Steel member damage image super-resolution reconstruction method based on corrosion evolution physical prior

Abstract

This invention discloses a super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution. The method includes: constructing a dedicated physical prior system for the entire corrosion evolution chain of coated steel, generating physical constraints and spatial distribution rules for corrosion characteristics; constructing a multimodal corrosion damage dataset integrating visible light images, electrochemical test data, and three-dimensional morphology data; constructing a PI-GAN model, designing a four-dimensional hybrid objective function, and performing phased training; constructing an extreme illumination adaptive preprocessing module and a cross-scene transfer learning module; and finally, based on the trained PI-GAN model, performing super-resolution reconstruction of the original corrosion image to be processed, outputting a high-resolution image that conforms to the physical laws of corrosion evolution. This invention significantly improves the reconstruction quality and engineering practicality in rare corrosion scenarios while maintaining millisecond-level inference speed, providing reliable support for intelligent detection and quantitative assessment of steel structure corrosion.

Description

Technical Field

[0001] This invention relates to the fields of disaster prevention and mitigation in civil engineering, structural health monitoring, computer vision, electrochemical corrosion and deep learning, and in particular to a super-resolution reconstruction method for steel component damage images based on the physical prior of corrosion evolution. It is applied to intelligent detection and quantitative assessment systems for corrosion damage in steel structure projects such as highway bridges, industrial plants, oil and gas pipelines, ports and docks, and offshore wind power, and realizes high-fidelity, physically self-consistent super-resolution reconstruction of low-quality on-site corrosion images. Background Technology

[0002] Coated steel components are widely used in various civil engineering structures such as highway bridges, industrial plants, oil and gas pipelines, ports and docks, and offshore wind power. Corrosion damage is a major cause of structural performance degradation and safety accidents. Among these, the failure to detect or misdetect hidden corrosion and early micro-corrosion under the coating is a core factor leading to significant safety hazards. Image-based corrosion detection technology has become the mainstream method for steel structure health monitoring due to its advantages of low cost, high efficiency, and scalability. However, corrosion images acquired on-site generally suffer from problems such as low resolution, uneven lighting, blurred details, and occlusion interference, resulting in the loss of key corrosion features such as early micro-bulges, pitting, and hairline cracks, which seriously affects the accuracy of downstream detection and assessment.

[0003] Current super-resolution reconstruction techniques for eroded images are mainly divided into two categories: mainstream techniques for engineering applications and advanced techniques for academic research. The most widely used techniques in engineering applications are traditional interpolation filtering methods and early convolutional neural network methods. Traditional interpolation filtering methods, such as bicubic interpolation and wavelet transform, can only achieve pixel-level smooth magnification and cannot recover high-frequency erosion details, resulting in extremely poor reconstruction effects. Super-resolution methods based on convolutional neural networks, such as super-resolution convolutional neural networks and enhanced deep super-resolution networks, can recover some texture details, but lack the ability to capture global features, resulting in poor reconstruction effects in low-contrast eroded areas and failing to meet the accuracy requirements of engineering detection. Current mainstream super-resolution methods based on Generative Adversarial Networks (GANs), represented by Enhanced Super-Resolution GANs and Improved Enhanced Super-Resolution GANs, generate more realistic textures through adversarial training, significantly improving the reconstruction quality of corrosion images. However, these methods are all purely data-driven and rely entirely on the distribution of training samples. When encountering rare corrosion scenarios not covered in the training set (such as early pitting corrosion under coatings, rust layer delamination in marine environments, and stress corrosion cracks in industrial atmospheres), they are prone to generating non-physical artifacts that do not conform to the corrosion evolution law, leading to misjudgments in downstream detection and failing to meet the reliability requirements of engineering detection.

[0004] In recent years, the development of Physical Information Neural Networks (PINNs) has provided a new approach to solving the physical consistency problem of purely data-driven models. However, existing PINN super-resolution technologies are mostly aimed at natural images or fluid dynamics scenarios, and have not yet incorporated the specific electrochemical evolution laws of coated steel corrosion. Existing research has not deeply integrated the coating failure process, the electrochemical kinetics of corrosion, the rust layer growth mechanism, and image super-resolution reconstruction, which cannot fundamentally guarantee the physical rationality of the reconstruction results, nor can it use physical laws to compensate for the lack of training data. At the same time, existing technologies have not effectively utilized multi-source engineering data such as electrochemical testing and three-dimensional morphology measurement, resulting in low data utilization and limited model generalization ability. In addition, existing models generally suffer from high computational cost and slow inference speed, have not achieved end-to-end integration with downstream corrosion detection tasks, and are difficult to deploy on resource-constrained portable field detection equipment. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to provide a super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution. This method significantly improves reconstruction quality and engineering practicality in rare corrosion scenarios while maintaining millisecond-level inference speed, providing reliable technical support for intelligent detection and quantitative assessment of steel structure corrosion.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0007] This invention provides a super-resolution reconstruction method for steel component damage images based on prior knowledge of corrosion evolution physics, comprising the following steps:

[0008] Step 1: Construct a dedicated physical prior system for the entire chain of corrosion evolution of coated steel, and generate physical rationality constraint boundaries and spatial distribution rules for corrosion features to constrain the super-resolution reconstruction process;

[0009] Step 2: Construct a multimodal corrosion damage dataset that integrates visible light images, electrochemical test data, and three-dimensional morphology data, covering the entire corrosion stage, multiple types of coatings, and multiple typical service environments, and complete the spatiotemporal alignment, annotation, and standardization preprocessing of multi-source data;

[0010] Step 3: Construct a physical information generative adversarial network super-resolution model, i.e., the PI-GAN model, which embeds multiple layers of physical constraints, and embed the physical prior of corrosion evolution into the entire forward and backward propagation process of the PI-GAN model;

[0011] Step 4: Design a four-dimensional hybrid objective function that integrates pixel loss, perceptual loss, spatial physical constraint loss and adversarial loss, and train the PI-GAN model in stages based on the multimodal corrosion damage dataset;

[0012] Step 5: Construct an extreme lighting adaptive preprocessing module and a cross-scene transfer learning module to improve the generalization ability of the PI-GAN model in complex field environments;

[0013] Step 6: Based on the trained PI-GAN model, perform super-resolution reconstruction on the original corrosion image to be processed, and output a target resolution corrosion image that conforms to the physical laws of corrosion evolution.

