Concrete structure damage image restoration method

Abstract

The invention relates to a concrete structure damage image restoration method, which adopts a dense multi-scale fusion network model and is applied to the field of concrete structure damage image restoration. According to the method, a special pixel-level image inpainting data set containing an occlusion image is collected and marked, the data set is marked and grouped, the data set is expanded through an image enhancement method, and an original data set for image inpainting is obtained. And then training the dense multi-scale fusion network model by using the model, and repairing the concrete structure damage image by using the trained model. Through the construction of a data set, the performance of the model in accurately removing shielding and pollution and repairing a shielded crack image is improved. Compared with the prior art, the method has the advantages of quickly, effectively and automatically cleaning complex shielding and pollution in the concrete structure image, repairing shielded cracks, enhancing the adaptability and generalization ability of an image repairing model and reducing manual intervention.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for repairing a damaged image of a concrete structure. Background Art

[0002] With the rapid development of computer vision and deep learning technologies, image analysis has become an important means of evaluating the safety of concrete structures. Concrete structures may suffer from damage such as concrete cracking during service or after natural disasters such as earthquakes. The correct identification of damage image signals such as cracks is crucial for quickly and accurately evaluating the structural performance status, guiding post-disaster rescue operations, and restoring and rebuilding work. However, during the image acquisition process, the surface of the concrete structure is often blocked by pipelines or caused by surface pollution.

[0003] There are existing technical methods for image enhancement, denoising and restoration in medical images, satellite images and some general images. However, when it comes to image restoration of concrete structures, especially in the assessment of the damage status of damaged structures, existing technologies often find it difficult to cope with complex occlusion and contamination problems.

[0004] Damaged building structure images usually contain complex structural features and possible crack distribution. In particular, the acquisition of crack images is often affected by occlusion and pollution in practical applications. Existing image restoration technologies are often unable to effectively identify and handle these specific problems, resulting in the inability of restored images to meet the needs of accurate structural damage assessment. In addition, existing technologies also have limitations in processing large-scale data sets and achieving real-time damage assessment. Summary of the invention

[0005] The purpose of the present invention is to provide a method for repairing a damaged image of a concrete structure in order to overcome the defects of the above-mentioned prior art, so as to realize the repair of the image of the concrete structure under occlusion and pollution.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to one aspect of the present invention, a method for repairing a damaged image of a concrete structure is provided. The method applies a dense multi-scale fusion network model designed based on a generative adversarial network architecture to the field of concrete structure image repair.

[0008] The method steps include:

[0009] S1, build a dense multi-scale fusion network architecture;

[0010] S2. Build a concrete component image restoration dataset;

[0011] S3, using the self-built concrete component image restoration dataset described in S2 to train the dense multi-scale fusion network architecture described in S1, to obtain a concrete structure damage image restoration model;

[0012] S4, inputting the damaged image of the concrete structure to be repaired into the damaged image repair model of the concrete structure for repair;

[0013] S5. Output the damaged image of the repaired concrete structure.

[0014] As a preferred technical solution, the process of obtaining the concrete component image restoration dataset in S2 is as follows:

[0015] Annotate the occlusion, pollution and crack parts in the original image pixel by pixel;

[0016] Convert the generated file after annotation into the mask image required for model training and save it;

[0017] Perform image enhancement, including cropping, rotation, and flipping operations;

[0018] The image set obtained after enhancement processing is obtained into a data set, and the data set is grouped into training set, validation set and test set.

[0019] Furthermore, the image to be enhanced includes an original image and a mask image.

[0020] Furthermore, the cropping operation adopts a sliding window method, that is, randomly selecting sub-regions in the original large-size image for cropping.

[0021] Furthermore, the original image refers to a picture of a concrete structural component with occluded and contaminated parts obtained from an actual structure, and the original image includes the entire process of the concrete structural component from the absence of cracks to the appearance of cracks to destruction.

[0022] Furthermore, the training set, the validation set, and the test set jointly train the dense multi-scale fusion network architecture described in S1; the training process includes: training the dense multi-scale fusion network architecture using the training set; validating the dense multi-scale fusion network architecture using the validation set; and testing the dense multi-scale fusion network architecture using the test set.

