Method and system for establishing hot-rolled common plate microstructure quantitative prediction model

Abstract

The invention discloses an establishment method and system of a hot-rolled common plate microstructure quantitative prediction model and electronic equipment, and belongs to the technical field of material science and artificial intelligence crossing. The method comprises the following steps: acquiring and preprocessing a microstructure image of a hot-rolled common plate, and performing data enhancement; the method comprises the following steps: constructing a deep learning model (ECA-U-Net) which is based on a U-Net architecture and is fused with an efficient channel attention mechanism; introducing a segmentation arbitrary model (SAM) to perform preliminary segmentation on an original image, and generating high-quality supplementary training data; and jointly training the ECA-U-Net and supplementary data to finally obtain a prediction model capable of quickly and accurately quantifying microstructure information. According to the method, the technical problems that traditional metallographic analysis is difficult to quantify and low in efficiency and methods such as EBSD are high in cost and long in consumed time are solved, the microstructure quantitative analysis speed and precision can be greatly improved, and powerful support is provided for precise prediction of the mechanical property of the hot-rolled plate.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of materials science and artificial intelligence, and particularly to a method for establishing a quantitative prediction model for the microstructure of hot-rolled plain steel. Specifically, it relates to a method and system for rapidly and accurately predicting the content of components such as grain size, ferrite, and pearlite in hot-rolled plain steel using deep learning technology. Background Technology

[0002] As a fundamental industrial material, the mechanical properties of hot-rolled steel sheets, such as yield strength, tensile strength, and elongation, are closely related to their internal microstructure. The microstructure mainly includes grain size, precipitates, and the content and distribution of various phases such as ferrite and pearlite. Therefore, accurate quantitative analysis of the microstructure is the prerequisite and foundation for establishing high-precision mechanical property prediction models and guiding and optimizing production processes.

[0003] However, there are currently many challenges in the quantitative analysis of microstructures:

[0004] (1) Although conventional metallographic analysis methods can obtain microstructure images, they mainly rely on manual observation and experience, making it difficult to quickly and accurately extract quantitative information such as grain size and the content of each component. This method is highly subjective and inefficient, and cannot meet the real-time requirements of quality control in modern industrial production.

[0005] (2) Although there are advanced quantitative characterization methods such as electron backscatter diffraction (EBSD) and chemical phase analysis, these methods also have inherent limitations. For example, the area analyzed by EBSD is usually very small, and its statistical representativeness is insufficient for large-size industrial plates; at the same time, sample preparation is complicated, testing is time-consuming, and equipment costs are high, making it difficult to apply to the rapid response process of large-scale industrial production.

[0006] (3) In recent years, some research institutions and enterprises have begun to try to use computer vision and machine learning technologies to process microscopic tissue images in order to achieve automated information extraction. However, these methods are often general-purpose research, limited by the depth of interdisciplinary knowledge integration, and fail to closely integrate with the product characteristics and process flow of specific production lines. Therefore, these models have poor versatility and cannot be directly applied to specific industrial scenarios, requiring in-depth customized development based on product characteristics.

[0007] In summary, traditional microstructure analysis techniques suffer from inefficiency, high levels of human intervention, and insufficient representativeness. Therefore, innovating existing methods for quantitative extraction of microstructures and developing a technology capable of rapidly and accurately providing quantitative information on the microstructure of hot-rolled steel sheets to effectively support and improve the prediction accuracy of mechanical property forecasting models, and resolving the contradiction between the inability of traditional models to simultaneously achieve high accuracy and adherence to physical metallurgical principles, has become one of the core technological bottlenecks in the development of intelligent manufacturing in the current steel industry. Summary of the Invention

[0008] The main objective of this invention is to provide a method and system for establishing a quantitative prediction model of the microstructure of hot-rolled steel plates. This aims to solve the technical problems of low efficiency, insufficient accuracy, and inability to meet the demands of rapid industrial production in existing quantitative microstructure analysis technologies. Through this invention, the microstructure information of ferrite, pearlite, and other components in hot-rolled steel plates can be quickly and accurately identified and quantified, thereby providing efficient and reliable data support for high-precision prediction models of the mechanical properties of hot-rolled steel plates, and meeting the urgent needs of current industrial production lines for intelligent manufacturing and precise quality control.

