A method and device for rapid evaluation of metal corrosion resistance based on deep learning

Abstract

The application discloses a kind of based on deep learning's metal corrosion resistance performance rapid evaluation method and device, mainly related to metal smelting technical field.It includes obtaining the original data of array droplet experimental metal sample collected by laser confocal microscope;Based on height scale and area scale, obtain the three-dimensional height image of corrosion sample;By pre-training UNet model, the metal matrix and corrosion product area are segmented;Based on the relative height difference of each pixel in corrosion product area, the arithmetic average roughness Ra is calculated, and the surface roughness parameter of corrosion product is obtained;Under the condition of same corrosion time, the height, area, volume and roughness of steel corrosion product are counted, and the rapid comparative evaluation of the corrosion resistance of metal sample is realized.The application has the beneficial effect that the area, volume and surface roughness parameter of corrosion product are quickly and accurately calculated, support multi-time point data analysis and model online self-learning, realize the automatic, quantitative and rapid evaluation of metal corrosion resistance.

Description

Technical Field

[0001] This invention relates to the field of metal smelting technology, specifically a method and apparatus for rapid evaluation of the corrosion resistance of metals based on deep learning. Background Technology

[0002] In the field of metallic materials, especially in the research and development of new materials such as high-end alloys and corrosion-resistant steels, evaluating their corrosion resistance is a crucial but extremely time-consuming step. Traditionally, researchers need to prepare a large number of standard samples and conduct long-term immersion experiments for hundreds or even thousands of hours in simulated or real harsh corrosive environments. Subsequently, they analyze the corrosion morphology and damage depth using a series of characterization methods such as metallographic microscopy and scanning electron microscopy. This process is not only lengthy and costly, but also heavily reliant on the experience of operators, highly subjective, and difficult to guarantee repeatability. This "trial and error" research and development model has become a key bottleneck restricting the rapid iteration and application of new materials, failing to meet the urgent needs of modern industry for efficient materials research and development.

[0003] To accelerate material screening processes, high-throughput experimental techniques have emerged. Among them, array droplet corrosion experiments have shown great potential. This method constructs an array of dozens to hundreds of independent micro-regions with varying compositions or processing techniques on a single, carefully treated sample surface using microdroplet distribution technology, and then accelerates corrosion under controlled conditions. This method can perform comparative experiments on a single sample in parallel, which previously required a large number of samples, greatly saving raw materials, reducing experimental variables, and significantly improving experimental efficiency and data output. Currently, this type of high-throughput method is being increasingly widely used in industrial R&D scenarios such as initial material selection, composition optimization, and rapid determination of process windows.

[0004] However, high-throughput experiments generate massive amounts of corrosion morphology image data. Traditional image analysis methods typically rely on manual feature extraction (such as the total volume of corrosion products, maximum height, and corrosion area), which is inefficient and struggles to comprehensively capture the deep corrosion information contained within complex three-dimensional morphologies. In recent years, with the development of deep learning technology, convolutional neural networks and semantic segmentation networks have been introduced into the field of steel corrosion image recognition to achieve automatic segmentation of corrosion areas and classification of corrosion levels. Existing deep learning methods have achieved good results in corrosion defect recognition based on ordinary visible light images. However, existing technologies mainly target two-dimensional color or grayscale images, only outputting the corrosion area mask or corrosion area ratio, failing to reflect three-dimensional information and unable to directly provide quantitative parameters such as the actual height, total volume, and surface roughness of corrosion products.

[0005] Furthermore, existing deep learning-based erosion segmentation methods typically employ offline training, requiring the pre-construction of large-scale manually labeled datasets, and the network parameters are fixed after training. When experimental conditions, imaging equipment, or erosion morphology change, the model often needs to be re-labeled and retrained to adapt to the new scenario, lacking online adaptive capabilities. Summary of the Invention

[0006] The purpose of this invention is to provide a method and apparatus for rapid evaluation of the corrosion resistance of metals based on deep learning. This method can quickly determine the quality of corrosion resistance among different metal samples, greatly shortening the test cycle.

