CN108229461A - A kind of tunnel slot method for quickly identifying based on deep learning - Google Patents
A kind of tunnel slot method for quickly identifying based on deep learning Download PDFInfo
- Publication number
- CN108229461A CN108229461A CN201810038939.5A CN201810038939A CN108229461A CN 108229461 A CN108229461 A CN 108229461A CN 201810038939 A CN201810038939 A CN 201810038939A CN 108229461 A CN108229461 A CN 108229461A
- Authority
- CN
- China
- Prior art keywords
- image
- crack
- denoted
- value
- tunnel
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
Abstract
The present invention is a kind of tunnel slot method for quickly identifying based on deep learning, mainly solves the problems, such as that the crack detection method currently based on deep learning can not directly acquire the length and width information in crack, the present invention includes step：S1, deep learning training set of images is created；S2, training depth convolutional neural networks model；S3, image to be detected is detected using trained convolutional neural networks model and exports prediction label image；S4, testing result, Pixel-level width value and length value including image category, the coordinate information in crack and crack are exported according to prediction label image；S5, output disease record as a result, if there are cracks in image to be detected, records the developed width value and length value of image name, the coordinate information in crack and crack according to testing result；If there is no crack in image to be detected, do not record.
Description
Technical field
The present invention relates to a kind of tunnel slot method for quickly identifying, and it is fast to disclose a kind of tunnel slot based on deep learning
Fast recognition methods, applied to underground engineering field.
Background technology
With advances in technology with the development of society, the construction scale of Tunnel Engineering is growing day by day, greatly facilitates people
Trip and life, such as subway tunnel, railway tunnel, vcehicular tunnel, and largely built Tunnel Engineering have been enter into it is foster
The repair stage is protected, therefore, as tunnel quantity continues to increase, the operation state of tunnel structure also becomes particularly to weigh with Defect inspection
It will.During the operation in tunnel, due to being vibrated by vehicle, the disturbing of all side loads, the influences such as pressure from surrounding rock variation, tunnel
Road surface will appear crack, and crack does not only result in concrete layer to internal reinforcing bar shielding failure, it is also possible to concrete can be caused to fall
It falls, causes Exploration on Train Operation Safety, serious crack is even more the omen of tunnel cave, therefore tunnel slot detection is the daily fortune in tunnel
One of battalion's state and the main project of Defect inspection.
Current tunnel slot detection method mainly has：Artificial observation, Digital Image Processing identification, detections of radar method etc..
Relative to other two methods, Digital Image Processing recognition methods is based especially on the tunnel slot detection method of deep learning
With the advantages that real-time height, precision is high, and robustness is high, it has also become tunnel slot detects the main stream approach of business.In recent years, have
Many scholars expand research to the Crack Detection based on deep learning.
The patent document of Publication No. CN106910186A discloses a kind of Bridge Crack inspection based on CNN deep learnings
Localization method is surveyed, this method trains one to include four convolutional layers and two using the image set of 55000 16*16 pixel sizes
The neural network of full articulamentum, and image to be detected is examined using sliding window and Gaussian image pyramid Downsapling method
It surveys, exports the boundary rectangle frame coordinate in crack and Pixel-level area.
The patent document of Publication No. CN106841216A discloses the tunnel defect based on panoramic picture CNN and knows automatically
Other device, this method obtains tunnel inner wall panoramic picture by panoramic vision sensor first, then using conventional digital image
Treatment technology extracts the image of doubtful disease, and the image of doubtful disease is detected finally by convolutional neural networks, and
The tested altimetric image of output is there are crackle, crack, lining cutting come off, the probability of four kinds of Damage Types of percolating water.
Crack detection method disclosed in document above, all without directly output crack segmentation as a result, therefore can not be direct
The length and width information in crack is obtained, but the length and width information in crack is the important finger of crack disease grade assessment
Mark.The length and width information that above scheme obtains crack needs to increase post-processing approach（As image is divided）, increase crack
Testing process and time cost, while the effect of post-processing approach directly affects the accurate of the length and width information in crack
Degree reduces the robustness of entire crack detection system.Therefore, it is necessary to it is directly exported by end-to-end convolutional neural networks
Crack segmentation as a result, to directly acquire the length and width information in crack, to improve efficiency and robustness.
Invention content
The length and width information in crack can not be directly acquired for the crack detection method currently based on deep learning
Problem, the present invention provides a kind of tunnel slot method for quickly identifying based on deep learning, are detected for tunnel slot, to carry
The efficiency of high tunnel slot detection work.
Technical scheme of the present invention：A kind of tunnel slot method for quickly identifying based on deep learning, which is characterized in that packet
Include following steps：
S1, deep learning training set of images is created.
The concrete operations of the step S1 are as follows：
（1）Acquire tunnel image：Tunnel surface is shot by mobile testing equipment and obtains single pass tunnel image, resolution ratio
General requirement is not less than 2,000,000 pixels, and the clarity requirement of image is sufficiently high, and the crack minimum widith in image is not less than 1 picture
Element；
（2）By artificial screening tunnel image, 50,000 tension fissure images and 50,000 non-crack images are selected, and it is big to adjust image
Small, picture format is consistent with original tunnel image.Original tunnel image is denoted as I, wide and high be denoted as（W,H）, the image after adjustment
It is denoted as I’, it is wide and high be denoted as（w’, H’）；
（3）Crack image after being sized using Photoshop softwares opening, and pass through " quick selection " tool and select crack
Region is denoted as, non-crack area is denoted as；
（4）The crack area selectedLabeled as 1, and crack area Fill Color is set for white.Non- crack areaLabeled as 0, and it is black to set non-crack area Fill Color；
（5）After the completion of crack and non-crack area color filling, training set label image is saved as, is denoted as, form with
Original tunnel image I is consistent, and size is（w’,H’）.
S2, training depth convolutional neural networks model.