[0009] To address the aforementioned technical problems, this invention provides a method for establishing a quantitative prediction model for the microstructure of hot-rolled plain steel plates, comprising the following steps: acquiring microstructure images of metallographic samples of hot-rolled plain steel plates; preprocessing the microstructure images to enhance the contrast between ferrite and pearlite structures and identify grain boundaries, obtaining a preprocessed image; performing data augmentation on the preprocessed image to expand the dataset size, obtaining an augmented dataset; normalizing the image sizes in the augmented dataset to obtain a training dataset; constructing a deep learning-based image segmentation model, wherein the image segmentation model uses a U-Net model as its basic architecture and integrates an efficient channel attention mechanism in the upsampling path of this architecture; training the image segmentation model using the training dataset until the model converges, thereby obtaining the quantitative prediction model for the microstructure of hot-rolled plain steel plates.

[0010] As a preferred technical solution, the step of obtaining microstructure images of hot-rolled plain steel sheet metallographic samples may include: selecting hot-rolled plain steel sheets of specific specifications, such as those with a thickness of 2.0-25.0 mm, and cutting full-thickness cross-sectional samples that can reflect the overall microstructure characteristics; then mounting and polishing the samples according to standard metallographic sample preparation procedures to prepare metallographic samples; using a specific etching solution, such as 3-4% nitric acid alcohol reagent, to etch and polish the polished surface of the metallographic sample, with the etching time adjusted within 5-50 seconds according to the alloy composition of the material to clearly display the microstructure boundaries; placing the etched sample under an optical microscope and observing it at a magnification sufficient to resolve grains and phase structures (e.g., 50× objective lens, 10× eyepiece lens, total magnification 500x), and acquiring digitized microstructure images. This step ensures the high quality and representativeness of the initial image data.

[0011] As a preferred technical solution, the preprocessing step for the microstructure image may include: first, binarizing the image using the Otsu's method (OTSU), which automatically finds the optimal threshold and effectively enhances the grayscale contrast between ferrite (bright areas) and pearlite (dark areas); then, using the Canny algorithm to perform edge detection on the OTSU-processed image to accurately extract grain boundaries; finally, purifying the image using a connected component area thresholding algorithm, which can identify and remove tiny black spots (noise) remaining on the ferrite matrix surface due to improper polishing and etching. The area threshold can be adaptively set based on the average area of ​​the microstructure or finely adjusted manually. This series of preprocessing operations lays a solid foundation for the accurate identification of subsequent models.

[0012] As a preferred technical solution, the data augmentation process may include using automated data augmentation toolkits such as Augmentor to perform various geometric transformations on the preprocessed image, including but not limited to rotation, flipping, segmentation, and stitching. In this way, the limited original image dataset can be expanded to thousands (e.g., ≥4000 images) without introducing additional experimental costs. This greatly enriches the diversity of the samples, effectively avoids overfitting that may occur during the training of deep learning models, and improves the model's generalization ability.

[0013] As a preferred technical solution, the steps for constructing a deep learning-based image segmentation model can be specifically designed as follows: The U-Net model, which performs exceptionally well in medical image segmentation, is selected as the basic network architecture. To further improve model performance, the classic VGG16 network is chosen as the encoder (backbone network) of the U-Net model, leveraging its powerful feature extraction capabilities. Furthermore, an efficient channel attention (ECA) mechanism is innovatively introduced into the decoder (upsampling path) of the U-Net model, constructing an ECA-Net model. The ECA mechanism is a lightweight channel attention module that adaptively adjusts the weights of the convolution kernels by performing one-dimensional convolutions on the globally average pooled features, thereby allowing the model to focus on feature channels with richer information, suppressing invalid information, and significantly improving segmentation accuracy. This design optimizes the model's feature learning ability without significantly increasing computational complexity.

[0014] As a preferred technical solution, the kernel size in the efficient channel attention mechanism can be determined adaptively using a function related to the channel dimension C. Specifically, first, the logarithm of the channel dimension C to base 2 is calculated, and then the first hyperparameter b is added. Next, the sum is divided by the second hyperparameter γ. Finally, the entire calculation result is rounded to obtain the kernel size, preferably taking the nearest odd number. This method achieves dynamic matching between the kernel coverage and the channel dimension, improving the flexibility and efficiency of information exchange between channels.

