A gan data enhancement and peeling severity classification method and system based on adaptive network topology grey wolf optimization
By optimizing the training control parameters of the generative adversarial network using the adaptive network topology gray wolf optimizer and combining it with the small-world network topology mechanism, high-quality spalling images are generated and an enhanced dataset is constructed. This solves the problems of sample imbalance and training instability in concrete spalling recognition, and improves the accuracy and classification performance of severe spalling recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI POLYTECHNIC UNIV
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-14
Smart Images
Figure CN122391726A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computer vision and artificial intelligence, and in particular to a method and system for data augmentation and peeling severity classification of GANs with adaptive network topology gray wolf optimization. Background Technology
[0002] Concrete structures, as the most widely used foundation components in civil engineering, directly affect the stability and service life of engineering structures through their service safety. During long-term service, concrete surfaces are prone to defects such as spalling and cracking due to environmental erosion, load effects, and material aging, with spalling being particularly common and causing significant damage. With the development of computer vision and artificial intelligence technologies, image-based automatic identification methods for concrete defects have gradually become a research hotspot. Existing technologies typically employ deep convolutional neural networks to extract features and classify concrete surface images to achieve automatic identification of spalling defects. However, these methods are highly dependent on the scale and distribution of training data. In practical engineering scenarios, the difficulty in obtaining severe spalling samples and the high cost of data collection often result in a long-tailed sample distribution, meaning a small number of samples in the severe spalling category. This leads to insufficient recognition ability of the classification model for a few categories, easily resulting in missed detections or misjudgments.
[0003] To alleviate the problem of data imbalance, existing methods typically employ traditional data augmentation techniques such as image flipping, rotation, and cropping, or utilize Generative Adversarial Networks (GANs) to generate synthetic samples to expand the dataset. However, traditional augmentation methods struggle to generate peeling images with complex texture features, while GANs are highly sensitive to parameter settings during training, easily leading to training instability or mode collapse, resulting in low-quality generated images and failing to effectively improve the performance of classification models.
[0004] Furthermore, the optimization of GAN training parameters often relies on empirical settings or conventional swarm intelligence optimization algorithms for parameter tuning. However, existing optimization methods are prone to getting stuck in local optima during the search process and lack an effective global and local search balancing mechanism. This makes it difficult to obtain the optimal parameter combination for complex texture generation tasks, thus limiting the application effect of generative adversarial networks in engineering image enhancement.
[0005] Therefore, how to construct a data augmentation and classification method that can stably generate high-quality spalling images and effectively improve the ability to identify severe spalling when concrete spalling samples are unevenly distributed has become a technical problem that urgently needs to be solved in this field. Summary of the Invention
[0006] In view of the above-mentioned problems, the present invention provides a GAN data augmentation and spalling severity classification method and system with adaptive network topology gray wolf optimization to solve the problems of sample imbalance, unstable training of generation model and easy getting trapped in local optima in concrete spalling identification in the prior art, thereby improving the quality of spalling image generation and the accuracy of identification of severe spalling disease.
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0008] In a first aspect, embodiments of the present invention provide a GAN data augmentation and peeling severity classification method with adaptive network topology gray wolf optimization, comprising: acquiring an original image of a concrete structure surface, extracting physical morphological features of the peeling area, establishing a severity level criterion based on the physical morphological features, dividing the original image into different peeling levels, and constructing an original sample dataset.
[0009] Generative Adversarial Network (GAN) is constructed based on the original sample dataset, and an adaptive network topology gray wolf optimizer is established. The training control parameters of the GAN are mapped to the optimization space of the adaptive network topology gray wolf optimizer, and the training control parameters are adaptively optimized. The training control parameters are iteratively optimized using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters, and a generative adversarial network is trained based on the optimal training control parameters to generate synthetic images corresponding to the peeling level.
[0010] The synthesized image is fused with the original sample dataset to construct an enhanced dataset;
[0011] A deep convolutional neural network classification model is constructed, and the deep convolutional neural network classification model is trained based on the augmented dataset to output the assessment results of the severity level of concrete spalling.