The concrete operations of the step S2 are as follows：
（1）Build convolutional neural networks structure：Convolutional neural networks are realized by improving AlexNet structures, 3 full connections
Layer changes 3 convolutional layers into, increases by 1 warp lamination, improved network is by 10 convolutional layers, 5 pond layers, 1 Dropout
Layer, 1 warp lamination, 10 convolutional layers are denoted as C1 ~ C10, and 5 pond layers are denoted as P1 ~ P5, and 1 warp lamination is denoted as DC1.Volume
The convolution kernel size of lamination C1 ~ C10（Width, high, port number）It is followed successively by（11,11,1）、（5,5,96）、（3,3,256）、（3,3,
384）、（3,3,384）、（1,1,256）、（1, Isosorbide-5-Nitrae 096）、（1, Isosorbide-5-Nitrae 096）、（1,1,384）、（1,1,384）.Warp lamination
DC1 is up-sampled and is integrated to C8 ~ C10 thermodynamic charts exported using interpolation method, and wherein interpolation method is initialized as two-wire
Property interpolation method, parameter can be learnt by backpropagation.The cost function selection softmax loss of convolutional neural networks
Function, activation primitive selection correction linear unit（Rectified linear unit, ReLU）Function.Convolution god in order to prevent
Through network model over-fitting, weights attenuation is added in cost function（weight decay）Regularization term, and in the 6th convolution
Dropout layers are added in after layer C6, Dropout ratios are set as 0.5；
（2）Select Training strategy：Convolutional neural networks training optimizes solution, implementation model using stochastic gradient descent method
Parameter updates, and uses momentum method（momentum）, batch regularization method accelerate learning process；
（3）Selected deep learning database：Above-described convolutional neural networks structure, and root are realized using deep learning library Caffe
It is trained according to the Training strategy and training set of images that have selected.
S3, image to be detected is detected using trained convolutional neural networks model and exports prediction label figure
Picture is denoted as。
The concrete operations of the step S3 are as follows：
（1）A tunnel image is selected as image to be detected, and using bilinear interpolation method the big ditty of image to be detected
It is whole extremely（w’,H’）The acquisition requirement of pixel, wherein image to be detected is consistent with deep learning training set image；
（2）The C++ interfaces of Caffe is called to load trained convolutional neural networks model, image to be detected are made inferences pre-
It surveys, exports prediction label image。
S4, according to prediction label imageTesting result is exported, including image category（There is free from flaw）, crack coordinate
Information and the Pixel-level width value and length value in crack, wherein image category are denoted as Class, and the coordinate information in crack is denoted as, the Pixel-level width value and length value in crack are denoted as W respectivelypixel、Lpixel。
The concrete operations of the step S4 are as follows：
（1）Prediction label image size is adjusted to tunnel original image size using bicubic interpolation method（W,H）, after adjustment
Prediction label image be denoted as, form is consistent with I；
（2）Traverse prediction label imageIn all connected domains, i.e. crack area, using connected domain area minimize strategy carry
The boundary rectangle of crack area is taken, and calculates the ratio of width to height of boundary rectangle, is denoted as.IfMore than or equal to linear decision threshold
Value, labeled as effective fracture region；IfLess than linear decision threshold value, then labeled as invalid crack area, wherein linear decision
Threshold value is denoted as.If prediction label imageMiddle no effective fracture region, image category Class are set as 0；Otherwise image class
Other Class is set as 1；
（3）Effective fracture region is traversed, its profile point coordinate set is extracted, is denoted as, the number of profile point coordinates is denoted as, and
The angle of its boundary rectangle long side is calculated, is denoted as.Computational methods are as follows：
（1）
Wherein,WithIt is two extreme coordinates of boundary rectangle long side；
（4）Calculate the pixel level length in crack.The profile point coordinate set in fracture regionIt is sampled, the sampling interval is denoted as, the number of the profile point coordinates after sampling is denoted as.According to the profile point coordinate set after samplingCalculate the Pixel-level in crack
Length value, computational methods are as follows：
(2)
（5）Calculate the pixel level width in crack.The profile point coordinate set in fracture regionIt is sampled, the sampling interval is denoted as, the profile point coordinate set after sampling is denoted as, the number of profile point coordinates is denoted as.According to the profile point coordinates after sampling
CollectionCalculate the Pixel-level width value in crack, it is as follows specifically to calculate step：
（a）The crack pixel level width value of each profile point position after sampling is initialized as 0；
（b）Profile point coordinate set after traversal sampling, adjacent 2 points of angle value is calculated, is denoted as.Compare
WithIf Error Absolute Value between the two is greater than or equal to angular deviation threshold value, then abandons calculating current outline point position
The Pixel-level width value in crack, traverses next profile point, and angular deviation threshold value is denoted as.If Error Absolute Value between the two is small
In angular deviation threshold value, then the Pixel-level width value in the crack of current outline point position is calculated.The crack of each profile point position
Pixel-level width value be denoted as, computational methods are as follows：
Wherein,。
（c）The Pixel-level width value post processing in crack.It removes firstIn zero, then calculateBe averaged
Value, takes Pixel-level width value of this average value as final crack。
S5, output disease record as a result, if there are cracks in image to be detected, records Image Name according to testing result
Claim, the coordinate information in crackAnd the developed width value and length value in crack；If there is no crack in image to be detected, no
Record.Wherein the developed width value in crack and length value are denoted as w respectivelyreal、Lreal, computational methods are as follows：
Wherein,It is change of scale coefficient, represents the corresponding actual distance value of single pixel in image, generally desirable 0.2 mm/
pixel。
Compared with prior art, the present invention has the following advantages：（1）The present invention can directly export crack segmentation prediction knot
Fruit simplifies crack disease testing process, more efficient.（2）The present invention is realized end-to-end using improved convolutional neural networks
Crack identification, robustness and accuracy rate higher.（3）The present invention can directly acquire the length and width information in crack, and essence
Degree is high, and the assessment of this fracture disease grade is most important.