[0015] As a preferred technical solution, a powerful foundational model—Segmentation Arbitrary Model (SAM)—can be introduced for auxiliary processing before constructing the image segmentation model. Specifically, the SAM model is used to perform a preliminary, high-quality segmentation of the original microscopic tissue image, generating detailed mask images. These segmentation results generated by the SAM model can be considered as improved, higher-quality labeled data. Subsequently, these mask images are processed through cropping and enhancement, and then used as a supplementary dataset, along with the aforementioned training dataset, and input into the backbone ECA-U-Net model for joint training. This "teacher-student" training strategy, utilizing the prior knowledge of the large model to guide and optimize the learning process of the target model, can significantly improve the segmentation accuracy and robustness of the final model.

[0016] As a preferred technical solution, to further improve the automation and prompting efficiency of the SAM model, a customized prompt generator class can be developed. This class integrates Canny edge detection and Otsu thresholding, enabling it to automatically generate high-quality region-aware prompt points on images. Furthermore, the generator can dynamically adjust the density of the prompt grid points based on the overall complexity of the image (e.g., edge density) by calculating the centroid of the image contour. This intelligent prompting mechanism reduces the need for manual interaction, making the application of the SAM model more efficient and intelligent.

[0017] Another aspect of the present invention provides a system for establishing a quantitative prediction model of the microstructure of hot-rolled plain steel plate, comprising: an image acquisition module for acquiring microstructure images of metallographic samples of hot-rolled plain steel plate; an image preprocessing module for preprocessing the microstructure images to obtain preprocessed images; a data augmentation module for performing data augmentation processing on the preprocessed images to obtain an augmented dataset; a data normalization module for normalizing the image sizes in the augmented dataset to obtain a training dataset; a model building module for constructing an image segmentation model based on deep learning; and a model training module for training the image segmentation model using the training dataset to obtain the quantitative prediction model of the microstructure of hot-rolled plain steel plate.

[0018] Compared with the prior art, the technical solution proposed in this invention has the following significant advantages:

[0019] (1) Rapid and accurate quantification of microstructure information: This invention proposes a complete deep learning-based quantitative analysis process for microstructure. By constructing a composite model combining ECA-U-Net and SAM models, it can automatically, quickly, and accurately segment ferrite, pearlite, and other microstructures from conventional metallographic images and calculate key quantitative information such as grain size and volume fraction of each phase. Compared with traditional manual analysis methods, efficiency is greatly improved; compared with expensive and time-consuming methods such as EBSD, cost is significantly reduced and speed is faster, meeting the real-time requirements of industrial production.

[0020] (2) Significantly improved accuracy of tissue prediction and performance forecasting: By introducing the SAM large model to assist in generating high-quality labeled data and combining it with the efficient channel attention (ECA) mechanism to optimize the U-Net model, the segmentation accuracy and robustness of this invention are greatly guaranteed. Experimental data show that the error between the model-predicted parameters such as grain size and phase content and the EBSD experimental measurements can be controlled within 5%, with an average relative error of only 2.25%. This high-precision quantitative information can be directly used as input for downstream mechanical property prediction models, thereby effectively improving the accuracy of performance prediction and providing a reliable data foundation for process optimization and quality control.

[0021] (3) It has good prospects for industrial application and scalability: The method proposed in this invention has a clear process, advanced technical route, and a high level of automation and intelligence. It is easy to deploy in industrial production lines and meets the development needs of smart manufacturing. At the same time, the method framework has good scalability. By adjusting the training dataset, it can be easily applied to the quantitative analysis of the microstructure of other grades of steel or other metal materials, and has broad application value. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments disclosed in this invention, the accompanying drawings of the embodiments will be briefly described below. These drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0023] Figure 1 This is a schematic diagram of the overall process for establishing a quantitative prediction model of the microstructure of hot-rolled steel plate in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the U-shaped network (ECA-U-Net) structure with an embedded efficient channel attention mechanism in an embodiment of the present invention.

[0025] Figure 3 Examples of microstructure images of hot-rolled ordinary steel products at different metallographic magnifications.

[0026] Figure 4 These are comparison images of microscopic tissue images before and after cleaning, including: (a) the original micrograph, (b) after OTSU processing, and (c) after cleaning using the connected area thresholding algorithm.

[0027] Figure 5 This is a schematic diagram of the dataset processing flow, showing the process from the original photo to enhancement and scaling, where: (a) the original photomicrograph, (b) the dataset enhancement and (c) the pixel scaling.

[0028] Figure 6 This is a schematic diagram of the grain boundary marking process, showing the process from the original photograph to Canny processing and artificial restoration, where: (a) original micrograph, (b) Canny processing, and (c) artificial restoration.

