Bridge disease detection method and system based on deep transfer learning

Abstract

The invention discloses a bridge disease detection method and system based on deep transfer learning, and relates to the technical field of pattern recognition, and the method comprises the following steps: S1, collecting a bridge disease CCD image and laser point cloud data, and building a bridge disease data set; s2, constructing a bridge disease detection model; s3, designing network parameters of the bridge disease detection model, and pre-training the bridge disease detection model by adopting a transfer learning method; and S4, dividing the bridge disease data set into a training set and a test set according to a preset proportion, performing secondary training on the pre-trained bridge disease detection model by using the training set, and detecting the test set by using the bridge disease detection model subjected to secondary training to obtain disease types and disease positions. According to the invention, the bridge disease condition can be rapidly and accurately detected, and the safety and reliability of the bridge can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of pattern recognition technology, and more specifically to a bridge defect detection method and system based on deep transfer learning. Background Art

[0002] Bridges are important links connecting cities, regions and various types of infrastructure, but due to the long-term influence of natural environment and human factors, bridges will have various diseases. These diseases not only affect the normal use of bridges, but may also threaten traffic safety and the surrounding environment. Bridge diseases mainly include two categories: structural diseases and non-structural diseases. Structural diseases include cracks, rust, concrete spalling, etc. in the bridge structure, while non-structural diseases mainly include damage to the bridge deck, damage to guardrails, blockage of the drainage system, etc. Although these diseases do not directly threaten the structural safety of the bridge, they will affect the bearing capacity and safety of the bridge, and thus affect the normal use and maintenance of the bridge.

[0003] Modern technology provides a variety of solutions to bridge defects. Among them, non-destructive testing technology is a commonly used method. Through sound waves, ultrasonic waves, magnetic particle testing and other methods, the bridge structure is comprehensively and accurately tested, potential defects are discovered early and corresponding measures are taken. In addition, the advancement of material technology has also provided more options for the repair and reinforcement of bridges. The application of new high-strength materials, nanomaterials, etc. can effectively improve the weathering and compression resistance of bridges. Many experts and scholars at home and abroad have conducted research on bridge defect detection methods. However, the environment for bridge defect detection is relatively complex and image acquisition is difficult, so bridge defect detection is a difficult point.

[0004] Therefore, how to improve the accuracy of bridge detection, shorten the detection time, reduce the missed detection rate, and effectively improve the safety and reliability of bridges is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0005] In view of this, the present invention provides a bridge defect detection method and system based on deep transfer learning, which solves the problems existing in the background technology.

[0006] In order to achieve the above object, the present invention provides the following technical solutions: A bridge disease detection method based on deep transfer learning includes the following steps: S1: Collect bridge damage CCD images and laser point cloud data to establish a bridge damage dataset; S2: Construct bridge disease detection model; S3: Design the network parameters of the bridge disease detection model and use the transfer learning method to pre-train the bridge disease detection model; S4: Divide the bridge disease data set into a training set and a test set according to a preset ratio, use the training set to perform secondary training on the pre-trained bridge disease detection model, and use the secondary trained bridge disease detection model to detect the test set to obtain the type and location of the disease.

[0007] Optionally, in S1, establishing a bridge disease dataset specifically includes the following steps: S11: Use laser scanner and CCD camera to obtain laser point cloud data and CCD images of bridge defects; S12: Set laser point cloud dataset , any measuring point The neighborhood point set and its unit normal vector are and , then the measuring point New points after bilateral filtering The values ​​are:

[0008] in, is the bilateral filtering factor, which is defined as:

[0009] In the formula, express The unit normal vector of Represents the spatial domain weight, which is used to control the smoothness of the measurement points; Represents the feature domain weight, which is used to control the degree of feature preservation; S13: Using an image restoration algorithm based on structure clustering and centralized regularization to restore the CCD image; S14: Establishing a mapping relationship between the laser point cloud data and the CCD image pixel points, and completing the fusion of the laser point cloud data and the CCD image data.

[0010] Optionally, in S13, the method for restoring the CCD image is specifically: The k-means method is used to perform structural clustering on the degraded CCD images, and the sparse K-SVD method is used for dictionary learning. A sparse restoration model based on structural clustering and non-local regularization is established and centralized regularization is performed to optimize the sparse coefficient and obtain high-quality restored images.