[0023] As a preferred technical solution, the training in S3 is controlled by the overall loss, and the training is terminated after the overall loss becomes stable.

[0024] Furthermore, the overall loss is the weighted sum of perceptual loss, reconstruction loss, and adversarial loss;

[0025] The perceptual loss is used to measure the difference between the generated image and the target image in terms of high-level feature representation;

[0026] The reconstruction loss is used to measure the difference between the generated image and the target image at the pixel level;

[0027] The adversarial loss is used to train the generator to generate data that approximates the true data distribution.

[0028] As a preferred technical solution, the best model is saved after iterating according to the preset maximum number of iterations during training; the training progress and loss value are displayed in the command line interface, and recorded every preset iteration cycle; according to the overall loss function, the loss reduction process during model training is recorded.

[0029] As a preferred technical solution, the repair in S4 includes removing occlusions and pollutions in the damaged image of the concrete structure to be repaired, and repairing the image after removing the occlusions, pollutions and cracks.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] 1. Effectively handle occlusion and pollution problems in concrete structure images. The dense multi-scale fusion network model based on the generative adversarial network is applied to the repair of damaged concrete structure images. Traditional image restoration technology cannot effectively handle occlusion and pollution problems in concrete structure images. On the one hand, the present invention obtains a data processing network with crack distribution characteristics of concrete structural components by constructing a data set, data enhancement technology, and data training; in addition, by fine-tuning the dense multi-scale fusion network model, it can adapt to the characteristics of concrete structure images and improve the repair effect;

[0032] 2. The model has good adaptability and strong generalization ability. In order to train and fine-tune the dense multi-scale fusion network model, the present invention establishes a high-precision pixel-level image restoration dataset containing contaminated images and their corresponding annotations. This dataset is formed by annotating the high-definition images collected in the concrete component structure engineering pixel by pixel and undergoing necessary data enhancement. The dataset provides a wealth of training samples for the model, which helps to improve the generalization ability and restoration accuracy of the model. By adjusting the loss function and training strategy of the model, the model can better handle occlusion and contamination problems when repairing concrete structure images, and generate more realistic and accurate restoration results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Flow chart of the concrete structure damage image repair method of the present invention

[0034] Figure 2 Flowchart for generating a trained dense multi-scale fusion network model;

[0035] Figure 3 is a schematic diagram of the original image;

[0036] Figure 4 This is a schematic diagram of the image after marking;

[0037] Figure 5 The function code diagram for converting the file generated after annotation into the mask image required for model training;

[0038] Figure 6 is a schematic diagram of a mask image;

[0039] Figure 7 Schematic diagram of the change of perceptual loss with the number of iterations;

[0040] Figure 8 Schematic diagram of how the image reconstruction loss changes with the number of iterations;

[0041] Fig. 9 Schematic diagram of the change of overall loss with the number of iterations;

[0042] Fig.10 Schematic diagram for comparing the original image and the restored image;

[0043] Fig.11 This is the DMFN architecture diagram;

[0044] Fig.12 This is a schematic diagram of the DMFB module structure. DETAILED DESCRIPTION

[0045] 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 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 should fall within the scope of protection of the present invention.

[0046] In this invention, a dense multi-scale fusion network model is used and applied to a new field: the restoration of concrete structure images. In order to enable the dense multi-scale fusion network model to effectively handle the occlusion and pollution problems in concrete structure images, a special pixel-level image restoration dataset is collected and annotated, and the dense multi-scale fusion network model is trained based on the dataset to obtain a concrete structure building damage image restoration model. Through this process, not only the applicability of the dense multi-scale fusion network model in concrete structure image restoration is verified, but also the performance of the model is further improved through the dataset, and the adaptability and generalization ability of the model are enhanced.

[0047] Example

[0048] In this embodiment, a method for repairing a damaged image of a concrete structure is invented. The process of the method for repairing a damaged image of a concrete structure is as follows: Figure 1 As shown in Figure 2, the process of generating a trained dense multi-scale fusion network model is as follows: Figure 2 shown.

[0049] In this embodiment, regarding the generation of adversarial networks, the details are as follows:

[0050] Generative Adversarial Networks (GAN) is a neural network architecture. GAN is an unsupervised learning method based on game theory, which optimizes model performance through the confrontation process of two neural networks: generator and discriminator. This structure introduces a unique training mechanism, in which the two models compete with each other during the training process to improve their respective performance.