[0029] Figure 7 These are the results of microstructure segmentation using the SAM large model, including: (a) the original image, (b) microstructure segmentation, and (c) grain boundary segmentation.

[0030] Figure 8 This is a functional module block diagram of a system for establishing a quantitative prediction model of the microstructure of hot-rolled steel plate in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions (including preferred technical solutions) of the present invention will be further described in detail below with reference to the accompanying drawings and by way of listing some optional embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] Example 1

[0033] This embodiment provides a complete process for establishing a quantitative prediction model for the microstructure of hot-rolled steel plates based on deep learning. The entire process is as follows: Figure 1 As shown, it mainly includes three stages: dataset construction and processing, model construction and training, and model evaluation and application.

[0034] Phase 1: Dataset Construction and Processing

[0035] The goal of this phase is to prepare high-quality, large-scale datasets for subsequent deep learning model training.

[0036] Step 101: Sample Screening and Preparation

[0037] First, a representative hot-rolled plain steel plate was selected as the research object. In this embodiment, low-alloy high-strength structural steel of grade Q355B was selected. This steel grade has a typical ferrite-pearlite microstructure. Its chemical composition (mass fraction %) is:

[0038] element C Si Mn P Cu Al Ni Cr Ti Nb content 0.147 0.139 0.494 0.018 0.022 0.031 0.010 0.038 0.021 0.003

[0039] Samples were taken from the steel plates (thickness range 2.0-25.0 mm) of this batch, ensuring that the samples were full-thickness cross-sections with a length of not less than 10 mm to guarantee representativeness. A total of 83 independent samples were selected. These samples underwent standard metallographic preparation processes such as mounting, rough grinding, fine grinding, and polishing to obtain metallographic samples with a mirror-like surface and no obvious scratches.

[0040] Step 102: Image Acquisition

[0041] Chemical etching is used to reveal the microstructure of the samples. A 3-4% (v / v) nitric acid-alcohol solution is used as the etchant to etch the polished sample surface. The etching time is fine-tuned according to the alloy element content and microstructure, generally controlled between 5-50 seconds, until clear and complete grain boundaries and phase boundaries can be observed under a microscope. After etching, the sample surface is immediately rinsed with distilled water and alcohol, and then dried with clean compressed air to avoid residual liquid affecting observation.

[0042] The prepared metallographic sample was placed on the stage of an optical microscope, and the focus was adjusted. A 50x objective lens and a 10x eyepiece were selected, resulting in a total magnification of 500x. This magnification was sufficient to clearly distinguish ferrite grains and pearlite clusters. Figure 3 Images (a) and (b) show typical tissue morphology at 500x magnification. For each sample, at least three non-overlapping representative fields of view were randomly selected on its cross-section for imaging to ensure the statistical significance of the data. Ultimately, 166 raw optical micrographs were collected from 83 samples, forming the original dataset. Figure 5 As shown in (a).

[0043] Step 103: Image Preprocessing and Cleansing

[0044] Due to factors such as uneven illumination and different corrosion, the original microscopic images require a series of preprocessing steps to improve image quality.

[0045] First, each grayscale image is binarized using OTSU (Octagonal Inter-Class Variance). The OTSU algorithm automatically calculates an optimal global threshold, segmenting the image into foreground (pearlite, dark) and background (ferrite, light). Figure 4 As shown in (b), the contrast between ferrite and pearlite is significantly enhanced after OTSU treatment.

[0046] Secondly, to accurately identify grain boundaries, the Canny edge detection algorithm is used to process the binarized image. The Canny algorithm can effectively detect edges in the image, i.e., grain boundary locations, such as... Figure 6 As shown in (b).

[0047] Furthermore, during the corrosion process, small black spots sometimes remain on the surface of the ferrite matrix. These are noise and can interfere with subsequent analysis. Therefore, a connected component area thresholding algorithm is used for image cleansing. This algorithm calculates the area of ​​all black regions (connected components) in the image and removes regions with areas smaller than a preset threshold. In this embodiment, the threshold is set by default to the average area of ​​the microstructure according to the algorithm, and can be manually fine-tuned according to actual conditions. Figure 4 As shown in (c), after purification, most of the isolated black spots were effectively removed, and the image became cleaner.

