Bridge disease identification method and device based on unsupervised learning, equipment and medium

Abstract

The invention relates to the technical field of computers, in particular to a bridge disease identification method and device based on unsupervised learning, equipment and a medium. And the global information auxiliary module is arranged, so that the learning ability of the model on image semantic knowledge is effectively enhanced. When the global information auxiliary module receives an input image, the image is segmented into a plurality of image blocks with the same size. The image blocks are converted into one-dimensional vectors, and the one-dimensional vectors are embedded into a high-dimensional feature space through linear projection, so that the modeling of the dependency relationship between the image blocks is realized, and the global information association of the image is accurately captured. And through the operation of the attention module, modeling is further carried out on the dependency relationship between the image blocks, and the capture of global information is enhanced. And fusing the output of the self-attention mechanism with a convolution operation result, and enabling the model to obtain image feature expression enhanced by global information on the basis of keeping original feature information. The implicit features contain rich image semantics and spatial information, and the accuracy of bridge disease recognition is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a bridge defect identification method, device, equipment and medium based on unsupervised learning. Background Art

[0002] With increasing service life and the impact of natural disasters, bridges are prone to typical surface defects such as exposed rebar and cracks, which seriously threaten bridge safety and can even cause a series of traffic accidents such as collapse and rollover. Therefore, automatic and accurate identification of bridge defect information is of great significance to ensuring bridge safety.

[0003] With the continuous advancement of computer technology and the deepening of machine learning theory, deep learning, with its powerful feature extraction capabilities, has been widely used in highway inspections. However, in the context of bridge defect identification, existing recognition models still face technical bottlenecks and fail to meet practical needs, primarily due to poor recognition accuracy. This problem not only affects the efficiency of highway bridge inspections but also poses potential risks to bridge structural safety assessments.

[0004] Based on this, the present invention proposes a bridge defect identification method and device based on unsupervised learning to solve the above technical problems. Summary of the Invention

[0005] The present invention describes a bridge defect identification method, device, equipment and medium based on unsupervised learning, which can effectively improve the identification accuracy of bridge defects.

[0006] According to a first aspect, the present invention provides a bridge defect identification method based on unsupervised learning, comprising:

[0007] Acquire a bridge image to be identified;

[0008] Inputting the bridge image to be identified into a trained bridge defect recognition model to obtain a bridge defect recognition result; wherein the bridge defect recognition model is obtained by training a preset deep learning network using unsupervised learning;

[0009] The deep learning network includes an input module, an output module, multiple convolution modules connected in sequence, multiple deconvolution modules connected in sequence, an attention module and a global information auxiliary module. The input module is connected to the input end of the first convolution module, the output end of the last convolution module is connected to the input end of the first deconvolution module, the output end of the last deconvolution module is connected to the output module, the output end of the input module is connected to the input end of the global information auxiliary module, the output end of the global information auxiliary module is connected to the input end of the attention module, and the output end of the attention module is connected to the input end of the first convolution module; the input module is used to receive the bridge image to be identified, the attention module and the convolution module are used to extract the features of the bridge image to be identified, the deconvolution module is used to decode the features of the bridge image to be identified, the output module is used to output the recognition result of the bridge disease, and the global information auxiliary module is used to enhance the learning ability of the bridge disease recognition model for image semantic knowledge.

[0010] According to a second aspect, the present invention provides a bridge defect identification device based on unsupervised learning, comprising:

[0011] an acquisition unit configured to acquire an image of a bridge to be identified;

[0012] The recognition unit is configured to input the bridge image to be recognized into a trained bridge defect recognition model to obtain a bridge defect recognition result; wherein the bridge defect recognition model is obtained by training a preset deep learning network using unsupervised learning;

[0013] The deep learning network includes an input module, an output module, multiple convolution modules connected in sequence, multiple deconvolution modules connected in sequence, an attention module and a global information auxiliary module. The input module is connected to the input end of the first convolution module, the output end of the last convolution module is connected to the input end of the first deconvolution module, the output end of the last deconvolution module is connected to the output module, the output end of the input module is connected to the input end of the global information auxiliary module, the output end of the global information auxiliary module is connected to the input end of the attention module, and the output end of the attention module is connected to the input end of the first convolution module; the input module is used to receive the bridge image to be identified, the attention module and the convolution module are used to extract the features of the bridge image to be identified, the deconvolution module is used to decode the features of the bridge image to be identified, the output module is used to output the recognition result of the bridge disease, and the global information auxiliary module is used to enhance the learning ability of the bridge disease recognition model for image semantic knowledge.