[0014] Preferably, in step 1, the specific steps of constructing a dedicated physical prior system for the entire corrosion evolution chain of coated steel include:

[0015] Step 1.1: Construct a coating failure kinetic model. Based on Fick's diffusion law and the degradation law of coating adhesion, derive the evolution equations for coating blistering, cracking, and peeling, and establish the quantitative relationship between coating failure time and ambient temperature, humidity, and salt spray concentration. The expression is as follows:

[0016] (1);

[0017] In the formula, t f k represents the coating failure time. f C is the coating failure rate constant. salt The concentration of salt spray in the environment is given by E, where a is the concentration index. f R is the activation energy for coating failure, R is the gas constant, and T is the absolute temperature.

[0018] Step 1.2: Construct an electrochemical kinetic model for the corrosion of coated steel. Based on Faraday's law and the Tafel equation, derive the quantitative relationship between corrosion current density and corrosion rate. Simultaneously, introduce a coating failure rate correction coefficient to consider the impact of localized coating failure on the corrosion rate. The expression is:

[0019] (2);

[0020] (3);

[0021] In the formula, β is the coating damage rate correction coefficient, with a value ranging from 0 to 1, i corr Let i be the corrosion current density, i0 be the exchange current density, α be the transfer coefficient, F be the Faraday constant, η be the overpotential, and v be the voltage. corr Here, M is the corrosion rate, n is the molar mass of iron, n is the number of electrons reacted, and ρ is the density of iron.

[0022] Step 1.3: Construct a rust layer growth kinetic model. Based on the power-law growth law and the rust layer porosity evolution equation, establish the mapping relationship between rust layer thickness, porosity, and corrosion time. The expression is:

[0023] (4);

[0024] (5);

[0025] In the formula, d(t) is the thickness of the rust layer at time t, and k d Let be the rust growth rate constant, e be the growth exponent, and its value range be 0.3 to 0.8; ε(t) be the rust porosity at time t, ε0 be the initial porosity, and k be the rust growth rate constant. ε λ and λ are porosity evolution parameters;

[0026] Step 1.4: Establish a quantitative mapping relationship between corrosion morphology features and multiple physical parameters. Define the corrosion morphology feature set F={f1,f2,f3,f4,f5,f6}, where f1 is the corrosion area ratio, f2 is the pitting density, f3 is the average pitting depth, f4 is the rust texture roughness, f5 is the average size of coating bulges, and f6 is the coating peeling rate. Based on multiple sets of accelerated corrosion test and field measurement data, fit the quantitative mapping functions between each morphology feature and coating failure time, corrosion rate, and rust thickness.

[0027] Step 1.5: Generate the physical rationality constraint boundary and spatial distribution rules of corrosion features. Based on the physical laws of corrosion evolution, define the value range of each morphological feature; at the same time, define the spatial distribution rules of corrosion features to provide a full-dimensional physical constraint basis for super-resolution reconstruction.

[0028] Preferably, in step 2, the step of constructing the multimodal corrosion damage dataset is as follows:

[0029] Step 2.1: Design a multi-condition accelerated corrosion test scheme, select a variety of commonly used coating types in engineering, set a variety of typical service conditions, carry out long-term accelerated corrosion tests covering the initial to mid-to-late stage of corrosion, and collect sample data at a preset time period.

[0030] Step 2.2: Establish a multi-source data synchronous acquisition standard, and synchronously acquire multiple types of data for each experimental sample, including:

[0031] Visible light images; orthogonal images taken under various lighting intensities using an industrial camera with original resolution that meets the input requirements of the PI-GAN model, including images with partial occlusion and dirt interference;

[0032] Electrochemical data; corrosion current density, polarization curves, and AC impedance spectra were obtained using an electrochemical workstation at a predetermined sampling frequency.

[0033] Physical morphological parameters, including rust layer thickness measured using a high-precision thickness gauge, and surface roughness and pitting depth measured using a high-precision three-dimensional morphology instrument.

[0034] Step 2.3: Supplement the collection of on-site engineering data, and obtain corrosion images and corresponding detection data of steel structure engineering sites from multiple geographical regions and various service years, in order to supplement rare corrosion samples under extreme service environments, so that the multimodal corrosion damage dataset covers diverse actual engineering scenarios.

[0035] Step 2.4: Label and classify the multimodal corrosion damage dataset. Based on the corrosion evaluation criteria, the corrosion damage is divided into five levels: no corrosion, slight corrosion, moderate corrosion, severe corrosion, and extremely severe corrosion. For each image, label the corrosion area outline, corrosion level, coating type, service environment, and corresponding electrochemical and physical data and three-dimensional morphology data.

[0036] Step 2.5: Preprocess and enhance the data. Crop and normalize the original images and perform data enhancement using various geometric transformations. Standardize the electrochemical and physical data and remove outliers. Finally, construct a multimodal corrosion damage dataset containing a large number of multimodal samples. Divide the dataset into training, validation, and test sets according to a predetermined ratio, and ensure that the samples of each corrosion level, coating type, and service environment are evenly distributed.

[0037] Preferably, in step 3, the PI-GAN model includes a generator, a dual discriminator architecture, and an online physical consistency verification module. The construction process of the PI-GAN model specifically includes:

[0038] Step 3.1: Construct a generator architecture with embedded multi-layer physical constraint layers. The generator adopts an encoder-decoder structure, and physical constraint layers are inserted between the 3rd, 4th, and 5th layers of the encoder and the 2nd and 3rd layers of the decoder, respectively.

[0039] The encoder consists of 6 convolutional blocks, each containing a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function, progressively extracting multi-scale features from the original erosion image;

[0040] The physical constraint layer is based on the exclusive physical prior system for the entire chain of corrosion evolution of coated steel constructed in step 1. It performs channel-by-channel physical rationality correction on the feature map and filters out features that do not conform to physical laws. The decoder consists of 6 deconvolution blocks, which are used to upsample the corrected feature map to a high-resolution image with a preset multiple resolution.

[0041] Step 3.2: Design the internal structure of the physical constraint layer. Each physical constraint layer contains three sub-modules:

[0042] The feature extraction submodule is used to extract the erosion morphology features from the current feature map;

[0043] The physical parameter inversion submodule is used to invert the corresponding corrosion physical parameters based on the extracted morphological features;

[0044] The feature correction submodule is used to correct features that exceed the reasonable range based on the physical rationality constraint boundary of corrosion features, and generate feature maps that conform to physical laws.