Description of the drawings
Fig. 1 is the crack identification process flow schematic diagram of the method for the present invention.
Fig. 2 is the convolutional neural networks structure chart of the method for the present invention.
Fig. 3 is the calculatingcrackswidth principle schematic of the method for the present invention.
Specific embodiment
The present invention is described in further detail below in conjunction with the accompanying drawings.
With reference to the accompanying drawings 1, the present invention is a kind of tunnel slot method for quickly identifying based on deep learning, which is characterized in that
The tunnel slot method for quickly identifying includes the following steps：
S1, deep learning training set of images is created.
The concrete operations of the step S1 are as follows：
（1）Acquire tunnel image.Tunnel surface is shot by mobile testing equipment and obtains single pass tunnel image, resolution ratio
General requirement is not less than 2,000,000 pixels, and the clarity requirement of image is sufficiently high, and the crack minimum widith in image is not less than 1 picture
Element；
（2）By artificial screening tunnel image, 50,000 tension fissure images and 50,000 non-crack images are selected, and it is big to adjust image
Small, picture format is consistent with original tunnel image.Original tunnel image is denoted as I, wide and high be denoted as（W,H）, the image after adjustment
It is denoted as I’, it is wide and high be denoted as（w’, H’）；
Wherein the width of image and coca are set according to the resolution ratio of camera,（W,H）Generally (2560,2048),（w’, H’）It is set as
(2048,2048)；
（3）Crack image after being sized using Photoshop softwares opening, and pass through " quick selection " tool and select crack
Region is denoted as, non-crack area is denoted as；
（4）The crack area selectedLabeled as 1, and crack area Fill Color is set for white.Non- crack areaLabeled as 0, and it is black to set non-crack area Fill Color；
（5）After the completion of crack and non-crack area color filling, training set label image is saved as, is denoted as, form with
Original tunnel image I is consistent, and size is（w’, H’）；
Training set of images is divided into training set, verification collection and three parts of test set, allocation proportion and is set as 0.7:0.2:0.1.
S2, training depth convolutional neural networks model.
The concrete operations of the step S2 are as follows：
（1）With reference to the accompanying drawings 2, build convolutional neural networks structure：Convolutional neural networks are realized by improving AlexNet structures,
3 full articulamentums are changed into 3 convolutional layers, increase by 1 warp lamination, improved network is by 10 convolutional layers, 5 ponds
Layer, 1 Dropout layers, 1 warp lamination, 10 convolutional layers are denoted as C1 ~ C10, and 5 pond layers are denoted as P1 ~ P5,1 deconvolution
Layer is denoted as DC1.The convolution kernel size of convolutional layer C1 ~ C10（Width, high, port number）It is followed successively by（11,11,1）、（5,5,96）、（3,
3,256）、（3,3,384）、（3,3,384）、（1,1,256）、（1, Isosorbide-5-Nitrae 096）、（1, Isosorbide-5-Nitrae 096）、（1,1,384）、（1,1,
384）.Warp lamination DC1 is up-sampled and is integrated to C8 ~ C10 thermodynamic charts exported using interpolation method, wherein interpolation method
Bilinear interpolation is initialized as, parameter can be learnt by backpropagation.The cost function selection of convolutional neural networks
Softmax loss functions, activation primitive selection correction linear unit（Rectified linear unit, ReLU）Function.For
It prevents convolutional neural networks model over-fitting, weights attenuation is added in cost function（weight decay）Regularization term,
And Dropout layers are added in after the 6th convolutional layer C6, Dropout ratios are set as 0.5；
（2）Select Training strategy：Convolutional neural networks training optimizes solution, implementation model using stochastic gradient descent method
Parameter updates, and uses momentum method（momentum）, batch regularization method accelerate learning process；
（3）Selected deep learning database：Above-described convolutional neural networks structure, and root are realized using deep learning library Caffe
It is trained according to the Training strategy and training set of images that have selected；
The Caffe hyper parameters file of convolutional neural networks mainly sets as follows：
test_iter: 10000
test_interval: 40000
base_lr: 0.01
momentum: 0.9
weight_decay: 0.0001
lr_policy: "inv"
gamma: 0.0001
power: 0.8
max_iter: 2000000
snapshot: 40000
snapshot_prefix: "./Crack_Detection_model"
solver_mode: GPU
type: "SGD"
Most of hyper parameter about convolutional neural networks is set, and has no general selection method at present, is all according to specific instruction
Depending on the situation of white silk.Wherein：
（1）Test_iter is to have traversed the iterations that all test sets need, and setting value is test set size divided by mini-
Batch sizes, mini-batch size are set as 2；
（2）Test_interval is test iteration interval, i.e., is tested per iteration how many times, and setting value is training set size
Divided by mini-batch sizes；
（3）Base_lr is initial learning rate, is generally set to 0.01, lr_policy as learning rate more new strategy, gamma, power
It is the relevant parameter of selected learning rate more new strategy；
（4）Momentum is the hyper parameter of momentum method, is generally set to 0.9；
（5）Weight_decay is the hyper parameter of weights attenuation, is generally set to 0.0001；
（6）Max_iter is maximum iteration, and setting wants sufficiently large to ensure that model can restrain, generally test_
30 ~ 100 times of interval；
（7）Snapshot be preserve training snapshot iteration interval, can generally set it is consistent with test_interval,
Snapshot_prefix is the path for preserving training snapshot；
（8）Solver_mode is operational mode, and GPU patterns are than CPU faster；
（9）Type is selected for optimization method, and depending on training, stochastic gradient descent method may be selected（SGD）.