[0014] According to a third aspect, the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method of the first aspect is implemented.

[0015] According to a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method of the first aspect.

[0016] The present invention provides a method, apparatus, device, and medium for bridge defect identification based on unsupervised learning. The invention aims to achieve accurate bridge defect identification through an unsupervised learning-based bridge defect identification model. An image of the bridge to be identified is obtained and input into a well-trained bridge defect identification model, which then generates a corresponding identification result. This bridge defect identification model uses an unsupervised learning strategy to train a pre-set deep learning network. The deep learning network architecture comprises an input module, an output module, and multiple interconnected convolutional and deconvolutional modules. Furthermore, it incorporates an attention module and a global information auxiliary module. The input module is connected to the input of the first convolutional module, the output of the last convolutional module is connected to the input of the first deconvolutional module, and the output of the last deconvolutional module is connected to the output module. Furthermore, the output of the input module is connected to the input of the global information auxiliary module, the output of the global information auxiliary module is connected to the input of the attention module, and the output of the attention module is associated with the outputs of the multiple connected convolutional modules. The input module receives the bridge image to be identified; the attention module and multiple sequentially connected convolution modules perform feature extraction of the bridge image; the multiple sequentially connected deconvolution modules decode the image features; and the output module outputs the bridge defect identification results. Notably, the inclusion of the global information auxiliary module effectively enhances the model's ability to learn image semantics. During training, given that the model is based on unsupervised learning and is configured before the attention module, the global information auxiliary module segments the input image into multiple uniformly sized patches. These patches are flattened to convert them into one-dimensional vectors and embedded into a high-dimensional feature space via linear projection. Within this space, the module calculates self-attention scores based on the interaction of query, key, and value operations, modeling the dependencies between patches and accurately capturing global information associations within the image. Subsequently, the attention module further models the dependencies between patches, enhancing the capture of global information. After completion, the output of the self-attention mechanism is fused with the convolution results via residual connections, allowing the model to retain the original feature information while obtaining a globally enhanced image feature representation. This implicit feature contains rich image semantics and spatial information, greatly improving the accuracy of bridge defect identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A schematic flow chart of a bridge defect identification method based on unsupervised learning according to one embodiment is shown;

[0019] Figure 2 A schematic block diagram of a bridge defect identification device based on unsupervised learning according to one embodiment is shown;

[0020] Figure 3 A schematic diagram of pixel shielding according to one embodiment is shown. DETAILED DESCRIPTION

[0021] The solution provided by the present invention is described below with reference to the accompanying drawings.

[0022] Figure 1 A flow chart of a bridge defect identification method based on unsupervised learning according to one embodiment is shown. It is understood that the method can be executed by any device, equipment, platform, or equipment cluster with computing and processing capabilities. Figure 1 As shown, the method includes:

[0023] Step 101: Acquire a bridge image to be identified;

[0024] Step 102: Input the bridge image to be identified into a trained bridge defect identification model to obtain a bridge defect identification result; wherein the bridge defect identification model is obtained by training a preset deep learning network using unsupervised learning;

[0025] The deep learning network includes an input module, an output module, multiple convolution modules connected in sequence, multiple deconvolution modules connected in sequence, an attention module and a global information auxiliary module. The input module is connected to the input end of the first convolution module, the output end of the last convolution module is connected to the input end of the first deconvolution module, the output end of the last deconvolution module is connected to the output module, the output end of the input module is connected to the input end of the global information auxiliary module, the output end of the global information auxiliary module is connected to the input end of the attention module, and the output end of the attention module is connected to the input end of the first convolution module; the input module is used to receive the bridge image to be identified, the attention module and the convolution module are used to extract the features of the bridge image to be identified, the deconvolution module is used to decode the features of the bridge image to be identified, the output module is used to output the recognition result of the bridge disease, and the global information auxiliary module is used to enhance the learning ability of the bridge disease recognition model for image semantic knowledge.