[0045] Step 3.3: Construct a dual discriminator architecture, which includes an image discriminator and a physical consistency discriminator; the image discriminator adopts a fully convolutional network structure, consisting of 5 convolutional blocks, and outputs a 16×16 discrimination matrix to discriminate the local visual realism of the generated image;

[0046] The physical consistency discriminator consists of four fully connected layers, which are used to take as input the erosion morphology features and corresponding physical parameters extracted from the generated image, and output a physical consistency score to determine whether the erosion morphology of the generated image conforms to the physical evolution law.

[0047] Step 3.4: Construct an online physical consistency verification module to extract the erosion morphology features and spatial distribution features of the generated image in real time, and perform global verification based on the physical rationality constraint boundary and spatial distribution rules in Step 1. Local corrections are made for areas that exceed the constraint range to ensure the overall physical rationality of the final output image.

[0048] Step 3.5: The PI-GAN model is designed to be lightweight. Depthwise separable convolution is used to replace standard convolution. Combined with channel pruning and integer quantization techniques, the number of parameters of the PI-GAN model is compressed to the range of memory capacity suitable for portable field detection equipment, and the inference time of a single image meets the requirements of real-time operation.

[0049] Preferably, step 4 specifically includes:

[0050] Step 4.1: Construct a four-dimensional hybrid objective function, the expression of which is:

[0051] (6);

[0052] In the formula, L pixel For pixel loss, L1 loss is used to measure the pixel-level difference between the generated image and the real image; L percep To measure the perceptual loss, the feature differences of layers 3, 4, and 5 of a pre-trained VGG16 network (a 16-layer network for visual geometry) were used; L phy The spatial physical constraint loss consists of two parts: eigenvalue constraint loss and spatial distribution constraint loss; L adv To counteract the loss, the values ​​are calculated jointly by the outputs of the image discriminator and the physical consistency discriminator; λ1~λ4 are weighting coefficients, which are 0.01, 1.0, 0.8 and 0.005 respectively.

[0053] Step 4.2: Calculate the space physical constraint loss, the expression of which is:

[0054] (7);

[0055] In the formula, f i,gen f is the i-th erosion morphology feature extracted from the generated image. i,phy For a reasonable value of this characteristic calculated based on prior physical calculations of corrosion evolution, w i L represents the weight coefficients for each feature; spatial The spatial distribution constraint loss measures the deviation between the spatial distribution of erosion features in the generated image and the physical laws; λ s The spatial distribution constraint weight coefficient is set to 0.3.

[0056] Step 4.3: Perform phased progressive training on the PI-GAN model, including at least three phases:

[0057] In the first pre-training stage, the generator is trained using only pixel loss and perceptual loss, iterating through a preset number of rounds, using an initial learning rate and batch size, until the validation set loss meets the convergence condition.

[0058] In the second adversarial training phase, adversarial loss is added, and the generator and dual discriminator architecture are trained together, iterating through a preset number of rounds, and a learning rate decay strategy is adopted.

[0059] In the third physical constraint reinforcement stage, the image discriminator is frozen, and only the generator and the physical consistency discriminator are trained. The learning rate is further reduced by iterating a preset number of rounds to strengthen the physical rationality of the generated image.

[0060] Step 4.4: Optimize the PI-GAN model parameters by using the AdamW optimizer and setting the weight decay coefficient; employ gradient pruning techniques and set a pruning threshold to prevent gradient explosion; and introduce an early stopping mechanism to terminate training early when the physical consistency score on the validation set no longer improves for a predetermined number of consecutive rounds to prevent overfitting.

[0061] Preferably, step 5 specifically includes:

[0062] Step 5.1: Construct an extreme lighting adaptive preprocessing module. Based on Retinex theory, the image is decomposed into illumination and erosion components. Adaptive gamma correction and contrast enhancement are performed on the illumination component, and noise suppression is performed on the erosion component to eliminate image degradation under extreme lighting conditions.

[0063] Step 5.2: Construct a cross-scene transfer learning module, which combines parameter transfer and feature transfer to transfer the parameters of the pre-trained PI-GAN model to the target scene, and only fine-tunes the local layer parameters of the PI-GAN model; at the same time, introduce domain adaptive loss to reduce the feature distribution difference between the source domain and the target domain, so that the PI-GAN model can adapt to different coating types and service environments.

[0064] Step 5.3: Small sample learning optimization. A meta-learning method is adopted to learn the general features of different corrosion scenarios during the pre-training stage of the PI-GAN model. This allows the PI-GAN model to be fine-tuned with only a small number of target scene samples, thereby improving the reconstruction effect in rare corrosion scenarios.

[0065] Preferably, step 6 specifically includes:

[0066] Step 6.1: Construct an end-to-end super-resolution-detection joint model, using the generator in the trained PI-GAN model as the super-resolution module, and sharing the feature extraction layer with the newly added corrosion detection module to avoid redundant calculations and improve inference efficiency; the corrosion detection module adopts a lightweight semantic segmentation network to realize corrosion region segmentation, corrosion level determination and corrosion rate prediction.

[0067] Step 6.2: Optimize the deployment of the super-resolution-detection joint model by using the TensorRT inference acceleration engine to accelerate the super-resolution-detection joint model and deploying it on portable field detection equipment, UAV inspection equipment, or edge computing nodes; at the same time, design an offline operation mode to support field detection in environments without network access.

[0068] Step 6.3: Based on the super-resolution reconstruction results and corrosion detection data, generate a structural health assessment report. The structural health assessment report includes information on corrosion location, corrosion level, corrosion rate, and remaining service life prediction, providing data support for steel structure maintenance decisions.

[0069] This invention also provides a super-resolution reconstruction system for steel component damage images based on prior knowledge of corrosion evolution physics: implemented using the above-mentioned method for reconstructing damaged images of coated steel components based on prior knowledge of corrosion evolution physics; including:

[0070] The physical prior construction module is used to construct a coating failure kinetic model based on Fick's diffusion law and the coating adhesion degradation law, a coating steel corrosion electrochemical kinetic model based on Faraday's law and Tafel equation, and a rust layer growth kinetic model based on power-law growth law and rust layer porosity evolution equation. It also establishes a quantitative mapping relationship between corrosion morphology characteristics and multiple physical parameters, thereby generating physical rationality constraint boundaries and spatial distribution rules for corrosion characteristics.

[0071] The multimodal dataset construction module is used to simultaneously acquire visible light images, electrochemical test data and three-dimensional morphology data for each test sample, and complete the spatiotemporal alignment, annotation and standardization preprocessing of multi-source data to form a multimodal corrosion damage dataset.

