CN109408985A - The accurate recognition methods in bridge steel structure crack based on computer vision - Google Patents
The accurate recognition methods in bridge steel structure crack based on computer vision Download PDFInfo
- Publication number
- CN109408985A CN109408985A CN201811295283.1A CN201811295283A CN109408985A CN 109408985 A CN109408985 A CN 109408985A CN 201811295283 A CN201811295283 A CN 201811295283A CN 109408985 A CN109408985 A CN 109408985A
- Authority
- CN
- China
- Prior art keywords
- width
- layer operation
- depth
- input
- layers
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computer Hardware Design (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Geometry (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
Abstract
The present invention relates to a kind of accurate recognition methods in bridge steel structure crack based on computer vision, solve the disadvantage that existing bridge steel box girder crack identification the degree of automation is too low and corresponding intelligent algorithm, method includes: to use steel box-girder common site crack image as initial data, the artificial mark of pixel scale is carried out, obtaining indicates different classes of mark image using different colours；Original image and mark image are split, distribution different digital is different classes of to do label label after mark image is extracted according to different colours；Training set is input in the full convolutional neural networks of depth and is trained, will be input in trained neural network after image to be identified cutting, carry out the Pixel-level recognition result that splicing returns to original image size after obtaining Pixel-level recognition result.The present invention is convenient, accurate, improves the accuracy and stability of steel box-girder crack identification result.
Description
Technical field
The present invention relates to science of bridge building health monitoring and detection care fields, and in particular to one kind is based on computer vision
The accurate recognition methods in bridge steel structure crack.
Background technique
Bridge is the throat in communications and transportation main artery.In recent years, bridge construction steps into climax, more and more to build up bridge
Beam is on active service, to improve China's traffic transportation efficiency, the economic rapid growth in China being pulled to play considerable effect.Steel box-girder
As the chief component of bridge especially long-span bridge, had been widely used in China and in the world.So
And since the reciprocation cycle of vehicular load acts on, the U-shaped rib of steel box-girder and top plate junction are easier to generate minute crack, this
Carrying out for a little cracks can generate huge adverse effect to the normal service of steel box-girder, therefore the detection in steel box-girder crack becomes
The hot issue of scientific circles and industrial circle research in recent years.Recently as the development of Bridge Health Monitoring Technology, researcher
A variety of non destructive detection mothods based on vibration are developed to be used to detect steel box girder surface and internal damage.However,
These methods are haveed the shortcomings that some inevitable: first is that sphere of action is too small, since these methods are by environment and structure sheet
Body vibration influence is excessive, therefore can only be detected in smaller range；Second is that required hardware device is excessively complicated, cost
It is excessively high and not strong for environmental robustness.
Now with the burning hot development of computer vision, using computer vision technique carry out steel box-girder disease detection and
The considerations of identification has entered researcher range.However, most of crack detection method based on computer vision is all
Based on image classification principle, it can not be accurate to pixel scale, there is huge hidden danger for subsequent crack quantitative judge for this.
Therefore, how it to be directed to steel box-girder damageable zone original image, the crack identification method of an effective robust is proposed, by steel box-girder
The precision improvement of crack identification is a urgently open question to Pixel-level.
Summary of the invention
The purpose of the present invention is to solve existing bridge steel box girder crack identification the degree of automation is too low and corresponding
The shortcomings that intelligent algorithm, proposes a kind of accurate recognition methods in bridge steel structure crack based on computer vision, including walks as follows
It is rapid:
Step 1: using common steel box-girder common site crack image as initial data, the artificial of pixel scale is carried out
Mark, obtaining indicates different classes of Pixel-level mark image using different colours；
Step 2: original image and mark image are split according to certain size, to reduce calculating cost, will mark
Distribution different digital is different classes of to do label label after image is extracted according to different colours；
It is trained Step 3: training set is input in the full convolutional neural networks of depth, is damaged used in training process
Mistake function is Dice related coefficient loss function, and optimization algorithm is adaptive moments estimation optimization algorithm；
Step 4: by trained neural network is input to after bridge steel box girder region of interest image to be identified cutting
In, the Pixel-level recognition result that splicing returns to original image size is carried out after obtaining Pixel-level recognition result.
The present invention also has following technical characteristic:
1, step 1 as described above specifically includes that steps are as follows:
One, using steel box-girder crack disease image data base, more than 4000 × 3000 pixel of picture size；
Two, different types of mark is carried out using pixel annotation tool to original image, covered in different colors not of the same race
Class region；
Three, original image and mark image are subjected to flip horizontal, flip vertical simultaneously, or original image is carried out
Tri- channels BGR apply 10% interference, the original image and its corresponding mark image after respectively obtaining overturning or interference.