[0011] Optionally, in S14, establishing a mapping relationship between laser point cloud data and CCD image pixel points specifically includes the following steps: Points in physical space Corresponding image space coordinate system coordinates , and its spatial transformation relationship is:

[0012] Where: represents the relative rotation matrix between image space coordinates and world coordinates, represents the relative translation matrix, represents the zero vector; Image space coordinate point The coordinates corresponding to the camera coordinate system are , the relationship is as follows:

[0013] Where: s represents the vertical distance from point W to the optical center of the camera, represents the focal length of the camera, Represents the coordinates of the camera optical center in the camera coordinate system; The pixel coordinates of the center point of the image are Pixels , whose coordinates in the camera coordinate system are ,but:

[0014] Where: , Indicates the physical length of each pixel in the camera coordinate system; Establishing world coordinate points The pixels of the camera image The relationship is:

[0015] In the formula, , represents the intrinsic parameter matrix of the camera; According to the intrinsic parameter matrix K, the relative rotation matrix R and the relative translation matrix T, the pixel coordinates of the point mapping in the world coordinate system are confirmed, and finally the mapping relationship between the laser point cloud data points and the CCD image pixel points is established to realize the fusion of the laser point cloud data and the CCD image data.

[0016] Optionally, in S2, a bridge disease detection model is constructed, which specifically includes the following steps: S21: Based on the detection framework of Faster R-CNN and the network design concept of FPN, the corresponding feature maps are extracted from the Conv2, Conv3, Conv4, and Conv5 layers of the ResNet backbone network, a multi-scale feature pyramid is constructed, and the feature maps of different scales are gradually fused through upsampling to achieve the fusion of deep features and shallow features; S22: Adding attention mechanism module CBAM in convolutional layer to multi-scale feature pyramid; S23: Using the RPN network to obtain the interest region containing multi-scale damaged targets, and then pooling the obtained interest regions respectively; through the Softmax classifier and The regressor classifies and regresses the domain of interest, obtains the disease type and location information, and fuses the output results of the detection subnet; S24: The detected damage location information is integrated with the depth information mapped from the damaged pixel location to obtain complete multi-scale damage features of the bridge.

[0017] Optionally, in S22, the attention mechanism module CBAM is composed of a channel attention mechanism module and a spatial attention mechanism module; wherein the channel attention mechanism module is used to adaptively adjust the weight of the channel information to achieve effective expression of damage characteristics; the spatial attention mechanism module is used to assign weights to each spatial position of the feature map to highlight the significant features of the damage.

[0018] Optionally, in S3, network parameters of the bridge defect detection model are designed, and the bridge defect detection model is pre-trained using a transfer learning method, specifically including the following steps: S31: Design objective function based on the constructed bridge defect detection model; The ReLU function is used as the network activation function, which is defined as follows:

[0019] The network bias and network weight are initialized using the MSRA initialization method. The network bias is set to 0, and the network weight is set to obey a Gaussian distribution with a mean of 0 and a variance of A random variable of , where k represents the size of the convolution kernel in the convolution operation, c represents the number of feature channels of the convolution layer corresponding to k, Represented as a numerical value; The Adam learning method is used to optimize the network parameters, as follows:

[0020]

[0021] Where: , Respectively represent the updated values ​​of the gradient and the exponentially weighted average of the square of the gradient at the tth iteration; Indicates t The parameter value of the iteration, represents the parameter value of the t+1th iteration; represents the learning rate; represents the learning rate of the tth iteration, represents the initial learning rate, represents the attenuation rate, is a constant; S32: Select a dataset of defect images with the same material and expand the data by translation, scaling, rotation and contrast transformation. Use the expanded data to pre-train a bridge defect detection model. Tune the network by changing network parameters to obtain a pre-trained bridge defect detection model.

[0022] Optionally, in S4, the pre-trained bridge disease detection model is trained again using the training set, specifically: S41: Randomly select 70% of the samples from the bridge disease dataset as the training set, and the remaining 30% as the test set; S42: Based on the pre-trained bridge disease detection model, the data output by Softmax is adjusted to the number of bridge disease classifications, and the network parameter values ​​of the pre-trained bridge disease detection model are used as the initial parameter values ​​of the secondary training, and the network is trained until the model converges, thereby obtaining a bridge disease detection model that has been completed with the secondary training; S43: Input the test set into the bridge defect detection model that has completed the secondary training for detection, and obtain the defect type and defect location.