[0007] To achieve the above objectives, the present invention employs the following technical solution: A rapid evaluation method for the corrosion resistance of metals based on deep learning includes the following steps: To obtain raw data of array droplet experimental metal samples acquired by laser confocal microscopy; Region cropping is performed on the original data image based on the height and area scales to obtain a three-dimensional height image containing the height and area scales; The 3D height image is input into the pre-trained UNet semantic segmentation network to perform pixel-level segmentation of the metal matrix region and corrosion product region in the image, and obtain the corresponding semantic segmentation mask. Pixels or regions with confidence scores greater than a first threshold in the segmentation results are filtered out. The high-confidence masks obtained from the filtering are added to the incremental training set along with the corresponding original images. When the preset triggering conditions are met, the UNet network is trained online using the incremental training set. The arithmetic mean roughness Ra is calculated based on the relative height difference of each pixel within the corrosion product area, thus obtaining the surface roughness parameters of the corrosion product. By statistically analyzing the height, area, volume, and roughness of corrosion products on steel under the same corrosion time conditions, a rapid comparative evaluation of the corrosion resistance of metal samples can be achieved.

[0008] Preferably, the array droplet experiment on the metal sample includes: The metal to be tested is pretreated; An array droplet experiment was conducted on a metal sample using an experimental setup.

[0009] Preferably, the pretreatment includes first cutting the metal to be tested into a metal sample of size 10×10×3 mm using a wire cutting machine, grinding the surface of the metal sample to 1500# with sandpaper, cleaning with anhydrous ethanol, and drying for later use; then mounting the metal sample using a metallographic mounting machine, and grinding and polishing it.

[0010] Preferably, the array droplet experiment includes using a pipette to uniformly drop a pre-prepared solution onto the polished sample surface, ensuring that the droplets are of uniform volume. Then, the sample is placed in a test chamber and subjected to a static corrosion environment of 25 °C and 100% humidity for 2 hours. After corrosion, the sample is removed and placed in anhydrous ethanol for 30 seconds to stop the corrosion and then dried for later use.

[0011] Preferably, obtaining the three-dimensional height image includes: after region cropping, performing grayscale conversion and contrast enhancement processing on the cropped region; using a pre-trained Tesseract + LSTM character recognition model to perform optical character recognition on the image, extracting the scale text information in the height and area scales and their corresponding pixel color parameters in the image; based on the scale text information and pixel color parameters, establishing a mapping relationship between image pixels and actual height values ​​to obtain the height scale; and establishing a mapping relationship between image pixels and actual area or length to obtain the area scale.

[0012] Preferably, the first threshold is 0.72.

[0013] The incremental training includes: (1) Statistically analyze the classification probability of each pixel output by the UNet network, and mark pixels with a classification probability greater than the first threshold as high confidence pixels; (2) Cluster the high-confidence pixels according to the connected components to form high-confidence corrosion product regions and high-confidence metal matrix regions; (3) The high-confidence region mask is used as a pseudo-label and added to the incremental training set along with the corresponding original image; (4) When the number of samples in the incremental training set reaches the preset number or the interval time reaches the preset duration, start online incremental training to update the UNet network parameters.

[0014] Preferably, the arithmetic mean roughness Ra is obtained through the following steps: using the average height of each pixel in the metal substrate region or the height of the fitted plane as the reference height; calculating the relative height difference of each pixel in the corrosion product region and multiplying it by the corresponding pixel area; multiplying the height and pixel area to obtain the pixel volume; summing all pixel volumes to obtain the total volume of the corrosion product; and calculating the arithmetic mean roughness Ra by the relative height difference of each pixel in the corrosion product region.