[0072] The PI-GAN model construction module is used to construct a super-resolution model of physical information generative adversarial network (PI-GAN) with embedded multiple layers of physical constraint layers. The PI-GAN model includes a generator, a dual discriminator architecture, and an online physical consistency verification module. The generator adopts an encoder-decoder structure, with physical constraint layers inserted between the 3rd, 4th, and 5th layers of the encoder and the 2nd and 3rd layers of the decoder. Each physical constraint layer includes a feature extraction submodule, a physical parameter inversion submodule, and a feature correction submodule.

[0073] The training module is used to design a four-dimensional hybrid objective function that integrates pixel loss, perceptual loss, spatial physical constraint loss and adversarial loss, and to train the PI-GAN model in stages based on the multimodal corrosion damage dataset.

[0074] The reconstruction module is used to perform super-resolution reconstruction of the original corrosion image to be processed based on the trained PI-GAN model, and output a target resolution corrosion image that conforms to the physical laws of corrosion evolution.

[0075] Beneficial effects:

[0076] 1. This invention, for the first time, systematically constructs a dedicated physical prior system covering the entire chain of corrosion evolution in coated steel, encompassing coating failure dynamics, electrochemical corrosion dynamics of coated steel, and rust layer growth dynamics. Based on this, it defines the physical rationality constraint boundaries and spatial distribution rules for corrosion characteristics; constructs a PI-GAN model embedded with multiple layers of physical constraints; and designs a four-dimensional hybrid objective function including spatial physical constraint loss. Through these techniques, the physical laws of corrosion evolution are embedded within the entire forward and backward propagation process of the PI-GAN model, ensuring that the super-resolution reconstruction results are strictly constrained at both the pixel level and the physical law level. This fundamentally suppresses the generation of non-physical artifacts, exhibiting significant reconstruction advantages, especially for rare corrosion scenarios with scarce training samples.

[0077] 2. This invention constructs a multimodal corrosion damage dataset that integrates visible light images, electrochemical test data, and three-dimensional morphology data, covering all corrosion stages, multiple coating types, and various typical service environments. It also completes the spatiotemporal alignment, annotation, and standardization preprocessing of the multi-source data; and constructs an extreme illumination adaptive preprocessing module and a cross-scene transfer learning module. Through these techniques, the PI-GAN model can learn the cross-modal mapping relationship between image features and underlying electrochemical physical data and three-dimensional morphology data. Furthermore, by combining illumination correction and domain-adaptive transfer, it significantly improves the generalization ability for different coating types, different service environments, and complex on-site illumination conditions.

[0078] 3. This invention designs an encoder-decoder structure and multiple physical constraint layers in the generator of the PI-GAN model. Each physical constraint layer includes a feature extraction submodule, a physical parameter inversion submodule, and a feature correction submodule. A four-dimensional hybrid objective function fusing pixel loss, perceptual loss, spatial physical constraint loss, and adversarial loss is designed, and a phased progressive training strategy is adopted. Through the above technical means, while maintaining the global structural authenticity, it can recover key corrosion morphology features such as corrosion area ratio, pitting density, rust texture roughness, average size of coating bulges, and coating peeling rate with high fidelity. This effectively solves the engineering problem of severe loss of high-frequency details such as early micro-bulges, pitting, and hairline cracks in low-quality field corrosion images.

[0079] 4. This invention features a lightweight design for the PI-GAN model, replacing standard convolution with depthwise separable convolution, and combining channel pruning and integer quantization techniques to compress the number of PI-GAN model parameters to a level suitable for the memory capacity of portable field inspection devices. A joint super-resolution-detection model is built, sharing the feature extraction layer between the super-resolution module and the corrosion detection module, and deployment optimization is achieved using the TensorRT inference acceleration engine. Through these techniques, the PI-GAN model achieves millisecond-level inference time per image, supports offline operation, and can be deployed on portable field inspection devices, UAV inspection devices, or edge computing nodes, solving the problems of high computational load, slow inference speed, and difficulty in engineering deployment of existing super-resolution models.

[0080] 5. This invention constructs an end-to-end super-resolution-detection joint model, sharing a feature extraction layer between the super-resolution module and the corrosion detection module. The corrosion detection module employs a lightweight semantic segmentation network to achieve corrosion region segmentation, corrosion level determination, and corrosion rate prediction. Based on the super-resolution reconstruction results and corrosion detection data, a structural health assessment report is generated, containing information on corrosion location, corrosion level, corrosion rate, and predicted remaining service life. Through these technical means, a complete technical closed loop is formed from low-quality on-site corrosion image input to engineering maintenance decision output, avoiding redundant calculations and information loss caused by the separation of super-resolution reconstruction and downstream detection, and significantly improving the automation and intelligence level of intelligent corrosion detection for steel structures. Attached Figure Description

[0081] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0082] Figure 1 A flowchart of a super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution is provided in this embodiment of the invention.

[0083] Figure 2 A framework diagram of a dedicated physical prior system for the entire chain of corrosion evolution of coated steel provided in this embodiment of the invention;

[0084] Figure 3 This is a diagram of the PI-GAN model architecture with embedded multi-layer physical constraint layers provided in an embodiment of the present invention;

[0085] Figure 4 This is a block diagram illustrating the working principle of the dual discriminator architecture provided in an embodiment of the present invention. Detailed Implementation

[0086] 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.

[0087] like Figure 1 As shown, this embodiment of the invention provides a super-resolution reconstruction method for steel component damage images based on prior knowledge of corrosion evolution physics, including the following steps:

[0088] Step 1: Construct a dedicated physical prior system for the entire corrosion evolution chain of coated steel (see...) Figure 2 ), generating physical rationality constraint boundaries and spatial distribution rules for corrosion features used to constrain the super-resolution reconstruction process; specifically including:

[0089] Step 1.1: Construct a coating failure kinetic model. Based on Fick's diffusion law and the degradation law of coating adhesion, derive the evolution equations for coating blistering, cracking, and peeling, and establish the quantitative relationship between coating failure time and ambient temperature, humidity, and salt spray concentration. The expression is as follows:

[0090] (1);

[0091] In the formula, t f k represents the coating failure time. f C is the coating failure rate constant. salt The concentration of salt spray in the environment is given by E, where a is the concentration index. f R is the activation energy for coating failure, R is the gas constant, and T is the absolute temperature.

[0092] Step 1.2: Construct an electrochemical kinetic model for the corrosion of coated steel. Based on Faraday's law and the Tafel equation, derive the quantitative relationship between corrosion current density and corrosion rate. Simultaneously, introduce a coating failure rate correction coefficient to consider the impact of localized coating failure on the corrosion rate. The expression is:

[0093] (2);

[0094] (3);

[0095] In the formula, β is the coating damage rate correction coefficient, with a value ranging from 0 to 1, i corr Let i be the corrosion current density, i0 be the exchange current density, α be the transfer coefficient, F be the Faraday constant, η be the overpotential, and v be the voltage. corr Here, M is the corrosion rate, n is the molar mass of iron, n is the number of electrons reacted, and ρ is the density of iron.

