Pavement disease data enhancement method based on multi-modal dynamic balance

Abstract

The invention discloses a pavement disease data enhancement method based on multi-modal dynamic balance. The method comprises the following steps: S10, acquiring a data set; s20, classifying and calibrating the image data; s30, carrying out image preprocessing; s40, dividing a data set; s50, carrying out feature decoupling; s60, building a generative network; s70, performing model training; s80, setting a loss function; s90, testing the model; and S100, carrying out balanced screening. According to the method, the number of rare disease samples is effectively increased, the quality and diversity of the generated image are ensured, the generalization ability of the model under different illumination conditions is improved, and thus the recognition precision and reliability of the pavement disease detection model are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of computer vision and intelligent transportation infrastructure inspection, and in particular to a method for enhancing pavement distress data based on multimodal dynamic equilibrium. Background Technology

[0002] With the increasingly widespread application of deep learning technology in road surface defect detection, its importance in improving road maintenance efficiency and reducing costs is becoming increasingly prominent. However, the application of this technology is limited by the quality and quantity of training datasets. In particular, the probability of serious defects occurring in actual road networks is extremely low, such as the probability of network cracks being less than 0.3%, resulting in an average daily effective collection of less than 20 cases per device. Furthermore, the samples of the head and tail categories in mainstream datasets are extremely imbalanced. For example, in the CRACK500 dataset, the ratio of transverse cracks to alligator cracks is as high as 120:1, causing the model to achieve an F1-score of less than 0.35 when processing tail categories.

[0003] To address these challenges, researchers have attempted to augment the training dataset using various methods, including oversampling techniques and ACGAN-based conditional generation methods. However, these methods have significant drawbacks: oversampling leads to feature space collapse, reducing the diversity of generated images; and ACGAN-based methods are prone to pattern collapse in long-tailed scenarios, resulting in a class confusion rate as high as 42% for generated images.

[0004] In addition, traditional data augmentation methods cannot simulate the differences in road surface characteristics under complex environments, such as nighttime reflections or rain and fog interference, which limits the further improvement of model performance.

[0005] Therefore, the key to the technical solution of this invention is to provide a method that enables multimodal controllable generation of disease morphology, texture, and lighting, ensuring that the FID value of the generated image is ≤25 and increasing the effective proportion of long-tailed category samples to more than 80% of the balanced distribution; at the same time, it supports domain-adaptive generation of cross-material (asphalt / cement) pavement features to enhance the generalization ability of the model on different pavement materials. Summary of the Invention

[0006] In view of this, the present invention provides a method for enhancing pavement distress data based on multimodal dynamic equilibrium.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for enhancing pavement distress data based on multimodal dynamic equilibrium includes the following steps:

[0009] S10. Obtain datasets: Collect images of normal road surfaces and images of road defects;

[0010] S20. Image data classification and labeling: Based on the dataset obtained in step S10, the diseases are classified and labeled.

[0011] S30. Image preprocessing: Preprocess the collected dataset;

[0012] S40. Dataset partitioning: Divide the dataset into stratified sampling and spatiotemporal isolation;

[0013] S50, Feature Decoupling: By combining semantic mask generation and illumination feature encoding, the generator is given the ability to independently control different features, thereby achieving fine control over the generation process;

[0014] S60. Constructing the generative network: Constructing the MD-CycleGAN network model and using the MD-CycleGAN network to generate images;

[0015] S70. Model Training: The model equipped with the MD-CycleGAN network is initialized, and the received normal road surface images and disease semantic masks are input into the model for training.

[0016] S80. Setting the loss function: The loss function during the training of the MD-CycleGAN network consists of three parts: adversarial loss, balanced cycle consistency loss, and optional identity loss.

[0017] S90. Model Testing: Introduce test set images for model testing;

[0018] S100, Balanced Screening: A lightweight ResNet-18 model is used to focus on the classification of diseased areas, constructing a high-quality balanced dataset, and evaluating it through quantitative and qualitative analysis.

[0019] Preferably, in step S10, the road surface image library includes asphalt roads and cement roads;

[0020] The methods for obtaining the dataset include: collecting images on-site using vehicle-mounted cameras to build a database with 10,000 images.