[0015] Preferably, the calculation of the total volume of corrosion products and surface roughness parameters includes: (1) The average height of each pixel in the metal substrate area or the height of the fitted plane is used as the reference height; (2) For each pixel within the corrosion product region, calculate its relative height difference and multiply it by the corresponding pixel area to obtain the pixel volume; sum all pixel volumes to obtain the total volume of the corrosion product; the calculation formula includes:

[0016] (3) Among them, A collection of pixels representing corrosion products. For pixel height, Let the pixel area be denoted by . The arithmetic mean roughness Ra is calculated based on the relative height difference between pixels within the corrosion product region, yielding the surface roughness parameters of the corrosion product. The calculation formula includes:

[0017] in, A collection of pixels representing corrosion products. For pixel height, For pixel area, The area represents the corrosion products.

[0018] As another aspect of the present invention, a rapid evaluation device for the corrosion resistance of metals based on deep learning includes: The acquisition module is configured to acquire raw data of the array droplet experimental metal sample obtained by laser confocal microscopy. The processing module is configured to perform region cropping on the image of the original data based on the height and area scales to obtain a three-dimensional height image containing the height and area scales; The segmentation module is configured to input a 3D height image into a pre-trained UNet semantic segmentation network to perform pixel-level segmentation of the metal matrix region and corrosion product region in the image, thereby obtaining the corresponding semantic segmentation mask. The training module is configured to filter pixels or region masks with confidence scores greater than a first threshold in the segmentation results, add the filtered high-confidence masks along with the corresponding original images to the incremental training set, and use the incremental training set to perform online incremental training on the UNet network when a preset trigger condition is met. The calculation module is configured to calculate the arithmetic mean roughness Ra based on the relative height difference of each pixel in the corrosion product area, and obtain the surface roughness parameters of the corrosion product. The evaluation module is configured to statistically analyze the height, area, volume, and roughness of corrosion products on steel under the same corrosion time conditions, enabling rapid comparative evaluation of the corrosion resistance of metal samples.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This method employs high-throughput array droplet experiments, which can rapidly acquire large amounts of data, ensuring the reliability of the results.

[0020] 2. This method provides a novel approach to processing data results when high-throughput experiments are applied in the field of materials, which greatly promotes the development of high-throughput experiments in the field of materials.

[0021] 3. This invention can simultaneously obtain multi-dimensional quantitative indicators such as the coverage of corrosion products, the accumulation volume, and the degree of morphological deterioration, providing a more comprehensive, detailed, and quantitative basis for evaluating the corrosion resistance of metals; 4. The method and apparatus of the present invention can automatically process corrosion test images over a long period of time and at multiple time points in batches, and quickly output graphs showing the changes in height, area, volume and roughness of corrosion products over time. This significantly improves the efficiency of experimental data processing and material screening, and is suitable for rapid evaluation and comparison of corrosion resistance in laboratory research and engineering field.

[0022] In summary, this method, by combining high-throughput experiments with deep learning, provides a rapid and reliable approach to evaluating the corrosion resistance of metals, offering important reference and technical support for the development of new corrosion-resistant metals. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the principle of the present invention.

[0024] Figure 2 This is a flowchart of the program processing of the present invention.

[0025] Figure 3 This is a diagram of the experimental apparatus of the present invention.

[0026] Figure 4 This is a diagram of the original data of the corrosion points in Example 1.

[0027] Figure 5 The graph shows the calculation results of the program in Example 1.

[0028] Figure 6 The image shows the results of manual calculations for Example 1. Detailed Implementation

[0029] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0030] Unless otherwise specified, the instruments, reagents, and materials used in the following embodiments are all conventional instruments, reagents, and materials already available in the prior art and can be obtained through legitimate commercial channels. Unless otherwise specified, the experimental methods and detection methods used in the following embodiments are all conventional experimental methods and detection methods already available in the prior art.