[0096] Step 1.3: Construct a rust layer growth kinetic model. Based on the power-law growth law and the rust layer porosity evolution equation, establish the mapping relationship between rust layer thickness, porosity, and corrosion time. The expression is:

[0097] (4);

[0098] (5);

[0099] In the formula, d(t) is the thickness of the rust layer at time t, and k dLet be the rust growth rate constant, e be the growth exponent, and its value range be 0.3 to 0.8; ε(t) be the rust porosity at time t, ε0 be the initial porosity, and k be the rust growth rate constant. ε λ and λ are porosity evolution parameters;

[0100] Step 1.4: Establish a quantitative mapping relationship between corrosion morphology features and multiple physical parameters. Define the corrosion morphology feature set F={f1,f2,f3,f4,f5,f6}, where f1 is the corrosion area ratio, f2 is the pitting density, f3 is the average pitting depth, f4 is the rust texture roughness, f5 is the average size of coating bulges, and f6 is the coating peeling rate. Based on multiple sets of accelerated corrosion test and field measurement data, fit the quantitative mapping functions between each morphology feature and coating failure time, corrosion rate, and rust thickness.

[0101] Step 1.5: Generate physical rationality constraint boundaries and spatial distribution rules for corrosion features. Based on the physical laws of corrosion evolution, define the value range of each morphological feature; for example, the diameter of pitting corrosion shall not exceed 3 times the thickness of the rust layer, and the distance between adjacent bulges shall not be less than 1.5 times the diameter of the bulge. At the same time, define the spatial distribution rules of corrosion features, for example, pitting corrosion is preferentially distributed on the edge of the coating bulge, and the roughness of the rust layer texture increases with the increase of corrosion depth, providing a full-dimensional physical constraint basis for super-resolution reconstruction.

[0102] Step 2: Construct a multimodal corrosion damage dataset integrating visible light images, electrochemical test data, and 3D morphology data, covering the entire corrosion stage, multiple coating types, and multiple typical service environments. Complete the spatiotemporal alignment, annotation, and standardization preprocessing of the multi-source data; the specific steps are as follows:

[0103] Step 2.1: Design a multi-condition accelerated corrosion test scheme, select six commonly used engineering coatings such as epoxy coating, polyurethane coating, fluorocarbon coating, chlorinated rubber coating, acrylic coating and inorganic zinc-rich coating, set up four typical service conditions such as salt spray environment, dry and wet alternating environment, industrial atmospheric environment and marine atmospheric environment, and carry out accelerated corrosion test for 0 to 720 days, and collect sample data every 15 days.

[0104] Step 2.2: Establish a multi-source data synchronous acquisition standard, and synchronously acquire multiple types of data for each experimental sample, including:

[0105] Visible light images; orthogonal shooting was performed using a 2448×2048 pixel industrial camera, simultaneously acquiring images under different lighting conditions of 0~1000 lux, as well as images with partial occlusion and dirt interference;

[0106] Electrochemical data: Corrosion current density, polarization curves, and AC impedance spectroscopy were measured using an electrochemical workstation at a sampling frequency of 1 Hz.

[0107] Physical morphological parameters: The thickness of the rust layer was measured using an ultrasonic thickness gauge with an accuracy of 0.01 mm; the surface roughness and pitting depth of the corroded surface were measured using a laser three-dimensional morphology instrument with an accuracy of 0.1 μm.

[0108] Step 2.3: Supplement the collection of on-site engineering data. Collect on-site corrosion images and corresponding detection data of steel structure bridges, factories and pipelines with service life of 3 to 30 years in 12 regions across the country. Supplement rare corrosion samples under extreme service environments to make the multimodal corrosion damage dataset cover diverse actual engineering scenarios.

[0109] Step 2.4: Label and classify the multimodal corrosion damage dataset. According to GB / T6461-2002 and ISO12944 standards, the corrosion damage is divided into five levels: no corrosion, slight corrosion, moderate corrosion, severe corrosion and extremely severe corrosion. For each image, label the corrosion area outline, corrosion level, coating type, service environment, and corresponding electrochemical and physical data and three-dimensional morphology data.

[0110] Step 2.5: Data preprocessing and enhancement are performed. The original images are cropped and normalized, and various geometric transformations are used for data enhancement, including rotation by 90°, 180°, 270°, horizontal and vertical flipping, and scaling by 0.8 to 1.2 times. Electrochemical physical data and three-dimensional morphology data are standardized and outliers are removed. Finally, a multimodal corrosion damage dataset containing 12,000 multimodal samples is constructed, which is divided into training set, validation set and test set in an 8:1:1 ratio, and the samples of each corrosion level, coating type and service environment are evenly distributed.

[0111] Step 3: Construct a physical information generative adversarial network super-resolution model, i.e., the PI-GAN model, which embeds multiple layers of physical constraints, and embed the physical prior of corrosion evolution into the entire forward and backward propagation process of the PI-GAN model;

[0112] See Figure 3 The PI-GAN model includes a generator, a dual discriminator architecture, and an online physical consistency verification module. The construction process of the PI-GAN model specifically includes:

[0113] Step 3.1: Construct a generator architecture with embedded multi-layer physical constraint layers. The generator adopts an encoder-decoder structure, and physical constraint layers are inserted between the 3rd, 4th, and 5th layers of the encoder and the 2nd and 3rd layers of the decoder, respectively.

[0114] The encoder consists of 6 convolutional blocks, each containing a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function, progressively extracting multi-scale features from the original erosion image;

[0115] The physical constraint layer is based on the exclusive physical prior system for the entire chain of corrosion evolution of coated steel constructed in step 1. It performs channel-by-channel physical rationality correction on the feature map and filters out features that do not conform to physical laws. The decoder consists of 6 deconvolution blocks, which are used to upsample the corrected feature map to a high-resolution image with a preset multiple resolution.

[0116] Step 3.2: Design the internal structure of the physical constraint layer. Each physical constraint layer contains three sub-modules:

[0117] The feature extraction submodule is used to extract the erosion morphology features from the current feature map;

[0118] The physical parameter inversion submodule is used to invert the corresponding corrosion physical parameters based on the extracted morphological features;

[0119] The feature correction submodule is used to correct features that exceed the reasonable range based on the physical rationality constraint boundary of corrosion features, and generate feature maps that conform to physical laws.

