Electrochemical constraint-based visual recognition method for corrosion of organic coating

Abstract

The present application belongs to the technical field of intelligent detection of material corrosion, and relates to a visual recognition method for organic coating corrosion based on electrochemical constraint. The method is based on a constructed corrosion visual dataset, extracts color features, texture features and defect distribution features of coating images, constructs a visual feature representation model, and introduces electrochemical parameters as physical constraint information in the model training process. By constructing a feature evolution consistency loss function, the change trend of the visual feature vector is constrained to be consistent with the evolution law of the low-frequency impedance modulus and other electrochemical parameters, so that the model is classified and learned at the same time, and its internal feature representation is guided by the corrosion kinetics law, thereby establishing a physical consistency mapping relationship between visual information and corrosion state, and realizing intelligent recognition of the corrosion grade of organic coating. The method can improve the physical consistency and generalization ability of the recognition result, and is suitable for rapid evaluation of the corrosion state of organic coating in complex environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification technology for material corrosion, and in particular to a visual identification method for organic coating corrosion based on electrochemical constraints, which is particularly suitable for rapid and accurate assessment of the corrosion status of metal structures coated with organic coatings in complex environments. Background Technology

[0002] Organic coatings are the primary means of preventing corrosion in metal structures and are widely used in aerospace, shipbuilding, bridge construction, energy, and chemical industries. As equipment service life increases, coatings inevitably age and fail, ultimately leading to corrosion of the base metal and seriously threatening structural safety. Therefore, timely and accurate identification and assessment of coating corrosion status is crucial for implementing preventative maintenance and ensuring equipment uptime.

[0003] In recent years, with the rapid development of artificial intelligence technology, image-based intelligent corrosion recognition methods have gradually become a research hotspot. These methods acquire images of coating surfaces, automatically extract visual features using algorithms such as deep learning, and establish a mapping relationship between features and corrosion levels, thereby achieving rapid determination of the corrosion state. Compared with manual visual inspection, intelligent recognition methods have the potential advantages of high efficiency and objectivity. However, existing image-based recognition methods still face many challenges in practical applications.

[0004] First, purely data-driven visual recognition models often lack an understanding of the physical nature of corrosion. These models only learn the statistical correlation between image pixels and labels; their decision-making process is a "black box," easily influenced by non-target factors such as changes in lighting, surface contaminants, and image noise, leading to unstable recognition results and poor generalization ability. For example, in insufficient lighting or with salt deposits on the surface, the model may misclassify a normal surface as corroded or miss minor corrosion. Second, the relationship between visual features and the actual corrosion state is not strictly linear, especially in the early stages of corrosion. Medium penetration may have already occurred within the coating, significantly reducing coating resistance, but surface morphology changes may not be obvious. Traditional visual models struggle to capture this "invisible" corrosion evolution, often missing early warnings and the optimal repair window. Furthermore, the training of existing models heavily relies on large amounts of manually labeled data, and manual labeling itself is subjective and inconsistent, further limiting the model's performance ceiling.

[0005] Therefore, there is an urgent need for a method that can integrate the physical knowledge of the corrosion process into a visual recognition model, so that the model can not only "see" the surface, but also "understand" the underlying corrosion nature, thereby significantly improving the accuracy, robustness and interpretability of the recognition. Summary of the Invention

[0006] The purpose of this invention is to provide a visual recognition method for organic coating corrosion based on electrochemical constraints, which aims to overcome the limitations of existing pure visual recognition methods that lack physical basis and are susceptible to interference. By introducing electrochemical parameters as physical constraints, the recognition results of the model are made consistent with the corrosion kinetic laws.

[0007] To achieve the above objectives, the core idea of ​​this invention is to use electrochemical parameters (such as low-frequency impedance modulus and coating resistance) that accurately reflect the internal corrosion state of the coating as prior knowledge. These parameters are then used as constraints during the training of the visual recognition model, forcing the model to ensure that the changing trends of the extracted visual features align with the evolution of the electrochemical parameters (such as a monotonically decreasing impedance over corrosion time). In this way, the model no longer merely fits the surface association between the image and the label, but learns the physical essence of the corrosion process, thus achieving stronger robustness and generalization ability. Specifically, this invention is achieved through the following technical solution: The first step is to acquire image data of the organic coating surface and the corresponding electrochemical parameters, and then perform preprocessing. Input: Original image of the coating surface, synchronously acquired electrochemical impedance spectroscopy data (or extracted electrochemical parameter values).