[0023] A bridge defect detection system based on deep transfer learning for implementing the above method comprises: Data acquisition module, used to collect bridge damage CCD images and laser point cloud data to establish bridge damage data set; Model building module, used to build bridge disease detection model; Parameter design and optimization module, which is used to design the network parameters of the bridge defect detection model and pre-train the bridge defect detection model using transfer learning method; A model training module is used to divide the bridge disease data set into a training set and a test set according to a preset ratio, and use the training set to perform secondary training on the pre-trained bridge disease detection model; The detection module is used to detect the test set through the bridge disease detection model completed by secondary training to obtain the type and location of the disease.

[0024] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides a bridge defect detection method and system based on deep transfer learning, which has the following beneficial effects: (1) In order to achieve accurate and rapid detection of bridge defects, the present invention combines the characteristics of bridges, bridge defects, the characteristics of laser point clouds collected by laser scanners, and the characteristics of images collected by CCD cameras to propose a fusion algorithm for laser point cloud data of bridge defects and CCD image data, which can improve the detection accuracy and speed of bridge defects; (2) In order to solve the problem of difficulty in detecting bridge defects due to the large number of types and large scale variations of bridge defects, the present invention proposes an algorithm model suitable for multi-scale bridge defect detection from the perspectives of detection network framework, multi-scale feature learning network, and convolutional network structure. This algorithm model can be combined with transfer learning to further improve the detection accuracy and speed of bridge defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0026] Figure 1 A flow chart of a bridge defect detection method based on deep transfer learning provided by the present invention; Figure 2 A diagram showing the relationship between the CCD camera coordinate system, the image space coordinate system and the world coordinate system provided by the present invention; Figure 3 A schematic diagram of the structure of the bridge disease detection model provided by the present invention; Figure 4 A schematic diagram of a channel attention mechanism module and a spatial attention mechanism module provided by the present invention; Figure 5 This is a training flow chart of the bridge disease detection model provided by the present invention. DETAILED DESCRIPTION

[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The embodiment of the present invention discloses a bridge disease detection method based on deep transfer learning, such as Figure 1 As shown, the following steps are included: S1: Collect bridge damage CCD images and laser point cloud data to establish a bridge damage dataset; S2: Construct bridge disease detection model; S3: Design the network parameters of the bridge disease detection model and use the transfer learning method to pre-train the bridge disease detection model; S4: Divide the bridge disease data set into a training set and a test set according to a preset ratio, use the training set to perform secondary training on the pre-trained bridge disease detection model, and use the secondary trained bridge disease detection model to detect the test set to obtain the type and location of the disease.

[0029] Next, taking the bridge as an example, Figure 1 The specific implementation of the bridge defect detection method shown is further described in detail.

[0030] 1. Dataset preparation S11: Use laser scanner and CCD camera to obtain laser point cloud data and CCD images of bridge defects; S12: Set laser point cloud dataset , any measuring point The neighborhood point set and its unit normal vector are and , then the measuring point New points after bilateral filtering The values ​​are:

[0031] Among them, the bilateral filtering algorithm can be used to denoise the laser point cloud. is the bilateral filtering factor, which is defined as:

[0032] In the formula, express The unit normal vector of Represents the spatial domain weight, which is used to control the smoothness of the measurement points; Represents the feature domain weight, which is used to control the degree of feature preservation; S13: Using an image restoration algorithm based on structure clustering and centralized regularization to restore the CCD image; Specifically, the k-means method is used to perform structural clustering on the degraded CCD image, and the sparse K-SVD method is used for dictionary learning; a sparse restoration model based on structural clustering and non-local regularization is established and centralized regularization is performed to optimize the sparse coefficient and obtain a high-quality restored image; S14: Establish a mapping relationship between the laser point cloud data and the CCD image pixel points, and establish the relationship between the CCD camera coordinate system, the image space coordinate system and the world coordinate system.