[0012] As a preferred embodiment of the adaptive network topology gray wolf optimization GAN data augmentation and peeling severity classification method described in this invention, the method includes: extracting the physical morphological features of the peeled region, establishing a severity level criterion based on the physical morphological features, dividing the original image into different peeling levels, and constructing an original sample dataset, including:
[0013] The extraction of physical morphological features of the peeled area includes: calculating the area ratio of the peeled area and texture feature parameters that characterize the degree of surface texture damage;
[0014] The method of establishing a severity level criterion based on the physical morphological features includes: setting corresponding threshold ranges according to the proportion of the peeling area and the degree of texture damage, and classifying the severity of peeling according to the combination relationship of the threshold ranges;
[0015] Specifically, when the proportion of peeling area and the degree of texture damage both exceed the corresponding preset threshold, the corresponding sample is judged as a severe peeling level; when both are below the corresponding threshold, it is judged as a no-peeling level; and in other cases, it is judged as a slight peeling level.
[0016] As a preferred embodiment of the adaptive network topology gray wolf optimization GAN data augmentation and peeling severity classification method described in this invention, the method involves mapping the training control parameters of the generative adversarial network to the optimization space of the adaptive network topology gray wolf optimizer, and adaptively optimizing the training control parameters, including:
[0017] The training control parameters of the generative adversarial network are constructed into a multidimensional parameter vector, and the multidimensional parameter vector is used as the position representation of the individual gray wolf to form an optimization space corresponding to the training control parameters.
[0018] An adaptive driving mechanism is set up, with the rate of change of the convergence speed of the FID score of the generated image during the training of the adversarial network as the trigger signal. When the trigger signal meets the preset conditions, a topology embedded search mechanism is introduced in the optimization iteration process of the gray wolf individuals. By dynamically defining and reorganizing the information interaction path between gray wolf individuals, the multidimensional parameter vector is iteratively updated to obtain the optimized training control parameters.
[0019] As a preferred embodiment of the adaptive network topology gray wolf optimizer for GAN data augmentation and peeling severity classification described in this invention, the method includes: iteratively optimizing the training control parameters using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters, including:
[0020] In the optimization space of the adaptive network topology gray wolf optimizer, boundary constraints of the training control parameters are set, upper and lower thresholds are set for continuous parameters, and a mapping rule for discrete parameters is constructed to round to the nearest integer.
[0021] Under the condition of satisfying the adaptive driving mechanism, the search path of individual gray wolves is dynamically adjusted by using the small-world network topology mechanism, and the FID score of the generated image is used as the fitness index for iterative update.
[0022] The optimal training control parameters are obtained by gradually approximating the global optimal solution through iterative optimization.
[0023] As a preferred embodiment of the adaptive network topology gray wolf optimization GAN data augmentation and stripping severity classification method described in this invention, the small-world network topology mechanism includes:
[0024] Initially, a regular ring network topology is established, and each optimization individual only interacts with its K neighboring individuals locally.
[0025] By introducing reconnection probability, the connection relationships between individuals are randomly reorganized to form a small-world topology with a high clustering coefficient and a short average path length;
[0026] During the optimization iteration process, the individual position is updated in conjunction with the influence of topological neighbors. The specific update formula is as follows:
[0027]
[0028] in, The optimal individual location in the neighborhood is determined by small-world topological connections. Let i be the current position of the i-th individual in t iterations. This represents the current globally optimal individual position. , These are the weighting coefficients.
[0029] As a preferred embodiment of the adaptive network topology gray wolf optimization GAN data augmentation and peeling severity classification method described in this invention, the method involves fusing the synthesized image with the original sample dataset to construct an augmented dataset, including: The synthesized stripped image is combined with the original image at the pixel level or feature vector level to form a new sample input;
[0030] The number of samples for each peeling level is counted; for cases where there are insufficient samples in the severe peeling category, the proportion of synthetic samples is increased to achieve a preset balance ratio for the number of samples in each category; and a confidence weight label lower than that of real samples is assigned to the synthetic peeling images to identify their generation characteristics.
