based on the trained number
According to, the training VGG16 model obtains, it can also be loaded directly into the VGG16 model trained in ImageNet data set, thus
Improve the efficiency of model training.
After the VGG16 model for obtaining pre-training, and the VGG16 model based on the pre-training, construct initial VGG-U-net
Model.
In accordance with an embodiment of the present disclosure, the initial VGG-U-net model includes coded portion and decoded portion.
The coded portion includes the input layer and preceding n coding convolutional coding structure of the VGG16 model, wherein n be less than
Positive integer equal to m.The coded portion is used to capture the contextual information in image during multiple convolution and pond, from
And obtain characteristics of image.
The decoded portion includes n decoding convolutional coding structure and export structure, and each decoding convolutional coding structure includes extremely
Few two convolutional layers and a up-sampling layer, the export structure include at least two convolutional layers and an output layer, are used for root
Characteristics of image is extracted according to the contextual information of coded portion capture, realizes and is accurately positioned.
Fig. 3 is the structural schematic diagram of VGG-U-net model according to an embodiment of the present invention, as shown in figure 3, in this hair
In a bright embodiment, the coded portion of the initial VGG-U-net model M 2 may include the VGG16 model of the pre-training
The input layer of M1 and 4 coding convolutional coding structures, i.e., the coded portion of the described VGG-U-net model M 2 are the pre-training
Include preceding 15 layer model including input layer in VGG16 model M 1.The decoded portion of the initial VGG-U-net model M 2 includes
4 decoding convolutional coding structures and export structure, each decoding convolutional coding structure includes 2~3 convolutional layers and 1 up-sampling layer, described
Export structure includes 2 convolutional layers and 1 output layer.
In an embodiment of the present invention, in addition to the convolution kernel of the last one convolutional layer in export structure is 1 × 1 matrix,
The convolution kernel of remaining convolutional layer is 3 × 3 matrixes；In same decoding convolutional coding structure, the convolution nuclear volume phase of each convolutional layer
Together, so that the port number of characteristic image determined by each convolutional layer is identical.
In an embodiment of the present invention, the pond window size phase of the sampling window and the pond layer of the up-sampling layer
Together.So that in i-th of up-sampling layer by the port number of the characteristic image of up-sampling and a pond layers of the n-th-i+1 by pond
The port number of characteristic image afterwards is identical, wherein i is the positive integer less than or equal to n.
In an embodiment of the present invention, the pond layer uses maximum Chi Huafa, and sampling window is 2 × 2, i.e., every time
The maximum value in window that record size is 2 × 2 has for carrying out dimensionality reduction to the matrix after convolution so that dimension be avoided to explode
Conducive to acquisition characteristics of image；The up-sampling layer uses warp area method, and the matrix that deconvolution core is 2 × 2, after to convolution
Matrix carry out liter dimension, advantageously reduce number of channels, realize precise positioning.
In an embodiment of the present invention, the up-sampling layer carries out up-sampling meter to the image that previous convolutional layer exports
It calculates.Fig. 4 is the flow chart that up-sampling according to an embodiment of the present invention calculates, as shown in figure 4, up-sampling calculating includes
Following steps:
Step S21, up-sampling are based on deconvolution core, determine the image of the previous convolutional layer output of the up-sampling layer
Fisrt feature image, the fisrt feature image include p channel, wherein p is the integer greater than 2；
Step S22, image duplication replicate the second feature image that port number in the coded portion is p；
The second feature image cropping is and the fisrt feature image pixel quantity phase by step S23, image cropping
Same third feature image；
Step S24, image merge, the third feature image and the fisrt feature image are merged, and obtain the
Four characteristic images.