[0120] Step 3.3: Construct a dual discriminator architecture (see...) Figure 4 It includes an image discriminator and a physical consistency discriminator; the image discriminator adopts a fully convolutional network structure, consists of 5 convolutional blocks, and outputs a 16×16 discrimination matrix to discriminate the local visual realism of the generated image;

[0121] The physical consistency discriminator consists of four fully connected layers, which are used to take as input the erosion morphology features and corresponding physical parameters extracted from the generated image, and output a physical consistency score to determine whether the erosion morphology of the generated image conforms to the physical evolution law.

[0122] Step 3.4: Construct an online physical consistency verification module to extract the erosion morphology features and spatial distribution features of the generated image in real time, and perform global verification based on the physical rationality constraint boundary and spatial distribution rules in Step 1. Local corrections are made for areas that exceed the constraint range to ensure the overall physical rationality of the final output image.

[0123] Step 3.5: The PI-GAN model is designed to be lightweight. Depthwise separable convolution is used to replace standard convolution. Combined with channel pruning and 8-bit integer quantization, the number of parameters of the PI-GAN model is compressed to less than 12MB. The inference time of a single 512×512 image is less than 40ms, which meets the real-time operation requirements of portable field detection equipment.

[0124] Step 4: Design a four-dimensional hybrid objective function that integrates pixel loss, perceptual loss, spatial physical constraint loss, and adversarial loss, and train the PI-GAN model in stages based on the multimodal corrosion damage dataset; specifically including:

[0125] Step 4.1: Construct a four-dimensional hybrid objective function, the expression of which is:

[0126] (6);

[0127] In the formula, L pixel For pixel loss, L1 loss is used to measure the pixel-level difference between the generated image and the real image; L percep To measure the perceptual loss, the feature differences of layers 3, 4, and 5 of the pre-trained VGG16 network are used; L phy The spatial physical constraint loss consists of two parts: eigenvalue constraint loss and spatial distribution constraint loss; L adv To counteract the loss, the values ​​are calculated jointly by the outputs of the image discriminator and the physical consistency discriminator; λ1~λ4 are weighting coefficients, which are 0.01, 1.0, 0.8 and 0.005 respectively.

[0128] Step 4.2: Calculate the space physical constraint loss, the expression of which is:

[0129] (7);

[0130] In the formula, f i,gen f is the i-th erosion morphology feature extracted from the generated image. i,phy For a reasonable value of this characteristic calculated based on prior physical calculations of corrosion evolution, w i L represents the weight coefficients for each feature; spatial The spatial distribution constraint loss measures the deviation between the spatial distribution of erosion features in the generated image and the physical laws; λ s The spatial distribution constraint weight coefficient is set to 0.3.

[0131] Step 4.3: Perform phased progressive training on the PI-GAN model, including at least three phases:

[0132] In the first pre-training stage, the generator is trained using only pixel loss and perceptual loss for 120 iterations. The initial learning rate is set to 0.0002 and the batch size is set to 16, allowing the generator to learn basic image mapping relationships. When the validation set loss no longer decreases for 10 consecutive iterations, the next stage begins.

[0133] In the second adversarial training phase, adversarial loss is added, and the generator and dual discriminator architecture are trained together. After 250 iterations, the learning rate is reduced to 0.5 of the original value, and the batch size is set to 8, which improves the visual realism and physical consistency of the generated images.

[0134] In the third physical constraint strengthening stage, the image discriminator is frozen, and only the generator and the physical consistency discriminator are trained. After 180 iterations, the learning rate is further reduced to 0.2, and the batch size is set to 8 to strengthen the physical rationality of the generated images and suppress the generation of non-physical artifacts.

[0135] Step 4.4: Optimize the PI-GAN model parameters by using the AdamW optimizer with a weight decay coefficient of 0.0001; employ gradient pruning to prevent gradient explosion with a pruning threshold of 1.0; and introduce an early stopping mechanism to terminate training early when the physical consistency score on the validation set no longer improves after 15 consecutive rounds, thus avoiding overfitting.

[0136] Step 5: Construct an extreme lighting adaptive preprocessing module and a cross-scene transfer learning module to improve the generalization ability of the PI-GAN model in complex field environments; specifically including:

[0137] Step 5.1: Construct an extreme lighting adaptive preprocessing module. Based on Retinex theory, the image is decomposed into illumination component and erosion component. Adaptive gamma correction and contrast enhancement are performed on the illumination component, and noise suppression is performed on the erosion component to solve the image degradation problem under extreme lighting conditions such as shadows, overexposure, and backlighting.

[0138] Step 5.2: Construct a cross-scene transfer learning module, which combines parameter transfer and feature transfer to transfer the parameters of the pre-trained PI-GAN model to the target scene, and only fine-tunes the local layer parameters of the PI-GAN model; at the same time, introduce domain adaptive loss to reduce the feature distribution difference between the source domain and the target domain, so that the PI-GAN model can adapt to different coating types and service environments.

[0139] Step 5.3: Small sample learning optimization. To address the problem of insufficient samples in rare corrosion scenarios, a meta-learning method is adopted to learn the general features of different corrosion scenarios during the pre-training stage of the PI-GAN model. This allows the PI-GAN model to be fine-tuned with only a small number of target scene samples, thereby improving the reconstruction effect in rare corrosion scenarios.

[0140] Step 6: Based on the trained PI-GAN model, perform super-resolution reconstruction on the original corrosion image to be processed, and output a target resolution corrosion image that conforms to the physical laws of corrosion evolution; specifically including:

[0141] Step 6.1: Construct an end-to-end super-resolution-detection joint model, using the generator in the trained PI-GAN model as the super-resolution module, and sharing the feature extraction layer with the newly added corrosion detection module to avoid redundant calculations and improve inference efficiency; the corrosion detection module adopts a lightweight semantic segmentation network to realize corrosion region segmentation, corrosion level determination and corrosion rate prediction.

[0142] Step 6.2: Optimize the deployment of the super-resolution-detection joint model by using the TensorRT inference acceleration engine to accelerate the super-resolution-detection joint model and deploying it on portable field detection equipment, UAV inspection equipment, or edge computing nodes; at the same time, design an offline operation mode to support field detection in environments without network access.

[0143] Step 6.3: Based on the super-resolution reconstruction results and corrosion detection data, generate a structural health assessment report. The structural health assessment report includes information on corrosion location, corrosion level, corrosion rate, and remaining service life prediction, providing data support for steel structure maintenance decisions.