[0031] The processed synthetic images are merged with the original sample dataset to form an augmented dataset containing balanced categories and confidence labels.
[0032] As a preferred embodiment of the adaptive network topology gray wolf optimization GAN data augmentation and spalling severity classification method described in this invention, the method includes: constructing a deep convolutional neural network classification model, training the deep convolutional neural network classification model based on the augmented dataset, and outputting the evaluation result of the concrete spalling severity level, including:
[0033] The deep convolutional neural network classification model is trained in a fully supervised manner using the constructed augmented dataset. During the training process, the loss function is adjusted by weighting the confidence weights of the synthetic images in the augmented dataset. The network parameters are updated using gradient descent or Adam optimization algorithms until the training converges.
[0034] Input an image of the concrete surface to be tested, and after passing through a trained deep convolutional neural network classification model, output the assessment result of the spalling severity level of the area.
[0035] Secondly, embodiments of the present invention provide a GAN data augmentation and stripping severity classification system with adaptive network topology gray wolf optimization, comprising:
[0036] The data acquisition and feature construction module is used to acquire the original image of the concrete structure surface, extract the physical morphological features of the spalled area, establish a severity level criterion based on the physical morphological features, divide the original image into different spalling levels, and construct the original sample dataset.
[0037] The model building and parameter optimization module is used to build a generative adversarial network based on the original sample dataset and establish an adaptive network topology gray wolf optimizer. The training control parameters of the generative adversarial network are mapped to the optimization space of the adaptive network topology gray wolf optimizer, and the training control parameters are adaptively optimized. The parameter optimization and data generation module is used to iteratively optimize the training control parameters using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters, and to train a generative adversarial network based on the optimal training control parameters to generate a synthetic image corresponding to the peeling level.
[0038] The data fusion and enhancement module is used to fuse the synthesized image with the original sample dataset to construct an enhanced dataset;
[0039] The classification and evaluation module is used to construct a deep convolutional neural network classification model, train the deep convolutional neural network classification model based on the augmented dataset, and output the evaluation result of the severity level of concrete spalling.
[0040] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, they implement the steps of the adaptive network topology gray wolf optimization GAN data augmentation and peeling severity classification method as described in the first aspect of the present invention.
[0041] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the adaptive network topology gray wolf optimization GAN data augmentation and peeling severity classification method as described in the first aspect of the present invention.
[0042] The beneficial effects of this invention are as follows: By constructing a physical feature space of concrete spalling and establishing a severity level criterion, this invention effectively organizes the original sample data and alleviates the problem of insufficient severe spalling samples; it uses an adaptive network topology gray wolf optimizer to dynamically optimize the training parameters of the generative adversarial network, and combines a small-world network topology mechanism to improve the quality and diversity of GAN-generated images, ensuring the stable generation of complex texture features; it fuses the generated synthetic spalling images with the original samples to form an enhanced dataset, supplementing the features of severe spalling samples, and improving the recognition rate of deep convolutional neural network classification models for a few categories and the overall classification performance. Attached Figure Description
[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0044] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation
[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0048] Example 1
[0049] Reference Figures 1-2 This is the first embodiment of the present invention, which provides a GAN data augmentation and peeling severity classification method based on adaptive network topology gray wolf optimization, including:
[0050] S1: Obtain the original image of the concrete structure surface, extract the physical morphological features of the spalled area, establish a severity level criterion based on the physical morphological features, divide the original image into different spalling levels, and construct the original sample dataset.
[0051] Furthermore, the extraction of the physical morphological features of the peeled area includes: calculating the area ratio of the peeled area and texture feature parameters that characterize the degree of surface texture damage;
[0052] The method of establishing a severity level criterion based on the physical morphological features includes: setting corresponding threshold ranges according to the proportion of the peeling area and the degree of texture damage, and classifying the severity of peeling according to the combination relationship of the threshold ranges;
[0053] Specifically, when the proportion of peeling area and the degree of texture damage both exceed the corresponding preset threshold, the corresponding sample is judged as a severe peeling level; when both are below the corresponding threshold, it is judged as a no-peeling level; and in other cases, it is judged as a slight peeling level.