[0033] Specifically, a point in physical space Corresponding image space coordinate system coordinates , and its spatial transformation relationship is:

[0034] Where: represents the relative rotation matrix between image space coordinates and world coordinates, represents the relative translation matrix, represents the zero vector; Image space coordinate point The coordinates corresponding to the camera coordinate system are , the relationship is as follows:

[0035] Where: s represents the vertical distance from point W to the optical center of the camera, represents the focal length of the camera, Represents the coordinates of the camera optical center in the camera coordinate system; The pixel coordinates of the center point of the image are Pixels , whose coordinates in the camera coordinate system are ,but:

[0036] Where: , Indicates the physical length of each pixel in the camera coordinate system; According to the above three formulas, the world coordinate point can be established The pixels of the camera image The relationship is:

[0037] In the formula, , represents the intrinsic parameter matrix of the camera, which is only related to the camera focal length and optical center position; According to the internal parameter matrix K, the relative rotation matrix R and the relative translation matrix T, the pixel coordinates of the point mapping in the world coordinate system are confirmed, and finally the mapping relationship between the laser point cloud data points and the CCD image pixel points is established to realize the fusion of the laser point cloud data and the CCD image data. The relationship between the CCD camera coordinate system, the image space coordinate system and the world coordinate system mentioned above is as follows: Figure 2 shown.

[0038] 2. Model construction S21: Reference Figure 3 Based on the detection framework of Faster R-CNN and the network design concept of FPN, the corresponding feature maps are extracted from the Conv2, Conv3, Conv4, and Conv5 layers of the ResNet backbone network, a multi-scale feature pyramid is constructed, and the feature maps of different scales are gradually fused through upsampling to achieve the fusion of deep features and shallow features; S22: Add the attention mechanism module CBAM in the convolution layer to the multi-scale feature pyramid; Figure 4 As shown in the figure, the attention mechanism module CBAM consists of a channel attention mechanism module and a spatial attention mechanism module; the channel attention mechanism module is used to adaptively adjust the weight of channel information to achieve effective expression of damage features; the spatial attention mechanism module is used to assign weights to each spatial position of the feature map to highlight the significant features of the damage, which can further promote the extraction of damage texture and detail features; S23: Using the RPN network to obtain the interest region containing multi-scale damaged targets, and then pooling the obtained interest regions respectively; through the Softmax classifier and The regressor classifies and regresses the domain of interest, obtains the disease type and location information, and fuses the output results of the detection subnet; S24: The detected defect location information is integrated with the depth information mapped from the damaged pixel location to achieve a complete description of the multi-scale defect characteristics of the bridge, and finally complete the bridge defect detection.

[0039] 3. Parameter design and optimization S31: Based on the constructed bridge defect detection model, the objective function is designed to improve the detection accuracy and speed of the algorithm; In order to solve the gradient diffusion and model overfitting problems in the back-propagation process, the ReLU (Rectified Linear Units) function is used as the network activation function, which is defined as follows:

[0040] The MSRA initialization method is used to initialize parameters such as network bias and network weight. The network bias is set to 0, and the network weight is set to obey the Gaussian distribution with a mean of 0 and a variance of A random variable of , where k represents the size of the convolution kernel in the convolution operation, c represents the number of feature channels of the convolution layer corresponding to k, Represented as a numerical value; The Adam learning method is used to optimize the network parameters, as follows:

[0041]

[0042] Where: , Respectively represent the updated values ​​of the gradient and the exponentially weighted average of the square of the gradient at the tth iteration; Indicates t The parameter value of the iteration, represents the parameter value of the t+1th iteration; represents the learning rate; represents the learning rate of the tth iteration, represents the initial learning rate, represents the attenuation rate, is a constant; S32: After determining the objective function and activation function optimization strategy, in order to further improve the detection performance of the bridge disease detection network, a disease image dataset with the same material is selected. This dataset contains 3 target categories and nearly 5,000 images, which can fully train the bridge disease detection model. In addition, data expansion can be performed by translation, scaling, rotation, and contrast transformation. The bridge disease detection model is pre-trained using the expanded data. The network is tuned by changing network parameters such as learning rate, convolution kernel size, and convolution step size to obtain a pre-trained bridge disease detection model. With the help of transfer learning, the knowledge learned by the ResNet model on the disease image dataset of the same material is transferred to the bridge disease recognition task, which can solve the overfitting problem caused by insufficient samples in the training of the bridge disease detection network and optimize the network detection performance.

[0043] 4. Model training and disease detection S41: Randomly select 70% of the samples from the bridge disease dataset as the training set, and the remaining 30% as the test set; S42: Figure 5 As shown in the figure, based on the pre-trained bridge disease detection model, the data output by Softmax is adjusted to the number of bridge disease classifications, and the network parameter values ​​of the pre-trained bridge disease detection model are used as the initial parameter values ​​of the secondary training. The network is trained until the model converges, and the bridge disease detection model completed by the secondary training is obtained. S43: Input the test set into the bridge defect detection model that has completed the secondary training for detection, and obtain the defect type and defect location.