[0026] In this embodiment, the solution aims to achieve accurate identification of bridge defects through a bridge defect identification model based on unsupervised learning. An image of the bridge to be identified is obtained and input into a well-trained bridge defect identification model, which then gives the corresponding identification result. The bridge defect identification model uses an unsupervised learning strategy to train a preset deep learning network. The deep learning network architecture includes an input module, an output module, and multiple interconnected convolution modules and deconvolution modules. In addition, an attention module and a global information auxiliary module are also incorporated. The input module is connected to the input end of the first convolution module, the output end of the last convolution module is connected to the input end of the first deconvolution module, and the output end of the last deconvolution module is connected to the output module. At the same time, the output end of the input module is connected to the input end of the global information auxiliary module, the output end of the global information auxiliary module is connected to the input end of the attention module, and the output end of the attention module is associated with the output ends of multiple connected convolution modules. The input module receives the bridge image to be identified; the attention module and convolution module extract features from the image; the deconvolution module decodes the image features; and the output module outputs the bridge defect identification results. Notably, the inclusion of the global information auxiliary module effectively enhances the model's ability to learn image semantics. During training, given that the model is based on unsupervised learning and is placed before the attention module, the global information auxiliary module segments the input image into multiple uniformly sized patches. These patches are flattened into one-dimensional vectors and embedded into a high-dimensional feature space via linear projection. Within this space, the module operates on the interaction of query, key, and value. The query represents the query information, the key is a vector that determines the similarity between features, and the value represents the similarity between the query and key. The module then calculates a self-attention score, modeling the dependencies between patches and accurately capturing the global information associations within the image. Subsequently, the attention module further models the dependencies between patches, enhancing the capture of global information. After the operation is complete, the output of the self-attention mechanism is fused with the convolution result using residual connections, allowing the model to obtain image feature representations enhanced by global information while retaining the original feature information. This implicit feature contains rich image semantics and spatial information, significantly improving the accuracy of bridge defect identification.

[0027] In one embodiment of the present invention, the bridge defect recognition model adopts the following strategy during the training process:

[0028] I m =I⊙[M s ]

[0029] M S(X,Y)∈{0,1}

[0030] Where, I m represents the image of the bridge disease sample after pixel masking processing, I represents the input image of the bridge disease sample, ⊙ represents the multiplication operation, M s Stands for pixel masking.

[0031] like Figure 3 As shown, in this embodiment, an image consists of eight sub-images, which are clearly divided into two categories: horizontal pixel masking and vertical pixel masking. When training a bridge defect recognition model, horizontal or vertical pixel masking is performed on the training samples. This greatly enriches the diversity of the training data. At the same time, it effectively improves the model's generalization ability in complex and changing scenarios, enabling the model to handle bridge image recognition tasks in diverse environments. This training strategy, through precise parameter control, generates a series of stripe images of varying scales and orientations. From the perspective of the model's learning mechanism, these images guide the model to conduct a detailed analysis of image texture details and edge features, helping the model build a global contextual information framework and achieve a comprehensive understanding of the image. It is worth noting that bridge defects such as exposed rebar and cracks have typical stripe morphological characteristics. Model training based on the aforementioned masking strategy enables the unsupervised network to accurately identify and locate abnormal structural features such as stripes. Therefore, this method significantly improves the model's effectiveness in bridge defect recognition, providing reliable and efficient technical support for bridge structural health monitoring and possessing important application value in bridge infrastructure maintenance.