[0054] It should be noted that, to further improve the accuracy of the spalling level classification, this embodiment uses multi-scale image analysis and advanced texture description operators to extract features from the spalled areas of the concrete surface. Specifically, in addition to calculating the proportion of the spalled area, texture feature parameters such as gray-level co-occurrence matrix, local binary mode, and wavelet energy are extracted to more comprehensively characterize the degree of texture damage in the spalled areas.
[0055] Based on the aforementioned multidimensional physical morphological features, an adaptive criterion for peeling severity is established. This is achieved by setting threshold ranges for the area proportion and various texture feature parameters, and classifying the level according to the combination of these thresholds. When the peeling area proportion and multiple texture indicators simultaneously exceed the corresponding preset thresholds, the sample is classified as severely peeling; when all indicators are below the corresponding thresholds, it is classified as having no peeling; and in other cases, it is classified as slightly peeling.
[0056] This embodiment can effectively improve the labeling accuracy and diversity of the original sample dataset, providing a more reliable data foundation for subsequent GAN data augmentation and peeling severity classification.
[0057] S2: Construct a generative adversarial network based on the original sample dataset, and establish an adaptive network topology gray wolf optimizer. Map the training control parameters of the generative adversarial network to the optimization space of the adaptive network topology gray wolf optimizer, and adaptively optimize the training control parameters.
[0058] Furthermore, the training control parameters of the generative adversarial network are constructed into multidimensional parameter vectors, and the multidimensional parameter vectors are used as the position representation of individual gray wolves to form an optimization space corresponding to the training control parameters.
[0059] An adaptive driving mechanism is set up, with the rate of change of the convergence speed of the FID score of the generated image during the training of the adversarial network as the trigger signal. When the trigger signal meets the preset conditions, a topology embedded search mechanism is introduced in the optimization iteration process of the gray wolf individuals. By dynamically defining and reorganizing the information interaction path between gray wolf individuals, the multidimensional parameter vector is iteratively updated to obtain the optimized training control parameters.
[0060] It should be noted that in this embodiment, a generative adversarial network (GAN) is first constructed to generate augmented data for concrete spalling images. The GAN consists of two core modules: a generator and a discriminator. The generator is responsible for generating images with a specified severity level, and the discriminator is used to determine the authenticity of the images.
[0061] Subsequently, an adaptive network topology gray wolf optimizer is constructed to adaptively optimize the training control parameters of the generative adversarial network. The specific method is as follows:
[0062] The generator learning rate, discriminator learning rate, first moment estimation decay rate, second moment estimation decay rate, and batch size are constructed into a five-dimensional parameter vector. This vector is used as the position representation of individual gray wolves, forming an optimization space corresponding to the training control parameters.
[0063] The rate of change of the convergence speed of the FID score of the generated images during the training of the generative adversarial network is used as the trigger signal. When the rate of change of the convergence speed meets the preset threshold, the position update and topology reorganization operation of the gray wolf individuals are triggered.
[0064] In the individual optimization iteration stage of the gray wolf, a small-world network topology mechanism is introduced. Initially, a regular ring network is established, and each individual only interacts with its K neighboring neighbors locally. Reconnection probability is introduced to randomly reorganize the connection relationships between individuals, forming a topology with a high clustering coefficient and a short average path length.
[0065] S3: The training control parameters are iteratively optimized using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters, and a generative adversarial network is trained based on the optimal training control parameters to generate a synthetic image corresponding to the peeling level.
[0066] Furthermore, in the optimization space of the adaptive network topology gray wolf optimizer, boundary constraints of the training control parameters are set, upper and lower thresholds are set for continuous parameters, and a mapping rule for discrete parameters is constructed to round to the nearest integer.
[0067] Under the condition of satisfying the adaptive driving mechanism, the search path of individual gray wolves is dynamically adjusted by using the small-world network topology mechanism, and the FID score of the generated image is used as the fitness index for iterative update.
[0068] The optimal training control parameters are obtained by gradually approximating the global optimal solution through iterative optimization.