[0021] Preferably, step S20 specifically includes the following steps:

[0022] S21. Collect transverse cracks, longitudinal cracks, pits, and crazing from the disease image library obtained in step S10 at a ratio of 4:1:1:0.5.

[0023] S22. Use the Supervisely platform to annotate the acquired images.

[0024] Preferably, step S30 specifically includes the following steps:

[0025] S31. The input image is scaled to 1024 pixels, and the Lanczos interpolation algorithm is used to preserve high-frequency details;

[0026] S32. Apply CLAHE to ensure a large difference between the brightness of darker and brighter areas. Divide the image into blocks, calculate the histogram for each block, then trim the histogram, and finally equalize the histogram calculation method.

[0027] Preferably, step S40 specifically includes the following steps:

[0028] S41. Stratified sampling: Stratify according to disease category to ensure that the category distribution of each subset is consistent with the original data, of which the training set is 80%, the validation set is 10%, and the test set is 10%.

[0029] S42, Spatiotemporal Isolation: The test set data collection time is more than 6 months later than the training set.

[0030] Preferably, step S50 specifically includes the following steps:

[0031] S51. Semantic Mask Generation: Build a core mask generation model, referred to as DeepLab-CBAM network, and use the DeepLab-CBAM network to generate training sets to generate masks that are adapted to specific disease features.

[0032] S52. Training set mask generation: After adding CBAM to the ASPP layer on the basic model DeepLabv3+with ResNet-101 backbone, it is adapted to specific disease features; where the loss function is: Dice Loss+Focal Loss (γ=2.0, α=0.25).

[0033] S53, Validation / Test Set Mask Generation: Generated using the model fine-tuned from the training set, ensuring consistent mask quality standards while avoiding overfitting;

[0034] S54. Illumination Feature Coding: Introducing HSV histogram coding into the generation of pavement defects;

[0035] S55. Through HSV spatial analysis: extract the brightness Value channel histogram and quantify the light intensity distribution;

[0036] S56. Use histogram matching: Align the illumination distribution of the generated image to the target domain using Earth Mover's Distance.

[0037] Preferably, step S60 specifically includes the following steps:

[0038] S61. Hierarchical adversarial training mechanism: Construct a dual discriminator architecture with a global discriminator D_global and a local discriminator D_local; where the global discriminator D_global is used to verify the overall formality of the image; the local discriminator D_local is used to focus on the fine-grained features of the disease area ROI-Aligned Attention.

[0039] The loss function calculation formulas for the global discriminator D_global and the local discriminator D_local are as follows:

[0040]

[0041] The gradient difference loss function is introduced as follows:

[0042]

[0043] In the formula The gradient magnitude is the result of inputting a real image into the discriminator. The gradient magnitude is generated after the image is input into the discriminator;

[0044] S62, Dynamic Class Balancing: Introducing Multi-Scale Structural Similarity (MS-SSIM), structural similarity is calculated at 5 scales, downsampled from the original resolution to 1 / 16; γ is linearly decayed from 1.0 to 0.5 to balance pixel-level and structural-level losses; the number of cracked category generated is increased by 8 times, and the recall rate t of the detection model is increased by 2 times.

[0045] S63, Cross-domain Feature Alignment Module: Replaces the first three layers of the generator with Discrete Wavelet Transform (DWT) convolutional layers. High-frequency components directly correspond to the geometric edges of cracks and potholes. By weighted retention and wavelet domain separation of different frequency features, the generator maintains the consistency of the high-frequency structure of the defects when converting asphalt / cement pavement.

[0046] Preferably, step S70 specifically includes the following steps:

[0047] S71, Generator Optimizer

[0048] Learning rate (lr): 2 × 10 -4 ;

[0049] Momentum parameters: β1 = 0.5, β2 = 0.999;

[0050] Weight decay: 0;

[0051] S72, Discriminator Optimizer

[0052] Learning rate (lr): 2 × 10 -3 ;

[0053] Momentum parameter: 0.9;

[0054] Weight decay: 1×10 -4 ;

[0055] S73, Phased Training

[0056] Phase 1: Texture generation module training:

[0057] Freeze layer settings, determine loss function weights, and adjust learning rate;

[0058] Phase 2: Dynamic Balance Training

[0059] Unfreezing and adding modules, adjusting the loss function, scheduling the learning rate, and step-wise decreasing the discriminator learning rate.