[0008] Processing Method: Image data can be derived from accelerated corrosion tests in the laboratory, outdoor exposure tests, or actual field inspections. To ensure data quality, the original images are first preprocessed, including illumination normalization to eliminate the influence of different light sources and angles, filtering algorithms to remove image noise, and appropriate contrast enhancement to highlight corrosion features. The preprocessed images are used as input for subsequent feature extraction. Simultaneously, synchronously acquired electrochemical parameters (such as the low-frequency impedance modulus |Z| 0.01 Hz) are used as additional input information for each sample for subsequent physical constraint construction.

[0009] Output: Preprocessed standard image data and corresponding electrochemical parameter values.

[0010] Technical effect: It ensures the consistency of input image quality and provides the necessary physical parameter quantification basis for the subsequent introduction of electrochemical constraints.

[0011] The second step is to extract the visual feature parameters of the image. Input: Preprocessed standard image.

[0012] Processing Methods: To comprehensively quantify the surface condition of the coating, three types of representative visual features are extracted from the preprocessed image. The first type is color distribution features, including the mean, standard deviation, and higher-order moments of each channel in the RGB or HSV color space, used to characterize color differences and hue changes caused by aging or corrosion products. The second type is texture statistical features, calculated using methods such as gray-level co-occurrence matrix or local binary mode to measure parameters such as entropy, energy, contrast, and correlation, used to quantify changes in surface roughness and structural uniformity caused by microcracks, blistering, etc. The third type is spatial distribution features of surface defects, using image segmentation techniques to identify and statistically analyze the area proportion, number density, and spatial distribution patterns of defect areas such as rust spots, blistering, and peeling, used to objectively evaluate the severity of macroscopic damage.

[0013] Output: A visual feature vector that quantifies the visual state of an image.

[0014] Technical effect: It transforms image information into computable numerical vectors, providing multi-dimensional input representations for the model and comprehensively covering different visual manifestations of erosion.

[0015] The third step is to construct a corrosion state identification model (technical implementation path). Input: Visual feature vector and corresponding corrosion level label.

[0016] Processing Method: A multi-layer feature extraction model based on a convolutional neural network or its variants is constructed as the erosion state recognition model. This model employs a multi-layer feature extraction structure, including an input layer, multiple hidden layers, and an output layer. It can automatically learn and fuse the hand-designed visual features mentioned above, establishing an end-to-end mapping from image to erosion level. The output layer corresponds to the classification result of the erosion level (e.g., 4 levels). To avoid dependence on a specific network architecture, this step does not limit the specific network type; any multi-layer model capable of feature extraction and classification is applicable.

[0017] Output: Preliminary corrosion level prediction results.

[0018] Technical effect: A basic mapping relationship between visual features and corrosion levels was established, providing a model framework for the subsequent introduction of physical constraints.

[0019] The fourth step involves introducing electrochemical parameters as constraint information for model training—a core innovation (technical implementation path). Input: visual feature vector, corresponding corrosion level label, and synchronously acquired electrochemical parameter values ​​(e.g., |Z| 0.01 Hz).

[0020] Solution: During model training, in addition to the conventional classification loss function (such as cross-entropy loss Lcls), a feature evolution consistency loss function Lec based on electrochemical constraints is added. The specific construction and calculation process of this constraint term are as follows: Constraint Principle: Based on prior knowledge of corrosion kinetics—as corrosion intensifies, electrochemical parameters characterizing coating protective performance (such as the low-frequency impedance modulus |Z| 0.01 Hz) exhibit a monotonically decreasing trend. Correspondingly, the evolution of the deep visual feature vector extracted by the model should also be consistent with this physical trend.