[0051] In the architecture of the Generative Adversarial Network, the task of the Generator (G) is to capture the distribution of the training data and try to generate data that looks real. These generated data are then fed into the Discriminator (D). The Discriminator acts like a binary classifier, and its task is to distinguish whether the input data comes from the actual training set (labeled as true) or is generated by the Generator (labeled as false). Through the continuous confrontation between the Generator and the Discriminator, the Generator learns to create more and more realistic data, while the Discriminator strives to improve its ability to distinguish true and false data.

[0052] In this embodiment, the training process of the generative adversarial network can be described as a two-party zero-sum game, in which the generator tries to "cheat" the discriminator so that it misjudges false data as real data. At the same time, the discriminator continuously adjusts its own parameters to better identify true and false data. This process continues until a state of equilibrium is reached, that is, the discriminator can no longer distinguish between real data and generated data. At this time, we believe that the training of the generative adversarial network has converged.

[0053] In this embodiment, the typical architecture of the generative adversarial network is as follows:

[0054] The design of GAN not only performs well in the field of image generation, but also shows wide application potential in many fields such as art creation, video generation and audio generation. In image processing, GAN can generate high-quality pictures for tasks such as image restoration, style transfer and virtual reality. In addition, it also shows excellent ability in processing missing data reconstruction, especially when facing complex scenes and detailed textures, surpassing traditional methods.

[0055] Regarding the DMFN model, the details are as follows:

[0056] DMFN is a dense multi-scale fusion network (DMFN) for image fine restoration. The network combines self-guided regression loss and geometric alignment constraints to significantly improve the quality of generated images.

[0057] Compared with other image restoration networks, DMFN fuses multi-scale features with different dilation rates and adopts dense multi-scale fusion blocks (DMFB) as basic building blocks. Compared with traditional sparse multi-scale feature fusion methods, it can extract features more effectively while maintaining a reasonable parameter scale. In order to generate images with realistic and semantic structures, a self-guided regression loss is proposed, which constrains the low-level features of the generated content through normalized difference maps (the difference between the output and the target) to enhance semantic details. In addition, a geometric alignment constraint is introduced to compensate for the pixel distance between the predicted features and the real features. A discriminator with local and global branches is adopted to ensure the consistency of local-global content. By introducing discriminator feature matching on the local branch, the similarity of intermediate features between the synthetic image and the real image patch is dynamically minimized.

[0058] Extensive experiments on multiple public datasets show that DMFN is able to generate images with global semantic structures and fine textures, and especially for architectural and facial images, the generated images are closer to real images.

[0059] The architecture of DMFN is as follows Fig.11 As shown in Figure 2, the framework of the model consists of a generator and a discriminator with two branches. The generator produces reasonable drawing results, and the discriminator performs adversarial training.

[0060] Among them, Conv-3 represents a convolutional layer with a convolution kernel size of 3×3; Conv-5 represents a convolutional layer with a convolution kernel size of 3×3;

[0061] In this embodiment, the DMFN architecture is as follows:

[0062] The core module of DMFN: Dense Multi-scale Fusion Block (DMFB) is mainly used to connect the convolutional layers in the generator in the entire model to achieve the function of generating multi-scale features. Fig.11 In the generator, DMFB is marked with a yellow module.

[0063] In this embodiment, the DMFB module structure is as follows:

[0064] Among them, Conv-3-8 represents a 3×3 convolutional layer with a dilation rate of 8; ⊕ represents the calculation of element-by-element summation.

[0065] Depend on Fig.12 As can be seen, the input image will first undergo a 3×3 convolution to reduce the dimension of the input image; then the image will undergo convolutions with four different dilation rates to obtain features of different scales. The slightly smaller scale features are fused two by two, and then a simple 1×1 convolution is used to fuse the concatenated features. DMFB specifically enhances the general dilated convolution, allowing the network to obtain dense multi-scale features, while having fewer parameters than large kernel convolution networks.