[0048] Step 104: Data Augmentation

[0049] The performance of deep learning models largely depends on the size of the training data. To prevent overfitting and improve generalization ability, data augmentation is needed on the existing 166 images. This example uses a Python library called Augmentor, which can easily perform a series of transformations on the image dataset. Specific operations include random rotation, horizontal / vertical flipping, elastic deformation, and segmentation followed by stitching. Through these operations, the dataset size is expanded from 166 images to over 4000 images. Figure 5 As shown in (b).

[0050] Step 105: Data Normalization

[0051] All enhanced images were resized to a uniform 128×128 pixels, such as... Figure 5 As shown in (c). This size reduces the computational cost of model training while preserving sufficient organizational detail. Simultaneously, the pixel values ​​of the image are linearly scaled from the range of [range 1] to the range of [range 2] to suit the input requirements of the neural network. At this point, a high-quality dataset for model training and validation is ready.

[0052] Phase Two: Model Building and Training

[0053] The core of this stage is to design and train a high-precision image segmentation model.

[0054] Step 201: Model Architecture Design (ECA-U-Net)

[0055] This embodiment uses U-Net as the basic model architecture. Due to its unique encoder-decoder structure and skip connection design, U-Net is very suitable for the segmentation of metallographic structures.

[0056] To further enhance performance, two key improvements were made:

[0057] Backbone Network: The VGG16 network was chosen as the encoder part (i.e., downsampling path) of U-Net. VGG16 consists of 13 convolutional layers and 3 fully connected layers. All convolutional layers use small 3×3 filters, and the network is deepened by stacking, which has a very strong feature extraction capability.

[0058] Attention mechanism: In each stage of the decoder part of U-Net (i.e., the upsampling path), an efficient channel attention (ECA) module is embedded, constructing a mechanism such as... Figure 2The ECA-U-Net model shown is an example. The ECA module is a lightweight attention mechanism that captures cross-channel interaction information by performing global average pooling on the feature maps, followed by a one-dimensional convolution, and generates weights for each channel. The kernel size k of this one-dimensional convolution is adaptive, determined by the channel dimension C. Its calculation method is as follows:

[0059]

[0060] k is the final calculated size of the one-dimensional convolution kernel; C is the number of channels in the input feature map; b and γ are two hyperparameters used to adjust the mapping relationship between the number of channels and the convolution kernel size.

[0061] In this embodiment, the hyperparameters b=1 and γ=2 are set. Using this formula, the model can dynamically and adaptively adjust the size of the convolutional kernel based on the number of channels in the feature map, thereby efficiently learning the dependencies between channels and improving segmentation accuracy without significantly increasing computational complexity.

[0062] This adaptive computation method ensures that the range of cross-channel interactions dynamically matches the channel dimension. This is achieved by multiplying the learned channel weights s by the original feature map X ( The model can dynamically enhance feature channels that are beneficial to the segmentation task and suppress irrelevant or noisy channels, thereby improving segmentation accuracy.

[0063] In some embodiments of the present invention, after global average pooling, the feature map z∈R1×1×C generates weights through one-dimensional convolution: s = σ(Conv1D(z)); the final weighted feature map is: .

[0064] Step 202: Introduce the SAM large model to assist training

[0065] To obtain more refined label data than manual annotation, this embodiment introduces the Segment Anything Model (SAM). SAM is a powerful visual foundational model with excellent zero-shot segmentation capabilities.

[0066] The specific process is as follows: The original 166 microscopic tissue images are input into the SAM model. Through automated cue point generation (using a developed cue generator that combines Canny and Otsu algorithms to automatically generate cue points), SAM can perform high-quality initial segmentation of the images. For example... Figure 7 As shown, (a) is the original image, (b) is the segmentation result of SAM, and (c) is the extracted grain boundary.

[0067] The mask images (i.e., segmentation results) output by the SAM model undergo a series of morphological processing steps, including cropping, skeletonization, expansion, and erosion, to better conform to metallographic standards. These refined images are used as a "supplementary dataset".

[0068] Step 203: Model Training

[0069] The dataset prepared in the first stage is divided into training and validation sets at a ratio of 80% and 20%, respectively. The training set and the supplementary dataset generated in step 202 are used as input to train the constructed ECA-U-Net+SAM composite model. During training, an appropriate loss function (such as cross-entropy loss) and optimizer (such as Adam) are used, and the network weights are continuously adjusted through backpropagation to minimize the difference between the model's output segmentation results and the true labels (or the high-quality labels generated by SAM). The training process is performed on a high-performance GPU until the model's performance on the validation set no longer improves, i.e., the model converges.