[0021] Constraint Implementation: For any two samples i and j in the training set, obtain their corresponding electrochemical parameter values ​​Zi and Zj. If Zi > Zj (indicating that the corrosion degree of sample i is lower than that of sample j), then the feature vector fi output by the last hidden layer of the model is required to satisfy a certain metric (e.g., vector norm ||f||, or eigenvalues ​​of a specific dimension) in the feature space with fj, i.e., it is expected that ||fi|| < ||fj|| (if the metric is positively correlated with the corrosion degree) or ||fi|| > ||fj|| (if the metric is negatively correlated with the corrosion degree).

[0022] Consistency judgment mechanism: Construct a ranking loss function. For all sample pairs satisfying Zi > Zj, check the relationship between feature metrics M(fi) and M(fj). If the actual relationship does not conform to the expected direction (e.g., expected M(fi) < M(fj) but actual M(fi) ≥ M(fj)), a penalty term is generated. Lec is the cumulative sum of penalties for all sample pairs that violate the expected relationship.

[0023] Constraints in training: The total loss function is L = Lcls + λ·Lec, where λ is the balance coefficient, and Lec is used to constrain the consistency between the visual feature ranking relationship and the electrochemical parameter ranking relationship. During each gradient backpropagation, the gradient of Lec simultaneously updates the model parameters, guiding the model to adjust the direction of feature extraction. This ensures that the finally learned feature vectors are not only accurate in classification, but their inherent evolutionary trend also conforms to the physical order defined by the electrochemical parameters.

[0024] Output: Feature representations with consistent erosion evolution trained by physical constraints and optimized model parameters.

[0025] Technical effects: By guiding the model's learning process with the laws of corrosion kinetics, the interpretability of the model and the physical consistency of the recognition results are significantly improved, avoiding misjudgments that may occur in purely data-driven models due to violations of physical laws. At the same time, this constraint enhances the model's robustness to non-target interference such as lighting and stains, because the model's core discrimination criterion is the evolution of visual features that conforms to electrochemical laws, rather than accidental changes in the surface.

[0026] The fifth step is to identify the erosion state of the input image based on the trained model (technical implementation path). Input: Image of the coating surface to be tested (after the same preprocessing).

[0027] Processing method: The image to be tested is input into the trained model. After forward propagation, the classification result of the output layer is obtained. At the same time, feature vectors can be extracted from the intermediate layers of the model and mapped to a predicted equivalent electrochemical state information based on the feature-electrochemical parameter correlation established during training, as an auxiliary decision reference.

[0028] Output: Corrosion level of the image under test (and optional electrochemical state prediction).

[0029] Technical effect: It enables rapid, objective, and physically consistent determination of corrosion level from on-site images, providing a reliable technical means for engineering applications.

[0030] Advantages of this invention: First, by introducing electrochemical parameters as physical constraints, the learning process of the visual recognition model is guided by the laws of corrosion kinetics, significantly improving the model's interpretability and the physical consistency of the recognition results, avoiding misjudgments that might occur with purely data-driven models due to violations of physical laws. Second, electrochemical constraints enhance the model's robustness to non-target interference such as light and stains, because the model's core discrimination criterion is the evolution of visual features that conform to electrochemical laws, rather than accidental changes in the surface. Third, this method is particularly beneficial for identifying early corrosion, because even in stages where image changes are not obvious, electrochemical parameters have already changed, and the model can learn this implicit mapping relationship under constraints, thus providing early warning. Finally, the method of this invention does not rely on a specific network architecture, has good versatility and portability, and provides key technical support for building a high-precision, high-reliability intelligent recognition system for organic coating corrosion. Attached Figure Description

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments: Figure 1This is an overall flowchart of the visual recognition method for organic coating corrosion based on electrochemical constraints of the present invention, which shows the complete technical route from image preprocessing, feature extraction, model construction, introduction of electrochemical constraints to final corrosion level recognition; Figure 2 This is a schematic diagram of the core electrochemical constraint mechanism in this invention, which exemplarily demonstrates how the decay law of electrochemical parameters (such as low-frequency impedance modulus) can be transformed into a constraint on the evolution of visual features, so that the feature space learned by the model is aligned with the physical space. Figure 3 This is a comparison chart of the recognition results on the test set with those of a traditional unconstrained visual model in one embodiment of the present invention, demonstrating the significant improvement in accuracy and robustness of the method of the present invention. Detailed Implementation