[0032] In one embodiment of the present invention, the attention module includes an embedding layer submodule, an intermediate processing submodule, and a multi-layer perception submodule connected in sequence; the embedding layer submodule is used to extract features of the bridge image to be identified, the intermediate processing submodule is used to normalize and capture features of the bridge image to be identified after feature extraction, and the multi-layer perception submodule is used to integrate the features of the bridge image to be identified after normalization and feature capture.

[0033] In this embodiment, the attention module includes an embedding layer submodule, an intermediate processing submodule, and a multi-layer perception submodule connected in series. The embedding layer submodule realizes preliminary feature extraction of the bridge image to be identified by constructing an adaptive convolutional neural network structure, and converts the image information into a feature vector that is easy for the model to process. The intermediate processing submodule uses a layer normalization algorithm to standardize the extracted feature vectors, balance the scales of features in each dimension, and improve the stability of model training. At the same time, with the help of the self-attention mechanism, the model performs weighted processing on the features of different areas of the image, captures key features in a targeted manner, and explores the intrinsic correlation between features. The multi-layer perception submodule is based on a fully connected neural network architecture to further integrate the feature vectors that have been normalized and feature captured. Through nonlinear transformation, the expressive ability of the features is enhanced, and feature representations with higher recognition and semantic information are generated, providing strong support for the precise analysis and decision-making of subsequent models.

[0034] In one embodiment of the present invention, the operation logic of the global information auxiliary module is expressed by the following formula:

[0035] F O =D(E(F I ))

[0036] D=BN(Relu(TransConv(E(F I )))) m

[0037] E=BN(LeakyRelu(Conv(F I ))) m +Transformer(F I )

[0038] Where, F I ,F O Represent the input image and reconstructed image respectively, E and D represent encoding operation and decoding operation respectively, BN, Relu, LeakyRelu, Conv, Transformer, Transformer represent batch normalization processing, Relu activation function, LeakyRelu activation function, convolution, deconvolution and Transformer self-attention mechanism operation respectively, and m represents the number of operations.

[0039] In this embodiment, taking full advantage of the self-attention mechanism's significant advantages in building long-range dependencies, the present invention utilizes the basic convolutional unit to simultaneously learn local spatial information about the image while also implementing a global information auxiliary module centered around the self-attention mechanism during image feature encoding. This approach is intended to significantly improve the model's learning efficiency of image semantics and effectively address the problem of incorrectly identifying some disease information due to the model's limited global perception capabilities. In terms of its operating principle, upon receiving an input image, the global information auxiliary module divides the image into multiple uniformly sized image blocks. These blocks are converted into one-dimensional vectors through a flattening operation and embedded into a high-dimensional feature space via linear projection. Within this high-dimensional feature space, the module calculates self-attention scores based on the interactive operation of query, key, and value, thereby modeling the dependencies between image blocks and accurately capturing global information associations in the image. After the operation, the output of the self-attention mechanism is fused with the result of the convolution operation via a residual connection. This allows the model to obtain a globally enhanced image feature representation while preserving the original feature information. This implicit feature contains rich image semantic and spatial information, providing strong support for image decoding and reconstruction. In the decoding stage, using these high-value features for image reconstruction can obtain comprehensive and accurate image reconstruction results, laying a solid foundation for subsequent bridge disease identification tasks.

[0040] In one embodiment of the present invention, the total loss of the bridge defect identification model is determined by the following formula:

[0041] L total =λ joint L joint +λ adv L adv

[0042] Where, L total Represents the total loss, L joint ,L adv denote the joint loss and adversarial loss, respectively, joint ,λ adv denote the weight of the joint loss and the weight of the adversarial loss, respectively.

[0043] In this embodiment, the present invention constructs a total loss function to drive the backpropagation process. This multi-level joint optimization strategy systematically optimizes the model's loss propagation process from multiple dimensions. This enables accurate identification and effective extraction of bridge defects without labeled data, promoting the model's application in large-scale deployment and cross-scenario portability. In summary, the technical solution proposed in this invention can significantly improve the accuracy of bridge defect identification, effectively overcoming the shortcomings of existing technologies in terms of defect identification accuracy and model versatility.