For example, in the step s 21, it is assumed that the image of previous convolutional layer output includes the picture in 64 channels and each channel
Element is 16 × 16, then is based on 2 × 2 deconvolution cores, by the determining fisrt feature image of up-sampling be comprising 64 channels,
And the pixel in each channel is 32 × 32；Then in step S22, the second feature figure in 64 channels is included in replica code part
Picture；Assuming that the pixel in each channel of second feature image is 36 × 36, then in step S23, to the second feature figure
As being cut, third feature image is determined, so that the third feature image includes 64 channels, and the pixel in each channel
It is 32 × 32；Finally in step s 24, the fisrt feature image and the third feature image are merged, obtained
Fourth feature image includes 128 channels, and the pixel in each channel is 32 × 32.
In an embodiment of the present invention, decoded portion is merged by image by contextual information in coded portion and opposite
The decoded portion answered carries out feature combination, completes deep layer abstract characteristics (feature obtained in decoded portion) and shallow-layer feature
The fusion of (feature obtained in coded portion), increases the contextual information of original image in decoded portion, is conducive to mend
The boundary information lost entirely promotes the accuracy of marginal information prediction.In this way, learning semantic information and location information by convolution
Combination, obtain more features, improve the effect divided to Small object, keep the result of model prediction more accurate.
In an embodiment of the present invention, the coding convolutional coding structure further includes normalization layer, active coating and random inactivation
(dropout) layer.The decoding convolutional coding structure further includes active coating and random inactivation (dropout) layer.Wherein, described
Dropout layers by the zero setting in certain proportion of the pixel value in characteristic image, to avoid producing because excessively relying on local feature
Raw over-fitting enhances model generalization ability.For example, the 4th coding convolutional coding structure of the VGG-U-net model M 2
Including dropout layers, when characteristic image passes through described dropout layers, pixel value has 50% ratio that may be zeroed out.
In an embodiment of the present invention, the active coating using line rectification function (Rectified LinearUnit,
ReLU), the ReLU is common activation primitive in a kind of artificial neural network, generally refer to be with ramp function and its mutation
The nonlinear function of representative.
Step S4 is based on the training data, the training initial pavement crack detection model；
Fig. 5 is the training flow chart of VGG-U-net model according to an embodiment of the present invention, as shown in figure 5, the base
Include: in the step of training data, the training initial VGG-U-net model
Step S31 is based on the VGG16 model, carries out to the parameter of the coded portion of the VGG-U-net model initial
Change；
Step S32 determines the output characteristic pattern of pavement image in the training data by the VGG-U-net model；
The output characteristic pattern is converted to probability distribution graph by step S33；
In an embodiment of the present invention, the output characteristic pattern can be converted into probability distribution by sigmoid function
Figure.
Step S34 calculates the deviation between the probability distribution graph and the crack tag image；
In an embodiment of the present invention, the probability distribution graph and the crack tag image can be calculated by loss function
Deviation.
Step S35 updates the parameter of decoded portion in the VGG-U-net model according to the deviation, then based on more
VGG-U-net model after new parameter repeats the step S32, step S33 and step S34, until the deviation is small
In preset threshold, the VGG-U-net model of training completion is obtained, wherein the preset threshold can be according to the needs of practical application
It is configured, the present invention is not especially limited it.
In an embodiment of the present invention, the loss function can be cross entropy loss function, be shown below:
In formula, n indicates the quantity of pavement image in training data,For the corresponding crack label figure of i-th of pavement image
Picture, yiFor the corresponding probability distribution graph of i-th of pavement image.
In an embodiment of the present invention, described according to the deviation, update decoded portion in the VGG-U-net model
Parameter, including be based on Adam optimization algorithm, successively update the parameter in the VGG-U-net model, wherein Adam optimization calculate
Method is a kind of algorithm for optimizing the VGG-U-net model based on First-order Gradient, i.e., the described Adam optimization algorithm is according to loss
Single order moments estimation (First Moment Estimation, the i.e. mean value of gradient) and second order of the function to the gradient of each parameter
Moments estimation (Second Moment Estimation, the i.e. variance of the non-centralization of gradient) dynamic adjustment is directed to each parameter
Learning rate.
