CN110147772B - Migration learning-based underwater dam body surface crack identification method - Google Patents
Migration learning-based underwater dam body surface crack identification method Download PDFInfo
- Publication number
- CN110147772B CN110147772B CN201910434440.0A CN201910434440A CN110147772B CN 110147772 B CN110147772 B CN 110147772B CN 201910434440 A CN201910434440 A CN 201910434440A CN 110147772 B CN110147772 B CN 110147772B
- Authority
- CN
- China
- Prior art keywords
- underwater
- model
- alexnet
- network model
- dam body
- 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.)
- Active
Links
Images
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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an underwater dam body surface crack identification method based on transfer learning, and belongs to the technical field of underwater image target identification, and aims to solve the technical problem that in the prior art, a deep learning method is applied to crack detection in an underwater complex environment, and the identification accuracy is influenced because a large amount of underwater sample data is difficult to obtain. The method comprises the following steps: constructing a mixed sample set based on the surface crack image of the overwater dam and the surface crack image of the underwater dam; training the deep convolutional neural network model by using the mixed sample set to obtain a pre-training network model; acquiring a target network model by utilizing a deep convolutional neural network model and a pre-training network model; and inputting the underwater dam surface crack image into a target network model, and identifying the dam surface damage degree according to the label type output by the target network model.
Description
Technical Field
The invention relates to an underwater dam body surface crack identification method based on transfer learning, and belongs to the technical field of underwater image target identification.
Background
The detection and identification of the surface cracks of the underwater dam body are of great significance to the maintenance of the safety of the dam. In recent years, a method based on visual image processing is often used for health monitoring of underwater structures due to the advantages of intuition, convenience and the like. However, due to the complex underwater environment, the images shot by the underwater camera usually have the problems of low contrast, blurred edges, serious noise interference and the like. In order to solve the problems, the detection and identification of the surface crack of the underwater dam body through deep learning are favored by researchers. However, when the method based on deep learning is applied to crack detection in an underwater complex environment, the difficulty that a large amount of sample data is difficult to obtain is often encountered, and the identification accuracy is further influenced.
Disclosure of Invention
The invention aims to overcome the defects in the prior art and provides an underwater dam body surface crack identification method based on migration learning, which comprises the following steps:
constructing a mixed sample set based on the surface crack image of the overwater dam and the surface crack image of the underwater dam;
training the deep convolutional neural network model by using the mixed sample set to obtain a pre-training network model;
acquiring a target network model by utilizing a deep convolutional neural network model and a pre-training network model;
and inputting the underwater dam surface crack image into a target network model, and identifying the dam surface damage degree according to the label type output by the target network model.
Further, based on the surface crack image of the water dam and the surface crack image of the underwater dam, a mixed sample set is constructed, and the method comprises the following steps:
acquiring a surface crack image of the water dam body, and constructing a surface crack sample set of the water dam body;
acquiring an image of a crack on the surface of the underwater dam body, and constructing a crack sample set on the surface of the underwater dam body;
and expanding the sample set of the cracks on the surface of the water dam body to the sample set of the cracks on the surface of the underwater dam body, and taking the sample set of the cracks on the surface of the underwater dam body after expansion as a mixed sample set.
Further, constructing a surface crack sample set of the water dam body comprises the following steps: preprocessing the collected water dam body surface crack image;
constructing a surface crack sample set of an underwater dam body, comprising the following steps of: preprocessing the acquired crack image on the surface of the underwater dam body;
the pretreatment comprises the following steps: the image size or/and resolution is adjusted.
Further, collecting the surface crack image of the underwater dam body, and constructing a surface crack sample set of the underwater dam body, wherein the method comprises the following steps:
expanding the number of the surface crack images of the underwater dam body by using a data expansion technology;
constructing an underwater dam body surface crack sample set based on the expanded underwater dam body surface crack image;
the data augmentation techniques include: at least any one of random cropping, rotational transformation, perspective transformation, and scale transformation.
Further, the tag categories include: at least any one of no crack, a light crack, a medium crack, and a through crack.
Further, the deep convolutional neural network model is an AlexNet model, and the AlexNet model comprises a convolutional module, a pooling module and at least one full-connection module;
the pre-training network model is an AlexNet-Damcrcks model, and the AlexNet-Damcrcks model comprises a convolution module and a pooling module.
Further, acquiring a target network model by using the deep convolutional neural network model and the pre-training network model, wherein the acquiring comprises:
freezing parameters of a convolution module and a pooling module in the AlexNet-Damcracks model to serve as the parameters of the convolution module and the pooling module in the AlexNet model;
setting the number of output units of a full-connection module in the AlexNet model as a positive integer not greater than 4;
training an AlexNet model by using a mixed sample set, and optimizing parameters of a full-connection module in the AlexNet model through iteration of a loss function of the AlexNet model;
and taking the AlexNet model after parameter optimization based on the full-connection module as a target network model.