2, in step 3 as described above each layer of the full convolutional neural networks of depth structure are as follows:
L1-0 layers: the width of input is 512, depth 3；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L1-1 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-2 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-3 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L1-4 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-5 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-6 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L1-7 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-8 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-9 layers: the width of input is 512, and depth 32 executes pond layer operation；
L2-0 layers: the width of input is 256, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L2-1 layers: the width of input is 256, depth 32；Execute activation layer operation；
L2-2 layers: the width of input is 256, depth 32, executing rule layer operation；
L2-3 layers: the width of input is 256, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 64, step pitch 1, zero padding 1；
L2-4 layers: the width of input is 256, depth 64；Execute activation layer operation；
L2-5 layers: the width of input is 256, depth 64, executing rule layer operation；
L2-6 layers: the width of input is 256, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 64, step pitch 1, zero padding 1；
L2-7 layers: the width of input is 256, depth 64；Execute activation layer operation；
L2-8 layers: the width of input is 256, depth 64, executing rule layer operation；
L2-9 layers: the width of input is 256, and depth 64 executes pond layer operation；
L3-0 layers: the width of input is 128, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 64, step pitch 1, zero padding 1；
L3-1 layers: the width of input is 128, depth 64；Execute activation layer operation；
L3-2 layers: the width of input is 128, depth 64, executing rule layer operation；
L3-3 layers: the width of input is 128, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 128, step pitch 1, zero padding 1；
L3-4 layers: the width of input is 128, depth 128；Execute activation layer operation；
L3-5 layers: the width of input is 128, depth 128, executing rule layer operation；
L3-6 layers: the width of input is 128, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 128, step pitch 1, zero padding 1；
L3-7 layers: the width of input is 128, depth 128；Execute activation layer operation；
L3-8 layers: the width of input is 128, depth 128, executing rule layer operation；
L3-9 layers: the width of input is 128, and depth 128 executes pond layer operation；
L4-0 layers: the width of input is 64, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 128, step pitch 1, zero padding 1；
L4-1 layers: the width of input is 64, depth 128；Execute activation layer operation；
L4-2 layers: the width of input is 64, depth 128, executing rule layer operation；
L4-3 layers: the width of input is 64, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L4-4 layers: the width of input is 64, depth 256；Execute activation layer operation；
L4-5 layers: the width of input is 64, depth 256, executing rule layer operation；
L4-6 layers: the width of input is 64, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L4-7 layers: the width of input is 64, depth 256；Execute activation layer operation；
L4-8 layers: the width of input is 64, depth 256, executing rule layer operation；
L4-9 layers: the width of input is 64, and depth 256 executes pond layer operation；
L5-0 layers: the width of input is 32, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L5-1 layers: the width of input is 32, depth 256；Execute activation layer operation；
L5-2 layers: the width of input is 32, depth 256, executing rule layer operation；
L5-3 layers: the width of input is 32, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 512, step pitch 1, zero padding 1；
L5-4 layers: the width of input is 32, depth 512；Execute activation layer operation；
L5-5 layers: the width of input is 32, depth 512, executing rule layer operation；
L5-6 layers: the width of input is 32, depth 512；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 512, step pitch 1, zero padding 1；
L5-7 layers: the width of input is 32, depth 512；Execute activation layer operation；
L5-8 layers: the width of input is 32, depth 512, executing rule layer operation；
L5-9 layers: the width of input is 32, and depth 512 executes pond layer operation；
L6-0 layers: the width of input is 16, depth 512；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 1024, step pitch 1, zero padding 1；
L6-1 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-2 layers: the width of input is 16, depth 1024, executing rule layer operation；
L6-3 layers: the width of input is 16, depth 1024；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 1024, step pitch 1, zero padding 1；
L6-4 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-5 layers: the width of input is 16, depth 512, executing rule layer operation；
L6-6 layers: the width of input is 16, depth 1024；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 1024, step pitch 1, zero padding 1；
L6-7 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-8 layers: the width of input is 16, depth 1024, executing rule layer operation；
L7-0 layers: the width of input is 32, depth 1536；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 1536, step pitch 1, zero padding 1；
L7-1 layers: the width of input is 32, depth 1536；Execute activation layer operation；
L7-2 layers: the width of input is 32, depth 1536, executing rule layer operation；
L7-3 layers: the width of input is 32, depth 1536；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 512, step pitch 1, zero padding 1；
L7-4 layers: the width of input is 32, depth 512；Execute activation layer operation；
L7-5 layers: the width of input is 32, depth 512, executing rule layer operation；
L7-6 layers: the width of input is 32, depth 3；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 512, step pitch 1, zero padding 1；
L7-7 layers: the width of input is 32, depth 512；Execute activation layer operation；
L7-8 layers: the width of input is 32, depth 512, executing rule layer operation；
L8-0 layers: the width of input is 64, depth 768；Deconvolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 768, step pitch 1, zero padding 1；
L8-1 layers: the width of input is 64, depth 768；Execute activation layer operation；
L8-2 layers: the width of input is 64, depth 768, executing rule layer operation；
L8-3 layers: the width of input is 64, depth 768；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L8-4 layers: the width of input is 64, depth 256；Execute activation layer operation；
L8-5 layers: the width of input is 64, depth 256, executing rule layer operation；
L8-6 layers: the width of input is 64, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L8-7 layers: the width of input is 64, depth 256；Execute activation layer operation；
L8-8 layers: the width of input is 64, depth 256, executing rule layer operation；
L9-0 layers: the width of input is 128, depth 384；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 384, step pitch 1, zero padding 1；
L9-1 layers: the width of input is 128, depth 384；Execute activation layer operation；
L9-2 layers: the width of input is 128, depth 384, executing rule layer operation；
L9-3 layers: the width of input is 128, depth 384；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 128, step pitch 1, zero padding 1；
L9-4 layers: the width of input is 128, depth 128；Execute activation layer operation；
L9-5 layers: the width of input is 128, depth 128, executing rule layer operation；
L9-6 layers: the width of input is 128, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 128, step pitch 1, zero padding 1；
L9-7 layers: the width of input is 128, depth 128；Execute activation layer operation；
L9-8 layers: the width of input is 128, depth 128, executing rule layer operation；
L10-0 layers: the width of input is 256, depth 192；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 192, step pitch 1, zero padding 1；
L10-1 layers: the width of input is 256, depth 192；Execute activation layer operation；
L10-2 layers: the width of input is 256, depth 192, executing rule layer operation；
L10-3 layers: the width of input is 256, depth 192；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 64, step pitch 1, zero padding 1；
L10-4 layers: the width of input is 256, depth 64；Execute activation layer operation；
L10-5 layers: the width of input is 256, depth 64, executing rule layer operation；
L10-6 layers: the width of input is 256, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 64, step pitch 1, zero padding 1；
L10-7 layers: the width of input is 256, depth 64；Execute activation layer operation；
L10-8 layers: the width of input is 256, depth 64, executing rule layer operation；
L11-0 layers: the width of input is 512, depth 96；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 96, step pitch 1, zero padding 1；
L11-1 layers: the width of input is 512, depth 96；Execute activation layer operation；
L11-2 layers: the width of input is 512, depth 96, executing rule layer operation；
L11-3 layers: the width of input is 512, depth 96；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 32, step pitch 1, zero padding 1；
L11-4 layers: the width of input is 512, depth 32；Execute activation layer operation；
L11-5 layers: the width of input is 512, depth 32, executing rule layer operation；
L11-6 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 32, step pitch 1, zero padding 1；
L11-7 layers: the width of input is 512, depth 32；Execute activation layer operation；
L11-8 layers: the width of input is 512, depth 32, executing rule layer operation；
L12-0 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 4, step pitch 1, zero padding 1；
L12-1 layers: the width of input is 512, depth 4；Execute activation layer operation；
L13 layers: the width of input is 512, depth 4；Execute loss layer operation.