[0069] Furthermore, an initial rule-based ring network topology is established, where each optimization individual only interacts with its K neighboring individuals locally.
[0070] By introducing reconnection probability, the connection relationships between individuals are randomly reorganized to form a small-world topology with a high clustering coefficient and a short average path length;
[0071] During the optimization iteration process, the individual position is updated in conjunction with the influence of topological neighbors. The specific update formula is as follows:
[0072]
[0073] in, The optimal individual location in the neighborhood is determined by small-world topological connections. Let i be the current position of the i-th individual in t iterations. This represents the current globally optimal individual position. , These are the weighting coefficients.
[0074] It should be noted that the iterative optimization process terminates when any of the following conditions are met:
[0075] The FID score changes less than a preset threshold in several consecutive iterations; the maximum number of iterations is reached; the globally optimal individual shows no further improvement within its topological neighborhood.
[0076] The obtained optimal training control parameters are used to train the GAN to generate high-quality synthetic peeling images, which are then labeled according to the peeling severity level.
[0077] The generated images are evaluated using FID and IS metrics to ensure that the synthesized images closely resemble real samples in terms of texture, brightness, and structural features. More samples can be generated for severe peeling levels as needed to address the shortage of severe peeling category samples in the original data.
[0078] S4: Fuse the synthesized image with the original sample dataset to construct an enhanced dataset.
[0079] Furthermore, the synthesized stripped image is combined with the original image at the pixel level or feature vector level to form a new sample input;
[0080] The number of samples for each peeling level is counted; for cases where there are insufficient samples in the severe peeling category, the proportion of synthetic samples is increased to achieve a preset balance ratio for the number of samples in each category; and a confidence weight label lower than that of real samples is assigned to the synthetic peeling images to identify their generation characteristics.
[0081] The processed synthetic images are merged with the original sample dataset to form an augmented dataset containing balanced categories and confidence labels.
[0082] It should be noted that pixel-level or feature vector-level fusion is performed between the synthetic stripped image and the original image:
[0083] New training samples are formed by combining synthetic images with original images through direct stitching, image overlay, or alpha channel weighting. Feature vectors are extracted from the synthetic and original images using a pre-trained convolutional network, and then stitched or weighted and fused in the feature space to form a high-dimensional feature input. The number of samples for each peeling level in the augmented dataset is statistically analyzed, with particular attention paid to the severe peeling category.
[0084] When the number of samples in the severely stripped category is insufficient, the proportion of synthetic images of that category in the augmented dataset is increased to achieve a preset balance in the number of samples in each category, thereby alleviating the long-tail distribution problem.
[0085] Assign a confidence weight label to the synthetic spalling images, lower than that of the real samples, to indicate their generation characteristics. For example, the confidence can be set in the range of 0.7 to 0.9, while the real samples are set to 1.0. This weight can be used in the weighted loss function during subsequent training to coordinate the model's learning ability on real and synthetic samples. The processed synthetic images are merged with the original sample dataset to form the final augmented dataset. The output dataset contains class-balanced concrete spalling images with confidence labels, which are used for subsequent training of a deep convolutional neural network classifier.
[0086] By employing fusion operations and class balancing strategies, augmented datasets can effectively supplement training samples with severely stripped-out classes, improving the classifier's ability to identify scarce classes. The introduction of confidence labels enables model training to distinguish the confidence levels of real and synthetic samples, thereby enhancing overall classification accuracy and robustness.
[0087] S5: Construct a deep convolutional neural network classification model, train the deep convolutional neural network classification model based on the augmented dataset, and output the assessment result of the severity level of concrete spalling.
[0088] Furthermore, the constructed augmented dataset is used to perform fully supervised training on the deep convolutional neural network classification model; during the training process, the loss function is adjusted by weighting the confidence weights of the synthetic images in the augmented dataset; gradient descent or Adam optimization algorithms are used to update the network parameters until the training converges.
[0089] Input an image of the concrete surface to be tested, and after passing through a trained deep convolutional neural network classification model, output the assessment result of the spalling severity level of the area.