[0031] Example 1: The device designed in this example is as follows: Pipettes: minimum graduation is 0.1 µl; Ultrasonic humidifier, rated capacity: 3L, rated humidification output: 300 ml / h, rated power: 25 W, rated voltage: 220 V; Test chamber, chamber dimensions: 40×25×15 cm; Chamber functions: ① Temperature and humidity sensor: real-time monitoring of temperature and humidity inside the chamber; ② Insulation layer: adopts a double-layer structure to reduce temperature fluctuations inside the chamber, and has a built-in electric heating tube to maintain the temperature inside the chamber.

[0032] The specific steps used in this example are as follows: 1) Pre-treat the metal to be tested: First, small-sized specimens suitable for the test requirements, i.e., 10×10×3 mm metal specimens, are cut from the rolled large-size steel plate. Then, the surface of the metal specimen is successively polished to 1500# using SiC sandpaper, cleaned with anhydrous ethanol, and dried for later use. Finally, the metal specimen is mounted using a metallographic mounting machine, and then polished. When grinding on sandpaper of different grits, the specimen is rotated 90° perpendicular to the old marks each time the sandpaper is changed, and grinding is continued in this direction until the old marks completely disappear and the new marks are uniform.

[0033] This example selected three metals for experimental verification: Q420RE, Q420, and Q355. Their specific compositions are shown in the table below:

[0034] 2) Using experimental apparatus to test metal samples Array of droplets experiment: Using a pipette with a minimum graduation of 0.1 µl, evenly drop the pre-prepared solution onto the polished sample surface, ensuring that the droplets are of uniform volume. Then, place the sample in the test chamber and allow it to corrode for 2 hours at 25 °C and 100% humidity. Remove the sample and place it in anhydrous ethanol for 30 seconds to stop the corrosion. Dry the sample for later use.

[0035] The solutions used in this experiment were NaCl solutions of different concentrations, specifically 1.2% NaCl, 1.6% NaCl, 2.0% NaCl, 2.4% NaCl, 2.8% NaCl, 3.2% NaCl, 3.6% NaCl, and 4.0% NaCl.

[0036] 3) Obtaining data using a laser confocal microscope Raw data: The sample treated with the droplet experiment was placed horizontally under a microscope using modeling clay; a suitable lens was selected to ensure that the complete morphology of the corrosion sites could be captured; the image containing the corrosion information was exported and saved, which is the raw image data; 4) Input the raw data Network Models and Data Processing Procedures Corrosion data of the samples and comparative evaluation results were obtained.

[0037] Specifically, the following steps are included: 4.1) Based on the approximate positions of the height and area scales, crop the image to the appropriate regions.

[0038] 4.2) Perform grayscale conversion and contrast enhancement on the cropped area.

[0039] 4.3) Using pre-trained Tesseract + LSTM Character recognition The model performs optical character recognition on the image, extracts the scale text information in the height and area scales and their corresponding pixel color parameters in the image; based on the scale text information and pixel color parameters, it establishes a mapping relationship between image pixels and actual height values ​​to obtain a height scale; and establishes a mapping relationship between image pixels and actual area or length to obtain an area scale.

[0040] 4.4) The 3D height image is input into a pre-trained UNet semantic segmentation network to segment the metal substrate in the image. The region and the corrosion product region are segmented at the pixel level to obtain the corresponding semantic segmentation mask; The UNet network adopts an encoder-decoder structure and sets up multi-level skip connections to fuse multi-scale features, thereby improving the segmentation accuracy of eroded region edges and details.

[0041] A three-dimensional height image is an image that contains information on the height, area, volume, and roughness-related parameters of corrosion products.

[0042] 4.5) Filter the pixel or region masks with confidence greater than the first threshold in the segmentation results, add the filtered high-confidence masks along with the corresponding original images to the incremental training set, and use the incremental training set to perform online incremental training on the UNet network when the preset triggering conditions are met.

[0043] Specifically, the first threshold is set to 0.72.