3, Dice related coefficient loss function in step 3 as described above specifically:
Y in formulaiWithIt is the true value and predicted value of label respectively, M is the number of pixels of input picture, smooth=1.
4, adaptive moments estimation optimization algorithm in step 3 as described above specifically:
G in formulatFor the gradient of t step, mtFor the gradient first moment of t step, vtFor the gradient second moment of t step, β1It is one
Rank momentum attenuation coefficient, β2For second order momentum attenuation coefficient, ε is numerical stability amount, and η is learning rate, θtIndicate t step to excellent
Change parameter.
5, in step 4 as described above, images to be recognized is cut into the size cut in step 2, input trains
Neural network in, respectively identify after spliced.
Beneficial effects of the present invention and advantage are as follows: the present invention is realized for the steel case comprising complex background interference information
The overall process that beam crack precisely identifies that neural network model training, the crack identification of original image size and result are shown is automatic
Change processing.The present invention is convenient, accurate, improves the accuracy and stability of steel box-girder crack identification result.Entire positioning and knowledge
Other process is automatic processing, significantly reduces the artificial participation in detection process.The present invention is also able to satisfy steel box-girder and splits
It is sewn on line monitoring and warning and real time data processing demand, i.e., updates without training set, directly acquired image is known
Not, as a result output delay can be down to ten seconds or less.The present invention improves the automation of bridge steel box girder crack intelligent recognition, intelligence
Change, accuracy and robustness, the intelligent recognition for steel box-girder crack in science of bridge building provide solution.
Detailed description of the invention
Fig. 1 is flow diagram of the invention；
Fig. 2 is the structural schematic diagram of the full convolutional neural networks of depth of step 3 of the present invention；
Fig. 3 is the original image in the embodiment；
Fig. 4 is the mark image (i.e. true value) in the embodiment；
Fig. 5 is the intelligent recognition result figure of proposition method of the present invention.
Specific embodiment
Below according to Figure of description citing, the present invention will be further described:
Embodiment 1
As shown in Figure 1, a kind of accurate recognition methods in bridge steel structure crack based on computer vision, including walk as follows
Rapid: step 1: using common steel box-girder common site crack image as initial data, carry out the artificial mark of pixel scale,
Obtaining indicates different classes of Pixel-level mark image using different colours, specific as follows:
(1), using common steel box-girder crack disease image data base, picture size with 4000 × 3000 pixels more than；
(2), different types of mark is carried out using pixel annotation tool to original image, covering is different in different colors
Category area；
(3), original image and mark image are subjected to flip horizontal, flip vertical etc. simultaneously, or by original image into
Tri- channels row BGR apply 10% interference etc., the original image and its corresponding mark figure after respectively obtaining overturning or interference
Picture.
Step 2: original image and mark image are split according to certain size, calculate cost and adaptation to reduce
Full convolutional network input requirements, distribution different digital marks inhomogeneity to make label after mark image is extracted according to different colours
Not, and in a manner of array it is stored；
Step 3: training set being input in the full convolutional neural networks of depth and is trained, and is damaged used in training process
Mistake function is Dice related coefficient loss function, Dice related coefficient loss function specifically:
Y in formulaiWithIt is the true value and predicted value of label respectively, M is the number of pixels of input picture, and smooth is anti-
Only denominator is set, generally 1 for 0, and optimization algorithm is adaptive moments estimation optimization algorithm, and adaptive moments estimation optimization algorithm is specific
Are as follows:
G in formulatFor the gradient of t step, mtFor the gradient first moment of t step, vtFor the gradient second moment of t step, β1It is one
Rank momentum attenuation coefficient, β2For second order momentum attenuation coefficient, ε is numerical stability amount, and η is learning rate, θtIndicate t step to excellent
Change parameter；
The wherein structure of each layer of the full convolutional neural networks of depth are as follows:
L1-0 layers: the width of input is 512, depth 3；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L1-1 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-2 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-3 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L1-4 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-5 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-6 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L1-7 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-8 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-9 layers: the width of input is 512, and depth 32 executes pond layer operation；
L2-0 layers: the width of input is 256, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 32, step pitch 1, zero padding 1；
L2-1 layers: the width of input is 256, depth 32；Execute activation layer operation；
L2-2 layers: the width of input is 256, depth 32, executing rule layer operation；
L2-3 layers: the width of input is 256, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 64, step pitch 1, zero padding 1；
L2-4 layers: the width of input is 256, depth 64；Execute activation layer operation；
L2-5 layers: the width of input is 256, depth 64, executing rule layer operation；
L2-6 layers: the width of input is 256, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 64, step pitch 1, zero padding 1；
L2-7 layers: the width of input is 256, depth 64；Execute activation layer operation；
L2-8 layers: the width of input is 256, depth 64, executing rule layer operation；
L2-9 layers: the width of input is 256, and depth 64 executes pond layer operation；
L3-0 layers: the width of input is 128, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 64, step pitch 1, zero padding 1；
L3-1 layers: the width of input is 128, depth 64；Execute activation layer operation；
L3-2 layers: the width of input is 128, depth 64, executing rule layer operation；
L3-3 layers: the width of input is 128, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 128, step pitch 1, zero padding 1；
L3-4 layers: the width of input is 128, depth 128；Execute activation layer operation；
L3-5 layers: the width of input is 128, depth 128, executing rule layer operation；
L3-6 layers: the width of input is 128, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 128, step pitch 1, zero padding 1；
L3-7 layers: the width of input is 128, depth 128；Execute activation layer operation；
L3-8 layers: the width of input is 128, depth 128, executing rule layer operation；
L3-9 layers: the width of input is 128, and depth 128 executes pond layer operation；
L4-0 layers: the width of input is 64, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 128, step pitch 1, zero padding 1；