[0066] In the invention, the DMFN model is used and applied to a new field: the restoration of concrete structure images. In order to enable the DMFN model to effectively handle occlusion and pollution problems in concrete structure images, we collected and annotated a special pixel-level image restoration dataset, and trained a dense multi-scale fusion network model based on the dataset to obtain a concrete structure building damage image restoration model. Through this process, we not only verified the applicability of the DMFN model in concrete structure image restoration, but also further improved the performance of the model through our dataset, reflecting the innovation and practical application value of the present invention.

[0067] In this embodiment, a dense multi-scale fusion network architecture is first constructed; then a concrete component image restoration dataset is constructed.

[0068] In this embodiment, a pixel-level image restoration dataset containing concrete structure pollution and occlusion images is collected, and the dataset is accurately labeled and grouped. The dataset images are from photos of low-cycle repeated static loading projects of concrete opening shear walls. The photos were taken with a NIKON D90 camera, and the shooting angle was facing the wall. The original image is as follows Figure 3 As shown in the figure, the wires of the displacement meter and strain gauge are complicated. The resolution of the original image is 4288×2848.

[0069] In this embodiment, the original photo of the concrete shear wall contains multiple crack development states from no cracks to cracks to destruction. To facilitate further annotation work, the image processing tool is first used to crop the image, retain the shear wall part in the original image, and remove the external environment. The image resolution is adjusted to 2048×2048. The data set contains 231 pictures of the whole wall. After data enhancement, the number of pictures in the data set reaches 14784.

[0070] In this embodiment, a supervised learning algorithm based on a generative adversarial network is used, and the model needs to input a contaminated image and an annotation of the contaminated part for training. The contaminated image refers to a shear wall image, and the annotation of the contaminated part requires manually marking the obstructing wires in the image based on the shear wall photo.

[0071] In this embodiment, image annotation software is used to annotate the obstructing wires in the shear wall photo. In the annotation software, the image is annotated point by point with red polylines, and a polygon connected end to end is synthesized to frame the area where the wires are located. The image obtained after annotation is as follows: Figure 4 shown.

[0072] In this embodiment, the result of Labelme annotation is in the form of a .json file. The generated .json file needs to be converted into a mask image required for model training and saved through the json_to_mask function. The json_to_mask function code is as follows: Figure 5 As shown. The mask image is a grayscale image, such as Figure 6 As shown, there are only black and white in the image, where white represents the occluded and contaminated pixels in the original image, and black represents the background pixels.

[0073] In this embodiment, the data set is expanded by using image enhancement technology (such as rotation, flipping, etc.) to obtain the original data set for image restoration, so as to enrich the image samples required for model training.

[0074] In this embodiment, in order to facilitate subsequent image enhancement processing, including cropping and rotation operations, the corresponding mask image is also processed synchronously to ensure data consistency and accuracy. This method ensures the effectiveness of model training and helps the image restoration model to more accurately understand and process specific features in the shear wall image.

[0075] The image enhancement process includes image cropping, image rotation, and image flipping.

[0076] Large datasets are a prerequisite for the success of deep neural networks. Image enhancement is a commonly used technique in the field of computer vision that generates new, changed image data by applying a series of transformations to the original image. These transformations include rotation, flipping, scaling, cropping, color adjustment (such as brightness, contrast adjustment), adding noise, etc. In deep learning and machine learning, the performance of the model depends largely on the diversity and quantity of training data. For limited or small datasets, image enhancement is an effective way to increase the size of the dataset and thus provide more training samples. At the same time, by introducing various changes in images for graphic enhancement, image enhancement helps the model learn more general features, reduce overfitting, and improve the model's generalization ability in new scenarios.

[0077] In this embodiment, the original image size is 2048×2048 pixels. If it is used directly for model training, it will bring a huge computational burden. To solve this problem, we adjusted the image size to 512×512 pixel resolution, which the model can process more efficiently. The cropping process uses a sliding window method to randomly select sub-areas in the original large-size image for cropping, ensuring the diversity and richness of the training data.

[0078] In this embodiment, the image rotation is to ensure the diversity of the data set and improve the generalization ability of the data set. The specific operation is to rotate the image cropped by the sliding window and select images with rotation angles of 90 degrees, 180 degrees and 270 degrees for further processing.