[0044] In one embodiment of the present invention, the joint loss is determined by the following formula:

[0045] L joint =λ Mae L Mae +λ Ssim L Ssim +λ Style L Style

[0046]

[0047] Where λ Mae ,λ Ssim ,λ Style Represent the weight of each loss function, I U Represent the reconstructed image and the input image respectively, SSIM represents the SSIM value of the image centered at pixel (i, j), Respectively represent the calculation of Gram matrix for the i-th layer feature map φ of the generated image and the original image, E i Represents the expected value of the operation results of all i-layer feature matrices.

[0048] In this embodiment, joint loss: To measure the difference between the reconstructed image and the input image, the present invention constructs a joint loss (joint loss) to calculate the similarity between the input and reconstructed images from three aspects: pixel, structure, and style. The mean absolute error (Maeloss) is used to calculate the pixel-by-pixel difference between the input image and the reconstructed image, and the feature distribution of the original input data is learned from the pixel dimension. However, this method not only ignores the spatial connection between adjacent pixels, but also fails to take into account the high-level semantic information of the image, resulting in poor detail features and overall consistency of the reconstructed image. Based on this, the present invention adds a structural similarity loss function (Ssimloss), which calculates the local structural similarity between the reconstructed image and the original image by comparing the brightness, contrast and structural information between the images, thereby enhancing the perceptual quality and structural stability of the reconstructed image. However, the loss function that only considers pixels and structures still fails to measure image differences from a global perspective. When faced with complex scenes, it is difficult to capture the diverse global style features of the image. Based on this, the present invention jointly adds a style loss function (Styleloss) to perform feature modeling from a global perspective, helping the model learn complex style patterns and comprehensively improve the details, texture and overall perceptual quality of the reconstructed image.

[0049] In one embodiment of the present invention, the adversarial loss is determined by the following formula:

[0050]

[0051] Where, Respectively represent the discrimination of the original image and the generated image. U ) is close to 1, Close to 0.

[0052] In this embodiment, in order to further improve the image reconstruction quality and optimize the disease localization effect, the present invention introduces an adversarial loss mechanism, which is specifically achieved by adding discriminant constraints. As a key component of the adversarial loss, the discriminant constraint can guide the model to generate reconstructed images that are closer to the real sample. During the model training process, the discriminator will distinguish between the input real image and the generated reconstructed image. By continuously adjusting the parameters of the generator and the discriminator, the generator can learn a more effective image reconstruction strategy, thereby improving the quality of the reconstructed image. High-quality reconstructed images help to more accurately identify disease features, thereby optimizing the effect of disease localization and providing a more reliable foundation for subsequent disease analysis and treatment.

[0053] The foregoing description describes specific embodiments of the present invention. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0054] According to another embodiment, the present invention provides a bridge defect identification device based on unsupervised learning. Figure 2 A schematic block diagram of a bridge defect identification device based on unsupervised learning according to an embodiment is shown. It is understood that the device can be implemented by any device, equipment, platform and device cluster with computing and processing capabilities. Figure 2 As shown, the device includes: an acquisition unit 201 and an identification unit 202. The main functions of each component unit are as follows:

[0055] An acquisition unit 201 is configured to acquire an image of a bridge to be identified;

[0056] The recognition unit 202 is configured to input the bridge image to be recognized into a trained bridge defect recognition model to obtain a bridge defect recognition result; wherein the bridge defect recognition model is obtained by training a preset deep learning network using unsupervised learning;

[0057] The deep learning network includes an input module, an output module, multiple convolution modules connected in sequence, multiple deconvolution modules connected in sequence, an attention module and a global information auxiliary module. The input module is connected to the input end of the first convolution module, the output end of the last convolution module is connected to the input end of the first deconvolution module, the output end of the last deconvolution module is connected to the output module, the output end of the input module is connected to the input end of the global information auxiliary module, the output end of the global information auxiliary module is connected to the input end of the attention module, and the output end of the attention module is connected to the input end of the first convolution module; the input module is used to receive the bridge image to be identified, the attention module and the convolution module are used to extract the features of the bridge image to be identified, the deconvolution module is used to decode the features of the bridge image to be identified, the output module is used to output the recognition result of the bridge disease, and the global information auxiliary module is used to enhance the learning ability of the bridge disease recognition model for image semantic knowledge.