L4-1 layers: the width of input is 64, depth 128；Execute activation layer operation；
L4-2 layers: the width of input is 64, depth 128, executing rule layer operation；
L4-3 layers: the width of input is 64, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L4-4 layers: the width of input is 64, depth 256；Execute activation layer operation；
L4-5 layers: the width of input is 64, depth 256, executing rule layer operation；
L4-6 layers: the width of input is 64, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L4-7 layers: the width of input is 64, depth 256；Execute activation layer operation；
L4-8 layers: the width of input is 64, depth 256, executing rule layer operation；
L4-9 layers: the width of input is 64, and depth 256 executes pond layer operation；
L5-0 layers: the width of input is 32, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L5-1 layers: the width of input is 32, depth 256；Execute activation layer operation；
L5-2 layers: the width of input is 32, depth 256, executing rule layer operation；
L5-3 layers: the width of input is 32, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 512, step pitch 1, zero padding 1；
L5-4 layers: the width of input is 32, depth 512；Execute activation layer operation；
L5-5 layers: the width of input is 32, depth 512, executing rule layer operation；
L5-6 layers: the width of input is 32, depth 512；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 512, step pitch 1, zero padding 1；
L5-7 layers: the width of input is 32, depth 512；Execute activation layer operation；
L5-8 layers: the width of input is 32, depth 512, executing rule layer operation；
L5-9 layers: the width of input is 32, and depth 512 executes pond layer operation；
L6-0 layers: the width of input is 16, depth 512；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 1024, step pitch 1, zero padding 1；
L6-1 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-2 layers: the width of input is 16, depth 1024, executing rule layer operation；
L6-3 layers: the width of input is 16, depth 1024；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 1024, step pitch 1, zero padding 1；
L6-4 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-5 layers: the width of input is 16, depth 512, executing rule layer operation；
L6-6 layers: the width of input is 16, depth 1024；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 1024, step pitch 1, zero padding 1；
L6-7 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-8 layers: the width of input is 16, depth 1024, executing rule layer operation；
L7-0 layers: the width of input is 32, depth 1536；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 1536, step pitch 1, zero padding 1；
L7-1 layers: the width of input is 32, depth 1536；Execute activation layer operation；
L7-2 layers: the width of input is 32, depth 1536, executing rule layer operation；
L7-3 layers: the width of input is 32, depth 1536；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 512, step pitch 1, zero padding 1；
L7-4 layers: the width of input is 32, depth 512；Execute activation layer operation；
L7-5 layers: the width of input is 32, depth 512, executing rule layer operation；
L7-6 layers: the width of input is 32, depth 3；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 512, step pitch 1, zero padding 1；
L7-7 layers: the width of input is 32, depth 512；Execute activation layer operation；
L7-8 layers: the width of input is 32, depth 512, executing rule layer operation；
L8-0 layers: the width of input is 64, depth 768；Deconvolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 768, step pitch 1, zero padding 1；
L8-1 layers: the width of input is 64, depth 768；Execute activation layer operation；
L8-2 layers: the width of input is 64, depth 768, executing rule layer operation；
L8-3 layers: the width of input is 64, depth 768；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L8-4 layers: the width of input is 64, depth 256；Execute activation layer operation；
L8-5 layers: the width of input is 64, depth 256, executing rule layer operation；
L8-6 layers: the width of input is 64, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 256, step pitch 1, zero padding 1；
L8-7 layers: the width of input is 64, depth 256；Execute activation layer operation；
L8-8 layers: the width of input is 64, depth 256, executing rule layer operation；
L9-0 layers: the width of input is 128, depth 384；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 384, step pitch 1, zero padding 1；
L9-1 layers: the width of input is 128, depth 384；Execute activation layer operation；
L9-2 layers: the width of input is 128, depth 384, executing rule layer operation；
L9-3 layers: the width of input is 128, depth 384；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 128, step pitch 1, zero padding 1；
L9-4 layers: the width of input is 128, depth 128；Execute activation layer operation；
L9-5 layers: the width of input is 128, depth 128, executing rule layer operation；
L9-6 layers: the width of input is 128, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 128, step pitch 1, zero padding 1；
L9-7 layers: the width of input is 128, depth 128；Execute activation layer operation；
L9-8 layers: the width of input is 128, depth 128, executing rule layer operation；
L10-0 layers: the width of input is 256, depth 192；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 192, step pitch 1, zero padding 1；
L10-1 layers: the width of input is 256, depth 192；Execute activation layer operation；
L10-2 layers: the width of input is 256, depth 192, executing rule layer operation；
L10-3 layers: the width of input is 256, depth 192；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 64, step pitch 1, zero padding 1；
L10-4 layers: the width of input is 256, depth 64；Execute activation layer operation；
L10-5 layers: the width of input is 256, depth 64, executing rule layer operation；
L10-6 layers: the width of input is 256, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 64, step pitch 1, zero padding 1；
L10-7 layers: the width of input is 256, depth 64；Execute activation layer operation；
L10-8 layers: the width of input is 256, depth 64, executing rule layer operation；
L11-0 layers: the width of input is 512, depth 96；Deconvolution layer operation is executed, the width of convolution layer operation is
3, quantity 96, step pitch 1, zero padding 1；
L11-1 layers: the width of input is 512, depth 96；Execute activation layer operation；
L11-2 layers: the width of input is 512, depth 96, executing rule layer operation；
L11-3 layers: the width of input is 512, depth 96；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 32, step pitch 1, zero padding 1；
L11-4 layers: the width of input is 512, depth 32；Execute activation layer operation；
L11-5 layers: the width of input is 512, depth 32, executing rule layer operation；
L11-6 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 32, step pitch 1, zero padding 1；
L11-7 layers: the width of input is 512, depth 32；Execute activation layer operation；
L11-8 layers: the width of input is 512, depth 32, executing rule layer operation；
L12-0 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3,
Quantity is 4, step pitch 1, zero padding 1；
L12-1 layers: the width of input is 512, depth 4；Execute activation layer operation；
L13 layers: the width of input is 512, depth 4；Execute loss layer operation；
Step 4: after bridge steel box girder region of interest image to be identified is cut into the size cut in step 2, it is defeated
Enter into trained neural network, carries out the Pixel-level knowledge that splicing returns to original image size after obtaining Pixel-level recognition result
Other result.