[0090] It should be noted that the deep convolutional neural network classification model includes an input layer, several convolutional layers, pooling layers, fully connected layers, and an output layer. The output is the level of peeling severity, including no peeling, slight peeling, and severe peeling.
[0091] The network structure can be adjusted according to the complexity of the task, such as the size of the convolutional kernel, the number of layers, and the activation function.
[0092] The input data is a constructed augmented dataset, including the original samples and the generated synthetic peel-off images. The input images are normalized, resized, and augmented to improve the model's generalization ability.
[0093] Cross-entropy loss is used as the basic loss metric. Confidence weights are introduced for the synthetic images in the augmented dataset, and the loss function is weighted during training.
[0094]
[0095] in, The confidence weights for samples are 1.0 for real samples and less than 1.0 for synthetic samples. For real labels, Output predictions for the model.
[0096] Network parameters are updated using gradient descent or Adam optimization algorithms, and the learning rate is dynamically adjusted to accelerate convergence. During training, the validation set accuracy and loss function are monitored, and early stopping or learning rate decay strategies are employed to prevent overfitting.
[0097] After training, the image of the concrete surface to be tested is input into the trained deep convolutional neural network classification model. After forward propagation, the predicted probability of each category is calculated, and the assessment result of the spalling severity level of the area is output.
[0098] By augmenting the dataset with severely spalled samples and introducing confidence weights, the model effectively improves the accuracy of identifying rare categories. Utilizing the feature extraction capabilities of deep convolutional neural networks, automated classification of spalling textures and regional features on concrete surfaces is achieved, improving classification efficiency and robustness.
[0099] This embodiment also provides an adaptive network topology gray wolf optimized GAN data augmentation and peeling severity classification system, including:
[0100] The data acquisition and feature construction module is used to acquire the original image of the concrete structure surface, extract the physical morphological features of the spalled area, establish a severity level criterion based on the physical morphological features, divide the original image into different spalling levels, and construct the original sample dataset.
[0101] The model building and parameter optimization module is used to build a generative adversarial network based on the original sample dataset and establish an adaptive network topology gray wolf optimizer. The training control parameters of the generative adversarial network are mapped to the optimization space of the adaptive network topology gray wolf optimizer, and the training control parameters are adaptively optimized. The parameter optimization and data generation module is used to iteratively optimize the training control parameters using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters, and to train a generative adversarial network based on the optimal training control parameters to generate a synthetic image corresponding to the peeling level.
[0102] The data fusion and enhancement module is used to fuse the synthesized image with the original sample dataset to construct an enhanced dataset;
[0103] The classification and evaluation module is used to construct a deep convolutional neural network classification model, train the deep convolutional neural network classification model based on the augmented dataset, and output the evaluation result of the severity level of concrete spalling.
[0104] Example 2
[0105] This embodiment is the second embodiment of the present invention. This embodiment provides a GAN data augmentation and peeling severity classification method with adaptive network topology gray wolf optimization. In order to verify the beneficial effects of the present invention, simulation experiments are used for scientific demonstration.
[0106] Data augmentation and classification performance validation: The SVM was trained using the augmented dataset generated by this system, and the classification accuracy was compared with other optimization algorithms. Table 1. Methods for comparing generated image quality (FID index)
[0107]
[0108] As shown in Table 1, the FID (Fréchet Inception Distance) metric measures the difference in distribution between the generated image and the real peeled image in the feature space; a smaller value indicates that the generated image is closer to the real peeled texture distribution. The IS (Inception Score) metric reflects the class consistency and texture stability of the generated image. Combining the experimental data in Table 1, it can be seen that the proposed adaptive network topology Grey Wolf optimized generative adversarial network model achieves the lowest FID value of 5.26 and the highest IS value of 9.55. Compared with other optimization algorithms, the FID and IS of this invention are both superior. This indicates that by introducing an adaptive small-world network topology mechanism, the algorithm effectively escapes local extrema, and the optimized control parameters make the severely peeled images generated by GAN most closely resemble the data distribution of real samples in terms of texture and physical morphology, effectively overcoming the mode collapse problem of generative adversarial networks and ensuring the high quality and diversity of the generated images.