[0032] The present invention will be further explained below with reference to specific implementation schemes, but it is not limited to the present invention. The structures, proportions, sizes, etc. shown in the accompanying drawings are only used to complement the content disclosed in the specification, so as to enable those skilled in the art to understand and read, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modification of the structure, change of the proportion relationship or adjustment of the size, without affecting the effect and purpose that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to two specific embodiments. Embodiment 1 will elaborate on training the model of this invention using a pre-constructed high-quality dataset in a laboratory environment and verifying its performance. Embodiment 2 will demonstrate the real-world application effect of this method on a coastal steel structure tower, verifying its feasibility and superiority in complex engineering environments.

[0034] Example 1: Model Training and Validation in a Laboratory Environment Data sources and preparation.

[0035] This embodiment utilizes existing datasets in the field or constructs datasets through acquisition. The dataset is based on a 2A12 aluminum alloy / epoxy polyurethane coating system. Under cyclic salt spray / drying / humid heat / UV accelerated corrosion test conditions, surface images and electrochemical impedance spectroscopy were simultaneously acquired. After automatic labeling and consistency verification of electrochemical parameters, approximately 2000 high-quality data samples were ultimately generated. Each sample includes a coating surface image taken under a standard light source, the corresponding key electrochemical parameters (0.01Hz low-frequency impedance modulus |Z|0.01Hz), and a manually verified corrosion level label (Level 0: Intact, Level 1: Initial Penetration, Level 2: Mid-term Failure, Level 3: Severe Corrosion). The dataset was randomly divided into training, validation, and test sets in a 7:2:1 ratio to ensure a consistent distribution of corrosion levels across all sets.

[0036] Image preprocessing and feature extraction.

[0037] All images in the training set were preprocessed. First, an adaptive histogram equalization algorithm was used for illumination normalization to eliminate brightness differences between batches of images. Then, a mean filter was applied to remove salt-and-pepper noise that might have been introduced during image acquisition. Finally, linear stretching was used to adjust the image contrast to a uniform range. After preprocessing, hand-designed visual features were extracted from each image as part of the model input. Color features were extracted from the mean, standard deviation, and skewness of the three RGB channels, resulting in 9 dimensions. Texture features were calculated based on the gray-level co-occurrence matrix, selecting the mean values ​​in four directions (0°, 45°, 90°, 135°) to obtain 5 dimensions: contrast, correlation, energy, homogeneity, and entropy. Defect distribution features were obtained using a pre-trained U-Net semantic segmentation model, automatically segmenting blister and rust spot regions in the image, calculating the defect area ratio, defect density, and average defect area, resulting in 3 dimensions. Finally, each image was represented as a 17-dimensional hand-designed feature vector.

[0038] Model building and introduction of electrochemical constraints.

[0039] A multi-layer feature extraction model constructed using a convolutional neural network or its variants is used as the corrosion state recognition model. This model comprises an input layer (accepting 17-dimensional handcrafted features), two hidden layers (64 and 32 neurons respectively, with ReLU activation), and an output layer (4 neurons, corresponding to four corrosion levels, with Softmax activation). The model's total loss function consists of two parts: the first is the standard cross-entropy classification loss Lcls, used to measure the difference between the model's predicted corrosion level and the true label; the second is the feature evolution consistency loss Lec based on electrochemical constraints. Lec is constructed based on the following physical constraint: for any two samples i and j in the training set, if the low-frequency impedance modulus |Z|i corresponding to sample i is greater than |Z|j corresponding to sample j (i.e., the corrosion level of i is lower than that of j), then the norm ||f|| of the feature vectors fi and fj output by the last hidden layer of the model should satisfy ||fi|| < ||fj|| (because the feature norm is positively correlated with the corrosion level in this embodiment). In practice, a ranking loss function is used. For all sample pairs that satisfy |Z|i > |Z|j, a penalty term is generated if the actual ||fi|| ≥ ||fj||. Lec is the sum of penalties for all sample pairs that violate the expected relationship. The total loss L = Lcls + λ·Lec, where λ is the balance coefficient, determined through validation set tuning (λ is 0.3 in this embodiment).