Embodiment 2
A kind of Pixel-level accurate intelligent recognition methods in bridge steel box girder crack based on computer vision, as shown in Figure 1, packet
It includes:
Step 1: using common steel box-girder common site crack image as initial data, the artificial of pixel scale is carried out
Mark, obtaining indicates different classes of Pixel-level mark image using different colours, in which:
(1), the steel box-girder crack database that the present embodiment uses is APESS2018Steel Girder Crack ID
Dataset, (https: //github.com/dawnnao/APESS2018_Steel_Girder_Crack_ID_datas et).
(2), original image and mark image are subjected to flip horizontal, flip vertical etc. simultaneously, or by original image into
Tri- channels row BGR apply 10% interference etc., the original image and its corresponding mark figure after respectively obtaining overturning or interference
Picture.
Step 2: original image and mark image are split according to certain size, to reduce calculating cost, will mark
Distribution different digital is different classes of to do label label after image is extracted according to different colours；
In one embodiment, the size of segmentation be 512 × 512 pixels, to adapt to network inputs, can be used MATLAB or
Person Python carries out color extraction and distributes number；
It is trained Step 3: training set is input in the full convolutional neural networks of depth, is damaged used in training process
Mistake function is Dice related coefficient loss function, and optimization algorithm is adaptive moments estimation optimization algorithm；
It is trained using the full convolutional network of depth, wherein optimization algorithm is using the initiation parameter defaulted；
Step 4: by trained neural network is input to after bridge steel box girder region of interest image to be identified cutting
In, the Pixel-level recognition result that splicing returns to original image size is carried out after obtaining Pixel-level recognition result, in which:
(1), image to be identified is cut, in the present embodiment, cut lengths are 512 × 512 pixels；
(2), the subgraph of cutting is input in trained neural network model, obtains the classification feelings of each pixel
Condition according to classification and matching corresponding color and is drawn；
(3), subgraph recognition result is subjected to image mosaic, obtains the crack identification result of final original size.
In one embodiment, respective algorithms can be developed under Python environment, can be directly applied for general with consumer level
The bridge steel box girder of logical camera shooting is also easy to produce location of cracks image, does not need special shooting or detection device.
Claims (6)
1. a kind of accurate recognition methods in bridge steel structure crack based on computer vision, which is characterized in that method includes as follows
Step:
Step 1: using steel box-girder common site crack image as initial data, the artificial mark of pixel scale is carried out, is obtained
Indicate that different classes of Pixel-level marks image using different colours；
Step 2: original image and mark image are split according to certain size, mark image is mentioned according to different colours
It is different classes of to do label label that different digital is distributed after taking, and is stored in a manner of array；
It is trained Step 3: training set is input in the full convolutional neural networks of depth, letter is lost used in training process
Number is Dice related coefficient loss function, and optimization algorithm is adaptive moments estimation optimization algorithm；
Step 4: being obtained being input in trained neural network after bridge steel box girder region of interest image to be identified cutting
The Pixel-level recognition result that splicing returns to original image size is carried out after to Pixel-level recognition result.
2. a kind of accurate recognition methods in bridge steel structure crack based on computer vision according to claim 1, special
Sign is that step 1 specifically includes:
Step 1 one, using steel box-girder crack disease image data base, it is more than 4000 × 3000 pixel of picture size；
Step 1 two carries out different types of mark using pixel annotation tool to original image, and covering is different in different colors
Category area；
Original image and mark image are carried out flip horizontal, flip vertical simultaneously, or original image are carried out by step 1 three
Tri- channels BGR apply 10% interference, the original image and its corresponding mark image after respectively obtaining overturning or interference.