[0044] and Figure 1 Corresponding to the method described above, the embodiment of the present invention also provides a bridge disease detection system based on deep transfer learning, which is used to Figure 1 The specific implementation of the method in the embodiment of the present invention is a bridge disease detection system based on deep transfer learning, which can be applied to computer terminals or various mobile devices, and specifically includes: Data acquisition module, used to collect bridge damage CCD images and laser point cloud data to establish bridge damage data set; Model building module, used to build bridge disease detection model; Parameter design and optimization module, which is used to design the network parameters of the bridge defect detection model and pre-train the bridge defect detection model using transfer learning method; A model training module is used to divide the bridge disease data set into a training set and a test set according to a preset ratio, and use the training set to perform secondary training on the pre-trained bridge disease detection model; The detection module is used to detect the test set through the bridge disease detection model completed by secondary training to obtain the type and location of the disease.

[0045] In summary, this embodiment aims to solve the problems of complex data collection scenarios and insufficient sample data in the bridge disease detection process, and studies the fusion algorithm of laser point cloud data and CCD image data to achieve effective extraction of bridge disease image features; based on the theoretical framework of target detection based on deep learning, studies the multi-scale disease detection algorithm to achieve accurate and rapid detection of bridge diseases. After the implementation of this solution, it will provide theoretical basis and technical support for bridge disease detection, and can effectively improve the safety and reliability of bridges.

[0046] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0047] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bridge disease detection method based on deep transfer learning, characterized in that: The following steps are involved: S1: Collect bridge damage CCD images and laser point cloud data to establish a bridge damage dataset; S2: Construct bridge disease detection model; S3: Design the network parameters of the bridge disease detection model and use the transfer learning method to pre-train the bridge disease detection model; S4: Divide the bridge disease data set into a training set and a test set according to a preset ratio, use the training set to perform secondary training on the pre-trained bridge disease detection model, and use the secondary trained bridge disease detection model to detect the test set to obtain the type and location of the disease.

2. A bridge defect detection method based on deep transfer learning according to claim 1, characterized in that: In S1, a bridge disease dataset is established, which specifically includes the following steps: S11: Use laser scanner and CCD camera to obtain laser point cloud data and CCD images of bridge defects; S12: Set laser point cloud dataset , any measuring point The neighborhood point set and its unit normal vector are and , then the measuring point New points after bilateral filtering The values ​​are: ， in, is the bilateral filtering factor, which is defined as: ， In the formula, express The unit normal vector of Represents the spatial domain weight, which is used to control the smoothness of the measurement points; Represents the feature domain weight, which is used to control the degree of feature preservation; S13: Using an image restoration algorithm based on structure clustering and centralized regularization to restore the CCD image; S14: Establishing a mapping relationship between the laser point cloud data and the CCD image pixel points, and completing the fusion of the laser point cloud data and the CCD image data.

3. A bridge defect detection method based on deep transfer learning according to claim 2, characterized in that: In S13, the method for restoring the CCD image is specifically as follows: The k-means method is used to perform structural clustering on the degraded CCD images, and the sparse K-SVD method is used for dictionary learning. A sparse restoration model based on structural clustering and non-local regularization is established and centralized regularization is performed to optimize the sparse coefficient and obtain high-quality restored images.

4. The bridge defect detection method based on deep transfer learning according to claim 2 is characterized in that: In S14, a mapping relationship between the laser point cloud data and the CCD image pixels is established, which specifically includes the following steps: Points in physical space Corresponding image space coordinate system coordinates , and its spatial transformation relationship is: ， Where: represents the relative rotation matrix between image space coordinates and world coordinates, represents the relative translation matrix, represents the zero vector; Image space coordinate point The coordinates corresponding to the camera coordinate system are , the relationship is as follows: ， Where: s represents the vertical distance from point W to the optical center of the camera, represents the focal length of the camera, Represents the coordinates of the camera optical center in the camera coordinate system; The pixel coordinates of the center point of the image are Pixels , whose coordinates in the camera coordinate system are ,but: ， Where: , Indicates the physical length of each pixel in the camera coordinate system; Establishing world coordinate points The pixels of the camera image The relationship is: ， In the formula, , represents the intrinsic parameter matrix of the camera; According to the intrinsic parameter matrix K, the relative rotation matrix R and the relative translation matrix T, the pixel coordinates of the point mapping in the world coordinate system are confirmed, and finally the mapping relationship between the laser point cloud data points and the CCD image pixel points is established to realize the fusion of the laser point cloud data and the CCD image data.

