CN110147772A - A kind of underwater dam surface crack recognition methods based on transfer learning - Google Patents
A kind of underwater dam surface crack recognition methods based on transfer learning Download PDFInfo
- Publication number
- CN110147772A CN110147772A CN201910434440.0A CN201910434440A CN110147772A CN 110147772 A CN110147772 A CN 110147772A CN 201910434440 A CN201910434440 A CN 201910434440A CN 110147772 A CN110147772 A CN 110147772A
- Authority
- CN
- China
- Prior art keywords
- surface crack
- dam surface
- model
- underwater
- alexnet
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
Abstract
The invention discloses a kind of underwater dam surface crack recognition methods based on transfer learning in underwater picture target identification technology field, aiming to solve the problem that in the prior art will be in the Crack Detection of the approach application of deep learning to underwater complex environment, since underwater sample data is difficult to largely obtain, and then the technical issues of influence recognition accuracy.Described method includes following steps: being based on dam surface crack waterborne image and underwater dam surface crack image, constructs mixing sample collection;Using mixing sample collection training depth convolutional neural networks model, pre-training network model is obtained;Using depth convolutional neural networks model and pre-training network model, target network model is obtained;Underwater dam surface crack image is inputted into target network model, according to the label classification that target network model exports, identification dam surface damages degree.
Description
Technical field
The underwater dam surface crack recognition methods based on transfer learning that the present invention relates to a kind of, belongs to underwater picture target
Identification technology field.
Background technique
The detection of underwater dam surface crack and identification are of great significance for maintenance dam safety.In recent years, based on view
Feel the advantages such as intuitive and convenient that the method for image procossing has by it, is commonly used for underwater works health monitoring.But due to water
Lower environment is complicated, and it is serious etc. that the image for causing underwater camera to shoot usually there will be lower contrast, edge blurry, noise jamming
Problem.To solve these problems, the detection of underwater dam surface crack and identification are solved by deep learning, are increasingly ground
The favor for the person of studying carefully.However, can often encounter sample in the Crack Detection based on the approach application of deep learning to underwater complex environment
Data are difficult to the largely problems such as acquisition, and then influence recognition accuracy.
Summary of the invention
It is an object of the invention to overcome deficiency in the prior art, a kind of underwater dam based on transfer learning is provided
Surface crack recognition methods, includes the following steps:
Based on dam surface crack waterborne image and underwater dam surface crack image, mixing sample collection is constructed;
Using mixing sample collection training depth convolutional neural networks model, pre-training network model is obtained;
Using depth convolutional neural networks model and pre-training network model, target network model is obtained;
Underwater dam surface crack image is inputted into target network model, the tag class exported according to target network model
Not, identification dam surface damages degree.
Further, it is based on dam surface crack waterborne image and underwater dam surface crack image, constructs mixing sample
Collection, comprising:
Dam surface crack waterborne image is acquired, dam surface crack waterborne sample set is constructed;
Underwater dam surface crack image is acquired, underwater dam surface crack sample set is constructed;
Dam surface crack waterborne sample set is extended in underwater dam surface crack sample set, based on the water after expansion
Lower dam surface crack sample set is as mixing sample collection.
Further, dam surface crack waterborne sample set is constructed, comprising: to collected dam surface crack pattern waterborne
As being pre-processed;
Construct underwater dam surface crack sample set, comprising: carry out to collected underwater dam surface crack image pre-
Processing;
Pretreatment includes: adjustment image size or/and resolution ratio.
Further, underwater dam surface crack image is acquired, underwater dam surface crack sample set is constructed, comprising:
Maintenance data extended technology expands underwater dam surface crack amount of images;
Based on the underwater dam surface crack image after expansion, underwater dam surface crack sample set is constructed;
The data extending technology includes: at least any in random cropping, rotation transformation, perspective transform and change of scale
?.
Further, the label classification include: in free from flaw, slight crack, moderate crack and through crack at least
Any one.
Further, the depth convolutional neural networks model is AlexNet model, and AlexNet model includes convolution mould
Block, pond module and at least one full link block;
The pre-training network model is AlexNet-Damcracks model, and AlexNet-Damcracks model includes volume
Volume module and pond module.