3. a kind of Pixel-level accurate intelligent identification in bridge steel box girder crack based on computer vision according to claim 1
Method, which is characterized in that the structure of each layer of the full convolutional neural networks of depth in step 3 are as follows:
L1-0 layers: the width of input is 512, depth 3；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
32, step pitch 1, zero padding 1；
L1-1 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-2 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-3 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
32, step pitch 1, zero padding 1；
L1-4 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-5 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-6 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
32, step pitch 1, zero padding 1；
L1-7 layers: the width of input is 512, depth 32；Execute activation layer operation；
L1-8 layers: the width of input is 512, depth 32, executing rule layer operation；
L1-9 layers: the width of input is 512, and depth 32 executes pond layer operation；
L2-0 layers: the width of input is 256, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
32, step pitch 1, zero padding 1；
L2-1 layers: the width of input is 256, depth 32；Execute activation layer operation；
L2-2 layers: the width of input is 256, depth 32, executing rule layer operation；
L2-3 layers: the width of input is 256, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
64, step pitch 1, zero padding 1；
L2-4 layers: the width of input is 256, depth 64；Execute activation layer operation；
L2-5 layers: the width of input is 256, depth 64, executing rule layer operation；
L2-6 layers: the width of input is 256, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
64, step pitch 1, zero padding 1；
L2-7 layers: the width of input is 256, depth 64；Execute activation layer operation；
L2-8 layers: the width of input is 256, depth 64, executing rule layer operation；
L2-9 layers: the width of input is 256, and depth 64 executes pond layer operation；
L3-0 layers: the width of input is 128, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
64, step pitch 1, zero padding 1；
L3-1 layers: the width of input is 128, depth 64；Execute activation layer operation；
L3-2 layers: the width of input is 128, depth 64, executing rule layer operation；
L3-3 layers: the width of input is 128, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
128, step pitch 1, zero padding 1；
L3-4 layers: the width of input is 128, depth 128；Execute activation layer operation；
L3-5 layers: the width of input is 128, depth 128, executing rule layer operation；
L3-6 layers: the width of input is 128, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 128, step pitch 1, zero padding 1；
L3-7 layers: the width of input is 128, depth 128；Execute activation layer operation；
L3-8 layers: the width of input is 128, depth 128, executing rule layer operation；
L3-9 layers: the width of input is 128, and depth 128 executes pond layer operation；
L4-0 layers: the width of input is 64, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
128, step pitch 1, zero padding 1；
L4-1 layers: the width of input is 64, depth 128；Execute activation layer operation；
L4-2 layers: the width of input is 64, depth 128, executing rule layer operation；
L4-3 layers: the width of input is 64, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
256, step pitch 1, zero padding 1；
L4-4 layers: the width of input is 64, depth 256；Execute activation layer operation；
L4-5 layers: the width of input is 64, depth 256, executing rule layer operation；
L4-6 layers: the width of input is 64, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
256, step pitch 1, zero padding 1；
L4-7 layers: the width of input is 64, depth 256；Execute activation layer operation；
L4-8 layers: the width of input is 64, depth 256, executing rule layer operation；
L4-9 layers: the width of input is 64, and depth 256 executes pond layer operation；
L5-0 layers: the width of input is 32, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
256, step pitch 1, zero padding 1；
L5-1 layers: the width of input is 32, depth 256；Execute activation layer operation；
L5-2 layers: the width of input is 32, depth 256, executing rule layer operation；
L5-3 layers: the width of input is 32, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
512, step pitch 1, zero padding 1；
L5-4 layers: the width of input is 32, depth 512；Execute activation layer operation；
L5-5 layers: the width of input is 32, depth 512, executing rule layer operation；
L5-6 layers: the width of input is 32, depth 512；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
512, step pitch 1, zero padding 1；
L5-7 layers: the width of input is 32, depth 512；Execute activation layer operation；
L5-8 layers: the width of input is 32, depth 512, executing rule layer operation；
L5-9 layers: the width of input is 32, and depth 512 executes pond layer operation；
L6-0 layers: the width of input is 16, depth 512；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
1024, step pitch 1, zero padding 1；
L6-1 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-2 layers: the width of input is 16, depth 1024, executing rule layer operation；
L6-3 layers: the width of input is 16, depth 1024；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 1024, step pitch 1, zero padding 1；
L6-4 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-5 layers: the width of input is 16, depth 512, executing rule layer operation；
L6-6 layers: the width of input is 16, depth 1024；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 1024, step pitch 1, zero padding 1；
L6-7 layers: the width of input is 16, depth 1024；Execute activation layer operation；
L6-8 layers: the width of input is 16, depth 1024, executing rule layer operation；
L7-0 layers: the width of input is 32, depth 1536；Deconvolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 1536, step pitch 1, zero padding 1；
L7-1 layers: the width of input is 32, depth 1536；Execute activation layer operation；
L7-2 layers: the width of input is 32, depth 1536, executing rule layer operation；
L7-3 layers: the width of input is 32, depth 1536；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 512, step pitch 1, zero padding 1；
L7-4 layers: the width of input is 32, depth 512；Execute activation layer operation；
L7-5 layers: the width of input is 32, depth 512, executing rule layer operation；
L7-6 layers: the width of input is 32, depth 3；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
512, step pitch 1, zero padding 1；
L7-7 layers: the width of input is 32, depth 512；Execute activation layer operation；
L7-8 layers: the width of input is 32, depth 512, executing rule layer operation；