S3, image to be detected is detected using trained convolutional neural networks model and exports prediction label figure
Picture is denoted as。
The concrete operations of the step S3 are as follows：
（1）A tunnel image is selected as image to be detected, and using bilinear interpolation method the big ditty of image to be detected
It is whole extremely（w’, H’）The acquisition requirement of pixel, wherein image to be detected is consistent with deep learning training set image；
（2）The C++ interfaces of Caffe is called to load trained convolutional neural networks model, image to be detected are made inferences pre-
It surveys, exports prediction label image。
S4, according to prediction label imageTesting result is exported, including image category（There is free from flaw）, crack coordinate
Information and the Pixel-level width value and length value in crack, wherein image category are denoted as Class, and the coordinate information in crack is denoted as, the Pixel-level width value and length value in crack are denoted as W respectivelypixel、Lpixel。
The concrete operations of the step S4 are as follows：
（1）Prediction label image size is adjusted to tunnel original image size using bicubic interpolation method（W,H）, after adjustment
Prediction label image be denoted as, form is consistent with I；
（2）Traverse prediction label imageIn all connected domains, i.e. crack area, using connected domain area minimize strategy carry
The boundary rectangle of crack area is taken, and calculates the ratio of width to height of boundary rectangle, is denoted as.IfMore than or equal to linear decision threshold
Value, labeled as effective fracture region；IfLess than linear decision threshold value, then labeled as invalid crack area, wherein linear decision
Threshold value is denoted as, can generally be set as 2.If prediction label imageMiddle no effective fracture region, image category Class are set as
0；Otherwise image category Class is set as 1；
（3）Effective fracture region is traversed, its profile point coordinate set is extracted, is denoted as, the number of profile point coordinates is denoted as, and
The angle of its boundary rectangle long side is calculated, is denoted as.Computational methods are as follows：
(1)
Wherein,WithIt is two extreme coordinates of boundary rectangle long side；
（4）Calculate the Pixel-level length value in crack.The profile point coordinate set in fracture regionIt is sampled, sampling interval note
For, the number of the profile point coordinates after sampling is denoted as.According to the profile point coordinate set after samplingCalculate the pixel in crack
Level length value, computational methods are as follows：
(2)
Wherein,30 can be generally set as, the value is bigger, and length computation accuracy rate is higher, and efficiency is lower, and the value is smaller, length gauge
Calculation accuracy rate is lower, and efficiency is higher；
（5）With reference to the accompanying drawings 3, calculate the Pixel-level width value in crack.The profile point coordinate set in fracture regionIt is sampled,
Sampling interval is denoted as, the profile point coordinate set after sampling is denoted as, the number of profile point coordinates is denoted as.After sampling
Profile point coordinate setCalculate the Pixel-level width value in crack, it is as follows specifically to calculate step：
（a）The crack pixel level width value of each profile point position after sampling is initialized as 0；
（b）Profile point coordinate set after traversal sampling, adjacent 2 points of angle value is calculated, is denoted as.Compare
WithIf Error Absolute Value between the two is greater than or equal to angular deviation threshold value, then abandons calculating current outline point position
The Pixel-level width value in crack, traverses next profile point, and angular deviation threshold value is denoted as.If Error Absolute Value between the two is small
In angular deviation threshold value, then the Pixel-level width value in the crack of current outline point position is calculated.The crack of each profile point position
Pixel-level width value be denoted as, computational methods are as follows：
Wherein,10 can be set as, the value is bigger, and width gauge calculation accuracy rate is higher, efficiency
Lower, the value is smaller, and width gauge calculation accuracy rate is lower, and efficiency is higher.It can be set as 300, the value is excessive or too small can all lead to width
Degree calculating accuracy rate is relatively low, can be adjusted according to practical fracture width value test case；
（c）The Pixel-level width value post processing in crack.It removes firstIn zero, then calculateAverage value, take
Pixel-level width value of this average value as final crack。
S5, output disease record as a result, if there are cracks in image to be detected, records Image Name according to testing result
Claim, the coordinate information in crackAnd the developed width value and length value in crack；If there is no crack in image to be detected, no
Record.Wherein the developed width value in crack and length value are denoted as w respectivelyreal、Lreal, computational methods are as follows：
Wherein,It is change of scale coefficient, the corresponding actual distance value of single pixel in image is represented, according to the resolution ratio of camera
Adjustment, generally desirable 0.2 mm/pixel；
The control parameter that fracture length value and width value calculate can be defined as follows：
struct CrackGeoCtlPara
{
int nSamLen;The sampling interval that // setting length value calculates
int nSamNumWid;The sample point number that // setting width value calculates
double dLineTH;// setting linear decision threshold value
double dScaleTran;// setting change of scale coefficient
double dAngleBias;// set angle deviation threshold
}。
The preferred embodiment of the present invention described in detail above.It should be appreciated that those of ordinary skill in the art without
Creative work is needed according to the present invention can to conceive and makes many modifications and variations.Therefore, all technologies in the art
Personnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea
Technical solution, all should be in the protection domain being defined in the patent claims.
Claims (6)
1. a kind of tunnel slot method for quickly identifying based on deep learning, which is characterized in that include the following steps：
S1, deep learning training set of images is created；
S2, training depth convolutional neural networks model；
S3, image to be detected is detected using trained convolutional neural networks model and exports prediction label image；
S4, testing result is exported according to prediction label image, includes the pixel of image category, the coordinate information in crack and crack
Level width value and length value；
S5, output disease record as a result, if there are cracks in image to be detected, records image name, splits according to testing result
The coordinate information of seam and the developed width value and length value in crack；If there is no crack in image to be detected, do not record.
A kind of 2. tunnel slot method for quickly identifying based on deep learning according to claim 1, which is characterized in that institute
The step S1 stated includes：
（1）Acquire tunnel image：Tunnel surface is shot by mobile testing equipment and obtains single pass tunnel image；
（2）By artificial screening tunnel image, sufficient amount of tension fissure image and non-crack image are selected, and it is big to adjust image
Small, picture format is consistent with original tunnel image；Original tunnel image is denoted as I, wide and high be denoted as（W,H）, the image after adjustment
It is denoted as I’, it is wide and high be denoted as（w’, H’）；
（3）Crack image after being sized using Photoshop softwares opening, and pass through " quick selection " tool and select crack
Region is denoted as, non-crack area is denoted as；
（4）The crack area selectedLabeled as 1, and crack area Fill Color is set for white；Non- crack area
Labeled as 0, and it is black to set non-crack area Fill Color；
（5）After the completion of crack and non-crack area color filling, training set label image is saved as, is denoted as, form with it is former
Beginning, image I in tunnel was consistent, and size is（w’,H’）.