[0058] As a preferred embodiment, the bridge defect recognition model adopts the following strategy during the training process:

[0059] I m =I⊙[M s ]

[0060] M S (X,Y)∈{0,1}

[0061] Where, I m represents the image of the bridge disease sample after pixel masking processing, I represents the input image of the bridge disease sample, ⊙ represents the multiplication operation, M s Stands for pixel masking.

[0062] As a preferred embodiment, the attention module includes an embedding layer submodule, an intermediate processing submodule and a multi-layer perception submodule connected in sequence. The embedding layer submodule is used to extract features of the bridge image to be identified, the intermediate processing submodule is used to normalize and capture features of the bridge image to be identified after feature extraction, and the multi-layer perception submodule is used to integrate the features of the bridge image to be identified after normalization and feature capture.

[0063] As a preferred embodiment, the global information auxiliary module is processed by the following formula:

[0064] F O =D(E(F I ))

[0065] D=BN(Relu(TransConv(E(F I )))) m

[0066] E=BN(LeakyRelu(Conv(F I ))) m +Transformer(F I )

[0067] Where, F I ,F O Represent the input image and reconstructed image respectively, E and D represent encoding operation and decoding operation respectively, BN, Relu, LeakyRelu, Conv, Transformer, Transformer represent batch normalization processing, Relu activation function, LeakyRelu activation function, convolution, deconvolution and Transformer self-attention mechanism operation respectively, and m represents the number of operations.

[0068] As a preferred embodiment, the total loss of the bridge defect identification model is determined by the following formula:

[0069] L total =λ joint L joint +λ adv L adv

[0070] Where, L total Denotes the total loss, L joint ,L adv Represent the joint loss and the adversarial loss, λ joint ,λ adv denote the weight of the joint loss and the weight of the adversarial loss, respectively.

[0071] As a preferred embodiment, the combined loss is determined by the following formula:

[0072] L joint =λ Mae L Mae +λ Ssim L Ssim +λ Style L Style

[0073]

[0074]

[0075] Where λ Mae ,λ Ssim ,λ Style Represent the weight of each loss function, I URepresent the reconstructed image and the input image respectively, SSIM represents the SSIM value of the image centered at pixel (i, j), Respectively represent the calculation of Gram matrix for the i-th layer feature map φ of the generated image and the original image, E i Represents the expected value of the operation results of all i-layer feature matrices.

[0076] As a preferred implementation, the adversarial loss is determined by the following formula:

[0077]

[0078] Where, Respectively represent the discrimination of the original image and the generated image.

[0079] According to another embodiment, there is also provided a computer readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute a combination of Figure 1 The method described.

[0080] According to another embodiment, an electronic device is provided, including a memory and a processor, wherein the memory stores an executable code, and when the processor executes the executable code, the Figure 1 method.

[0081] The various embodiments of the present invention are described in a progressive manner. Similar portions between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the device embodiments are generally similar to the method embodiments, so their description is relatively simple. For relevant portions, refer to the description of the method embodiments.

[0082] Those skilled in the art will appreciate that, in one or more of the above examples, the functions described herein may be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions may be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0083] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. A bridge defect identification method based on unsupervised learning, characterized in that: include: Obtaining a bridge image to be identified; Inputting the bridge image to be identified into a trained bridge defect recognition model to obtain a bridge defect recognition result; wherein the bridge defect recognition model is obtained by training a preset deep learning network using unsupervised learning; The deep learning network includes an input module, an output module, multiple convolution modules connected in sequence, multiple deconvolution modules connected in sequence, an attention module and a global information auxiliary module. The input module is connected to the input end of the first convolution module, the output end of the last convolution module is connected to the input end of the first deconvolution module, the output end of the last deconvolution module is connected to the output module, the output end of the input module is connected to the input end of the global information auxiliary module, the output end of the global information auxiliary module is connected to the input end of the attention module, and the output end of the attention module is connected to the input end of the first convolution module; the input module is used to receive the bridge image to be identified, the attention module and the convolution module are used to extract the features of the bridge image to be identified, the deconvolution module is used to decode the features of the bridge image to be identified, the output module is used to output the recognition result of the bridge disease, and the global information auxiliary module is used to enhance the learning ability of the bridge disease recognition model for image semantic knowledge.