L8-0 layers: the width of input is 64, depth 768；Deconvolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 768, step pitch 1, zero padding 1；
L8-1 layers: the width of input is 64, depth 768；Execute activation layer operation；
L8-2 layers: the width of input is 64, depth 768, executing rule layer operation；
L8-3 layers: the width of input is 64, depth 768；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
256, step pitch 1, zero padding 1；
L8-4 layers: the width of input is 64, depth 256；Execute activation layer operation；
L8-5 layers: the width of input is 64, depth 256, executing rule layer operation；
L8-6 layers: the width of input is 64, depth 256；Convolution layer operation is executed, the width of convolution layer operation is 3, and quantity is
256, step pitch 1, zero padding 1；
L8-7 layers: the width of input is 64, depth 256；Execute activation layer operation；
L8-8 layers: the width of input is 64, depth 256, executing rule layer operation；
L9-0 layers: the width of input is 128, depth 384；Deconvolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 384, step pitch 1, zero padding 1；
L9-1 layers: the width of input is 128, depth 384；Execute activation layer operation；
L9-2 layers: the width of input is 128, depth 384, executing rule layer operation；
L9-3 layers: the width of input is 128, depth 384；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 128, step pitch 1, zero padding 1；
L9-4 layers: the width of input is 128, depth 128；Execute activation layer operation；
L9-5 layers: the width of input is 128, depth 128, executing rule layer operation；
L9-6 layers: the width of input is 128, depth 128；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 128, step pitch 1, zero padding 1；
L9-7 layers: the width of input is 128, depth 128；Execute activation layer operation；
L9-8 layers: the width of input is 128, depth 128, executing rule layer operation；
L10-0 layers: the width of input is 256, depth 192；Deconvolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 192, step pitch 1, zero padding 1；
L10-1 layers: the width of input is 256, depth 192；Execute activation layer operation；
L10-2 layers: the width of input is 256, depth 192, executing rule layer operation；
L10-3 layers: the width of input is 256, depth 192；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 64, step pitch 1, zero padding 1；
L10-4 layers: the width of input is 256, depth 64；Execute activation layer operation；
L10-5 layers: the width of input is 256, depth 64, executing rule layer operation；
L10-6 layers: the width of input is 256, depth 64；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 64, step pitch 1, zero padding 1；
L10-7 layers: the width of input is 256, depth 64；Execute activation layer operation；
L10-8 layers: the width of input is 256, depth 64, executing rule layer operation；
L11-0 layers: the width of input is 512, depth 96；Deconvolution layer operation is executed, the width of convolution layer operation is 3, number
Amount is 96, step pitch 1, zero padding 1；
L11-1 layers: the width of input is 512, depth 96；Execute activation layer operation；
L11-2 layers: the width of input is 512, depth 96, executing rule layer operation；
L11-3 layers: the width of input is 512, depth 96；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 32, step pitch 1, zero padding 1；
L11-4 layers: the width of input is 512, depth 32；Execute activation layer operation；
L11-5 layers: the width of input is 512, depth 32, executing rule layer operation；
L11-6 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 32, step pitch 1, zero padding 1；
L11-7 layers: the width of input is 512, depth 32；Execute activation layer operation；
L11-8 layers: the width of input is 512, depth 32, executing rule layer operation；
L12-0 layers: the width of input is 512, depth 32；Convolution layer operation is executed, the width of convolution layer operation is 3, quantity
It is 4, step pitch 1, zero padding 1；
L12-1 layers: the width of input is 512, depth 4；Execute activation layer operation；
L13 layers: the width of input is 512, depth 4；Execute loss layer operation.
4. a kind of accurate recognition methods in bridge steel structure crack based on computer vision according to claim 1, special
Sign is, Dice related coefficient loss function in step 3 specifically:
Y in formulaiWithIt is the true value and predicted value of label respectively, M is the number of pixels of input picture, smooth=1.
5. a kind of accurate recognition methods in bridge steel structure crack based on computer vision according to claim 1, special
Sign is, adaptive moments estimation optimization algorithm in step 3 specifically:
G in formulatFor the gradient of t step, mtFor the gradient first moment of t step, vtFor the gradient second moment of t step, β1It is dynamic for single order
Measure attenuation coefficient, β2For second order momentum attenuation coefficient, ε is numerical stability amount, and η is learning rate, θtIndicate the ginseng to be optimized of t step
Number.
6. a kind of accurate recognition methods in bridge steel structure crack based on computer vision according to claim 1, special
Sign is, in step 4, images to be recognized is cut into the size cut in step 2, is inputted in trained neural network,
Spliced after identifying respectively.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811295283.1A CN109408985A (en) | 2018-11-01 | 2018-11-01 | The accurate recognition methods in bridge steel structure crack based on computer vision |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811295283.1A CN109408985A (en) | 2018-11-01 | 2018-11-01 | The accurate recognition methods in bridge steel structure crack based on computer vision |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109408985A true CN109408985A (en) | 2019-03-01 |
Family
ID=65471256
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811295283.1A Pending CN109408985A (en) | 2018-11-01 | 2018-11-01 | The accurate recognition methods in bridge steel structure crack based on computer vision |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109408985A (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222701A (en) * | 2019-06-11 | 2019-09-10 | 北京新桥技术发展有限公司 | A kind of bridge defect automatic identifying method |
| CN110502972A (en) * | 2019-07-05 | 2019-11-26 | 广东工业大学 | A kind of pavement crack segmentation and recognition methods based on deep learning |
| CN110533629A (en) * | 2019-07-10 | 2019-12-03 | 湖南交工智能技术有限公司 | A kind of detection method and detection device of Bridge Crack |
| CN111507990A (en) * | 2020-04-20 | 2020-08-07 | 南京航空航天大学 | Tunnel surface defect segmentation method based on deep learning |
| CN111737909A (en) * | 2020-06-10 | 2020-10-02 | 哈尔滨工业大学 | Structural health monitoring data anomaly identification method based on space-time graph convolutional network |