Fig. 6 is VGG-U-net model training schematic diagram according to an embodiment of the present invention, as shown in Figure 6, it is assumed that be based on
The VGG16 model M 1 initializes the parameter of the coded portion of the VGG-U-net model M 2, by pavement image X and its right
The process that the known crack tag image R that answers carries out model optimization can be with are as follows: first passes through the initial VGG-U-net model
M2 determines the output characteristic pattern X1 of the pavement image X；Then by sigmoid function, the output characteristic pattern is converted to
Probability distribution graph X2, then the inclined of the probability distribution graph X2 and known crack tag image R is calculated by loss function
Difference b；Assuming that the deviation b is less than preset threshold, then step S4 terminates；Assuming that the deviation b is more than or equal to default threshold
Value is then based on Adam optimization algorithm, and the parameter updated in the initial VGG-U-net model M 2 obtains new VGG-U-net mould
Type M2 '.After obtaining new VGG-U-net model M 2 ', the output characteristic pattern X1 ' and probability point of the pavement image X are redefined
Butut X2 ', and calculate by loss function the deviation of the probability distribution graph X2 ' Yu the known crack tag image R
B ', then the relationship of the deviation b ' and preset threshold judge whether the ginseng for needing to update the VGG-U-net model M 2
Number.
In an embodiment of the present invention, the coded portion of the VGG-U-net model can be initialized by transfer learning
Parameter, so that in the training process, it is only necessary to partial parameters are finely adjusted, without all parameters of re -training.In this way
It can deepen model depth on the one hand, be conducive to improve model accuracy, be on the other hand based on transfer learning in the training process
The parameter of the VGG16 model of the pre-training initializes model, need to only be finely adjusted to the parameter of coded portion, and
The parameter for initializing simultaneously re -training decoded portion, improves the convergence rate and generalization ability of model, has saved the training time.
Step S5 obtains the crack of the pavement image to be detected based on the pavement crack detection model obtained after training
Tag image obtains pavement crack image detection result.
For example, the crack label figure of pavement image Y to be detected can be obtained by the VGG-U-net model M 2 ' after training
Picture marks in the crack tag image and is stated crack information included in pavement image Y to be detected, for example, crack
Position, width, length etc..
The VGG-U-net model optimization proposed by the present invention performance indicator of Crack Monitoring, for example, for illustrating cut section
The integrality in domain and the performance indicator Mean_IoU of positional accuracy solve traditional detection model relatively thin to a certain extent
Less divided phenomenon under crack, while having preferable performance in terms of real-time, be conducive to Crack Detection cutting procedure real time implementation
It realizes, can also preferably identify that the mankind may be ignored compared with small crack details in marking by hand, there is very high robust
Property.
Particular embodiments described above has carried out further in detail the purpose of the present invention, technical scheme and beneficial effects
It describes in detail bright, it should be understood that the above is only a specific embodiment of the present invention, is not intended to restrict the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution, improvement and etc. done should be included in guarantor of the invention
Within the scope of shield.
Claims (10)
1. a kind of pavement crack image detecting method characterized by comprising
Step S1 obtains pavement image to be detected；
Step S2 obtains training data, and the training data includes that multiple pavement images are corresponding with each pavement image
Crack tag image；
Step S3 obtains the depth model of pre-training, and constructs initial pavement crack detection model based on the depth model；
Step S4 is based on the training data, the training initial pavement crack detection model；
Step S5 obtains the crack label of the pavement image to be detected based on the pavement crack detection model obtained after training
Image obtains pavement crack image detection result.
2. the method as described in claim 1, which is characterized in that the depth model is taken as VGG16 model, the pavement crack
Detection model is VGG-U-net model.