Further, using depth convolutional neural networks model and pre-training network model, target network model, packet are obtained
It includes:
The parameter for freezing convolution module and pond module in AlexNet-Damcracks model, as in AlexNet model
The parameter of convolution module and pond module;
The positive integer no more than 4 is set by the output unit number of one in AlexNet model full link block;
Using mixing sample collection training AlexNet model, pass through the iteration of AlexNet model loss function, Lai Youhua
The parameter of full link block in AlexNet model;
AlexNet model after parameter optimization based on full link block, as target network model.
Further, output unit number is set as being not more than the full link block of 4 positive integer in AlexNet model,
Output unit respectively corresponds the label classification of dam surface damage degree.
Further, using mixing sample collection training AlexNet model, by the iteration of AlexNet model loss function,
To optimize the parameter of full link block in AlexNet model, comprising:
Using the parameter of full link block in random device initialization AlexNet model;
Using the parameter of full link block in stochastic gradient descent method optimization AlexNet model, in stochastic gradient descent method
Learning rate more new strategy use exponential attenuation method;
Training obtains the parameter of full link block when the reduction of loss function value and recognition accuracy increase;
Reduced based on loss function value and when recognition accuracy increases the parameter of full link block AlexNet model, press
Default the number of iterations repetition training process.
Further, learning rate includes following operational formula:
L=l0*γ[α/β],
In formula, l is the learning rate after decaying, l0For initial learning rate, γ is attenuation coefficient, and α is current the number of iterations,
β is attenuation steps, and [] indicates to be rounded downwards;
Loss function includes following operational formula:
In formula, J is training loss, and θ is weight parameter, and p is desired class probability, and x is batch training sample, and q is pre-
The class probability of survey, λ are regularization coefficient, | | θ | |1Indicate the sum of the absolute value of each element in weight parameter θ.
Compared with prior art, advantageous effects of the invention: being instructed using the mixing sample collection for expanding building
Practice, efficiently solves depth convolutional neural networks and be used in underwater dam surface crack image recognition processes, due to sample set mistake
Small the problem of causing recognition accuracy too low or even recognition failures, generalization ability and the identification for substantially increasing identification are accurate
Rate.
Detailed description of the invention
Fig. 1 is the method for the present invention flow chart;
Fig. 2 is AlexNet Construction of A Model figure of the present invention.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings.Following embodiment is only used for clearly illustrating the present invention
Technical solution, and not intended to limit the protection scope of the present invention.
As shown in Figure 1, being the method for the present invention flow chart, described method includes following steps:
Step 1 acquires the dam surface crack waterborne image of larger samples quantity, and carries out to acquired image pre-
Processing, the pretreatment include the size and resolution ratio of adjustment image.
Step 2 acquires the underwater dam surface crack image of small sample quantity, equally carries out to acquired image pre-
Processing, i.e. adjustment image size and resolution.
The data extendings technology such as random cropping, rotation transformation, perspective transform, change of scale is applied to obtain by step 3
Underwater dam surface crack image in, with achieve the purpose that expand underwater dam surface crack image.
Step 4, the underwater dam surface crack image and collected dam surface crack pattern waterborne obtained using expansion
As constituting mixing sample collection, using the mixing sample collection training AlexNet depth convolutional neural networks model, AlexNet- is obtained
Damcracks model.
Step 5 is still used under water in the building of dam surface crack identification model using the thought of transfer learning
AlexNet model structure is made of 5 convolution modules and 3 full link blocks, as shown in Fig. 2, being of the present invention
AlexNet Construction of A Model figure;
The parameter for freezing the convolution module and pond module in AlexNet-Damcracks model, as submarine bar body surface
Facial cleft stitches the convolution module of identification model and the parameter of pond module, and the output unit number of 3 full link blocks is respectively set
It is 4098,4098 and 4;Consider to convert four classification problems for crack identification problem, it may be assumed that 4 of the last one full articulamentum are defeated
Unit respectively corresponds four label classifications of dam surface damage degree out, and four label classifications include: free from flaw, slight
Crack, moderate crack and through crack;Then, the ginseng of full articulamentum in model is continued to optimize by the iteration of loss function
Number, obtains object module AlexNet-Damcracks-Migration.