[0044] The incremental training of the UNet semantic segmentation network specifically includes: (1) Statistically analyze the classification probability of each pixel output by the UNet network, and mark pixels with a classification probability greater than the first threshold as high confidence pixels; (2) Cluster the high-confidence pixels according to the connected components to form high-confidence corrosion product regions and high-confidence metal matrix regions; (3) The high-confidence region mask is used as a pseudo-label and added to the incremental training set along with the corresponding original image; (4) When the number of samples in the incremental training set reaches the preset number or the interval time reaches the preset duration, start online incremental training to update the UNet network parameters.

[0045] 4.6) Use the average height of each pixel within the metal substrate area or the height of the fitted plane as the reference height.

[0046] 4.7) For each pixel within the corrosion product region, calculate its relative height difference and compare it with the corresponding pixel area.

[0047] 4.8) Multiply the height and pixel area to obtain the pixel volume; sum all pixel volumes to obtain the total volume of the corrosion product; calculate the arithmetic mean roughness Ra by the relative height difference of each pixel in the corrosion product area to obtain the surface roughness parameter of the corrosion product.

[0048] The calculation of the total volume of corrosion products and surface roughness parameters includes: (1) The average height of each pixel in the metal substrate area or the height of the fitted plane is used as the reference height; (2) For each pixel within the corrosion product region, calculate its relative height difference and multiply it by the corresponding pixel area to obtain the pixel volume; sum all pixel volumes to obtain the total volume of the corrosion product; the calculation formula includes:

[0049] (3) Among them, A collection of pixels representing corrosion products. For pixel height, The pixel area; The surface roughness parameter of the corrosion product is obtained by calculating the arithmetic mean roughness Ra of the relative height difference of each pixel in the corrosion product area. The calculation formula includes:

[0050] in, A collection of pixels representing corrosion products. For pixel height, For pixel area, The area represents the corrosion products.

[0051] 4.9) Under the same corrosion time conditions, compare the height, area, volume and roughness evolution of corrosion products of steel to achieve rapid comparative evaluation of the corrosion resistance of various steels.

[0052] like Figure 4 The image shows the three-dimensional morphology of corrosion sites in NaCl electrolyte droplets of three steels after 2 hours of corrosion. In the electrolyte droplets, the main driving force for corrosion is the oxygen concentration cell formed by the difference in oxygen concentration between the droplet center and edge. The high-oxygen region at the droplet edge becomes the cathode of the oxygen concentration cell, while the low-oxygen region in the droplet center becomes the anode. (Fe) 2+ It forms from the center of the droplet and diffuses towards the edge, while OH - It forms from the edge of the droplet and diffuses towards the center. Fe 2+ With OH - The deposits combine with sediment to form ring-shaped corrosion products.

[0053] Depend on Figure 6 (a) The total volume curve of corrosion products shows that for Q420RE and Q420, when the NaCl solution concentration is higher than 2.8 wt.%, the total volume of corrosion products increases with increasing solution concentration, indicating that the corrosion degree of the sample is more severe. This indicates that when the solution concentration is lower than 2.8 wt.%, the effect of the oxygen concentration cell is not obvious or is suppressed, and the overall corrosion tends to be more uniform. The corrosion products produced by uniform corrosion will inhibit the diffusion of ions in the droplets, leading to a decrease in the overall corrosion degree. As can be seen from the figure, when the concentration is lower than 2.8 wt.%, the total volume of corrosion products of the two steels does not change much, while when the solution concentration is higher than 2.8 wt.%, the total volume of corrosion products of both increases significantly. Figure 4 The three-dimensional morphology of the corrosion products also shows that, with the continuous increase of solution concentration, the ring-shaped corrosion products become more and more obvious. Figure 6 (b) Cross-sectional area curves of corrosion products and Figure 6 The average height curve of corrosion products in (c) also shows a similar pattern. However, for Q355 steel, the total volume of corrosion products significantly increases after the NaCl solution concentration exceeds 2.0 wt.%, and from... Figure 6 As shown in (a), the total volume of corrosion products of Q420RE steel is significantly lower than that of Q420 and Q355 steels. This phenomenon is also reflected in figures (b) and (c). This indicates that Q420RE steel has significantly better corrosion resistance than Q420 and Q355 steels in a droplet corrosion environment.