[0144] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A super-resolution reconstruction method for damage images of steel components based on prior knowledge of corrosion evolution physics, characterized in that, Includes the following steps: Step 1: Construct a dedicated physical prior system for the entire chain of corrosion evolution of coated steel, and generate physical rationality constraint boundaries and spatial distribution rules for corrosion features to constrain the super-resolution reconstruction process; Step 2: Construct a multimodal corrosion damage dataset that integrates visible light images, electrochemical test data, and three-dimensional morphology data, covering the entire corrosion stage, multiple types of coatings, and multiple typical service environments, and complete the spatiotemporal alignment, annotation, and standardization preprocessing of multi-source data; Step 3: Construct a physical information generative adversarial network super-resolution model, i.e., the PI-GAN model, which embeds multiple layers of physical constraints, and embed the physical prior of corrosion evolution into the entire forward and backward propagation process of the PI-GAN model; Step 4: Design a four-dimensional hybrid objective function that integrates pixel loss, perceptual loss, spatial physical constraint loss and adversarial loss, and train the PI-GAN model in stages based on the multimodal corrosion damage dataset; Step 5: Construct an extreme lighting adaptive preprocessing module and a cross-scene transfer learning module to improve the generalization ability of the PI-GAN model in complex field environments; Step 6: Based on the trained PI-GAN model, perform super-resolution reconstruction on the original corrosion image to be processed, and output a target resolution corrosion image that conforms to the physical laws of corrosion evolution.

2. The super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution as described in claim 1, characterized in that, Step 1, specifically the steps for constructing a dedicated physical prior system for the entire corrosion evolution chain of coated steel, include: Step 1.1: Construct a coating failure kinetic model. Based on Fick's diffusion law and the degradation law of coating adhesion, derive the evolution equations for coating blistering, cracking, and peeling, and establish the quantitative relationship between coating failure time and ambient temperature, humidity, and salt spray concentration. The expression is as follows: （1）； In the formula, t f k represents the coating failure time. f C is the coating failure rate constant. salt The concentration of salt spray in the environment is given by E, where a is the concentration index. f R is the activation energy for coating failure, R is the gas constant, and T is the absolute temperature. Step 1.2: Construct an electrochemical kinetic model for the corrosion of coated steel. Based on Faraday's law and the Tafel equation, derive the quantitative relationship between corrosion current density and corrosion rate. Simultaneously, introduce a coating failure rate correction coefficient to consider the impact of localized coating failure on the corrosion rate. The expression is: （2）； （3）； In the formula, β is a coating damage rate correction coefficient, the value range is 0~1, i corr is the corrosion current density, i0 is the exchange current density, α is the transfer coefficient, F is the Faraday constant, η is the overpotential, v corr is the corrosion rate, M is the molar mass of iron, n is the number of reaction electrons, and ρ is the density of iron; Step 1.3: Construct a rust layer growth kinetic model. Based on the power-law growth law and the rust layer porosity evolution equation, establish the mapping relationship between rust layer thickness, porosity, and corrosion time. The expression is: （4）； （5）； In the formula, d(t) is the rust layer thickness at time t, k d is the rust layer growth rate constant, e is the growth index, and the value range is 0.3-0.8; ε(t) is the rust layer porosity at time t, ε0 is the initial porosity, k ε and λ are the porosity evolution parameters; Step 1.4: Establish a quantitative mapping relationship between corrosion morphology features and multiple physical parameters. Define the corrosion morphology feature set F={f1,f2,f3,f4,f5,f6}, where f1 is the corrosion area ratio, f2 is the pitting density, f3 is the average pitting depth, f4 is the rust texture roughness, f5 is the average size of coating bulges, and f6 is the coating peeling rate. Based on multiple sets of accelerated corrosion test and field measurement data, fit the quantitative mapping functions between each morphology feature and coating failure time, corrosion rate, and rust thickness. Step 1.5: Generate the physical rationality constraint boundary and spatial distribution rules of corrosion features. Based on the physical laws of corrosion evolution, define the value range of each morphological feature; at the same time, define the spatial distribution rules of corrosion features to provide a full-dimensional physical constraint basis for super-resolution reconstruction.

3. The super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution according to claim 1, characterized in that, Step 2 involves constructing a multimodal corrosion damage dataset as follows: Step 2.1: Design a multi-condition accelerated corrosion test scheme, select a variety of commonly used coating types in engineering, set a variety of typical service conditions, carry out long-term accelerated corrosion tests covering the initial to mid-to-late stage of corrosion, and collect sample data at a preset time period. Step 2.2: Establish a multi-source data synchronous acquisition standard, and synchronously acquire multiple types of data for each experimental sample, including: Visible light images; orthogonal images taken under various lighting intensities using an industrial camera with original resolution that meets the input requirements of the PI-GAN model, including images with partial occlusion and dirt interference; Electrochemical data; corrosion current density, polarization curves, and AC impedance spectra were obtained using an electrochemical workstation at a predetermined sampling frequency. Physical morphological parameters, including rust layer thickness measured by a thickness gauge, and surface roughness and pitting depth measured by a three-dimensional morphology instrument; Step 2.3: Supplement the collection of on-site engineering data, and obtain corrosion images and corresponding detection data of steel structure engineering sites from multiple geographical regions and various service years, in order to supplement rare corrosion samples under extreme service environments, so that the multimodal corrosion damage dataset covers diverse actual engineering scenarios. Step 2.4: Label and classify the multimodal corrosion damage dataset. Based on the corrosion evaluation criteria, the corrosion damage is divided into five levels: no corrosion, slight corrosion, moderate corrosion, severe corrosion, and extremely severe corrosion. For each image, label the corrosion area outline, corrosion level, coating type, service environment, and corresponding electrochemical and physical data and three-dimensional morphology data. Step 2.5: Preprocess and enhance the data. Crop and normalize the original images and perform data enhancement using various geometric transformations. Standardize the electrochemical and physical data and remove outliers. Finally, construct a multimodal corrosion damage dataset containing a large number of multimodal samples. Divide the dataset into training, validation, and test sets according to a predetermined ratio, and ensure that the samples of each corrosion level, coating type, and service environment are evenly distributed.