A kind of 3. tunnel slot method for quickly identifying based on deep learning according to claim 1, which is characterized in that institute
The step S2 stated includes：
1）Build convolutional neural networks structure：Convolutional neural networks are realized by improving AlexNet structures, 3 full connections
Layer changes 3 convolutional layers into, increases by 1 warp lamination, improved network is by 10 convolutional layers, 5 pond layers, 1 Dropout
Layer, 1 warp lamination, 10 convolutional layers are denoted as C1 ~ C10, and 5 pond layers are denoted as P1 ~ P5, and 1 warp lamination is denoted as DC1；Volume
The cost function selection softmax loss functions of product neural network, the activation primitive selection linear unit function of rectification；In order to anti-
Only convolutional neural networks model over-fitting adds in weights attenuation regularization term, and after the 6th convolutional layer C6 in cost function
Dropout layers are added in, Dropout ratios are set as 0.5；
（2）Select Training strategy：Convolutional neural networks training optimizes solution, implementation model using stochastic gradient descent method
Parameter updates, and accelerates learning process using momentum method, batch regularization method；
（3）Selected deep learning database：Above-described convolutional neural networks structure, and root are realized using deep learning library Caffe
It is trained according to the Training strategy and training set of images that have selected.
A kind of 4. tunnel slot method for quickly identifying based on deep learning according to claim 1, which is characterized in that institute
Step S3 is stated to include：
（1）A tunnel image is selected as image to be detected, and using bilinear interpolation method the big ditty of image to be detected
It is whole extremely（w’,H’）The acquisition requirement of pixel, wherein image to be detected is consistent with deep learning training set image；
（2）The C++ interfaces of Caffe is called to load trained convolutional neural networks model, image to be detected are made inferences pre-
It surveys, exports prediction label image。
A kind of 5. tunnel slot method for quickly identifying based on deep learning according to claim 1, which is characterized in that institute
Step S4 is stated to include：
（1）Prediction label image size is adjusted to tunnel original image size using bicubic interpolation method（W,H）, after adjustment
Prediction label image be denoted as, form is consistent with I；
（2）Traverse prediction label imageIn all connected domains, i.e. crack area, using connected domain area minimize strategy carry
The boundary rectangle of crack area is taken, and calculates the ratio of width to height of boundary rectangle, is denoted as；IfMore than or equal to linear decision threshold
Value, labeled as effective fracture region；IfLess than linear decision threshold value, then labeled as invalid crack area, wherein linear decision
Threshold value is denoted as；If prediction label imageMiddle no effective fracture region, image category Class are set as 0；Otherwise image class
Other Class is set as 1；
（3）Effective fracture region is traversed, its profile point coordinate set is extracted, is denoted as, the number of profile point coordinates is denoted as, and count
The angle of its boundary rectangle long side is calculated, is denoted as；Computational methods are as follows：
Wherein,WithIt is two extreme coordinates of boundary rectangle long side；
（4）Calculate the pixel level length in crack：The profile point coordinate set in fracture regionIt is sampled, the sampling interval is denoted as
, the number of the profile point coordinates after sampling is denoted as；According to the profile point coordinate set after samplingThe Pixel-level for calculating crack is long
Angle value, computational methods are as follows：
；
（5）Calculate the pixel level width in crack：The profile point coordinate set in fracture regionIt is sampled, the sampling interval is denoted as
, the profile point coordinate set after sampling is denoted as, the number of profile point coordinates is denoted as；According to the profile point coordinate set after samplingCalculate the Pixel-level width value in crack, it is as follows specifically to calculate step：
（a）The crack pixel level width value of each profile point position after sampling is initialized as 0；
（b）Profile point coordinate set after traversal sampling, adjacent 2 points of angle value is calculated, is denoted as；Compare
WithIf Error Absolute Value between the two is greater than or equal to angular deviation threshold value, then abandons calculating current outline point position
The Pixel-level width value in crack, traverses next profile point, and angular deviation threshold value is denoted as；If Error Absolute Value between the two is small
In angular deviation threshold value, then the Pixel-level width value in the crack of current outline point position is calculated；The crack of each profile point position
Pixel-level width value be denoted as, computational methods are as follows：
Wherein,；
（c）The Pixel-level width value post processing in crack：It removes firstIn zero, then calculateAverage value, take
Pixel-level width value of this average value as final crack。
A kind of 6. tunnel slot method for quickly identifying based on deep learning according to claim 1, which is characterized in that institute
Step S5 is stated to include：Output disease record records Image Name as a result, if there are cracks in image to be detected according to testing result
Claim, the coordinate information in crackAnd the developed width value and length value in crack；If there is no crack in image to be detected, no