2. The method according to claim 1, characterized in that The bridge defect recognition model adopts the following strategy during the training process: I m =I⊙[M s ] M S (X,Y)∈{0,1} Where, I m represents the image of the bridge disease sample after pixel masking processing, I represents the input image of the bridge disease sample, ⊙ represents the multiplication operation, M s Stands for pixel masking.

3. The method according to claim 2, characterized in that The attention module includes an embedding layer submodule, an intermediate processing submodule and a multi-layer perception submodule connected in sequence. The embedding layer submodule is used to extract features of the bridge image to be identified, the intermediate processing submodule is used to normalize and capture features of the bridge image to be identified after feature extraction, and the multi-layer perception submodule is used to integrate the features of the bridge image to be identified after normalization and feature capture.

4. The method according to claim 3, characterized in that The global information auxiliary module is processed by the following formula: F O =D(E(F I )) D=BN(Relu(TransConv(E(F) I )))) m E=BN(LeakyRelu(Conv(F I ))) m +Transformer(F I ) Where, F I ,F O Represent the input image and reconstructed image respectively, E and D represent encoding operation and decoding operation respectively, BN, Relu, LeakyRelu, Conv, Transformer, Transformer represent batch normalization processing, Relu activation function, LeakyRelu activation function, convolution, deconvolution and Transformer self-attention mechanism operation respectively, and m represents the number of operations.

5. The method according to claim 1, wherein The total loss of the bridge defect identification model is determined by the following formula: L total =λ joint L joint +λ adv L adv Where, L total Denotes the total loss, L joint ,L adv denote the joint loss and adversarial loss, respectively, joint ,λ adv denote the weight of the joint loss and the weight of the adversarial loss, respectively.

6. The method according to claim 5, characterized in that The combined loss is determined by the following formula: L joint =λ Mae L Mae +λ Ssim L Ssim +λ Style L Style Where λ Mae ,λ Ssim ,λ Style Represent the weight of each loss function, I U Represent the reconstructed image and the input image respectively, SSIM represents the SSIM value of the image centered at pixel (i, j), Respectively represent the calculation of Gram matrix for the i-th layer feature map φ of the generated image and the original image, E i Represents the expected value of the operation results of all i-layer feature matrices.

7. The method according to claim 6, characterized in that The adversarial loss is determined by the following formula: Where, Respectively represent the discrimination of the original image and the generated image.

8. A bridge defect identification device based on unsupervised learning, characterized in that: include: an acquisition unit configured to acquire an image of a bridge to be identified; The recognition unit is configured to input the bridge image to be recognized into a trained bridge defect recognition model to obtain a bridge defect recognition result; wherein the bridge defect recognition model is obtained by training a preset deep learning network using unsupervised learning; The deep learning network includes an input module, an output module, multiple convolution modules connected in sequence, multiple deconvolution modules connected in sequence, an attention module and a global information auxiliary module. The input module is connected to the input end of the first convolution module, the output end of the last convolution module is connected to the input end of the first deconvolution module, the output end of the last deconvolution module is connected to the output module, the output end of the input module is connected to the input end of the global information auxiliary module, the output end of the global information auxiliary module is connected to the input end of the attention module, and the output end of the attention module is connected to the input end of the first convolution module; the input module is used to receive the bridge image to be identified, the attention module and the convolution module are used to extract the features of the bridge image to be identified, the deconvolution module is used to decode the features of the bridge image to be identified, the output module is used to output the recognition result of the bridge disease, and the global information auxiliary module is used to enhance the learning ability of the bridge disease recognition model for image semantic knowledge.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

Images