[0070] Phase 3: Model Evaluation and Application

[0071] Step 301: Model Performance Evaluation

[0072] After model training, pixel accuracy (PA) was used as the primary evaluation metric. PA is defined as the proportion of correctly classified pixels out of the total number of pixels. Evaluation showed that the final model trained in this embodiment consistently achieved a PA value above 90% on the validation set, indicating very high segmentation accuracy.

[0073] Step 302: Compare and verify with EBSD experimental results

[0074] To further verify the practical application effect of the model, the quantitative analysis results of the model were rigorously compared with the results of the recognized accurate measurement method—electron backscatter diffraction (EBSD).

[0075] Seventeen test samples were selected and analyzed using the trained model and EBSD, respectively.

[0076] Quantitative analysis of grain size: As shown in Table 1, for the 17 samples, the difference between the grain size grade predicted by the model and the results measured by EBSD was less than 5%, with the difference for most samples being less than 2%, and the average relative error being only 2.25%. This indicates that the model has extremely high accuracy in grain size assessment.

[0077] Table 1 Quantitative Analysis of Grain Size

[0078] Sample number EBSD results Model prediction difference 1 11 10.91 0.79% 2 11 11.10 0.91% 3 10 10.94 9.40% 4 11 10.97 0.27% 5 10 10.95 9.56% 6 9.5 9.86 3.70% 7 11 10.89 1.00% 8 11 10.94 0.54% 9 11 10.97 0.20% 10 10 10.23 2.30% 11 10.5 10.68 1.70% 12 10 10.18 1.80% 13 11 10.97 0.27% 14 10 10.31 3.10% 15 11 10.96 0.36% 16 11 10.86 1.27% 17 11 10.88 1.09%

[0079] Quantitative analysis of ferrite grain size: As shown in Table 2, 10 samples were selected for more detailed comparison of grain size (unit: μm). The results show that the difference between the model prediction and the EBSD measurement is basically controlled within 10%, and the difference is even smaller for some samples.

[0080] Table 2 Quantitative Analysis of Ferrite Grain Size

[0081] Sample number Grain size Model prediction difference 1 4.68(5) 5.13(10) 4.30 7.97% 2 4.70(5) 5.46(10) 4.43 5.64% 3 4.53(5) 5.00(10) 4.69 3.68% 4 3.99(5) 4.47(10) 4.45 5.20% 5 6.40(5) 6.64(10) 5.98 8.28% 6 6.00(5) 6.84(10) 5.76 10.20% 7 6.20(5) 6.43(10) 6.07 3.87% 8 6.49(5) 6.73(10) 6.33 4.23% 9 4.94(5) 5.32(10) 4.99 2.72% 10 4.83(5) 5.21(10) 4.96 1.19%

[0082] Quantitative analysis of volume fraction: As shown in Table 3, the ferrite volume fraction of 14 samples was compared. The difference between the model segmentation results and the EBSD experimental results was basically controlled within 5%, demonstrating excellent statistical ability of phase content.

[0083] Table 3. Quantitative Analysis by Volume Fraction

[0084] Sample number Experimental results Segmentation results difference 1 83.5 85.9 2.40% 2 78.2 81.0 2.80% 3 68.9 76.0 7.10% 4 68.5 72.96 4.46% 5 76.4 81.63 5.23% 6 73.7 77.04 3.34% 7 76.6 78.08 1.48% 8 85.5 87.91 2.41% 9 80.3 87.07 3.81% 10 74.3 78.22 3.92% 11 81.3 82.51 1.21% 12 78.1 82.50 4.40% 13 75.5 81.71 4.14% 14 82.0 85.15 3.15%

[0085] The above comparison results fully demonstrate that the quantitative prediction model established in this invention is comparable to the EBSD method in accuracy, but requires significantly less time, reducing it from several hours to several minutes, showcasing enormous potential for industrial applications. The final model can be deployed on production lines to perform real-time analysis of online metallographic images, outputting quantitative microstructure parameters, and providing immediate and accurate data support for subsequent mechanical property prediction and process adjustments.

[0086] Example 2

[0087] This embodiment is largely the same as the previous one, except for some optional variations in the model building phase, to illustrate the flexibility and scalability of the method framework of the present invention.