Furthermore, the number of output units in the AlexNet model is set to be not more than 4, and the output units of the AlexNet model correspond to the label types of dam surface damage degrees respectively.
Further, training an AlexNet model by using a mixed sample set, and optimizing parameters of a full-connection module in the AlexNet model through iteration of a AlexNet model loss function, wherein the parameters comprise:
initializing parameters of a full-connection module in an AlexNet model by adopting a random method;
optimizing parameters of a full-connection module in the AlexNet model by adopting a random gradient descent method, wherein an exponential decay method is adopted as a learning rate updating strategy in the random gradient descent method;
training parameters of the full-connection module when the obtained loss function value is reduced and the recognition accuracy is increased;
and repeating the training process according to the preset iteration times based on the AlexNet model of the parameters of the full-connection module when the loss function value is reduced and the identification accuracy is increased.
Further, the learning rate includes the following operational formula:
l=l 0 *γ [α/β] ,
wherein l is the learning rate after attenuation, l 0 Is an initial learning rate, gamma is an attenuation coefficient, alpha is the current iteration number, beta is an attenuation step length, [ alpha ], [ beta ], []Represents rounding down;
the loss function includes the following operational formula:
where J is the training loss, θ is the weight parameter, p is the expected class probability, x is the batch of training samples, q is the predicted class probability, λ is the regularization coefficient, | θ | | Y 1 Representing the sum of the absolute values of the various elements in the weight parameter theta.
Compared with the prior art, the invention has the following beneficial effects: the extended and constructed mixed sample set is used for training, the problems that in the process that the deep convolutional neural network is used for identifying the crack image on the surface of the underwater dam body, the identification accuracy is too low and even the identification fails due to the fact that the sample set is too small are effectively solved, and the generalization capability and the identification accuracy of identification are greatly improved.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a structural diagram of the AlexNet model of the invention.
Detailed Description
The invention is further described below with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
As shown in fig. 1, is a flow chart of the method of the present invention, which comprises the following steps:
acquiring surface crack images of the water dam body with a large number of samples, and preprocessing the acquired images, wherein the preprocessing comprises the adjustment of the size and the resolution of the images.
And step two, collecting the crack images of the surface of the underwater dam body with small sample number, and preprocessing the collected images, namely adjusting the size and the resolution of the images.
And step three, applying data expansion technologies such as random cutting, rotation transformation, perspective transformation, scale transformation and the like to the obtained surface crack image of the underwater dam body so as to achieve the purpose of expanding the surface crack image of the underwater dam body.
And fourthly, forming a mixed sample set by using the underwater dam surface crack image obtained by expansion and the collected water dam surface crack image, training an AlexNet deep convolution neural network model by using the mixed sample set, and obtaining an AlexNet-Damcrcks model.
Step five, utilizing the idea of transfer learning, and still adopting an AlexNet model structure in the construction of the underwater dam body surface crack identification model, namely, the AlexNet model structure is composed of 5 convolution modules and 3 full-connection modules, and is a structural diagram of the AlexNet model disclosed by the invention as shown in FIG. 2;
freezing parameters of a convolution module and a pooling module in the AlexNet-Damcracks model, taking the parameters as the parameters of the convolution module and the pooling module of the underwater dam body surface crack identification model, and setting the number of output units of 3 full-connection modules to be 4098, 4098 and 4 respectively; consider the conversion of the fracture identification problem into a four classification problem, namely: the 4 output units of the last full-connection layer respectively correspond to four label categories of dam body surface damage degree, and the four label categories comprise: no cracks, slight cracks, medium cracks and through cracks; and then, continuously optimizing parameters of a full connection layer in the model through iteration of the loss function to obtain a target model AlexNet-Damcracks-Migration.
The method for optimizing the parameters of the full connection layer in the model comprises the following specific steps:
initializing parameters of the full-connection layer by a random method, optimizing by adopting a random gradient descent method, and updating the learning rate by adopting an exponential decay method, wherein the learning rate of the exponential decay method is updated according to the following formula:
l=l 0 *γ [α/β] ,
wherein l is the learning rate after attenuation, l 0 Is an initial learning rate, gamma is an attenuation coefficient, alpha is the current iteration number, beta is an attenuation step length, [ alpha ], [ beta ], []Meaning rounding down.
The loss function is expressed by adopting cross entropy, and a regularization coefficient is added to punish a weight parameter; the number of iterations is n (n is 500k, k is a positive integer), and the loss function is given by the following formula:
where J is the training loss, θ is the weight parameter, p is the expected class probability, x is the batch of training samples, q is the predicted class probability, λ is the regularization coefficient, | θ | | Y 1 Represents the sum of the absolute values of the respective elements in the weight parameter θ, wherein the class probability is calculated by the Softmax layer.