Wherein the parameter of full articulamentum comprises the following specific steps that in Optimized model:
The parameter that full articulamentum is initialized by random device is optimized using stochastic gradient descent method, and learning rate is more
New strategy uses exponential attenuation method, and following formula is shown in the learning rate update of exponential attenuation method:
L=l0*γ[α/β],
In formula, l is the learning rate after decaying, l0For initial learning rate, γ is attenuation coefficient, and α is current the number of iterations,
β is attenuation steps, and [] indicates to be rounded downwards.
Loss function is indicated using cross entropy, and additional regularization coefficient punishes weight parameter;The number of iterations
For n (n=500k, k are positive integer), loss function is shown in following formula:
In formula, J is training loss, and θ is weight parameter, and p is desired class probability, and x is batch training sample, and q is pre-
The class probability of survey, λ are regularization coefficient, | | θ | |1Indicate the sum of the absolute value of each element in weight parameter θ, wherein classification
Probability is calculated by Softmax layers.
Then it constantly trains to find and makes loss function value and recognition accuracy while optimal full connection layer parameter situation, so
Many experiments substitute into different value of K afterwards, and the optimized parameter in the case of best the number of iterations is underwater dam surface crack identification mould
The parameter of full articulamentum in type.
Step 6 utilizes the Optimal Parameters dam surface crack identification model under water of full articulamentum obtained in step 5
On the basis of constantly training obtain object module AlexNet-Damcracks-Migration, by underwater dam surface crack figure
As test set input object module AlexNet-Damcracks-Migration, accordingly exported, thus realize by
To object module AlexNet-Damcracks-Migration move to it is final in underwater dam surface crack image recognition
Task.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, without departing from the technical principles of the invention, several improvement and deformations can also be made, these improvement and deformations
Also it should be regarded as protection scope of the present invention.
Claims (10)
1. a kind of underwater dam surface crack recognition methods based on transfer learning, characterized in that include the following steps:
Based on dam surface crack waterborne image and underwater dam surface crack image, mixing sample collection is constructed;
Using mixing sample collection training depth convolutional neural networks model, pre-training network model is obtained;
Using depth convolutional neural networks model and pre-training network model, target network model is obtained;
Underwater dam surface crack image is inputted into target network model, according to the label classification that target network model exports, is known
Other dam surface damages degree.
2. the underwater dam surface crack recognition methods according to claim 1 based on transfer learning, characterized in that be based on
Dam surface crack waterborne image and underwater dam surface crack image construct mixing sample collection, comprising:
Dam surface crack waterborne image is acquired, dam surface crack waterborne sample set is constructed;
Underwater dam surface crack image is acquired, underwater dam surface crack sample set is constructed;
Dam surface crack waterborne sample set is extended in underwater dam surface crack sample set, based on the submarine bar after expansion
Body surface crack sample set is as mixing sample collection.
3. the underwater dam surface crack recognition methods according to claim 2 based on transfer learning, characterized in that
Construct dam surface crack waterborne sample set, comprising: pre-process to collected dam surface crack waterborne image;
Construct underwater dam surface crack sample set, comprising: pre-process to collected underwater dam surface crack image;
Pretreatment includes: adjustment image size or/and resolution ratio.
4. the underwater dam surface crack recognition methods according to claim 2 based on transfer learning, characterized in that acquisition
Underwater dam surface crack image constructs underwater dam surface crack sample set, comprising:
Maintenance data extended technology expands underwater dam surface crack amount of images;
Based on the underwater dam surface crack image after expansion, underwater dam surface crack sample set is constructed;
The data extending technology includes: at least any one in random cropping, rotation transformation, perspective transform and change of scale.
5. the underwater dam surface crack recognition methods according to claim 1 based on transfer learning, characterized in that described
Label classification includes: at least any one in free from flaw, slight crack, moderate crack and through crack.
6. the underwater dam surface crack recognition methods according to any one of claim 1 to 5 based on transfer learning,
It is characterized in, the depth convolutional neural networks model is AlexNet model, and AlexNet model includes convolution module, pond module
With at least one full link block;
The pre-training network model is AlexNet-Damcracks model, and AlexNet-Damcracks model includes convolution mould
Block and pond module.