[0088] Model architecture variations: In step 201, in addition to VGG16, other advanced convolutional neural networks can be selected as the backbone network of U-Net, such as the ResNet series (e.g., ResNet-34, ResNet-50) or DenseNet. These networks have deeper network structures and more optimized residual connections or dense connections, which may extract more robust features on certain specific datasets, thus bringing further performance improvements. The choice of which backbone network to use depends on the trade-off between computational resources and the requirements for model complexity.

[0089] Attention Mechanism Variations: In step 201, in addition to the ECA module, other types of attention mechanisms can be introduced, such as Squeeze-and-Excitation (SE) attention or Convolutional Block Attention Module (CBAM). The SE module is also a channel attention mechanism, while CBAM combines channel and spatial attention, allowing the model to focus not only on "what" features are important, but also on "where" features are important. These variations provide more possibilities for model optimization.

[0090] The model established in this embodiment is mainly for Q355B steel. However, the entire methodology of this invention, including data acquisition, preprocessing, data augmentation, and the composite model framework of ECA-U-Net+SAM, has strong versatility. Simply changing the training dataset—for example, acquiring microstructure images of other grades of steel (such as duplex steel, martensitic steel) or even other metallic materials (such as aluminum alloys, titanium alloys) and performing corresponding annotations or using SAM to generate labels—allows for retraining and the acquisition of a quantitative prediction model for the microstructure of new materials.

[0091] The variations of this embodiment demonstrate that the present invention provides a robust and flexible technical framework, rather than a fixed model. By replacing and adjusting specific modules, it can adapt to more diverse material analysis needs, and has broad application prospects.

[0092] Example 3

[0093] like Figure 8 As shown, this embodiment provides a system 100 for implementing the method described in Embodiment 1. The system 100 can be deployed on one or more computers, servers, or other electronic devices. In terms of hardware, it includes at least one processor and memory. In terms of software, the system 100 includes a series of functional modules, as follows:

[0094] Image acquisition module 10: This module is used to acquire microstructure images from hot-rolled steel plate samples. It can be connected to the digital imaging device of a metallurgical microscope to receive image data. The output raw image is split into two paths: one path is sent to the image preprocessing module 20 for main processing, and the other path is sent to the auxiliary segmentation module 70 for supplementary data generation.

[0095] Image preprocessing module 20: This module receives the original image from the image acquisition module 10 and integrates the OTSU algorithm, the Canny edge detection algorithm, and the connected component area thresholding algorithm to clean and enhance the original image to obtain a clear preprocessed image.

[0096] Data augmentation module 30: This module receives images from image preprocessing module 20 and augments the images using a variety of built-in image transformation algorithms (such as rotation, flipping, stitching, etc.) to construct a large-scale dataset.

[0097] Data Normalization Module 40: This module performs normalization processing on the enhanced large-scale dataset, including adjusting the size of all images to the fixed size required for model input (e.g., 128×128 pixels), and scaling the pixel values ​​of the images to generate a training dataset that can be directly used for training.

[0098] Auxiliary segmentation module 70: This module receives the raw image from the image acquisition module 10 and integrates a pre-trained SAM large model and an automated cue generator. It is capable of performing high-quality preliminary segmentation on the raw image and outputting a fine-grained mask image.

[0099] Supplementary data generation module 80: This module receives the mask image from the auxiliary segmentation module 70 and performs a series of post-processing operations such as skeletonization, morphological erosion and dilation on it to generate a supplementary dataset that is more in line with metallographic standards.

[0100] Model building module 50: This module is responsible for defining and instantiating the deep learning model. Its internal encoding implements the ECA-U-Net network structure, including a VGG16-based encoder, decoder, and an ECA module embedded in the decoder. Specifically, this module includes the adaptive convolutional kernel size calculation logic described in Example 1, calculated according to the formula k = ⌊ (log2(C) + b) / γ ⌋. This module provides the model architecture it constructs to the model training module 60.

[0101] Model Training Module 60: This module is the core computing unit of the system. It receives the model architecture from the model building module 50, the training dataset from the data normalization module 40, and the supplementary dataset from the supplementary data generation module 80. It is responsible for executing the entire model training process, including data loading, forward propagation, loss calculation, backpropagation, and weight updates, until the model converges, and finally outputs the final prediction model after training.

[0102] In actual operation, the above modules work together to complete the entire process from raw image acquisition to final model generation and application. The system design automates and automates the process, greatly reducing manual intervention and ensuring the objectivity and repeatability of the results.