3. method according to claim 2, which is characterized in that the VGG16 model includes input layer, m coding convolution knot
Structure, p full connection structures and output layer, wherein m and p is positive integer, and each coding convolutional coding structure includes volume at least two
Lamination and a pond layer, the full connection structure include multiple full articulamentums.
4. method as claimed in claim 3, which is characterized in that the VGG-U-net model includes:
Coded portion, the coded portion include the input layer and preceding n coding convolutional coding structure of the VGG16 model, wherein n
For the positive integer less than or equal to m；
Decoded portion, the decoded portion include n decoding convolutional coding structure and export structure, each decoding convolutional coding structure packet
At least two convolutional layers and a up-sampling layer are included, the export structure includes at least two convolutional layers and an output layer.
5. method as claimed in claim 4, which is characterized in that the image that the up-sampling layer exports previous convolutional layer into
Row up-sampling calculates.
6. method as claimed in claim 5, which is characterized in that the up-sampling calculate the following steps are included:
Step S21, up-sampling are based on deconvolution core, determine the first of the image of the previous convolutional layer output of the up-sampling layer
Characteristic image, the fisrt feature image include p channel, wherein p is the integer greater than 2；
Step S22, image duplication replicate the second feature image that port number in the coded portion is p；
The second feature image cropping is identical with the fisrt feature image pixel quantity by step S23, image cropping
Third feature image；
Step S24, image merge, the third feature image and the fisrt feature image are merged, and obtain the 4th spy
Levy image.
7. method as claimed in claim 4, which is characterized in that the pond layer uses maximum Chi Huafa, and sampling window is 2
×2；The up-sampling layer uses warp area method, and the matrix that deconvolution core is 2 × 2.
8. method as claimed in claim 4, which is characterized in that
The coding convolutional coding structure further includes normalization layer, active coating and random deactivating layer；
The decoding convolutional coding structure further includes active coating and random deactivating layer.
9. method according to claim 2, which is characterized in that it is described obtain pre-training VGG16 model include:
Based on the training data, the training VGG16 model；Or
The VGG16 model trained in load ImageNet data set.
10. method according to claim 2, which is characterized in that described to be based on the training data, the training initial VGG-
U-net model includes:
Step S31 is based on the VGG16 model, initializes to the parameter of the coded portion of the VGG-U-net model；
Step S32 determines the output characteristic pattern of pavement image in the training data by the VGG-U-net model；
The output characteristic pattern is converted to probability distribution graph by step S33；
Step S34 calculates the deviation between the probability distribution graph and the crack tag image；
Step S35 updates the parameter of decoded portion in the VGG-U-net model according to the deviation, then is joined based on updating
VGG-U-net model after number repeats the step S32, step S33 and step S34, until the deviation is less than in advance
If threshold value, the VGG-U-net model of training completion is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910399544.2A CN110120041A (en) | 2019-05-14 | 2019-05-14 | Pavement crack image detecting method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910399544.2A CN110120041A (en) | 2019-05-14 | 2019-05-14 | Pavement crack image detecting method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110120041A true CN110120041A (en) | 2019-08-13 |
Family
ID=67522401