And continuously training to find the condition of the parameters of the full-connection layer, which enables the loss function value and the recognition accuracy rate to be optimal at the same time, substituting multiple experiments into different k values, wherein the optimal parameters under the condition of the optimal iteration times are the parameters of the full-connection layer in the underwater dam body surface crack recognition model.
And step six, continuously training the optimized parameters of the full connection layer obtained in the step five on the basis of the underwater dam surface crack recognition model to obtain a target model AlexNet-Damcrcks-Migration, inputting the underwater dam surface crack image serving as a test set into the target model AlexNet-Damcrcks-Migration to obtain corresponding output, and accordingly achieving the final task of transferring the obtained target model AlexNet-Damcrcks-Migration to the underwater dam surface crack image recognition.
The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.
Claims (8)
1. A migration learning-based underwater dam body surface crack identification method is characterized by comprising the following steps:
constructing a mixed sample set based on the surface crack image of the overwater dam and the surface crack image of the underwater dam;
training the deep convolutional neural network model by using the mixed sample set to obtain a pre-training network model;
acquiring a target network model by utilizing a deep convolutional neural network model and a pre-training network model;
inputting the crack image of the surface of the underwater dam body into a target network model, and identifying the damage degree of the surface of the dam body according to the label type output by the target network model;
the deep convolutional neural network model is an AlexNet model, and the AlexNet model comprises a convolutional module, a pooling module and at least one full-connection module;
the pre-training network model is an AlexNet-Damcrcks model, and the AlexNet-Damcrcks model comprises a convolution module and a pooling module;
obtaining a target network model by utilizing a deep convolutional neural network model and a pre-training network model, wherein the method comprises the following steps:
freezing parameters of a convolution module and a pooling module in the AlexNet-Damcracks model to serve as the parameters of the convolution module and the pooling module in the AlexNet model;
setting the number of output units of a full-connection module in the AlexNet model as a positive integer not greater than 4;
training an AlexNet model by using a mixed sample set, and optimizing parameters of a full-connection module in the AlexNet model through iteration of a loss function of the AlexNet model;
and taking the AlexNet model after parameter optimization based on the full-connection module as a target network model.
2. The method for identifying the surface cracks of the underwater dam based on the transfer learning as claimed in claim 1, wherein a mixed sample set is constructed based on the surface crack images of the water dam and the surface crack images of the underwater dam, and comprises the following steps:
acquiring a surface crack image of the water dam body, and constructing a surface crack sample set of the water dam body;
acquiring an image of a surface crack of an underwater dam body, and constructing a sample set of the surface crack of the underwater dam body;
and expanding the sample set of the cracks on the surface of the water dam body to the sample set of the cracks on the surface of the underwater dam body, and taking the sample set of the cracks on the surface of the underwater dam body after expansion as a mixed sample set.
3. The method for identifying the surface cracks of the underwater dam based on the migration learning as claimed in claim 2,
constructing a sample set of surface cracks of the water dam body, comprising the following steps: preprocessing the collected water dam body surface crack image;
constructing a surface crack sample set of an underwater dam body, comprising the following steps of: preprocessing the collected surface crack image of the underwater dam body;
the pretreatment comprises the following steps: the image size or/and resolution is adjusted.
4. The method for identifying the surface cracks of the underwater dam based on the migration learning as claimed in claim 2, wherein the step of collecting the surface crack images of the underwater dam and constructing the surface crack sample set of the underwater dam comprises the following steps:
expanding the number of the surface crack images of the underwater dam body by using a data expansion technology;
constructing an underwater dam body surface crack sample set based on the expanded underwater dam body surface crack image;
the data augmentation technique includes: at least any one of random cropping, rotational transformation, perspective transformation, and scale transformation.
5. The method for identifying the surface cracks of the underwater dam based on the transfer learning as claimed in claim 1, wherein the label categories comprise: at least any one of no crack, a light crack, a medium crack, and a through crack.
6. The method for identifying the surface cracks of the underwater dam based on the transfer learning as claimed in claim 1, wherein the number of output units in the AlexNet model is set to be a full-connection module which is not more than a positive integer of 4, and the output units of the AlexNet model respectively correspond to the label types of the dam surface damage degrees.
7. The method for identifying the surface cracks of the underwater dam based on the transfer learning of claim 1, wherein a AlexNet model is trained by using a mixed sample set, and parameters of a full-connection module in the AlexNet model are optimized through iteration of a AlexNet model loss function, and the method comprises the following steps:
initializing parameters of a full-connection module in an AlexNet model by adopting a random method;
optimizing parameters of a full-connection module in the AlexNet model by adopting a random gradient descent method, wherein an exponential decay method is adopted as a learning rate updating strategy in the random gradient descent method;
training parameters of the full-connection module when the obtained loss function value is reduced and the recognition accuracy is increased;
and repeating the training process according to the preset iteration times based on the AlexNet model of the parameters of the full-connection module when the loss function value is reduced and the identification accuracy is increased.