[0040] Model training and performance validation.

[0041] The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. Loss was monitored on the validation set, and training was terminated early when the loss stopped decreasing after 10 consecutive epochs. During training, the classification accuracy on the validation set was calculated after each epoch. After training, performance was evaluated on an independent test set. Results showed that the proposed method achieved an accuracy of 87.3% in corrosion level identification on the test set. In contrast, the baseline model with the electrochemical constraint removed (i.e., λ=0) achieved an accuracy of only 77.8% on the same dataset. Further analysis of the confusion matrix revealed that the proposed method significantly improved the identification accuracy for levels 2 (intermediate failure) and 3 (severe corrosion), while reducing the proportion of early level 1 samples misclassified as level 0 from 18% in the baseline model to 9%. This result fully demonstrates that the introduction of electrochemical constraints effectively guides the model to learn the evolution direction of visual features that conforms to physical laws, improving the model's ability to capture the essence of corrosion, especially in the early stages and at level boundaries. Furthermore, robustness tests were conducted by artificially adding simulated light changes and salt contamination images to the test set. The accuracy decrease of the method of this invention (approximately 5%) was significantly smaller than that of the baseline model (approximately 12%). This indicates that the electrochemical constraint makes the model focus more on the evolution of features related to the nature of corrosion rather than random changes on the surface, which significantly enhances the model's resistance to non-target interference.

[0042] Example 2: Field Application Verification of Coastal Steel Structure Tower Application scenarios and data collection.

[0043] This embodiment selects a steel-structured broadcast television tower in a coastal area of ​​eastern my country that has been in service for 15 years as the verification object. The tower is 150 meters high, with a main structure of Q345B steel and a surface coating system of epoxy zinc-rich primer + epoxy micaceous iron oxide intermediate paint + acrylic polyurethane topcoat. It has been exposed to a marine atmospheric environment with high salt spray and strong sunlight for a long time. To verify the field applicability of the method of this invention, the project team selected 30 typical measuring points at different heights (30 meters, 60 meters, and 90 meters) and different orientations (southeast windward side, northwest leeward side) of the tower and conducted quarterly follow-up inspections for one year. During each inspection, inspectors arrived at the measuring point using an aerial work platform and took five surface images at the same location using a handheld industrial intelligent terminal (equipped with a ring standard light source). Simultaneously, a portable solid gel electrolyte probe and an electrochemical workstation were used to measure the low-frequency impedance modulus of the measuring point. The synchronously collected data were used to construct a field test set. In addition, images and impedance data of different corrosion stages were obtained from the furnace hangers exposed at the bottom of the tower at the same time, which were used for transfer learning of the model trained in the laboratory.

[0044] Model transfer and field application.

[0045] The model trained on the laboratory dataset in Example 1 was used as the base model. Considering the differences between the field coating system and the laboratory samples, as well as the complexity of the field environment, approximately 300 sets of field data obtained from the bottom of the tower were used to perform transfer learning and fine-tuning on the model. During fine-tuning, the parameters of the first two layers of the model were frozen, and only the last two fully connected layers were trained a small number of times. An electrochemically constrained feature evolution consistency loss (Lec) was also introduced to ensure that the model maintained physical consistency of feature evolution while adapting to the field data. After fine-tuning, the model was deployed to a portable terminal for real-time identification of field measurement points. The identification process was as follows: the images taken at the field were preprocessed and manually feature-extracted in the same way as during training, and then input into the model, which directly output the corrosion level. Simultaneously, the feature vector of the last hidden layer of the model was mapped to a predicted equivalent impedance value, which was compared with the measured impedance value as a reference for confidence assessment.

[0046] Identification results and comparative analysis.