| CN111795978A (en) * | 2020-09-08 | 2020-10-20 | 湖南大学 | Steel bridge structure health state assessment method, device, equipment and storage medium |
| 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 |
| CN113406088A (en) * | 2021-05-10 | 2021-09-17 | 同济大学 | Fixed point type steel box girder crack development observation device |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR100826153B1 (en) * | 2006-11-29 | 2008-04-30 | 한국표준과학연구원 | Width measurement method of the crack by using the depth value in histogram of image |
| CN106018411A (en) * | 2016-05-09 | 2016-10-12 | 广州市九州旗建筑科技有限公司 | Crack width measuring and computing method and measuring device |
| CN106910187A (en) * | 2017-01-13 | 2017-06-30 | 陕西师范大学 | A kind of artificial amplification method of image data set for Bridge Crack detection |
| CN108346144A (en) * | 2018-01-30 | 2018-07-31 | 哈尔滨工业大学 | Bridge Crack based on computer vision monitoring and recognition methods automatically |
-
2018
- 2018-11-01 CN CN201811295283.1A patent/CN109408985A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR100826153B1 (en) * | 2006-11-29 | 2008-04-30 | 한국표준과학연구원 | Width measurement method of the crack by using the depth value in histogram of image |
| CN106018411A (en) * | 2016-05-09 | 2016-10-12 | 广州市九州旗建筑科技有限公司 | Crack width measuring and computing method and measuring device |
| CN106910187A (en) * | 2017-01-13 | 2017-06-30 | 陕西师范大学 | A kind of artificial amplification method of image data set for Bridge Crack detection |
| CN108346144A (en) * | 2018-01-30 | 2018-07-31 | 哈尔滨工业大学 | Bridge Crack based on computer vision monitoring and recognition methods automatically |
Non-Patent Citations (3)
| Title |
|---|
| XINCONG YANG ET AL: ""Automatic Pixel-Level Crack Detection and Measurement Using Fully Convolutional Network"", 《HTTPS://DOI.ORG/10.1111/MICE.12412》 * |
| YANG XU ET AL: ""Identification framework for cracks on a steel structure surface by a restricted Boltzmann machines algorithm based on consumer‐grade camera images"", 《WILEY》 * |
| 黄宏伟 等: ""基于深度学习的盾构隧道渗漏水病害图像识别"", 《岩石力学与工程学报》 * |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110222701A (en) * | 2019-06-11 | 2019-09-10 | 北京新桥技术发展有限公司 | A kind of bridge defect automatic identifying method |
| CN110222701B (en) * | 2019-06-11 | 2019-12-27 | 北京新桥技术发展有限公司 | Automatic bridge disease identification method |
| CN110502972A (en) * | 2019-07-05 | 2019-11-26 | 广东工业大学 | A kind of pavement crack segmentation and recognition methods based on deep learning |
| CN110533629A (en) * | 2019-07-10 | 2019-12-03 | 湖南交工智能技术有限公司 | A kind of detection method and detection device of Bridge Crack |
| CN111507990A (en) * | 2020-04-20 | 2020-08-07 | 南京航空航天大学 | Tunnel surface defect segmentation method based on deep learning |
| CN111507990B (en) * | 2020-04-20 | 2022-02-11 | 南京航空航天大学 | Tunnel surface defect segmentation method based on deep learning |
| CN111737909A (en) * | 2020-06-10 | 2020-10-02 | 哈尔滨工业大学 | Structural health monitoring data anomaly identification method based on space-time graph convolutional network |
| CN111737909B (en) * | 2020-06-10 | 2021-02-09 | 哈尔滨工业大学 | Structural health monitoring data anomaly identification method based on space-time graph convolutional network |
| CN111795978A (en) * | 2020-09-08 | 2020-10-20 | 湖南大学 | Steel bridge structure health state assessment method, device, equipment and storage medium |
| CN111795978B (en) * | 2020-09-08 | 2020-12-04 | 湖南大学 | Steel bridge structure health state assessment method, device, equipment and storage medium |
| 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 |
| CN113406088A (en) * | 2021-05-10 | 2021-09-17 | 同济大学 | Fixed point type steel box girder crack development observation device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109408985A (en) | The accurate recognition methods in bridge steel structure crack based on computer vision | |
| CN107423698B (en) | A kind of gesture estimation method based on convolutional neural networks in parallel | |
| CN104573731B (en) | Fast target detection method based on convolutional neural networks | |
| CN110246181B (en) | Anchor point-based attitude estimation model training method, attitude estimation method and system | |
| CN108520516A (en) | A kind of bridge pavement Crack Detection and dividing method based on semantic segmentation | |
| CN108805070A (en) | A kind of deep learning pedestrian detection method based on built-in terminal | |
| CN110348376A (en) | A kind of pedestrian's real-time detection method neural network based | |
| CN106408015A (en) | Road fork identification and depth estimation method based on convolutional neural network | |
| CN107038416A (en) | A kind of pedestrian detection method based on bianry image modified HOG features | |
| CN110232379A (en) | A kind of vehicle attitude detection method and system | |
| CN108288269A (en) | Bridge pad disease automatic identifying method based on unmanned plane and convolutional neural networks | |
| CN106127205A (en) | A kind of recognition methods of the digital instrument image being applicable to indoor track machine people | |
| CN103020614B (en) | Based on the human motion identification method that space-time interest points detects | |
| CN111368637B (en) | Transfer robot target identification method based on multi-mask convolutional neural network | |
| CN109376676A (en) | Highway engineering site operation personnel safety method for early warning based on unmanned aerial vehicle platform | |
| CN109543676A (en) | A kind of print-wheel type water meter Number character recognition method based on image procossing | |
| CN107392251A (en) | A kind of method that target detection network performance is lifted using category images | |
| CN102880870B (en) | The extracting method of face characteristic and system | |
| CN105469111A (en) | Small sample set object classification method on basis of improved MFA and transfer learning | |
| CN114140665A (en) | Dense small target detection method based on improved YOLOv5 | |
| CN113762009A (en) | Crowd counting method based on multi-scale feature fusion and double-attention machine mechanism | |
| CN105405138A (en) | Water surface target tracking method based on saliency detection | |
| CN110334719A (en) | The method and system of object image are built in a kind of extraction remote sensing image | |
| CN109635726A (en) | A kind of landslide identification method based on the symmetrical multiple dimensioned pond of depth network integration | |
| CN106909881A (en) | The method and system of corn breeding base ridge number are extracted based on unmanned aerial vehicle remote sensing images |
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 |
Application publication date: 20190301 |
|
| RJ01 | Rejection of invention patent application after publication |