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910399544.2A Pending CN110120041A (en) | 2019-05-14 | 2019-05-14 | Pavement crack image detecting method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110120041A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110555831A (en) * | 2019-08-29 | 2019-12-10 | 天津大学 | Drainage pipeline defect segmentation method based on deep learning |
| CN110910343A (en) * | 2019-09-29 | 2020-03-24 | 北京建筑大学 | Method and device for detecting pavement cracks and computer equipment |
| CN111242017A (en) * | 2020-01-10 | 2020-06-05 | 北京邮电大学 | Multi-marking-line pavement crack identification method, device, equipment and storage medium |
| CN111257341A (en) * | 2020-03-30 | 2020-06-09 | 河海大学常州校区 | Underwater building crack detection method based on multi-scale features and stacked full convolution network |
| CN111445446A (en) * | 2020-03-16 | 2020-07-24 | 重庆邮电大学 | Concrete surface crack detection method based on improved U-net |
| CN111597932A (en) * | 2020-04-30 | 2020-08-28 | 汕头大学 | Road crack image identification method, device and system based on convolutional neural network |
| CN112233105A (en) * | 2020-10-27 | 2021-01-15 | 江苏科博空间信息科技有限公司 | Road crack detection method based on improved FCN |
| CN112257622A (en) * | 2020-10-28 | 2021-01-22 | 汕头大学 | Road crack segmentation method based on genetic algorithm and U-shaped neural network |
| CN112634195A (en) * | 2020-11-23 | 2021-04-09 | 清华大学 | Concrete structure crack prediction method, device and system |
| CN113870263A (en) * | 2021-12-02 | 2021-12-31 | 湖南大学 | Real-time monitoring method and system for pavement defect damage |
| CN114463597A (en) * | 2022-01-20 | 2022-05-10 | 广州市建筑科学研究院集团有限公司 | Bridge crack detection method, system and medium based on coding and decoding network |
| CN115345881A (en) * | 2022-10-18 | 2022-11-15 | 上海交强国通智能科技有限公司 | Pavement disease detection method based on computer vision |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107871133A (en) * | 2017-12-27 | 2018-04-03 | 中公高科养护科技股份有限公司 | The recognition methods of the optimization method, pavement disease of rim detection network and system |
| CN108416307A (en) * | 2018-03-13 | 2018-08-17 | 北京理工大学 | A kind of Aerial Images road surface crack detection method, device and equipment |
| KR101926561B1 (en) * | 2018-03-13 | 2018-12-07 | 연세대학교 산학협력단 | Road crack detection apparatus of patch unit and method thereof, and computer program for executing the same |
-
2019
- 2019-05-14 CN CN201910399544.2A patent/CN110120041A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107871133A (en) * | 2017-12-27 | 2018-04-03 | 中公高科养护科技股份有限公司 | The recognition methods of the optimization method, pavement disease of rim detection network and system |
| CN108416307A (en) * | 2018-03-13 | 2018-08-17 | 北京理工大学 | A kind of Aerial Images road surface crack detection method, device and equipment |
| KR101926561B1 (en) * | 2018-03-13 | 2018-12-07 | 연세대학교 산학협력단 | Road crack detection apparatus of patch unit and method thereof, and computer program for executing the same |
Non-Patent Citations (4)
| Title |
|---|
| 徐江川: "基于深度卷积神经网络的熟料颗粒方法研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
| 朱苏雅 等: "采用U-Net卷积网络的桥梁裂缝检测方法", 《西安电子科技大学学报.HTTP://KNS.CNKI.NET/KCMS/DETAIL/61.1076.TN.20190511.1239.008.HTML》 * |
| 李德毅 等: "《人工智能导论》", 30 September 2018 * |
| 胡志伟 等: "基于全卷积网络的生猪轮廓提取", 《华南农业大学学报》 * |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110555831A (en) * | 2019-08-29 | 2019-12-10 | 天津大学 | Drainage pipeline defect segmentation method based on deep learning |
| CN110555831B (en) * | 2019-08-29 | 2023-09-26 | 天津大学 | Deep learning-based drainage pipeline defect segmentation method |
| CN110910343A (en) * | 2019-09-29 | 2020-03-24 | 北京建筑大学 | Method and device for detecting pavement cracks and computer equipment |
| CN111242017A (en) * | 2020-01-10 | 2020-06-05 | 北京邮电大学 | Multi-marking-line pavement crack identification method, device, equipment and storage medium |
| CN111242017B (en) * | 2020-01-10 | 2024-01-23 | 北京邮电大学 | Multi-marking pavement crack identification method, device, equipment and storage medium |
| CN111445446A (en) * | 2020-03-16 | 2020-07-24 | 重庆邮电大学 | Concrete surface crack detection method based on improved U-net |