[0060] Preferably, in step S80, the calculation formulas for adversarial loss, balanced cycle consistency loss, and optional identity loss are as follows:

[0061]

[0062] In the formula: G and F refer to two mapping relationships. G is used to realize X->Y, and the training should make G(X) as close to Y as possible. The discriminator Dy is used to distinguish between real and fake samples. F is used to realize Y->X, and the training should make G(Y) as close to X as possible. The discriminator DX is used to distinguish between real and fake samples. G is the generator from domain A to domain B; F is the generator from domain B to domain A; X is the original image from domain A; Y is the original image from domain B; G(X) is the reconstructed image of X by generator G; F(Y) is the reconstructed image of Y by generator F. T is the total number of training steps; N c The number of samples in category c, T 3 / 4 Control the sensitivity of the training process to avoid early overfitting; N c 1 / 2 Give higher weight to tail categories.

[0063] Preferably, step 100 specifically includes the following steps:

[0064] S101. Quantitative Analysis: Comparative analysis of mainstream target detection algorithms YOLOv11, YOLOv12 and YOLOv8, and quality assessment of the generated pavement distress images using these algorithms;

[0065] S102. Qualitative Analysis: Comparative analysis of various image quality assessment indicators, selecting Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Multiscale Structural Similarity Index (MS-SSIM), and Fraser Distance (FID) as key evaluation indicators to conduct a comprehensive and accurate quantitative evaluation of the generated pavement distress images.

[0066] The present invention achieves the following technical effects compared to the prior art:

[0067] (1) This invention can not only effectively increase the number of rare disease samples, but also ensure the quality and diversity of generated images and improve the generalization ability of the model under different lighting conditions.

[0068] (2) The method of the present invention can accurately control the generation of specific disease morphology while maintaining the natural background of the road surface, thereby greatly improving the recognition accuracy and reliability of the road surface disease detection model. Attached Figure Description

[0069] Figure 1 This is a flowchart of a method for enhancing pavement distress data based on multimodal dynamic equilibrium according to the present invention;

[0070] Figure 2 This is the overall network structure of a pavement distress data enhancement method based on multimodal dynamic equilibrium according to the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] like Figure 1 As shown, this invention discloses a method for enhancing pavement distress data based on multimodal dynamic equilibrium, characterized by comprising the following steps:

[0073] S10. Obtain datasets: Collect images of normal road surfaces and images of road defects;

[0074] The road surface image library includes asphalt roads and cement roads;

[0075] Methods for obtaining the dataset include: acquiring images on-site using vehicle-mounted cameras to create a database with 10,000 images;

[0076] S20. Image data classification and labeling: Based on the dataset obtained in step S10, the diseases are classified and labeled.

[0077] S21. Collect transverse cracks, longitudinal cracks, pits, and crazing from the disease image library obtained in step S10 at a ratio of 4:1:1:0.5.

[0078] S22. Use the Supervisely platform to annotate the acquired images;

[0079] S30. Image preprocessing: Preprocess the collected dataset;

[0080] S31. The input image is scaled to 1024 pixels, and the Lanczos interpolation algorithm is used to preserve high-frequency details;

[0081] S32. Apply CLAHE (Contrast Limiting Adaptive Histogram Equalization) to ensure a large difference between the brightness of darker areas and brighter areas. Divide the image into blocks, calculate the histogram for each block, then trim the histogram, and finally equalize the histogram calculation method.

[0082] S40. Dataset partitioning: Divide the dataset into stratified sampling and spatiotemporal isolation;

[0083] S41. Stratified sampling: Stratify according to disease category to ensure that the category distribution of each subset is consistent with the original data, of which the training set is 80%, the validation set is 10%, and the test set is 10%.

[0084] S42, Spatiotemporal isolation: The test set data was collected more than 6 months later than the training set;

[0085] S50, Feature Decoupling: By combining semantic mask generation and illumination feature encoding, the generator is given the ability to independently control different features, thereby achieving fine control over the generation process;

[0086] S51. Semantic Mask Generation: Build a core mask generation model, referred to as DeepLab-CBAM network, and use the DeepLab-CBAM network to generate training sets to generate masks that are adapted to specific disease features, such as cracked mesh structures.