Record；Wherein the developed width value in crack and length value are denoted as w respectivelyreal、Lreal, computational methods are as follows：
Wherein,It is change of scale coefficient, represents the corresponding actual distance value of single pixel in image, take 0.2 mm/pixel.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810038939.5A CN108229461B (en) | 2018-01-16 | 2018-01-16 | Tunnel crack rapid identification method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810038939.5A CN108229461B (en) | 2018-01-16 | 2018-01-16 | Tunnel crack rapid identification method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108229461A true CN108229461A (en) | 2018-06-29 |
| CN108229461B CN108229461B (en) | 2021-12-28 |
Family
ID=62640776
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810038939.5A Active CN108229461B (en) | 2018-01-16 | 2018-01-16 | Tunnel crack rapid identification method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108229461B (en) |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109615653A (en) * | 2018-12-19 | 2019-04-12 | 上海大学 | Percolating water area detecting and recognition methods based on deep learning and visual field projection model |
| CN109712148A (en) * | 2018-12-19 | 2019-05-03 | 上海勘察设计研究院(集团)有限公司 | Segment joint position automatic identifying method based on shield tunnel image |
| CN109900706A (en) * | 2019-03-20 | 2019-06-18 | 易思维（杭州）科技有限公司 | A kind of weld seam and weld defect detection method based on deep learning |
| CN110097140A (en) * | 2019-05-13 | 2019-08-06 | 招商局重庆交通科研设计院有限公司 | A kind of data processing system for tunnel crusing robot |
| CN110163842A (en) * | 2019-04-15 | 2019-08-23 | 深圳高速工程检测有限公司 | Building cracks detection method, device, computer equipment and storage medium |
| CN110197483A (en) * | 2019-05-31 | 2019-09-03 | 南京深地智能建造技术研究院有限公司 | Deep basal pit crack detection method based on vision signal |
| CN110197203A (en) * | 2019-05-08 | 2019-09-03 | 湖北民族大学 | Bridge pavement classification of rifts recognition methods based on width learning neural network |
| CN110348342A (en) * | 2019-06-27 | 2019-10-18 | 广东技术师范大学天河学院 | A kind of piping disease image partition method based on full convolutional network |
| CN110414073A (en) * | 2019-07-04 | 2019-11-05 | 中国神华能源股份有限公司神朔铁路分公司 | State evaluating method, device, computer equipment and the storage medium in tunnel |
| CN110503627A (en) * | 2019-07-11 | 2019-11-26 | 平安科技（深圳）有限公司 | Building tear detection method, device, storage medium and computer equipment |
| CN110689042A (en) * | 2019-08-20 | 2020-01-14 | 中国矿业大学(北京) | Tunnel leakage grade identification method and device, storage medium and electronic device |
| CN110909657A (en) * | 2019-11-19 | 2020-03-24 | 上海勘察设计研究院(集团)有限公司 | Method for identifying apparent tunnel disease image |
| CN110952978A (en) * | 2019-12-20 | 2020-04-03 | 西南石油大学 | Drilling leakage fracture width prediction method based on neural network data mining |
| CN111489352A (en) * | 2020-04-29 | 2020-08-04 | 安徽国钜工程机械科技有限公司 | Tunnel gap detection and measurement method and device based on digital image processing |
| CN111563888A (en) * | 2020-05-06 | 2020-08-21 | 清华大学 | Quantitative crack growth monitoring method |
| CN111932548A (en) * | 2020-04-21 | 2020-11-13 | 南京航空航天大学 | Subway tunnel fine crack segmentation method based on deep learning |
| CN111967398A (en) * | 2020-08-18 | 2020-11-20 | 创新奇智(合肥)科技有限公司 | Tunnel detection method and device, electronic equipment and readable storage medium |
| CN112001359A (en) * | 2020-09-09 | 2020-11-27 | 郑州航空工业管理学院 | Heuristic multi-branch line repairing method based on pattern recognition and optimal distribution |
| CN112330593A (en) * | 2020-10-10 | 2021-02-05 | 南京理工大学 | Building surface crack detection method based on deep learning network |
| CN112488119A (en) * | 2020-12-18 | 2021-03-12 | 山西省信息产业技术研究院有限公司 | Tunnel block falling or water seepage detection and measurement method based on double-depth learning model |
| CN112508030A (en) * | 2020-12-18 | 2021-03-16 | 山西省信息产业技术研究院有限公司 | Tunnel crack detection and measurement method based on double-depth learning model |
| WO2021068848A1 (en) * | 2019-10-09 | 2021-04-15 | 山东大学 | Tunnel structure disease multi-scale measurement and intelligent diagnosis system and method |
| CN112924653A (en) * | 2021-01-15 | 2021-06-08 | 清华大学 | Non-contact concrete segregation degree intelligent real-time analysis method and device |
| CN112967266A (en) * | 2021-03-23 | 2021-06-15 | 武汉大学 | Laser directional energy deposition area calculation method of full convolution neural network |
| CN113033702A (en) * | 2021-04-21 | 2021-06-25 | 上海勘察设计研究院(集团)有限公司 | Structure safety early warning method based on space-time data |
| CN113607546A (en) * | 2021-08-10 | 2021-11-05 | 西北工业大学 | Crack measuring method |
| CN114004004A (en) * | 2022-01-04 | 2022-02-01 | 城云科技（中国）有限公司 | Tunnel crack change trend prediction method and device based on deep learning and application |
| CN114066891A (en) * | 2022-01-17 | 2022-02-18 | 北京市建设工程质量第一检测所有限责任公司 | Wall crack identification and judgment method, device, equipment and medium based on deep learning |
| CN114187313A (en) * | 2021-11-10 | 2022-03-15 | 泰瑞数创科技(北京)有限公司 | Artificial neural network building contour extraction method based on sliding edge detection |
| CN114384073A (en) * | 2021-11-30 | 2022-04-22 | 杭州申昊科技股份有限公司 | Crack detection method and system based on subway tunnel |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101169322A (en) * | 2006-10-25 | 2008-04-30 | 刘其伟 | Concrete surface crack long distance photoelectric type digital detection system |
| CN102297660A (en) * | 2011-05-20 | 2011-12-28 | 同济大学 | Measuring method of shield tunnel lining segment seam open width and apparatus thereof |