[0109] Table 2 Statistical performance of the optimization algorithm
[0110]
[0111] As shown in Table 2, this system successfully solved the engineering challenge of insufficient severely eroded samples by introducing adaptive network topology optimization, improving the identification rate of dangerous diseases to 96.39%. Analysis suggests that this improvement in classification performance mainly stems from the adaptive network topology mechanism's introduction of neighborhood cooperation and random reconstruction during parameter optimization. This allows the generative adversarial network to avoid falling into a single eroded texture pattern, thereby generating more diverse and realistically distributed severely eroded samples, providing more comprehensive feature support for the classification model.
[0112] This embodiment also provides a computer device applicable to an adaptive network topology gray wolf optimized GAN data augmentation and stripping severity classification method, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the adaptive network topology gray wolf optimized GAN data augmentation and stripping severity classification method as proposed in the above embodiment.
[0113] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0114] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements a GAN data augmentation and peeling severity classification method based on adaptive network topology gray wolf optimization as proposed in the above embodiment.
[0115] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for GAN data augmentation and peeling severity classification based on adaptive network topology gray wolf optimization, characterized in that, include: The original image of the concrete structure surface is acquired, the physical morphological features of the spalled area are extracted, and a severity level criterion is established based on the physical morphological features. The original image is divided into different spalling levels to construct an original sample dataset. Generative Adversarial Network (GAN) is constructed based on the original sample dataset, and an adaptive network topology gray wolf optimizer is established. The training control parameters of the GAN are mapped to the optimization space of the adaptive network topology gray wolf optimizer, and the training control parameters are adaptively optimized. The training control parameters are iteratively optimized using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters, and a generative adversarial network is trained based on the optimal training control parameters to generate synthetic images corresponding to the peeling level. The synthesized image is fused with the original sample dataset to construct an enhanced dataset; A deep convolutional neural network classification model is constructed, and the deep convolutional neural network classification model is trained based on the augmented dataset to output the assessment results of the severity level of concrete spalling.
2. The GAN data augmentation and stripping severity classification method with adaptive network topology gray wolf optimization as described in claim 1, characterized in that, The physical morphological features of the peeled areas are extracted, and a severity level criterion is established based on these features. The original image is then divided into different peeling levels to construct an original sample dataset, including: The extraction of physical morphological features of the peeled area includes: calculating the area ratio of the peeled area and texture feature parameters that characterize the degree of surface texture damage; The method of establishing a severity level criterion based on the physical morphological features includes: setting corresponding threshold ranges according to the proportion of the peeling area and the degree of texture damage, and classifying the severity of peeling according to the combination relationship of the threshold ranges; Specifically, when the proportion of peeling area and the degree of texture damage both exceed the corresponding preset threshold, the corresponding sample is judged as a severe peeling level; when both are below the corresponding threshold, it is judged as a no-peeling level; and in other cases, it is judged as a slight peeling level.
3. The GAN data augmentation and stripping severity classification method with adaptive network topology gray wolf optimization as described in claim 1, characterized in that, The step of mapping the training control parameters of the generative adversarial network to the optimization space of the adaptive network topology gray wolf optimizer, and adaptively optimizing the training control parameters, includes: The training control parameters of the generative adversarial network are constructed into a multidimensional parameter vector, and the multidimensional parameter vector is used as the position representation of the individual gray wolf to form an optimization space corresponding to the training control parameters. An adaptive driving mechanism is set up, with the rate of change of the convergence speed of the FID score of the generated image during the training of the adversarial network as the trigger signal. When the trigger signal meets the preset conditions, a topology embedded search mechanism is introduced in the optimization iteration process of the gray wolf individuals. By dynamically defining and reorganizing the information interaction path between gray wolf individuals, the multidimensional parameter vector is iteratively updated to obtain the optimized training control parameters.