| CN111445446B (en) * | 2020-03-16 | 2022-05-10 | 重庆邮电大学 | Concrete surface crack detection method based on improved U-net |
| CN111257341A (en) * | 2020-03-30 | 2020-06-09 | 河海大学常州校区 | Underwater building crack detection method based on multi-scale features and stacked full convolution network |
| CN111597932A (en) * | 2020-04-30 | 2020-08-28 | 汕头大学 | Road crack image identification method, device and system based on convolutional neural network |
| CN112233105A (en) * | 2020-10-27 | 2021-01-15 | 江苏科博空间信息科技有限公司 | Road crack detection method based on improved FCN |
| CN112257622B (en) * | 2020-10-28 | 2022-08-16 | 汕头大学 | Road crack segmentation method based on genetic algorithm and U-shaped neural network |
| CN112257622A (en) * | 2020-10-28 | 2021-01-22 | 汕头大学 | Road crack segmentation method based on genetic algorithm and U-shaped neural network |
| CN112634195A (en) * | 2020-11-23 | 2021-04-09 | 清华大学 | Concrete structure crack prediction method, device and system |
| CN113870263B (en) * | 2021-12-02 | 2022-02-25 | 湖南大学 | Real-time monitoring method and system for pavement defect damage |
| CN113870263A (en) * | 2021-12-02 | 2021-12-31 | 湖南大学 | Real-time monitoring method and system for pavement defect damage |
| CN114463597A (en) * | 2022-01-20 | 2022-05-10 | 广州市建筑科学研究院集团有限公司 | Bridge crack detection method, system and medium based on coding and decoding network |
| CN115345881A (en) * | 2022-10-18 | 2022-11-15 | 上海交强国通智能科技有限公司 | Pavement disease detection method based on computer vision |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110120041A (en) | Pavement crack image detecting method | |
| CN109829891B (en) | Magnetic shoe surface defect detection method based on dense generation of antagonistic neural network | |
| CN102902956B (en) | A kind of ground visible cloud image identifying processing method | |
| CN111832546B (en) | Lightweight natural scene text recognition method | |
| CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
| CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
| CN103778435A (en) | Pedestrian fast detection method based on videos | |
| CN110675462A (en) | Gray level image colorizing method based on convolutional neural network | |
| CN111127360B (en) | Gray image transfer learning method based on automatic encoder | |
| CN107203752A (en) | A kind of combined depth study and the face identification method of the norm constraint of feature two | |
| CN110675421B (en) | Depth image collaborative segmentation method based on few labeling frames | |
| CN110032952B (en) | Road boundary point detection method based on deep learning | |
| CN110309770A (en) | A kind of vehicle discrimination method again based on the study of four-tuple loss metric | |
| CN107038416A (en) | A kind of pedestrian detection method based on bianry image modified HOG features | |
| CN114419413A (en) | Method for constructing sensing field self-adaptive transformer substation insulator defect detection neural network | |
| CN109815864A (en) | A kind of facial image age recognition methods based on transfer learning | |
| CN106530247A (en) | Multi-scale image restoring algorithm based on structure information | |
| CN116704137B (en) | Reverse modeling method for point cloud deep learning of offshore oil drilling platform | |
| CN113298817A (en) | High-accuracy semantic segmentation method for remote sensing image | |
| Dong et al. | Pixel-level intelligent segmentation and measurement method for pavement multiple damages based on mobile deep learning | |
| Liu et al. | Residual YOLOX-based ship object detection method | |
| CN116977280A (en) | Rail surface defect detection method based on improved UPerNet and connected domain analysis | |
| CN112270289A (en) | Intelligent monitoring method based on graph convolution attention network | |
| CN116416161A (en) | Image restoration method for improving generation of countermeasure network | |
| CN115497006B (en) | Urban remote sensing image change depth monitoring method and system based on dynamic mixing strategy |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190813 |