7. the underwater dam surface crack recognition methods according to claim 6 based on transfer learning, characterized in that utilize
Depth convolutional neural networks model and pre-training network model obtain target network model, comprising:
The parameter for freezing convolution module and pond module in AlexNet-Damcracks model, as convolution in AlexNet model
The parameter of module and pond module;
The positive integer no more than 4 is set by the output unit number of one in AlexNet model full link block;
Using mixing sample collection training AlexNet model, pass through the iteration of AlexNet model loss function, Lai Youhua AlexNet
The parameter of full link block in model;
AlexNet model after parameter optimization based on full link block, as target network model.
8. the underwater dam surface crack recognition methods according to claim 7 based on transfer learning, characterized in that
Output unit number is set as the full link block of the positive integer no more than 4 in AlexNet model, and output unit respectively corresponds dam
The label classification of body surface face damage degree.
9. the underwater dam surface crack recognition methods according to claim 7 based on transfer learning, characterized in that utilize
Mixing sample collection training AlexNet model is complete in Lai Youhua AlexNet model by the iteration of AlexNet model loss function
The parameter of link block, comprising:
Using the parameter of full link block in random device initialization AlexNet model;
Optimize the parameter of full link block in AlexNet model, in stochastic gradient descent method using stochastic gradient descent method
Habit rate more new strategy uses exponential attenuation method;
Training obtains the parameter of full link block when the reduction of loss function value and recognition accuracy increase;
Reduced based on loss function value and when recognition accuracy increases the parameter of full link block AlexNet model, by default
The number of iterations repetition training process.
10. the underwater dam surface crack recognition methods according to claim 9 based on transfer learning, characterized in that
Learning rate includes following operational formula:
L=l0*γ[α/β],
In formula, l is the learning rate after decaying, l0For initial learning rate, γ is attenuation coefficient, and α is current the number of iterations, and β is to decline
Subtract step-length, [] indicates to be rounded downwards;
Loss function includes following operational formula:
In formula, J is training loss, and θ is weight parameter, and p is desired class probability, and x is batch training sample, and q is prediction
Class probability, λ are regularization coefficient, | | θ | |1Indicate the sum of the absolute value of each element in weight parameter θ.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910434440.0A CN110147772B (en) | 2019-05-23 | 2019-05-23 | Migration learning-based underwater dam body surface crack identification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910434440.0A CN110147772B (en) | 2019-05-23 | 2019-05-23 | Migration learning-based underwater dam body surface crack identification method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110147772A true CN110147772A (en) | 2019-08-20 |
CN110147772B CN110147772B (en) | 2022-09-20 |
Family
ID=67592912
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910434440.0A Active CN110147772B (en) | 2019-05-23 | 2019-05-23 | Migration learning-based underwater dam body surface crack identification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110147772B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111257341A (en) * | 2020-03-30 | 2020-06-09 | 河海大学常州校区 | Underwater building crack detection method based on multi-scale features and stacked full convolution network |
CN111691358A (en) * | 2020-06-18 | 2020-09-22 | 河海大学 | Gravity dam apparent crack risk prediction method and system |
CN112668631A (en) * | 2020-12-24 | 2021-04-16 | 哈尔滨理工大学 | Mobile terminal community pet identification method based on convolutional neural network |
CN112926669A (en) * | 2021-03-09 | 2021-06-08 | 杭州电子科技大学 | Tunnel crack rapid detection method based on feature enhancement |
CN113076959A (en) * | 2021-04-27 | 2021-07-06 | 中国矿业大学 | Concrete structure surface microcrack feature extraction method based on convolutional neural network |
CN113358582A (en) * | 2021-06-04 | 2021-09-07 | 山东国瑞新能源有限公司 | Method, equipment and medium for detecting concrete structure defects |
CN115953672A (en) * | 2023-03-13 | 2023-04-11 | 南昌工程学院 | Method for identifying surface cracks of underwater dam |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109345507A (en) * | 2018-08-24 | 2019-02-15 | 河海大学 | A kind of dam image crack detection method based on transfer learning |
CN109508650A (en) * | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
-
2019