5. The bridge defect detection method based on deep transfer learning according to claim 1 is characterized in that: In S2, a bridge disease detection model is constructed, which specifically includes the following steps: S21: Based on the detection framework of Faster R-CNN and the network design concept of FPN, the corresponding feature maps are extracted from the Conv2, Conv3, Conv4, and Conv5 layers of the ResNet backbone network, a multi-scale feature pyramid is constructed, and the feature maps of different scales are gradually fused through upsampling to achieve the fusion of deep features and shallow features; S22: Adding attention mechanism module CBAM in convolutional layer to multi-scale feature pyramid; S23: Using the RPN network, the interest region containing multi-scale damaged targets is obtained, and then the obtained interest regions are pooled respectively; through the Softmax classifier and The regressor classifies and regresses the domain of interest, obtains the disease type and location information, and fuses the output results of the detection subnet; S24: The detected damage location information is integrated with the depth information mapped from the damaged pixel location to obtain complete multi-scale damage features of the bridge.

6. A bridge defect detection method based on deep transfer learning according to claim 5, characterized in that: In S22, the attention mechanism module CBAM consists of a channel attention mechanism module and a spatial attention mechanism module; the channel attention mechanism module is used to adaptively adjust the weight of channel information to achieve effective expression of damage features; the spatial attention mechanism module is used to assign weights to each spatial position of the feature map to highlight the significant features of the damage.

7. The bridge defect detection method based on deep transfer learning according to claim 1 is characterized in that: In S3, the network parameters of the bridge disease detection model are designed, and the bridge disease detection model is pre-trained using the transfer learning method, which specifically includes the following steps: S31: Design objective function based on the constructed bridge defect detection model; The ReLU function is used as the network activation function, which is defined as follows: ， The network bias and network weight are initialized using the MSRA initialization method. The network bias is set to 0, and the network weight is set to obey a Gaussian distribution with a mean of 0 and a variance of A random variable of , where k represents the size of the convolution kernel in the convolution operation, c represents the number of feature channels of the convolution layer corresponding to k, is represented as a numerical value; The Adam learning method is used to optimize the network parameters, as follows: ， ， Where: , Respectively represent the updated values ​​of the gradient and the exponentially weighted average of the square of the gradient at the tth iteration; Indicates t The parameter value of the iteration, represents the parameter value of the t+1th iteration; represents the learning rate; represents the learning rate of the tth iteration, represents the initial learning rate, represents the attenuation rate, is a constant; S32: Select a dataset of defect images with the same material and expand the data by translation, scaling, rotation and contrast transformation. Use the expanded data to pre-train a bridge defect detection model. Tune the network by changing network parameters to obtain a pre-trained bridge defect detection model.

8. The bridge defect detection method based on deep transfer learning according to claim 1 is characterized in that: In S4, the pre-trained bridge disease detection model is trained again using the training set, specifically: S41: Randomly select 70% of the samples from the bridge disease dataset as the training set, and the remaining 30% as the test set; S42: Based on the pre-trained bridge disease detection model, the data output by Softmax is adjusted to the number of bridge disease classifications, and the network parameter values ​​of the pre-trained bridge disease detection model are used as the initial parameter values ​​of the secondary training, and the network is trained until the model converges, thereby obtaining a bridge disease detection model that has been completed with the secondary training; S43: Input the test set into the bridge defect detection model that has completed the secondary training for detection, and obtain the defect type and defect location.

9. A bridge defect detection system based on deep transfer learning that implements the method according to any one of claims 1 to 8, characterized in that: include: Data acquisition module, used to collect bridge damage CCD images and laser point cloud data to establish bridge damage data set; Model building module, used to build bridge disease detection model; Parameter design and optimization module, which is used to design the network parameters of the bridge defect detection model and pre-train the bridge defect detection model using transfer learning method; A model training module is used to divide the bridge disease data set into a training set and a test set according to a preset ratio, and use the training set to perform secondary training on the pre-trained bridge disease detection model; The detection module is used to detect the test set through the bridge disease detection model completed by secondary training to obtain the type and location of the disease.

Images