[0079] In this embodiment, the image flipping is to increase the diversity of the data set, thereby improving the adaptability and generalization ability of the model for images of different perspectives and directions. The specific steps are to use horizontal flipping and vertical flipping for image enhancement. By introducing the flipped image, the model learns that the direction of the occlusion is not a decisive factor in the recognition task, thereby improving the recognition ability of unseen occlusions with different directions.

[0080] The three image enhancement methods mentioned above expand the dataset to 14,784 images, and the same transformation operation is performed on the annotated masks to obtain the original dataset of image restoration. After image augmentation, the training set contains 11,840 images, and the validation set and test set each contain 1,472 images.

[0081] In this embodiment, after obtaining the original data set, the original data set is grouped according to the ratio of 80% training set, 10% validation set and 10% test set, and then the dense multi-scale fusion network architecture is trained using the self-built concrete component image restoration data set to obtain a concrete structure damage image restoration model;

[0082] In this embodiment, a dense multi-scale fusion network model is used to train the data set after data enhancement. The model can effectively extract feature information of concrete structure images through multi-scale feature fusion.

[0083] In this embodiment, a unified platform is used, Python is used as the main development language, and the model is built based on the PyTorch framework. The model training is based on a dense multi-scale fusion network architecture. The model training is performed on a cloud platform.

[0084] Table 1 Cloud platform configuration table used for training

[0085]

[0086] During the training process, the best model is saved after iterating according to the preset maximum number of iterations; the training progress and loss value are displayed in the command line interface, and recorded every preset iteration cycle; according to the overall loss function, the loss reduction process during model training is recorded.

[0087] In this embodiment, the preset maximum number of iterations is 10,000 times, and the preset iteration cycle is 50 times.

[0088] In this embodiment, the training process is controlled by the overall loss, and the best model is saved after every 10,000 iterations. The training progress and loss value are displayed in the command line interface, and are recorded every 50 iterations. According to the different loss functions, the loss reduction process during model training is recorded. Specifically, the training software and hardware configuration of the cloud platform is shown in Table 1.

[0089] The model training process went through nearly 80,000 iterations, and the training was terminated after the overall loss approached stability.

[0090] Specifically, the overall loss is a weighted combination of reconstruction loss, perceptual loss, and adversarial loss, which is used to control the training process. The purpose of the overall loss is to balance these different losses to generate images that are accurate at the pixel level, expressive at the feature level, and perform well in adversarial tasks. The overall loss during training decreases as Figure 7 The formula used to calculate the overall loss in this embodiment is as follows:

[0091] L total =L re +ηL prec +μL gan ;

[0092] Wherein, η and μ are weighting coefficients, which are 25 and 0.03 respectively in this embodiment.

[0093] The overall loss is made up of the following categories of losses:

[0094] Perceptual loss: The model uses perceptual loss to measure the difference between the generated image and the target image in high-level feature representation. This is achieved by forward propagating the generated image and the target image in the VGG network and calculating the difference between their feature maps on multiple layers. Perceptual loss is defined by the following formula:

[0095]

[0096] Among them, φ i represents the activation function in the i-th layer VGG network; I out Indicates the generated image; I g Represents the target image.

[0097] The perceptual loss helps the generated images to be more visually realistic, capturing texture and shape features that are more recognizable to the human eye. The perceptual loss during training is reduced as Figure 8 shown.

[0098] Reconstruction loss:

[0099] The model uses reconstruction loss (also called pixel loss) to measure the difference between the generated image and the target image at the pixel level. It is achieved by calculating the difference in pixel values ​​between the generated image and the real image. In this example, L1 loss is used to define the reconstruction loss of the model:

[0100] L re =‖I out -I g ‖1;

[0101] This loss helps ensure that the generated images are similar to the real images at the pixel level. The reconstruction loss during training is reduced as Fig. 9 shown.