[0087] S52. Training set mask generation: After adding CBAM (Convolutional Block Attention Module) to the ASPP layer on the basic model DeepLabv3+with ResNet-101 backbone, it is adapted to specific disease features (such as cracked mesh structure); where the loss function is: Dice Loss+Focal Loss (γ=2.0, α=0.25);

[0088] Specifically, it includes:

[0089] a) Determine training details:

[0090] Input size: 512×512 pixels, batch size: 16;

[0091] Optimizer: Adam (lr = 3e-4, weight_decay = 1e-4);

[0092] Data: The Cityscapes pre-trained model was fine-tuned for 150 epochs on our own dataset;

[0093] b) Perform post-processing;

[0094] 3) Perform morphological optimization:

[0095] Opening operation: 3×3 elliptic kernel, eliminating fine noise;

[0096] Hole filling: Scan connected regions and automatically fill holes with an area of ​​less than 50 pixels;

[0097] S53, Validation / Test Set Mask Generation: Generated using the model fine-tuned from the training set, ensuring consistent mask quality standards while avoiding overfitting;

[0098] S54. Illumination Feature Coding: Introducing HSV histogram coding into the generation of pavement defects;

[0099] S55. Through HSV spatial analysis: extract the brightness Value channel histogram and quantify the light intensity distribution;

[0100] Specifically, it includes:

[0101] a) HSV spatial analysis:

[0102] Extract the V channel histogram:

[0103] Histogram intervals: 64, with the interval [0, 255] evenly divided;

[0104] Normalization: L2 normalization to a unit-length vector;

[0105] Matching algorithm:

[0106] Target Histogram H target With source histogram H source Earth Mover's Distance calculation:

[0107]

[0108] Histogram transfer: transforming H... source CDF aligned to H target ;

[0109] 2) Feature fusion

[0110] Generator input layer design:

[0111] Expand the 64-dimensional histogram vector into a 128×128 feature map;

[0112] It is concatenated with the original image channels to form a 6-channel input (RGB+HSV-V+semantic mask);

[0113] S56. Using histogram matching: Align the illumination distribution of the generated image to the target domain using Earth Mover's Distance;

[0114] like Figure 2 As shown, S60, build the generative network: build the MD-CycleGAN network model and use the MD-CycleGAN network to generate images;

[0115] S61. Hierarchical adversarial training mechanism: Construct a dual discriminator architecture with a global discriminator D_global and a local discriminator D_local; where the global discriminator D_global is used to verify the overall formality of the image; the local discriminator D_local is used to focus on the fine-grained features of the disease area ROI-AlignedAttention.

[0116] The loss function calculation formulas for the global discriminator D_global and the local discriminator D_local are as follows:

[0117]

[0118] The gradient difference loss function is introduced as follows:

[0119]

[0120] In the formula The gradient magnitude is the result of inputting a real image into the discriminator. The gradient magnitude is generated after the image is input into the discriminator;

[0121] S62, Dynamic Class Balancing: Introducing Multi-Scale Structural Similarity (MS-SSIM), structural similarity is calculated at 5 scales, downsampled from the original resolution to 1 / 16; γ is linearly decayed from 1.0 to 0.5 to balance pixel-level and structural-level losses; the number of cracked category generated is increased by 8 times, and the recall rate t of the detection model is increased by 2 times.

[0122] S63, Cross-domain feature alignment module: Replace the first three layers of the generator with Discrete Wavelet Transform (DWT) convolutional layers. The high-frequency components directly correspond to the geometric edges of cracks and potholes. By weighted retention and wavelet domain separation of different frequency features, the generator maintains the consistency of the high-frequency structure of the defects when converting asphalt / cement pavement.

[0123] S70. Model Training: The model equipped with the MD-CycleGAN network is initialized, and the received normal road surface images and disease semantic masks are input into the model for training.

[0124] S71, Generator Optimizer (Adam: combines momentum and adaptive learning rate characteristics, suitable for generators to learn complex texture features)

[0125] Learning rate (lr): 2 × 10 -4 ;

[0126] Momentum parameter: (slows down the accumulation of historical gradients, preventing the generator from converging to a local optimum too early), (preserves long-term gradient variance information, and enhances the ability to capture low-frequency features).