[0103] It will be readily understood by those skilled in the art that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, combinations, substitutions, improvements, etc., made under the spirit and principles of the present invention are included within the protection scope of the present invention.

Claims

1. A method for establishing a quantitative prediction model for the microstructure of hot-rolled plain steel plates, characterized in that, Includes the following steps: Obtain microstructure images of hot-rolled sheet metallographic samples; The microstructure image is preprocessed to enhance the contrast between ferrite and pearlite structures and to identify grain boundaries, resulting in a preprocessed image. The preprocessed image is subjected to data augmentation processing to expand the dataset size, resulting in an augmented dataset; The image sizes in the augmented dataset are normalized to obtain the training dataset; A deep learning-based image segmentation model is constructed, which uses the U-Net model as its basic architecture and integrates an efficient channel attention mechanism in the upsampling path of the basic architecture. The image segmentation model is trained using the training dataset until the model converges, thus obtaining the quantitative prediction model for the microstructure of hot-rolled steel plate.

2. The method according to claim 1, characterized in that, The step of obtaining microstructure images of hot-rolled plain sheet metallographic samples includes: Hot-rolled plain plates with a thickness of 2.0 to 25.0 mm were selected, and full-thickness cross-sectional samples were cut. The samples were then mounted and prepared into metallographic samples. The metallographic sample was etched and polished. The etched and polished metallographic sample was placed under a microscope to acquire images of its microstructure.

3. The method according to claim 1, characterized in that, The step of preprocessing the microscopic tissue image includes: The microstructure image was processed using the Otsu's method to enhance the contrast between ferrite and pearlite. The Canny algorithm is used to extract boundaries from images processed by the Otsu's method. The image is cleaned using a connected region area thresholding algorithm to eliminate residual black spots on the ferrite matrix surface.

4. The method according to claim 1, characterized in that, The data augmentation process includes using the Augmentor tool to perform geometric transformation operations, including segmentation and stitching, on the preprocessed image.

5. The method according to claim 1, characterized in that, The step of constructing a deep learning-based image segmentation model also includes selecting the VGG16 architecture as the backbone network of the downsampling path of the U-Net model; An efficient channel attention mechanism is introduced into the upsampling path of the U-Net model to construct the ECA-Net model. The efficient channel attention mechanism improves channel attention by adaptively adjusting the convolution kernel weights.

6. The method according to claim 5, characterized in that, The efficient channel attention mechanism determines the kernel size by calculating it using a preset formula based on the channel dimension C, thereby achieving adaptive adjustment of the kernel size.

7. The method according to claim 1, characterized in that, Before constructing the image segmentation model, the following is also included: An arbitrary segmentation model is introduced to perform preliminary segmentation on the original microscopic tissue images to improve the labeled data; The mask image extracted by the arbitrary segmentation model is enhanced and used as a supplementary dataset, which is then input into the image segmentation model along with the training dataset for training.

8. A system for establishing a quantitative prediction model for the microstructure of hot-rolled plain steel plates, characterized in that, include: The image acquisition module is used to acquire microstructure images of hot-rolled plain metallographic samples; The image preprocessing module is used to preprocess the microstructure image to enhance the contrast of ferrite and pearlite structures and identify grain boundaries to obtain a preprocessed image. The data augmentation module is used to perform data augmentation processing on the preprocessed image to expand the dataset size and obtain an augmented dataset; The data normalization module is used to normalize the image sizes in the augmented dataset to obtain the training dataset; An efficient channel attention mechanism is integrated into the upsampling path of the architecture; The model training module is used to train the image segmentation model using the training dataset until the model converges, thereby obtaining the quantitative prediction model of the microstructure of the hot-rolled steel plate.

9. The system according to claim 8, characterized in that, The model building module is specifically used for: The VGG16 architecture was selected as the backbone network of the U-Net model. An efficient channel attention mechanism is introduced into the upsampling path of the U-Net model to construct the ECA-Net model. The efficient channel attention mechanism improves channel attention by adaptively adjusting the convolution kernel weights.

10. The system according to claim 8, characterized in that, Also includes: The auxiliary segmentation module is used to introduce arbitrary segmentation models to perform preliminary segmentation on the original microscopic tissue images in order to improve the labeled data; The supplementary data generation module is used to enhance the mask image extracted by the arbitrary segmentation model and use it as a supplementary dataset, which is then input into the image segmentation model for training along with the training dataset.

Images