[0102] Adversarial loss: Adversarial loss is a core concept in generative adversarial networks, which is used to train the generator to generate data that is as close to the real data distribution as possible. In the GAN framework, the generator tries to generate fake images to "fool" the discriminator, and the discriminator's task is to distinguish between real and fake images. Adversarial loss reflects the performance of the generator in this adversarial game, usually the degree to which the fake images generated by the generator are correctly identified by the discriminator. The adversarial loss of the model is calculated as follows:

[0103]

[0104] In this embodiment, the trained dense multi-scale fusion network model is used to repair the damaged image of the concrete structure. The model automatically removes pollution and occlusion in the image through multi-scale feature learning, and repairs the damaged part. The damaged image of the concrete structure to be repaired is input into the damaged image repair model of the concrete structure; the trained model is used to repair the image in the test set, and the repaired damaged image of the concrete structure is output.

[0105] Since the data set is segmented at the level of the original image, the generated test results can be pieced together into a complete image. Fig.10 shown. Fig.10 (a) is the original image, and (b) is the repaired image. It can be seen from the images that the wires and shadows in the image are cleaned up relatively cleanly, and the cracks blocked by the wires are also restored.

[0106] In this example, the applicability and superiority of the model in image restoration of concrete structures were verified through image comparison and performance index evaluation after restoration. In general, the restoration image was cleaned up in the main crack area, which met the expectations of image restoration of concrete structure damage.

[0107] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for repairing a damaged concrete structure image, characterized in that: The dense multi-scale fusion network model designed based on the generative adversarial network architecture is applied to the field of concrete structure image restoration, and the dense multi-scale fusion network model is trained using a concrete component image restoration dataset adapted to the crack characteristics of concrete structures. The method steps include: S1, build a dense multi-scale fusion network architecture; S2. Build a concrete component image restoration dataset; S3, using the self-built concrete component image restoration dataset in S2 to train the dense multi-scale fusion network architecture described in S1, to obtain a concrete structure damage image restoration model; S4, inputting the damaged image of the concrete structure to be repaired into the damaged image repair model of the concrete structure for repair; S5. Output the damaged image of the repaired concrete structure.

2. A concrete structure damage image restoration model according to claim 1, characterized in that: The process of obtaining the concrete component image restoration dataset in S2 is as follows: Annotate the occlusion, pollution and crack parts in the original image pixel by pixel; Convert the generated file after annotation into the mask image required for model training and save it; Perform image enhancement, including cropping, rotation, and flipping operations; The image set obtained after enhancement processing is obtained into a data set, and the data set is grouped into training set, validation set and test set.

3. A concrete structure damage image repair model according to claim 2, characterized in that: The image to be enhanced includes an original image and a mask image.

4. A concrete structure damage image repair model according to claim 2, characterized in that: The cropping operation adopts a sliding window method, that is, randomly selecting a sub-region in the original large-size image for cropping.

5. The concrete structure damage image repair model according to claim 2 is characterized in that: The original image refers to the picture of the concrete structure component with blocked and contaminated parts obtained during the experiment, including the whole process of the concrete structure component from the absence of cracks to the appearance of cracks and then to destruction.

6. A concrete structure damage image repair model according to claim 2, characterized in that: The training set, validation set, and test set jointly train the dense multi-scale fusion network architecture described in S1; The training process includes: using the training set to train the dense multi-scale fusion network architecture; using the validation set to verify the dense multi-scale fusion network architecture; and using the test set to test the dense multi-scale fusion network architecture.

7. The concrete structure damage image repair model according to claim 1 is characterized in that: The training in S3 is controlled by the overall loss, and the training is terminated after the overall loss becomes stable.

8. The concrete structure damage image repair model according to claim 7 is characterized in that: The overall loss is the weighted sum of perceptual loss, reconstruction loss and adversarial loss; The perceptual loss is used to measure the difference between the generated image and the target image in terms of high-level feature representation; The reconstruction loss is used to measure the difference between the generated image and the target image at the pixel level; The adversarial loss is used to train the generator to generate data that approximates the true data distribution.

9. A concrete structure damage image repair model according to claim 8, characterized in that: During the training process, the best model is saved after iterating according to a preset maximum number of iterations; the training progress and loss value are displayed in the command line interface, and are recorded every preset iteration cycle; according to the overall loss function, the loss reduction process during model training is recorded.

10. The concrete structure damage image restoration model according to claim 1, characterized in that: The repair in S4 includes removing the occlusion, pollution and crack parts in the damaged image of the concrete structure to be repaired and repairing the image after removing the occlusion, pollution and crack parts.

Images