[0047] During four quarterly inspections over a year, a total of 120 sets of on-site measurement data were obtained (30 measurement points × 4 times). The recognition results of the model of this invention were compared with the visual rating results of three senior corrosion engineers, with the expert's comprehensive opinion as the reference standard. Statistical results show that the average recognition accuracy of the model of this invention reached 88.3%, with an accuracy rate of over 90% for levels 0 and 3, and approximately 85% for levels 1 and 2. In contrast, another commercial visual recognition software deployed on-site, based solely on image training, had an average accuracy rate of only 75.2% on the same test set, and repeatedly misclassified normal surfaces in backlight as level 1 corrosion, or misclassified early slight salt stains as level 2 blistering. Specific recognition results are shown in the table below: In the early stage of corrosion, the actual on-site condition was slight aging, and the method of this invention correctly identified it, while the traditional method misidentified it due to light interference; in the middle stage of corrosion, obvious blistering appeared on the surface, and both methods could correctly identify it; in the later stage of corrosion, severe rust appeared, and the method of this invention also correctly identified it. Further analysis revealed that the method of this invention has a higher tolerance to interference from changes in lighting and slight surface salt deposits, and its false positive rate is significantly lower than that of traditional visual models. This is attributed to the fact that electrochemical constraints ensure the physical consistency of the feature evolution laws learned by the model, effectively distinguishing between real corrosion changes and accidental surface interference. Field inspection personnel reported that the method is easy to operate, with a single identification time of no more than 10 seconds, effectively assisting frontline maintenance personnel in rapid corrosion screening and severity determination, providing objective and reliable data support for targeted tower maintenance.

[0048] Summary of project results.

[0049] This embodiment fully verifies the feasibility and superiority of the method of the present invention in real and complex engineering environments. By introducing electrochemical constraints, the model possesses physical consistency and enhanced anti-interference ability, maintaining a high recognition accuracy under varying lighting and surface conditions on site. This method can not only accurately identify macroscopic corrosion that has already appeared, but also perceive the early degradation trend of coating performance to a certain extent. The final technical solution can be directly embedded into a portable detection terminal to achieve rapid field detection, which is of great significance for improving the intelligent operation and maintenance level of corrosion in large steel structure facilities.

[0050] In summary, the present invention provides a visual recognition method for organic coating corrosion based on electrochemical constraints. By introducing the evolution law of electrochemical parameters as a physical constraint into model training, it effectively overcomes the inherent defects of traditional pure visual recognition methods, significantly improves the accuracy, robustness and interpretability of recognition results, and has broad application prospects.

[0051] Matters not covered in this invention are common knowledge.

[0052] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A visual recognition method for corrosion of organic coatings based on electrochemical constraints, characterized in that, Includes the following steps: (1) Obtain image data of the organic coating surface and the corresponding electrochemical parameters, and preprocess the images; (2) Extract visual feature parameters of the image, including color features, texture features and defect distribution features; (3) Construct a corrosion state recognition model based on visual features; (4) Electrochemical parameters are introduced as physical constraint information. During the model training process, feature evolution consistency constraints are constructed to ensure that the changing trend of the visual feature vector extracted by the model is consistent with the evolution law of the electrochemical parameters. (5) Based on the trained model, the corrosion state of the input image is identified and the corresponding corrosion level is output.

2. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The image preprocessing includes illumination normalization, noise removal, and contrast enhancement.

3. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The visual features include color distribution features, texture statistical features, and surface defect spatial distribution features.

4. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The recognition model is a multi-layer feature extraction model constructed from a convolutional neural network or its variants.

5. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The electrochemical parameters include the low-frequency impedance modulus and the coating resistance.

6. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The specific implementation of the electrochemical constraint in step (4) is as follows: Construct a feature evolution consistency loss function. For any two samples in the training samples, constrain the relative distance or sorting relationship between the corresponding visual feature vectors according to the magnitude relationship of their electrochemical parameter values, so that the evolution direction of the feature vector is consistent with the physical evolution direction of the electrochemical parameters.

7. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The feature evolution consistency loss function is constructed based on the ranking constraints between sample pairs: for any two samples i and j, if the electrochemical parameter value of sample i is greater than that of sample j, it means that the corrosion degree of i is lower than that of j. Then, the feature vectors fi and fj extracted by the constraint model satisfy a certain measure (including norm and specific dimension value) in the feature space. When the actual relationship does not match the expectation, a penalty term is generated.

8. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The model output includes both corrosion level and corresponding electrochemical state information.

9. The method for visual recognition of organic coating corrosion based on electrochemical constraints according to claim 1, characterized in that, The corrosion levels include at least four levels.

Images