[0054] 4.10) Process the results obtained from the array droplet experiment using NaCl solutions of different concentrations using a program, as shown in Table 1; plot the results as graphs, such as... Figure 5 As shown.

[0055] Table 1. Machine-measured corrosion product parameter data Table 1-1. Total volume of corrosion products (×10) 6 μm 3 )

[0056] Table 1-2. Total area of ​​corrosion products (×10) 5 μm 2 )

[0057] Table 1-3. Average height of corrosion products (μm)

[0058] Table 1-4. Surface roughness (μm) of corrosion products

[0059] 4.11): Use 3D analysis software to analyze the height, area, and volume of the acquired images.

[0060] 4.12) The results obtained in 4.10) and 4.11) are analyzed to obtain the average relative error between the program measurement and the manual measurement results, as shown in Table 3.

[0061] As shown in Table 3, the average relative error between machine measurement and manual measurement is within 15%, proving that the method of this invention has high accuracy and reliability. Since the roughness calculation method in this program differs from the manual measurement method and is not comparable, it will not be analyzed here.

[0062] Table 2. Manually measured corrosion product parameter data (manual processing refers to processing the raw data images using the analysis and processing software attached to the laser confocal microscope to obtain relevant corrosion product data, such as its volume, cross-sectional area, height, etc.) Table 2-1. Total volume of corrosion products (×10) 6 μm 3 )

[0063] Table 2-2. Total area of ​​corrosion products (×10) 5 μm 2 )

[0064] Table 2-3. Average height of corrosion products (μm)

[0065] Table 3. Comparison of relative errors between machine measurement and manual measurement Table 3-1. Total volume of corrosion products (%)

[0066] Table 3-2. Total area of ​​corrosion products (%)

[0067] Table 3-3. Average height of corrosion products (%)

Claims

1. A rapid evaluation method for the corrosion resistance of metals based on deep learning, characterized in that, Includes the following steps: To obtain raw data of array droplet experimental metal samples acquired by laser confocal microscopy; Region cropping is performed on the original data image based on the height and area scales to obtain a three-dimensional height image containing the height and area scales; The 3D height image is input into the pre-trained UNet semantic segmentation network to perform pixel-level segmentation of the metal matrix region and corrosion product region in the image, and obtain the corresponding semantic segmentation mask. Pixels or regions with confidence scores greater than a first threshold in the segmentation results are filtered out. The high-confidence masks obtained from the filtering are added to the incremental training set along with the corresponding original images. When the preset triggering conditions are met, the UNet network is trained online using the incremental training set. The arithmetic mean roughness Ra is calculated based on the relative height difference of each pixel within the corrosion product area, thus obtaining the surface roughness parameters of the corrosion product. By statistically analyzing the height, area, volume, and surface roughness parameters of corrosion products of steel under the same corrosion time conditions, a rapid comparative evaluation of the corrosion resistance of metal samples can be achieved.

2. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 1, characterized in that, The array droplet experiment on the metal sample includes: The metal to be tested is pretreated; An array droplet experiment was conducted on a metal sample using an experimental setup.

3. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 2, characterized in that, The pretreatment includes first cutting the metal to be tested into a 10×10×3 mm metal sample using a wire cutting machine, then polishing the surface of the metal sample to 1500# using sandpaper, cleaning it with anhydrous ethanol, and drying it for later use; then mounting the metal sample using a metallographic mounting machine, followed by grinding and polishing.

4. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 2, characterized in that, The array droplet experiment involves using a pipette to uniformly drop a pre-prepared solution onto the polished sample surface, ensuring that the droplets are of uniform volume. The sample is then placed in a test chamber and subjected to a static etching process at 25 °C and 100% humidity for 2 hours. After etching, the sample is removed and placed in anhydrous ethanol for 30 seconds to stop etching and then dried for later use.

5. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 1, characterized in that, Obtaining the three-dimensional height image includes: after region cropping, performing grayscale conversion and contrast enhancement processing on the cropped region; using a pre-trained Tesseract + LSTM character recognition model to perform optical character recognition on the image, extracting the scale text information and its corresponding pixel color parameters in the image from the height and area scales; based on the scale text information and pixel color parameters, establishing a mapping relationship between image pixels and actual height values ​​to obtain the height scale; and establishing a mapping relationship between image pixels and actual area or length to obtain the area scale.

6. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 1, characterized in that, The first threshold is 0.

72.

7. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 6, characterized in that, The incremental training includes: (1) Statistically analyze the classification probability of each pixel output by the UNet network, and mark pixels with a classification probability greater than the first threshold as high confidence pixels; (2) Cluster the high-confidence pixels according to the connected components to form high-confidence corrosion product regions and high-confidence metal matrix regions; (3) The high-confidence region mask is used as a pseudo-label and added to the incremental training set along with the corresponding original image; (4) When the number of samples in the incremental training set reaches the preset number or the interval time reaches the preset duration, start online incremental training to update the UNet network parameters.

8. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 1, characterized in that, The arithmetic mean roughness Ra is obtained through the following steps: using the average height of each pixel in the metal substrate region or the height of the fitted plane as the reference height; calculating the relative height difference of each pixel in the corrosion product region and multiplying it by the corresponding pixel area; multiplying the height and pixel area to obtain the pixel volume; summing all pixel volumes to obtain the total volume of the corrosion product; and calculating the arithmetic mean roughness Ra by the relative height difference of each pixel in the corrosion product region.

9. The method for rapid evaluation of metal corrosion resistance based on deep learning according to claim 8, characterized in that, The calculation of the total volume of corrosion products and surface roughness parameters includes: (1) The average height of each pixel in the metal substrate area or the height of the fitted plane is used as the reference height; (2) For each pixel within the corrosion product region, calculate its relative height difference and multiply it by the corresponding pixel area to obtain the pixel volume; sum all pixel volumes to obtain the total volume of the corrosion product; the calculation formula includes: ； (3) Among them, A collection of pixels representing corrosion products. For pixel height, The pixel area; The surface roughness parameter of the corrosion product is obtained by calculating the arithmetic mean roughness Ra of the relative height difference of each pixel in the corrosion product area. The calculation formula includes: ； in, A collection of pixels representing corrosion products. For pixel height, For pixel area, The area represents the corrosion products.

10. A rapid evaluation device for the corrosion resistance of metals based on deep learning, characterized in that, include: The acquisition module is configured to acquire raw data of the array droplet experimental metal sample obtained by laser confocal microscopy. The processing module is configured to perform region cropping on the image of the original data based on the height and area scales to obtain a three-dimensional height image containing the height and area scales; The segmentation module is configured to input a 3D height image into a pre-trained UNet semantic segmentation network to perform pixel-level segmentation of the metal matrix region and corrosion product region in the image, thereby obtaining the corresponding semantic segmentation mask. The training module is configured to filter pixels or region masks with confidence scores greater than a first threshold in the segmentation results, add the filtered high-confidence masks along with the corresponding original images to the incremental training set, and use the incremental training set to perform online incremental training on the UNet network when a preset trigger condition is met. The calculation module is configured to calculate the arithmetic mean roughness Ra based on the relative height difference of each pixel in the corrosion product area, and obtain the surface roughness parameters of the corrosion product. The evaluation module is configured to statistically analyze the height, area, volume, and roughness of corrosion products on steel under the same corrosion time conditions, enabling rapid comparative evaluation of the corrosion resistance of metal samples.

Images