4. The super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution according to claim 1, characterized in that, In step 3, the PI-GAN model includes a generator, a dual discriminator architecture, and an online physical consistency verification module. The construction process of the PI-GAN model specifically includes: Step 3.1: Construct a generator architecture with embedded multi-layer physical constraint layers. The generator adopts an encoder-decoder structure, and physical constraint layers are inserted between the 3rd, 4th, and 5th layers of the encoder and the 2nd and 3rd layers of the decoder, respectively. The encoder consists of 6 convolutional blocks, each containing a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation function, progressively extracting multi-scale features from the original erosion image; The physical constraint layer is based on the exclusive physical prior system for the entire chain of corrosion evolution of coated steel constructed in step 1. It performs channel-by-channel physical rationality correction on the feature map and filters out features that do not conform to physical laws. The decoder consists of 6 deconvolution blocks, which are used to upsample the corrected feature map to a high-resolution image with a preset multiple resolution. Step 3.2: Design the internal structure of the physical constraint layer. Each physical constraint layer contains three sub-modules: The feature extraction submodule is used to extract the erosion morphology features from the current feature map; The physical parameter inversion submodule is used to invert the corresponding corrosion physical parameters based on the extracted morphological features; The feature correction submodule is used to correct features that exceed the reasonable range based on the physical rationality constraint boundary of corrosion features, and generate feature maps that conform to physical laws. Step 3.3: Construct a dual discriminator architecture, which includes an image discriminator and a physical consistency discriminator; the image discriminator adopts a fully convolutional network structure, consisting of 5 convolutional blocks, and outputs a 16×16 discrimination matrix to discriminate the local visual realism of the generated image; The physical consistency discriminator consists of four fully connected layers, which are used to take as input the erosion morphology features and corresponding physical parameters extracted from the generated image, and output a physical consistency score to determine whether the erosion morphology of the generated image conforms to the physical evolution law. Step 3.4: Construct an online physical consistency verification module to extract the erosion morphology features and spatial distribution features of the generated image in real time, and perform global verification based on the physical rationality constraint boundary and spatial distribution rules in Step 1. Local corrections are made for areas that exceed the constraint range to ensure the overall physical rationality of the final output image. Step 3.5: The PI-GAN model is designed to be lightweight. Depthwise separable convolution is used to replace standard convolution. Combined with channel pruning and integer quantization techniques, the number of parameters of the PI-GAN model is compressed to the range of memory capacity suitable for portable field detection equipment, and the inference time of a single image meets the requirements of real-time operation.

5. The super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution according to claim 4, characterized in that, Step 4 specifically includes: Step 4.1: Construct a four-dimensional hybrid objective function, the expression of which is: （6）； In the formula, L pixel For pixel loss, L1 loss is used to measure the pixel-level difference between the generated image and the real image; L percep To measure the perceptual loss, the feature differences of layers 3, 4, and 5 of the pre-trained VGG16 network are used; L phy The spatial physical constraint loss consists of two parts: eigenvalue constraint loss and spatial distribution constraint loss; L adv To counteract the loss, the values ​​are calculated jointly by the outputs of the image discriminator and the physical consistency discriminator; λ1~λ4 are weighting coefficients, which are 0.01, 1.0, 0.8 and 0.005 respectively. Step 4.2: Calculate the space physical constraint loss, the expression of which is: （7）； In the formula, f i,gen f is the i-th erosion morphology feature extracted from the generated image. i,phy For a reasonable value of this characteristic calculated based on prior physical calculations of corrosion evolution, w i L represents the weight coefficients for each feature; spatial The spatial distribution constraint loss measures the deviation between the spatial distribution of erosion features in the generated image and the physical laws; λ s The spatial distribution constraint weight coefficient is set to 0.

3. Step 4.3: Perform phased progressive training on the PI-GAN model, including at least three phases: In the first pre-training stage, the generator is trained using only pixel loss and perceptual loss, iterating through a preset number of rounds, using an initial learning rate and batch size, until the validation set loss meets the convergence condition. In the second adversarial training phase, adversarial loss is added, and the generator and dual discriminator architecture are trained together, iterating through a preset number of rounds, and a learning rate decay strategy is adopted. In the third physical constraint reinforcement stage, the image discriminator is frozen, and only the generator and the physical consistency discriminator are trained. The learning rate is further reduced by iterating a preset number of rounds to strengthen the physical rationality of the generated image. Step 4.4: Optimize the PI-GAN model parameters by using the AdamW optimizer and setting the weight decay coefficient; employ gradient pruning techniques and set a pruning threshold to prevent gradient explosion; and introduce an early stopping mechanism to terminate training early when the physical consistency score on the validation set no longer improves for a predetermined number of consecutive rounds to prevent overfitting.

6. The super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution according to any one of claims 1-5, characterized in that, Step 5 specifically includes: Step 5.1: Construct an extreme lighting adaptive preprocessing module. Based on Retinex theory, the image is decomposed into lighting components and erosion components. Adaptive gamma correction and contrast enhancement are performed on the lighting components, and noise suppression is performed on the erosion components to eliminate image degradation under extreme lighting conditions. Step 5.2: Construct a cross-scene transfer learning module, which combines parameter transfer and feature transfer to transfer the parameters of the pre-trained PI-GAN model to the target scene, and only fine-tunes the local layer parameters of the PI-GAN model; at the same time, introduce domain adaptive loss to reduce the feature distribution difference between the source domain and the target domain, so that the PI-GAN model can adapt to different coating types and service environments. Step 5.3: Small sample learning optimization. A meta-learning method is adopted to learn the general features of different corrosion scenarios during the pre-training stage of the PI-GAN model. This allows the PI-GAN model to be fine-tuned with only a small number of target scene samples, thereby improving the reconstruction effect in rare corrosion scenarios.

7. The super-resolution reconstruction method for steel component damage images based on prior physical knowledge of corrosion evolution according to claim 6, characterized in that, Step 6 specifically includes: Step 6.1: Construct an end-to-end super-resolution-detection joint model, using the generator in the trained PI-GAN model as the super-resolution module, and sharing the feature extraction layer with the newly added corrosion detection module to avoid redundant calculations and improve inference efficiency; the corrosion detection module adopts a lightweight semantic segmentation network to realize corrosion region segmentation, corrosion level determination and corrosion rate prediction. Step 6.2: Optimize the deployment of the super-resolution-detection joint model by using the TensorRT inference acceleration engine to accelerate the super-resolution-detection joint model and deploying it on portable field detection equipment, UAV inspection equipment, or edge computing nodes; at the same time, design an offline operation mode to support field detection in environments without network access. Step 6.3: Based on the super-resolution reconstruction results and corrosion detection data, generate a structural health assessment report. The structural health assessment report includes information on corrosion location, corrosion level, corrosion rate, and remaining service life prediction, providing data support for steel structure maintenance decisions.

Images