| CN103440657A (en) * | 2013-08-27 | 2013-12-11 | 武汉大学 | Method for online screening cracks of road |
| CN104483330A (en) * | 2014-11-11 | 2015-04-01 | 浙江大学 | Concrete surface crack real-time monitoring system and cracking risk dynamic assessment method |
| CN105719259A (en) * | 2016-02-19 | 2016-06-29 | 上海理工大学 | Pavement crack image detection method |
| CN107133960A (en) * | 2017-04-21 | 2017-09-05 | 武汉大学 | Image crack dividing method based on depth convolutional neural networks |
| CN107230202A (en) * | 2017-05-16 | 2017-10-03 | 淮阴工学院 | The automatic identifying method and system of pavement disease image |
| CN107240079A (en) * | 2017-06-09 | 2017-10-10 | 南京理工大学 | A kind of road surface crack detection method based on image procossing |
| CN107403197A (en) * | 2017-07-31 | 2017-11-28 | 武汉大学 | A kind of crack identification method based on deep learning |
-
2018
- 2018-01-16 CN CN201810038939.5A patent/CN108229461B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101169322A (en) * | 2006-10-25 | 2008-04-30 | 刘其伟 | Concrete surface crack long distance photoelectric type digital detection system |
| CN102297660A (en) * | 2011-05-20 | 2011-12-28 | 同济大学 | Measuring method of shield tunnel lining segment seam open width and apparatus thereof |
| CN103440657A (en) * | 2013-08-27 | 2013-12-11 | 武汉大学 | Method for online screening cracks of road |
| CN104483330A (en) * | 2014-11-11 | 2015-04-01 | 浙江大学 | Concrete surface crack real-time monitoring system and cracking risk dynamic assessment method |
| CN105719259A (en) * | 2016-02-19 | 2016-06-29 | 上海理工大学 | Pavement crack image detection method |
| CN107133960A (en) * | 2017-04-21 | 2017-09-05 | 武汉大学 | Image crack dividing method based on depth convolutional neural networks |
| CN107230202A (en) * | 2017-05-16 | 2017-10-03 | 淮阴工学院 | The automatic identifying method and system of pavement disease image |
| CN107240079A (en) * | 2017-06-09 | 2017-10-10 | 南京理工大学 | A kind of road surface crack detection method based on image procossing |
| CN107403197A (en) * | 2017-07-31 | 2017-11-28 | 武汉大学 | A kind of crack identification method based on deep learning |
Non-Patent Citations (2)
| Title |
|---|
| SHANSHAN ZHAO ET AL: "Road surface crack detection based on CNN", 《TRANSDUCER AND MICROSYSTEM TECHNOLOGIES》 * |
| 夏军: "基于图像分析技术的结构表面裂缝形态检测", 《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》 * |
Cited By (41)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109615653A (en) * | 2018-12-19 | 2019-04-12 | 上海大学 | Percolating water area detecting and recognition methods based on deep learning and visual field projection model |
| CN109712148A (en) * | 2018-12-19 | 2019-05-03 | 上海勘察设计研究院(集团)有限公司 | Segment joint position automatic identifying method based on shield tunnel image |
| CN109615653B (en) * | 2018-12-19 | 2023-05-05 | 上海大学 | Leakage water area detection and identification method based on deep learning and visual field projection model |
| CN109900706A (en) * | 2019-03-20 | 2019-06-18 | 易思维（杭州）科技有限公司 | A kind of weld seam and weld defect detection method based on deep learning |
| CN110163842A (en) * | 2019-04-15 | 2019-08-23 | 深圳高速工程检测有限公司 | Building cracks detection method, device, computer equipment and storage medium |
| CN110163842B (en) * | 2019-04-15 | 2021-06-25 | 深圳高速工程检测有限公司 | Building crack detection method and device, computer equipment and storage medium |
| CN110197203A (en) * | 2019-05-08 | 2019-09-03 | 湖北民族大学 | Bridge pavement classification of rifts recognition methods based on width learning neural network |
| CN110197203B (en) * | 2019-05-08 | 2022-06-10 | 湖北民族大学 | Bridge pavement crack classification and identification method based on width learning neural network |
| CN110097140A (en) * | 2019-05-13 | 2019-08-06 | 招商局重庆交通科研设计院有限公司 | A kind of data processing system for tunnel crusing robot |
| CN110197483A (en) * | 2019-05-31 | 2019-09-03 | 南京深地智能建造技术研究院有限公司 | Deep basal pit crack detection method based on vision signal |
| CN110348342A (en) * | 2019-06-27 | 2019-10-18 | 广东技术师范大学天河学院 | A kind of piping disease image partition method based on full convolutional network |
| CN110414073A (en) * | 2019-07-04 | 2019-11-05 | 中国神华能源股份有限公司神朔铁路分公司 | State evaluating method, device, computer equipment and the storage medium in tunnel |
| CN110414073B (en) * | 2019-07-04 | 2023-05-26 | 中国神华能源股份有限公司神朔铁路分公司 | Tunnel state evaluation method and device, computer equipment and storage medium |
| CN110503627A (en) * | 2019-07-11 | 2019-11-26 | 平安科技（深圳）有限公司 | Building tear detection method, device, storage medium and computer equipment |
| CN110689042A (en) * | 2019-08-20 | 2020-01-14 | 中国矿业大学(北京) | Tunnel leakage grade identification method and device, storage medium and electronic device |
| CN110689042B (en) * | 2019-08-20 | 2022-05-13 | 中国矿业大学(北京) | Tunnel leakage grade identification method and device, storage medium and electronic device |
| WO2021068848A1 (en) * | 2019-10-09 | 2021-04-15 | 山东大学 | Tunnel structure disease multi-scale measurement and intelligent diagnosis system and method |
| CN113424055A (en) * | 2019-10-09 | 2021-09-21 | 山东大学 | Multi-scale tunnel structure disease detection and intelligent diagnosis system and method |
| CN110909657A (en) * | 2019-11-19 | 2020-03-24 | 上海勘察设计研究院(集团)有限公司 | Method for identifying apparent tunnel disease image |
| CN110952978A (en) * | 2019-12-20 | 2020-04-03 | 西南石油大学 | Drilling leakage fracture width prediction method based on neural network data mining |
| CN111932548A (en) * | 2020-04-21 | 2020-11-13 | 南京航空航天大学 | Subway tunnel fine crack segmentation method based on deep learning |
| CN111489352B (en) * | 2020-04-29 | 2023-06-02 | 安徽国钜工程机械科技有限公司 | Tunnel gap detection and measurement method and device based on digital image processing |