[0127] Weight decay: 0 (no L2 regularization added);

[0128] S72, Discriminator Optimizer (SGD: provides a stable gradient update direction, avoiding overfitting of the discriminator to the instantaneous output mode of the generator);

[0129] Learning rate (lr): 2 × 10 -3 ;

[0130] Momentum parameter: 0.9; A high momentum value (0.9) enhances the discriminator's robustness to gradient noise, especially in the early stages of adversarial training;

[0131] Weight decay: 1×10 -4 (L2 regularization prevents the discriminator weights from becoming too large, maintaining the balance of generative adversarial processes);

[0132] S73, Phased Training

[0133] Phase 1: Texture Generation Module Training (First 100 Epochs)

[0134] 1. Objective: Prioritize learning the texture features of the road surface background (such as asphalt particles and cement joints) and the ability of light migration to avoid the shape control layer interfering with the generator convergence too early.

[0135] 2. Freeze layer settings:

[0136] The morph control layer parameters of the fixed generator are set.

[0137] Optimize only the following module parameters:

[0138] Feature encoder: ResNet residual blocks, extracting multi-scale texture features.

[0139] Light transfer module (light_trans): HSV histogram matching network (3-layer MLP)

[0140] Decoder: Transposed convolutional layer + CBAM attention module, reconstructing texture details.

[0141] Learning rate adjustment:

[0142] Linear Warmup is used: the learning rate is gradually increased from 0 to 2×10 for the first 10 epochs. -4

[0143] formula:

[0144]

[0145] Phase 2: Dynamic Balance Training (Last 50 Epochs)

[0146] 1. Objective: To thaw the morphology control layer and refine the geometric structure of diseases by combining a dynamic category balancing strategy, thereby improving the quality of long-tail category generation.

[0147] 2. Unfreezing and adding modules:

[0148] Thawing morphology control parameters, initiating a dynamic category balancing strategy:

[0149] Generator morph control layer (generator.morph_layer.requires_grad = True (code))

[0150] Local discriminator loss (D_local) is enabled, with weights increasing linearly. Initial weights λ local =0.1, increasing linearly by 0.018 per epoch, eventually reaching 1.0:

[0151]

[0152] The loss function adjustments are shown in Table 1.

[0153] Table 1:

[0154]

[0155]

[0156] Learning rate scheduling:

[0157] Generator learning rate cosine annealing:

[0158]

[0159] lr max =2×10 -4 ,lr min =5×10 -5

[0160] The discriminator learning rate decreases in steps: it decreases to 0.5 times the original rate every 15 epochs.

[0161] S80. Setting the loss function: The loss function during the training of the MD-CycleGAN network consists of three parts: adversarial loss, balanced cycle consistency loss, and optional identity loss.

[0162] The formulas for calculating adversarial loss, balanced cycle consistency loss, and optional identity loss are as follows:

[0163]

[0164] In the formula: G and F refer to two mapping relationships. G is used to realize X->Y, and the training should make G(X) as close to Y as possible. The discriminator Dy is used to distinguish between real and fake samples. F is used to realize Y->X, and the training should make G(Y) as close to X as possible. The discriminator DX is used to distinguish between real and fake samples. G is the generator from domain A to domain B; F is the generator from domain B to domain A; X is the original image from domain A; Y is the original image from domain B; G(X) is the reconstructed image of X by generator G; F(Y) is the reconstructed image of Y by generator F. T is the total number of training steps; N c The number of samples in category c, T 3 / 4 Control the sensitivity of the training process to avoid early overfitting; N c 1 / 2 Give higher weight to tail categories.

[0165] S90. Model Testing: Introduce test set images for model testing;

[0166] S100, Balanced Screening: A lightweight ResNet-18 model is used to focus on the classification of diseased areas, a high-quality balanced dataset is constructed, and it is evaluated through quantitative and qualitative analysis;

[0167] S101. Quantitative Analysis: Comparative analysis of mainstream target detection algorithms YOLOv11, YOLOv12 and YOLOv8, and quality assessment of the generated pavement distress images using these algorithms;

[0168] S102. Qualitative Analysis: Comparative analysis of various image quality assessment indicators, selecting Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Multi-Scale Structural Similarity Index (MS-SSIM), and Fraser Distance (FID) as key evaluation indicators to conduct a comprehensive and accurate quantitative evaluation of the generated pavement distress images.