4. The GAN data augmentation and stripping severity classification method with adaptive network topology gray wolf optimization as described in claim 1, characterized in that, The step of iteratively optimizing the training control parameters using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters includes: In the optimization space of the adaptive network topology gray wolf optimizer, boundary constraints of the training control parameters are set, upper and lower thresholds are set for continuous parameters, and a mapping rule for discrete parameters is constructed to round to the nearest integer. Under the condition of satisfying the adaptive driving mechanism, the search path of individual gray wolves is dynamically adjusted by using the small-world network topology mechanism, and the FID score of the generated image is used as the fitness index for iterative update. The optimal training control parameters are obtained by gradually approximating the global optimal solution through iterative optimization.
5. The GAN data augmentation and stripping severity classification method with adaptive network topology gray wolf optimization as described in claim 4, characterized in that, The small-world network topology mechanism includes: Initially, a regular ring network topology is established, and each optimization individual only interacts with its K neighboring individuals locally. By introducing reconnection probability, the connection relationships between individuals are randomly reorganized to form a small-world topology with a high clustering coefficient and a short average path length; During the optimization iteration process, the individual position is updated in conjunction with the influence of topological neighbors. The specific update formula is as follows: ； in, The optimal individual location in the neighborhood is determined by small-world topological connections. Let i be the current position of the i-th individual in t iterations. This represents the current globally optimal individual position. , These are the weighting coefficients.
6. The GAN data augmentation and stripping severity classification method with adaptive network topology gray wolf optimization as described in claim 1, characterized in that, The step of fusing the synthesized image with the original sample dataset to construct an enhanced dataset includes: The synthesized stripped image is combined with the original image at the pixel level or feature vector level to form a new sample input; The number of samples for each peeling level is counted; for cases where there are insufficient samples in the severe peeling category, the proportion of synthetic samples is increased to achieve a preset balance ratio for the number of samples in each category; and a confidence weight label lower than that of real samples is assigned to the synthetic peeling images to identify their generation characteristics. The processed synthetic images are merged with the original sample dataset to form an augmented dataset containing balanced categories and confidence labels.
7. The GAN data augmentation and stripping severity classification method with adaptive network topology gray wolf optimization as described in claim 1, characterized in that, The construction of a deep convolutional neural network classification model, training the deep convolutional neural network classification model based on the augmented dataset, and outputting the assessment result of the severity level of concrete spalling include: The deep convolutional neural network classification model is trained in a fully supervised manner using the constructed augmented dataset. During the training process, the loss function is adjusted by weighting the confidence weights of the synthetic images in the augmented dataset. The network parameters are updated using gradient descent or Adam optimization algorithms until the training converges. Input an image of the concrete surface to be tested, and after passing through a trained deep convolutional neural network classification model, output the assessment result of the spalling severity level of the area.
8. A GAN data augmentation and stripping severity classification system with adaptive network topology gray wolf optimization, characterized in that, include: The data acquisition and feature construction module is used to acquire the original image of the concrete structure surface, extract the physical morphological features of the spalled area, establish a severity level criterion based on the physical morphological features, divide the original image into different spalling levels, and construct the original sample dataset. The model building and parameter optimization module is used to build a generative adversarial network based on the original sample dataset and establish an adaptive network topology gray wolf optimizer. The training control parameters of the generative adversarial network are mapped to the optimization space of the adaptive network topology gray wolf optimizer, and the training control parameters are adaptively optimized. The parameter optimization and data generation module is used to iteratively optimize the training control parameters using the adaptive network topology gray wolf optimizer to obtain the optimal training control parameters, and to train a generative adversarial network based on the optimal training control parameters to generate a synthetic image corresponding to the peeling level. The data fusion and enhancement module is used to fuse the synthesized image with the original sample dataset to construct an enhanced dataset; The classification and evaluation module is used to construct a deep convolutional neural network classification model, train the deep convolutional neural network classification model based on the augmented dataset, and output the evaluation result of the severity level of concrete spalling.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the adaptive network topology gray wolf optimization GAN data augmentation and peeling severity classification method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the adaptive network topology gray wolf optimization GAN data augmentation and peeling severity classification method as described in any one of claims 1 to 7.