- 2019-05-23 CN CN201910434440.0A patent/CN110147772B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109345507A (en) * | 2018-08-24 | 2019-02-15 | 河海大学 | A kind of dam image crack detection method based on transfer learning |
CN109508650A (en) * | 2018-10-23 | 2019-03-22 | 浙江农林大学 | A kind of wood recognition method based on transfer learning |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111257341A (en) * | 2020-03-30 | 2020-06-09 | 河海大学常州校区 | Underwater building crack detection method based on multi-scale features and stacked full convolution network |
CN111691358A (en) * | 2020-06-18 | 2020-09-22 | 河海大学 | Gravity dam apparent crack risk prediction method and system |
CN112668631A (en) * | 2020-12-24 | 2021-04-16 | 哈尔滨理工大学 | Mobile terminal community pet identification method based on convolutional neural network |
CN112668631B (en) * | 2020-12-24 | 2022-06-24 | 哈尔滨理工大学 | Mobile terminal community pet identification method based on convolutional neural network |
CN112926669A (en) * | 2021-03-09 | 2021-06-08 | 杭州电子科技大学 | Tunnel crack rapid detection method based on feature enhancement |
CN113076959A (en) * | 2021-04-27 | 2021-07-06 | 中国矿业大学 | Concrete structure surface microcrack feature extraction method based on convolutional neural network |
CN113358582A (en) * | 2021-06-04 | 2021-09-07 | 山东国瑞新能源有限公司 | Method, equipment and medium for detecting concrete structure defects |
CN115953672A (en) * | 2023-03-13 | 2023-04-11 | 南昌工程学院 | Method for identifying surface cracks of underwater dam |
CN115953672B (en) * | 2023-03-13 | 2024-02-27 | 南昌工程学院 | Method for identifying surface cracks of underwater dam |
Also Published As
Publication number | Publication date |
---|---|
CN110147772B (en) | 2022-09-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110147772A (en) | A kind of underwater dam surface crack recognition methods based on transfer learning | |
CN113178255B (en) | GAN-based medical diagnosis model attack resistance method | |
CN106919909B (en) | The metric learning method and system that a kind of pedestrian identifies again | |
CN109241913A (en) | In conjunction with the ship detection method and system of conspicuousness detection and deep learning | |
CN107563411B (en) | Online SAR target detection method based on deep learning | |
CN109165660A (en) | A kind of obvious object detection method based on convolutional neural networks | |
CN111814875A (en) | Method for expanding ship samples in infrared image based on pattern generation countermeasure network | |
CN111368671A (en) | SAR image ship target detection and identification integrated method based on deep learning | |
CN100395777C (en) | Multisensor image fusion method based on optimized small wave filter set | |
CN106570485B (en) | A kind of raft culture remote sensing images scene mask method based on deep learning | |
CN109523013A (en) | A kind of air particle pollution level estimation method based on shallow-layer convolutional neural networks | |
CN105931264B (en) | A kind of sea infrared small target detection method | |
CN110427836A (en) | A kind of high-resolution remote sensing image Clean water withdraw method based on multi_dimension optimization | |
CN107730479A (en) | High dynamic range images based on compressed sensing go artifact fusion method | |
CN110991257B (en) | Polarized SAR oil spill detection method based on feature fusion and SVM | |
CN111611889A (en) | Miniature insect pest recognition device in farmland based on improved convolutional neural network | |
CN115841629A (en) | SAR image ship detection method based on convolutional neural network | |
CN113780242A (en) | Cross-scene underwater sound target classification method based on model transfer learning | |
CN114187522A (en) | Detection method for dangerous case of yellow river basin dam bank based on DETR model | |
CN116485885A (en) | Method for removing dynamic feature points at front end of visual SLAM based on deep learning | |
CN114818826A (en) | Fault diagnosis method based on lightweight Vision Transformer module | |
CN113205016B (en) | River and lake shoreline change detection method of constant residual error type Unet and remote sensing water body index | |
CN112818777B (en) | Remote sensing image target detection method based on dense connection and feature enhancement | |
CN113869157A (en) | Cloud classification method based on visible light and infrared cloud pictures | |
CN112418229A (en) | Unmanned ship marine scene image real-time segmentation method based on deep learning |
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 | ||
CB03 | Change of inventor or designer information | ||
CB03 | Change of inventor or designer information |
Inventor after: Shi Pengfei Inventor after: Xin Yuanxue Inventor after: Fan Xinnan Inventor after: Zhang Xuewu Inventor after: Yu Zihan Inventor before: Yu Zihan Inventor before: Shi Pengfei Inventor before: Xin Yuanxue Inventor before: Fan Xinnan Inventor before: Zhang Xuewu |
|
GR01 | Patent grant | ||
GR01 | Patent grant |