8. The method for identifying the surface cracks of the underwater dam based on the migration learning as claimed in claim 7,
the learning rate includes the following operational formula:
l=l 0 *γ [α/β] ,
wherein l is the learning rate after attenuation, l 0 Is an initial learning rate, gamma is an attenuation coefficient, alpha is the current iteration number, beta is an attenuation step length, [ alpha ], [ beta ], []Represents rounding down;
the loss function includes the following operational formula:
where J is the training loss, θ is the weight parameter, p is the expected class probability, x is the batch of training samples, q is the predicted class probability, λ is the regularization coefficient, | θ | | Y 1 Representing the sum of the absolute values of the individual elements in the weight parameter theta.
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 CN110147772A (en) | 2019-08-20 |
CN110147772B true 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) |
Families Citing this family (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111257341B (en) * | 2020-03-30 | 2023-06-16 | 河海大学常州校区 | Underwater building crack detection method based on multi-scale features and stacked full convolution network |
CN111691358B (en) * | 2020-06-18 | 2021-05-11 | 河海大学 | Gravity dam apparent crack risk prediction method and system |
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 |
CN112926669B (en) * | 2021-03-09 | 2022-03-08 | 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 |
CN113358582B (en) * | 2021-06-04 | 2022-08-26 | 山东国瑞新能源有限公司 | Method, equipment and medium for detecting concrete structure defects |
CN114596262A (en) * | 2022-01-27 | 2022-06-07 | 福建华电福瑞能源发展有限公司古田溪水力发电厂 | Dam monitoring and analyzing method and system based on image recognition technology |
CN114677601A (en) * | 2022-04-12 | 2022-06-28 | 雅砻江流域水电开发有限公司 | Dam crack detection method based on unmanned aerial vehicle inspection and combined with deep learning |
CN115953672B (en) * | 2023-03-13 | 2024-02-27 | 南昌工程学院 | Method for identifying surface cracks of underwater dam |
CN117249801B (en) * | 2023-09-21 | 2024-07-12 | 深圳市水务工程检测有限公司 | Dam deformation monitoring management system and method based on big data |
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 |
Also Published As
Publication number | Publication date |
---|---|
CN110147772A (en) | 2019-08-20 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110147772B (en) | Migration learning-based underwater dam body surface crack identification method | |
CN106971152B (en) | Method for detecting bird nest in power transmission line based on aerial images | |
CN111739075B (en) | Deep network lung texture recognition method combining multi-scale attention | |
CN109523013B (en) | Air particulate matter pollution degree estimation method based on shallow convolutional neural network | |
CN107909206B (en) | PM2.5 prediction method based on deep structure recurrent neural network | |
CN111178149B (en) | Remote sensing image water body automatic extraction method based on residual pyramid network | |
CN113486865A (en) | Power transmission line suspended foreign object target detection method based on deep learning | |
CN110991359A (en) | Satellite image target detection method based on multi-scale depth convolution neural network | |
CN113989257A (en) | Electric power comprehensive pipe gallery settlement crack identification method based on artificial intelligence technology | |
CN112258490A (en) | Low-emissivity coating intelligent damage detection method based on optical and infrared image fusion | |
CN111178438A (en) | ResNet 101-based weather type identification method | |
CN113780242A (en) | Cross-scene underwater sound target classification method based on model transfer learning | |
CN112749675A (en) | Potato disease identification method based on convolutional neural network | |
CN111028238A (en) | Robot vision-based three-dimensional segmentation method and system for complex special-shaped curved surface | |
CN114359702A (en) | Method and system for identifying building violation of remote sensing image of homestead based on Transformer | |
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 | |
CN117236201B (en) | Diffusion and ViT-based downscaling method | |
Tao et al. | Electric insulator detection of UAV images based on depth learning | |
CN117392065A (en) | Cloud edge cooperative solar panel ash covering condition autonomous assessment method | |
CN110766662B (en) | Forging surface crack detection method based on multi-scale and multi-layer feature learning | |
CN117292126A (en) | Building elevation analysis method and system using repeated texture constraint and electronic equipment | |
CN111275025A (en) | Parking space detection method based on deep learning | |
CN115294381B (en) | Small sample image classification method and device based on feature migration and orthogonal prior | |
CN116309387A (en) | Deep learning-based nuclear containment crack detection method and system |
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 |
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 |
|
CB03 | Change of inventor or designer information | ||
GR01 | Patent grant | ||
GR01 | Patent grant |