[0169] In summary, this invention, through learning and processing a large number of normal road surface images and defect images, enables the model to learn more robust feature representations, thereby increasing generalization ability and improving recognition accuracy. In terms of data processing, firstly, normal road surface image databases and defect image databases are collected. Defect images are classified, labeled, and preprocessed. Then, the dataset is partitioned through stratified sampling and spatiotemporal isolation to ensure data diversity and representativeness. Regarding feature decoupling, semantic mask generation and illumination feature encoding are combined to give the generator the ability to independently control different features, achieving fine-grained control over the generation process. In terms of generating network construction, an MD-CycleGAN network model is proposed, employing a hierarchical adversarial training mechanism, introducing gradient difference loss and dynamic class balancing strategies, and designing a cross-domain feature alignment module to improve the quality and diversity of generated images. In terms of model training, a staged training method is adopted, with reasonable setting of the loss function and optimization of the generator and discriminator parameters to ensure stable training and effective learning of the model. In terms of model testing and evaluation, test set images are introduced for model testing, and quantitative and qualitative analyses are used for evaluation to verify the accuracy and practicality of the generated images. The entire method effectively increases the number of rare disease samples by optimizing the generative adversarial network structure, ensuring the quality and diversity of generated images, and improving the model's generalization ability under different lighting conditions, thereby greatly improving the recognition accuracy and reliability of the road surface disease detection model.

[0170] The above description is merely a preferred embodiment of the present invention and does not constitute any limitation on the technical scope of the present invention. Therefore, any minor modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for enhancing pavement distress data based on multimodal dynamic equilibrium, characterized in that, Includes the following steps: S10. Obtain datasets: Collect images of normal road surfaces and images of road defects; S20. Image data classification and labeling: Based on the dataset obtained in step S10, the diseases are classified and labeled. S30. Image preprocessing: Preprocess the collected dataset; S40. Dataset partitioning: Divide the dataset into stratified sampling and spatiotemporal isolation; S50, Feature Decoupling: By combining semantic mask generation and illumination feature encoding, the generator is given the ability to independently control different features, thereby achieving fine control over the generation process; S60. Constructing the generative network: Constructing the MD-CycleGAN network model and using the MD-CycleGAN network to generate images; S70. Model Training: The model equipped with the MD-CycleGAN network is initialized, and the received normal road surface images and disease semantic masks are input into the model for training. S80. Setting the loss function: The loss function during the training of the MD-CycleGAN network consists of three parts: adversarial loss, balanced cycle consistency loss, and optional identity loss. S90. Model Testing: Introduce test set images for model testing; S100, Balanced Screening: A lightweight ResNet-18 model is used to focus on the classification of diseased areas, constructing a high-quality balanced dataset, and evaluating it through quantitative and qualitative analysis.

2. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, In step S10, the road surface image library includes asphalt roads and cement roads; The methods for obtaining the dataset include: collecting images on-site using vehicle-mounted cameras to build a database with 10,000 images.

3. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, Step S20 specifically includes the following steps: S21. Collect transverse cracks, longitudinal cracks, pits, and crazing from the disease image library obtained in step S10 at a ratio of 4:1:1:0.

5. S22. Use the Supervisely platform to annotate the acquired images.

4. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, Step S30 specifically includes the following steps: S31. The input image is scaled to 1024 pixels, and the Lanczos interpolation algorithm is used to preserve high-frequency details; S32. Apply CLAHE to ensure a large difference between the brightness of darker areas and brighter areas. Divide the image into blocks, calculate the histogram for each block, then trim the histogram, and finally equalize the histogram calculation method.

5. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, Step S40 specifically includes the following steps: S41. Stratified sampling: Stratify according to disease category to ensure that the category distribution of each subset is consistent with the original data, where the training set is 80%, the validation set is 10%, and the test set is 10%. S42, Spatiotemporal Isolation: The test set data collection time is more than 6 months later than the training set.

6. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, Step S50 specifically includes the following steps: S51. Semantic Mask Generation: Build a core mask generation model, referred to as DeepLab-CBAM network, and use the DeepLab-CBAM network to generate training sets to generate masks that are adapted to specific disease features. S52. Training set mask generation: After adding CBAM to the ASPP layer on the basic model DeepLabv3+with ResNet-101backbone, it is adapted to specific disease features; where the loss function is: Dice Loss+FocalLoss (γ=2.0, α=0.25). S53, Validation / Test Set Mask Generation: Generated using the model fine-tuned from the training set, ensuring consistent mask quality standards while avoiding overfitting; S54. Illumination Feature Coding: Introducing HSV histogram coding into the generation of pavement defects; S55. Through HSV spatial analysis: extract the brightness Value channel histogram and quantify the light intensity distribution; S56. Use histogram matching: Align the illumination distribution of the generated image to the target domain using Earth Mover's Distance.

7. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, Step S60 specifically includes the following steps: S61. Hierarchical adversarial training mechanism: Construct a dual discriminator architecture with a global discriminator D_global and a local discriminator D_local; where the global discriminator D_global is used to verify the overall formality of the image; the local discriminator D_local is used to focus on the fine-grained features of the disease area ROI-AlignedAttention. The loss function calculation formulas for the global discriminator D_global and the local discriminator D_local are as follows: The gradient difference loss function is introduced as follows: In the formula The gradient magnitude is the result of inputting a real image into the discriminator. The gradient magnitude is generated after the image is input into the discriminator; S62, Dynamic Class Balancing: Introducing Multi-Scale Structural Similarity (MS-SSIM), structural similarity is calculated at 5 scales, downsampled from the original resolution to 1 / 16; γ is linearly decayed from 1.0 to 0.5 to balance pixel-level and structural-level losses; the number of cracked category generated is increased by 8 times, and the recall rate t of the detection model is increased by 2 times. S63, Cross-domain Feature Alignment Module: Replaces the first three layers of the generator with Discrete Wavelet Transform (DWT) convolutional layers. High-frequency components directly correspond to the geometric edges of cracks and potholes. By weighted retention and wavelet domain separation of different frequency features, the generator maintains the consistency of the high-frequency structure of the defects when converting asphalt / cement pavement.

8. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, Step S70 specifically includes the following steps: S71, Generator Optimizer Learning rate (lr): 2 × 10 -4 ; Momentum parameters: β1 = 0.5, β2 = 0.999; Weight decay: 0; S72, Discriminator Optimizer Learning rate (lr): 2×10 -3 ; Momentum parameter: 0.9; Weight decay: 1×10 -4 ; S73, Phased Training Phase 1: Texture generation module training: Freeze layer settings, determine loss function weights, and adjust learning rate; Phase 2: Dynamic Balance Training Unfreezing and adding modules, adjusting the loss function, scheduling the learning rate, and step-wise decreasing the discriminator learning rate.

9. The method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, In step S80, the calculation formulas for adversarial loss, balanced cycle consistency loss, and optional identity loss are as follows: In the formula: G and F refer to two mapping relationships. G is used to realize X->Y, and the training should make G(X) as close to Y as possible. The discriminator Dy is used to distinguish between real and fake samples. F is used to realize Y->X, and the training should make G(Y) as close to X as possible. The discriminator DX is used to distinguish between real and fake samples. G is the generator from domain A to domain B; F is the generator from domain B to domain A; X is the original image from domain A; Y is the original image from domain B; G(X) is the reconstructed image of X by generator G; F(Y) is the reconstructed image of Y by generator F. T is the total number of training steps; N c The number of samples in category c, T 3 / 4 Control the sensitivity of the training process to avoid early overfitting; N c 1 / 2 Give higher weight to tail categories.

10. A method for enhancing pavement distress data based on multimodal dynamic equilibrium according to claim 1, characterized in that, Step 100 specifically includes the following steps: S101. Quantitative Analysis: Comparative analysis of mainstream target detection algorithms YOLOv11, YOLOv12 and YOLOv8, and quality assessment of the generated pavement distress images using these algorithms; S102. Qualitative Analysis: Comparative analysis of various image quality assessment indicators, selecting Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Multi-Scale Structural Similarity Index (MS-SSIM), and Fraser Distance (FID) as key evaluation indicators to conduct a comprehensive and accurate quantitative evaluation of the generated pavement distress images.

Images