| CN111489352A (en) * | 2020-04-29 | 2020-08-04 | 安徽国钜工程机械科技有限公司 | Tunnel gap detection and measurement method and device based on digital image processing |
| CN111563888A (en) * | 2020-05-06 | 2020-08-21 | 清华大学 | Quantitative crack growth monitoring method |
| CN111967398A (en) * | 2020-08-18 | 2020-11-20 | 创新奇智(合肥)科技有限公司 | Tunnel detection method and device, electronic equipment and readable storage medium |
| CN111967398B (en) * | 2020-08-18 | 2022-08-09 | 创新奇智(合肥)科技有限公司 | Tunnel detection method and device, electronic equipment and readable storage medium |
| CN112001359A (en) * | 2020-09-09 | 2020-11-27 | 郑州航空工业管理学院 | Heuristic multi-branch line repairing method based on pattern recognition and optimal distribution |
| CN112001359B (en) * | 2020-09-09 | 2023-07-18 | 郑州航空工业管理学院 | Heuristic multi-fork line repairing method based on pattern recognition and optimal allocation |
| CN112330593A (en) * | 2020-10-10 | 2021-02-05 | 南京理工大学 | Building surface crack detection method based on deep learning network |
| CN112508030A (en) * | 2020-12-18 | 2021-03-16 | 山西省信息产业技术研究院有限公司 | Tunnel crack detection and measurement method based on double-depth learning model |
| CN112488119A (en) * | 2020-12-18 | 2021-03-12 | 山西省信息产业技术研究院有限公司 | Tunnel block falling or water seepage detection and measurement method based on double-depth learning model |
| CN112924653A (en) * | 2021-01-15 | 2021-06-08 | 清华大学 | Non-contact concrete segregation degree intelligent real-time analysis method and device |
| CN112967266A (en) * | 2021-03-23 | 2021-06-15 | 武汉大学 | Laser directional energy deposition area calculation method of full convolution neural network |
| CN112967266B (en) * | 2021-03-23 | 2024-02-06 | 湖南珞佳智能科技有限公司 | Laser directional energy deposition area calculation method of full convolution neural network |
| CN113033702A (en) * | 2021-04-21 | 2021-06-25 | 上海勘察设计研究院(集团)有限公司 | Structure safety early warning method based on space-time data |
| CN113607546A (en) * | 2021-08-10 | 2021-11-05 | 西北工业大学 | Crack measuring method |
| CN114187313A (en) * | 2021-11-10 | 2022-03-15 | 泰瑞数创科技(北京)有限公司 | Artificial neural network building contour extraction method based on sliding edge detection |
| CN114384073A (en) * | 2021-11-30 | 2022-04-22 | 杭州申昊科技股份有限公司 | Crack detection method and system based on subway tunnel |
| CN114004004B (en) * | 2022-01-04 | 2022-04-01 | 城云科技（中国）有限公司 | Tunnel crack change trend prediction method and device based on deep learning and application |
| CN114004004A (en) * | 2022-01-04 | 2022-02-01 | 城云科技（中国）有限公司 | Tunnel crack change trend prediction method and device based on deep learning and application |
| CN114066891A (en) * | 2022-01-17 | 2022-02-18 | 北京市建设工程质量第一检测所有限责任公司 | Wall crack identification and judgment method, device, equipment and medium based on deep learning |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108229461B (en) | 2021-12-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108229461A (en) | A kind of tunnel slot method for quickly identifying based on deep learning | |
| US20210319561A1 (en) | Image segmentation method and system for pavement disease based on deep learning | |
| CN108960135B (en) | Dense ship target accurate detection method based on high-resolution remote sensing image | |
| CN107066995A (en) | A kind of remote sensing images Bridges Detection based on convolutional neural networks | |
| CN103714339B (en) | SAR image road damaging information extracting method based on vector data | |
| CN107451997A (en) | A kind of automatic identifying method of the welding line ultrasonic TOFD D scanning defect types based on deep learning | |
| CN111986099A (en) | Tillage monitoring method and system based on convolutional neural network with residual error correction fused | |
| CN109816012A (en) | A kind of multiscale target detection method of integrating context information | |
| CN107480611A (en) | A kind of crack identification method based on deep learning convolutional neural networks | |
| CN113392775A (en) | Sugarcane seedling automatic identification and counting method based on deep neural network | |
| CN114548278A (en) | In-service tunnel lining structure defect identification method and system based on deep learning | |
| CN108875794A (en) | Image visibility detection method based on transfer learning | |
| den Bieman et al. | Deep learning video analysis as measurement technique in physical models | |
| CN109785307A (en) | A kind of unmanned plane image road Damage assessment method based on vector guidance | |
| CN112801972A (en) | Bridge defect detection method, device, system and storage medium | |
| CN116863274A (en) | Semi-supervised learning-based steel plate surface defect detection method and system | |
| CN108229524A (en) | A kind of chimney and condensing tower detection method based on remote sensing images | |
| CN108520277A (en) | Reinforced concrete structure seismic Damage automatic identification based on computer vision and intelligent locating method | |
| CN110532914A (en) | Building analyte detection method based on fine-feature study | |
| CN114255250A (en) | River protection engineering slope collapse detection method based on deep learning | |
| Zhang et al. | Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learning | |
| CN109241867A (en) | Using the method and device of intelligent algorithm identification digital cores image | |
| CN111626169A (en) | Image-based railway dangerous falling rock size judgment method | |
| CN109993742A (en) | Bridge Crack method for quickly identifying based on diagonal operator reciprocal | |
| CN114494870A (en) | Double-time-phase remote sensing image change detection method, model construction method and device |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |