Quantitative characterization method of material microstructure by coupling image enhancement and machine learning

Abstract

The present invention belongs to the technical field of material microstructure state test characterization, and specifically provides a material microstructure quantitative characterization method that couples image enhancement and machine learning, including the following steps: obtaining a set of observed sample microstructure images through a microscopic observation method; performing image enhancement, multiple noise reduction and binarization on the microstructure image set; using a deep convolutional neural network algorithm to extract the features of the binarized image to a high-dimensional latent space, and then using a consistent manifold algorithm to reduce the dimension and quantitatively characterize the high-dimensional features; conducting a physical interpretability analysis of the results of the dimension reduction of material microstructure features, and evaluating the effect of the dimension reduction quantitative index on the characterization of the material microstructure features. The present invention realizes a unified description and quantitative characterization of various complex microstructure morphologies, sizes and their evolutions of materials, greatly reduces the cost and error of artificial statistical methods, and has strong invention innovation and good engineering application prospects.

Description

Technical Field

[0001] The present invention belongs to the technical field of material microstructure state testing and characterization, and specifically provides a material microstructure quantitative characterization method by coupling image enhancement and machine learning. Background Art

[0002] As a carrier of the intrinsic properties and state of materials, material microstructure images contain rich microscopic scenes and direct physical information related to material deformation, failure and damage mechanisms. The complex microstructure characteristics of materials directly determine their macroscopic mechanical properties. Therefore, identifying and quantifying microstructure characteristics from microstructure images is crucial for material failure analysis and multi-scale mechanical modeling.

[0003] With the further development of modern industry, the microstructure of materials will evolve during actual service, which is mainly reflected in the time-varying characteristics of microstructure morphology and size, which will affect the mechanical properties of the alloy. For nickel-based high-temperature alloys widely used in the manufacture of high-temperature components of advanced aircraft engines and gas turbines, the alloys often have a two-phase coherent structure of strengthening phase / matrix phase. Therefore, it is necessary to develop methods for extracting and quantifying the microstructure characteristics of such two-phase alloy materials.

[0004] The traditional method of quantitatively characterizing the two-phase microstructure of alloys is to import the alloy microscopic image into image processing software such as Image J, and use the software's built-in threshold segmentation and statistical functions to perform statistics on the two-phase structural parameters. However, the results of software characterization largely rely on the operator's experience and understanding, and have strong human subjectivity and high manpower and material costs. Therefore, how to establish a scientific and efficient method for quantitative characterization of the two-phase microstructure of materials is a key issue that needs to be urgently addressed in the fields of alloy design, preparation, and operation and maintenance.

[0005] In recent years, machine learning methods under the data-driven paradigm have attracted widespread attention in the fields of microscopic image characterization, material optimization design, and damage identification. The two-point correlation algorithm is one of the commonly used algorithms for microstructure feature extraction. The correlation map can be used to directly obtain the statistical description of the characteristic size of the microstructure, or to indirectly obtain the dimensionality reduction description of the microstructure in combination with the dimensionality reduction method. In addition, deep learning algorithms represented by convolutional neural networks have become one of the core technologies in the field of image processing and computer vision. Convolutional neural networks can identify high-dimensional structural information and low-dimensional pixel information from images, and are widely used in image recognition, segmentation, and dimensionality reduction. They have broad application prospects in the field of material microstructure characterization.

[0006] Chinese patent CN111738130B discloses a quantitative characterization method for rafting of nickel-based single crystal alloys based on a two-point correlation algorithm and a principal component analysis algorithm. The alloy microstructure size parameters are extracted by spatial variation of the cross-correlation function, and the evolution of the alloy rafting morphology and size parameters is quantitatively identified using the principal component analysis algorithm, thus realizing batch characterization of the material microstructure. However, the developed algorithm fails to uniformly consider the extraction and quantitative characterization of complex microstructure features such as alloy orientation characteristics, isotropic coarsening, and anisotropic rafting. In addition, how to use advanced image algorithms to effectively enhance and process original microscopic SEM images remains a difficult problem. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a material microstructure quantitative characterization method that couples image enhancement and machine learning. Based on the developed Gaussian distribution adaptive threshold segmentation method, deep convolutional neural network algorithm (deep CNN) and uniform manifold dimensionality reduction algorithm (UMAP algorithm), the complex microstructure characteristics of the two phases of the alloy are quantitatively characterized by performing high-quality binarization processing on the microstructure image, extracting high-dimensional features of the image, and performing dimensionality reduction quantitative characterization.

[0008] The present invention is implemented by providing a material microstructure quantitative characterization method coupled with image enhancement and machine learning, comprising the following steps:

[0009] S1: preparing an observation sample of the target alloy material, and obtaining a microstructure image set of the observation sample by a microscopic observation method;

[0010] S2: Image enhancement, multiple noise reduction and binarization processing of microscopic tissue image sets based on Gaussian distribution adaptive threshold segmentation, mask processing and adaptive histogram equalization;

[0011] S3: using a deep CNN algorithm to extract the features of the microstructure image set processed in step S2 to a high-dimensional latent space, obtaining high-dimensional latent space features, and then using the UMAP algorithm to reduce the dimension and quantitatively characterize the high-dimensional latent space features, that is, to obtain a reduced-dimensional quantitative characterization of the microstructure of the target alloy material;

[0012] S4: Conduct physical interpretability analysis of dimensionality reduction quantitative characterization of the target alloy material microstructure, and evaluate the effect of dimensionality reduction quantitative indicators on the quantitative characterization of the target alloy material microstructure.

[0013] Preferably, in step S1, sample blocks are cut from target alloy materials with different service histories or pre-treated, and prepared into microscopic observation samples after mechanical sample preparation or fine slicing; images of different microstructural states of the observation samples are captured using a microscopic observation method to form a microstructural image set including various microstructural features of the target alloy material.

[0014] More preferably, in step S1, different service histories or pre-treatments refer to target alloy materials that have undergone service histories or pre-tests at different temperatures, times and stresses.

[0015] Further preferably, in step S1, the microstructure image set includes microstructure images of different microstructure morphologies and sizes of the target alloy material.

[0016] Preferably, the target alloy material is a two-phase metal alloy material, and the step S2 takes a microstructure image set as input, and specifically comprises the following steps:

[0017] S2.1: Gaussian noise reduction and Otsu method binarization: Gaussian filter is used to process the noise in the microstructure image set, and Otsu method is used to perform preliminary binarization on the denoised microstructure image set. The black pixel value of 1 in the black and white binary image represents the I phase of the target alloy material, and the white pixel value of 0 represents the II phase of the target alloy material;

[0018] S2.2: Mask enhancement processing: The microstructure image set after binarization by Otsu method is used as a mask for mask processing, and is applied to the microstructure image set after Gaussian noise reduction, so that the phase boundary area in the material microstructure is more prominent, and a mask image set is obtained;

[0019] S2.3: Adaptive histogram equalization: The mask image set is processed using an adaptive histogram equalization method so that the grayscale histogram is nonlinearly mapped from the foreground interval to the entire grayscale range, thereby enhancing the contrast between different phases in the target alloy material and obtaining a balanced mask image set;

[0020] S2.4: Histogram Figure 2 Second-order Gaussian distribution fitting: The grayscale histogram of the mask image set after equalization is a bimodal distribution. The second-order Gaussian distribution is used to fit the grayscale histogram, and the grayscale corresponding to the left peak value is mainly corresponding to phase II.

[0021] S2.5: Image adaptive threshold binarization: The grayscale corresponding to the left peak of the second-order Gaussian distribution in step S2.4 is used as the threshold, and the pixels in the image that are smaller than the threshold are set to 0 and the pixels that are larger than the threshold are set to 1 using a fixed threshold function, thereby achieving the final binarization processing of the equalized mask image set;

[0022] S2.6: Opening operation denoising and connected domain denoising: Based on the opening operation denoising algorithm, the discrete points near phase I are expanded and eroded by pixels, and the boundary of phase I after binarization is smoothed. The connected domain denoising method is used to eliminate discrete foreground points scattered in phase II and discrete pixel blocks that affect the image quality. Finally, a set of microstructure images is obtained after image enhancement, multiple denoising and binarization processing.

[0023] Further preferably, in step S2.3, the mask processing can mark most of the phase II as 0, highlighting the phase I and the boundary area between the two phases, so as to facilitate further processing of the image and observation of image details in the area of ​​interest.

[0024] Preferably, the step S3 uses the microscopic tissue image set after image enhancement, multiple noise reduction and binarization processing in step S2 as input, and specifically includes the following steps:

[0025] S3.1: Deep CNN algorithm selection and structure design: Select the deep CNN algorithm, remove the classification layer, retain the convolution layer and pooling layer, and add a linear fully connected layer to average pool and flatten the output of the deep CNN algorithm;

[0026] S3.2: Image preprocessing and cropping: Based on the microscopic tissue image set obtained in step S2 after image enhancement, multiple noise reduction and binarization processing, select an image cropping algorithm, crop the microscopic tissue image set to the input dimension of the deep CNN algorithm selected in step S3.1, and expand the cropped microscopic tissue image set to 3 channels to form the deep CNN algorithm input image set;

[0027] S3.3: Image feature extraction and mapping: The deep CNN algorithm input image set in step S3.2 is input into the deep CNN algorithm one by one, the microscopic tissue image features are extracted and mapped to the high-dimensional latent space, and the high-dimensional latent space features of the microscopic tissue image set are obtained;

[0028] S3.4: UMAP algorithm parameter setting: setting the upper limit and selected lower limit parameters of the number of local connections of the UMAP algorithm data points;

[0029] S3.5: High-dimensional latent space manifold structure learning: The high-dimensional latent space features of the microscopic tissue image set obtained in step S3.3 are input into the UMAP algorithm model to construct a weighted neighborhood graph and a connected neighborhood graph, thereby learning the manifold structure of the microscopic tissue image set in the high-dimensional latent space;

[0030] S3.6: Optimal low-dimensional manifold representation and mapping: Project the approximate manifold structure of the high-dimensional latent space to the low-dimensional space, convert the variable distance of the approximate manifold structure into the standard Euclidean distance of the low-dimensional global coordinate system, take distance invariance as the optimization goal, and use cross entropy as the loss function. By adjusting the upper limit of the number of local connections of data points in the UMAP algorithm and the lower limit of the selection, find the optimal low-dimensional manifold representation, and finally obtain the dimensionality reduction quantitative representation of the microscopic tissue image set.

[0031] Further preferably, in step S3.2, a single original image in the microscopic tissue image set is cropped into multiple sub-images, each of which contains two-phase features of several cycle widths to retain basic features of the original image.

[0032] Preferably, in step S4, the UMAP components after dimensionality reduction by the UMAP algorithm in step S3 are used as quantitative characterization parameters of the microstructure characteristics of the target alloy material, and physical interpretability analysis of the dimensionality reduction results is carried out; global UMAP topological space mapping and local UMAP topological space mapping are performed on the microstructure image set to evaluate the characterization effect of the dimensionality reduction quantitative indicators, thereby achieving a comprehensive quantitative characterization of the microstructure of the target alloy material.

[0033] Preferably, in step S1, an optical microscope, a scanning electron microscope or a transmission electron microscope is used to perform microscopic observation of the sample.

[0034] Preferably, steps S2-S4 are implemented by programming in Python language or Java language or commercial computing software Matlab.

[0035] Compared with the prior art, the advantages of the present invention are:

[0036] 1. The present invention develops an image binarization method based on Gaussian distribution adaptive threshold segmentation, which realizes high-quality and high-fidelity enhancement and binarization processing of the original microscopic image of the material;

[0037] 2. The present invention uses an efficient deep CNN algorithm to extract and map features of microstructure images, scientifically and efficiently realizing the rapid characterization of the two-phase microstructure of the alloy, greatly reducing the cost and error of manual statistical methods;

[0038] 3. The present invention uses the advanced UMAP algorithm to reduce the dimensionality of the high-dimensional image features in the latent space, successfully obtains the quantitative indicators of the alloy microstructure characteristics after dimensionality reduction, and realizes the unified description and quantitative characterization of various complex microstructure characteristics of the alloy and their evolution. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flow chart of the method provided by the present invention on two-phase materials;

[0040] Figure 2 The original SEM images of a nickel-based single crystal superalloy, (a) is a positive cubic morphology, (b) is a positive bamboo raft morphology;

[0041] Figure 3 The SEM mask images after preliminary binarization using Gaussian noise reduction and Otsu method, (a) is the positive cube morphology, (b) is the positive bamboo raft morphology;

[0042] Figure 4 It is the grayscale histogram after adaptive histogram equalization and its second-order Gaussian distribution fitting result;

[0043] Figure 5The final SEM mask image after segmentation based on Gaussian distribution adaptive threshold, (a) is the positive cube morphology, (b) is the positive bamboo raft morphology;

[0044] Figure 6 It is the operation flow of image pre-processing and deep CNN algorithm;

[0045] Figure 7 The distribution of various microstructures of nickel-based single crystal high-temperature alloys in the global UMAP topological space;

[0046] Figure 8 The specific microstructure morphology of nickel-based single crystal high-temperature alloy is distributed in the local UMAP topological space. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the specific implementation methods of the present invention are further described in detail below by taking the quantitative characterization of complex microstructure characteristics of a nickel-based single crystal high-temperature alloy as an example in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0048] The present invention provides a method for quantitative characterization of material microstructure by coupling image enhancement and machine learning, see Figure 1 , is a simplified flow chart for quantitatively characterizing complex microstructure characteristics of a nickel-based high-temperature alloy provided in an embodiment of the present invention, and the specific steps are:

[0049] S1: Prepare observation samples of target alloy materials and obtain their microstructure image sets by microscopic observation method;

[0050] S2: Based on Gaussian distribution adaptive threshold segmentation, mask processing, and adaptive histogram equalization, the original microscopic tissue image set is subjected to image enhancement, multiple noise reduction, and binarization processing;

[0051] S3: Use the deep CNN algorithm to extract the features of the binary image into a high-dimensional latent space, and then use the UMAP algorithm to reduce the dimension and quantitatively characterize the high-dimensional latent space features;

[0052] S4: Conduct physical interpretability analysis on the dimensionality reduction results of material microstructure characteristics, and evaluate the effect of dimensionality reduction quantitative indicators on the characterization of material microstructure characteristics.

[0053] In a specific embodiment, the target alloy material described in step S1 is selected as a nickel-based single crystal high-temperature alloy.

[0054] In a specific embodiment, the microscopic observation means in step S1 is a scanning electron microscope.

[0055] In a specific embodiment, the deep CNN algorithm described in step S3 is selected as the DenseNet121 network.

[0056] In a specific embodiment, the quantitative analysis described in step S4 uses the first two order UMAP components as dimension reduction characterization parameters.

[0057] In a specific embodiment, steps S2-S4 are implemented by programming in Python.

[0058] In order to demonstrate the application of the technical means of the present invention more clearly, the present invention has actually applied the material microscopic SEM image set. Now, taking a selected nickel-based single crystal high-temperature alloy as an example, the above technical means are used to achieve quantitative characterization of its complex microstructural characteristics.

[0059] S1. Material preparation and microstructure image set acquisition, including the following three steps:

[0060] (1) Based on the pretreatment test of different temperature, stress and time, nickel-based single crystal high-temperature alloy parent materials with different microstructures were obtained, and 1cm×1cm×1cm sample blocks were cut from the original samples by wire cutting method, and metallographic samples were prepared after mechanical grinding, polishing, chemical etching, cleaning and drying. The chemical etching solution ratio is: 50mL HCl + 100mL H2O + 5mL H2SO4 + 20g CuSO4;

[0061] (2) SEM grayscale images of alloys in different microstructural states were taken using a scanning electron microscope, with a fixed magnification of 10,000 times, the alloy crystallographic orientation being

[001] or

[011] , and the original image size being 3072×2304. The scanning electron microscope was a ZEISS SUPRA55 field emission scanning electron microscope;

[0062] (3) A microscopic SEM image set containing various microstructural features of the alloy is formed. The microscopic morphology includes positive cubic morphology, bamboo raft morphology, needle morphology, oblique cubic morphology, bamboo raft morphology, and microscopic size includes different degrees of evolution from cubic morphology to bamboo raft and needle morphology, totaling 40 microstructural state characteristics. Figure 2 (a) (b) show the microscopic SEM images of the standard cube morphology and bamboo raft-like morphology. It can be seen that the microstructure contains a dark reinforcement phase and a light matrix phase.

[0063] S2. Microscopic image binarization and enhancement based on Gaussian distribution adaptive threshold segmentation, taking the original microscopic tissue image as input, includes the following six steps:

[0064] S2.1: Gaussian noise reduction and Otsu binarization: Use Gaussian filter to process the noise in the microstructure image to make the image clearer. Use Otsu method to perform preliminary binarization on the denoised image. The black pixel value of 1 in the black and white binary image represents the strengthening phase of the alloy, and the white pixel value of 0 represents the matrix phase of the alloy. The preliminary binarization mask image is as follows Figure 3 As shown in (a) and (b), labels 1 to 5 show the labeling errors and noise on the boundary between the two phases of the image after the initial binarization;

[0065] S2.2: Mask enhancement processing: The image binarized by Otsu method is used as the mask for mask processing, which is applied to the image after Gaussian noise reduction to make the phase boundary area in the microstructure more prominent. Mask processing can mark most of the matrix phase as 0, highlight the strengthening phase and the boundary area between the two phases, and facilitate further processing of the image and observation of image details in the area of ​​interest;

[0066] S2.3: Adaptive histogram equalization: The mask image is processed using the adaptive histogram equalization method so that the grayscale histogram is nonlinearly mapped from the foreground interval to the entire grayscale range, thereby enhancing the contrast between different phases in the alloy. The mapped grayscale histogram is as follows: Figure 4 As shown;

[0067] S2.4: Histogram Figure 2 Second-order Gaussian distribution fitting: The grayscale histogram of the mask image after equalization can be regarded as a bimodal distribution. The second-order Gaussian distribution is used to fit the grayscale histogram, and the grayscale corresponding to the left peak value is mainly corresponding to the white matrix phase, such as Figure 4 The figure shows the fitting result of the second-order Gaussian distribution, and the left and right peaks correspond to the two-phase organization respectively;

[0068] S2.5: Image adaptive threshold binarization: The grayscale corresponding to the left peak of the second-order Gaussian distribution in step S2.4 is used as the threshold, and the pixels in the image that are less than the threshold are set to 0 and the pixels that are greater than the threshold are set to 1 using a fixed threshold function, thereby achieving the final binarization of the image;

[0069] S2.6: Opening operation denoising and connected domain denoising: Based on the opening operation denoising algorithm, the expansion and corrosion of pixels in the image are used to eliminate the discrete points near the black phase I, and the boundary of the phase I after binarization is smoothed; the connected domain denoising method is used to eliminate the discrete foreground points scattered in the phase II and the discrete pixel blocks that affect the image quality. The final processed mask image is as follows Figure 5 As shown in (a) and (b), it can be found that after image enhancement and multiple noise reduction, the labeling errors and noise points at labels 1' to 5' of the image have been effectively removed and suppressed after the final binarization. The final output of step S2 is the microscopic tissue image after image enhancement, multiple noise reduction and binarization.

[0070] S3. Using deep convolutional neural network and consistent manifold algorithm to extract features and quantitatively characterize microscopic images, the microscopic tissue image set after image enhancement processing in step S2 is used as input, including the following six steps:

[0071] S3.1: Deep CNN selection and structure design: Select the DenseNet121 network as the deep CNN of the embodiment, remove its classification layer and only keep the convolution layer and pooling layer, and add a linear fully connected layer to average pool and flatten the output of the DenseNet121 network;

[0072] S3.2: Image preprocessing and cropping: Based on the binary image of the alloy microstructure obtained in step S2, select a suitable image cropping algorithm, crop the original image (3072×2304) of each microstructure state to the DenseNet121 network input dimension (512×512), and expand the cropped image to 3 channels. The embodiment uses a random cropping method to randomly crop 25 sub-images of 512×512 size from each original image, so a total of 40×25=1000 images are included to form a CNN input image set;

[0073] S3.3: Image feature extraction and mapping: The image set in step S3.2 is input into the deep CNN one by one, and the microscopic tissue image features are extracted based on the DenseNet121 network and mapped to the latent space. The operation process of the deep CNN is as follows Figure 6 As shown;

[0074] S3.4: UMAP model parameter setting: setting the core parameters of the UMAP algorithm, that is, setting the upper limit of the number of local connections of data points and the lower limit of the selection. In the embodiment, the upper limit is set to 10 and the lower limit is set to 1;

[0075] S3.5: Learning the manifold structure in high-dimensional space: Input the high-dimensional latent space features of the microscopic tissue image set obtained in step S3.3 into the UMAP model to construct a weighted neighborhood graph and a connected neighborhood graph of the data set, thereby learning the manifold structure of the data set in high-dimensional space;

[0076] S3.6: Optimal low-dimensional manifold representation and mapping: Project the approximate manifold of the high-dimensional space to the low-dimensional space, convert the variable distance of the manifold space into the standard Euclidean distance of the low-dimensional global coordinate system, take the distance invariance as the optimization goal, take the cross entropy as the loss function, and find the optimal low-dimensional popular representation by adjusting the core parameters of the algorithm. Considering the powerful nonlinear dimensionality reduction capability of the UMAP algorithm, the embodiment selects the dimension after dimensionality reduction as 2 dimensions.

[0077] The final output of step S3 is a dimensionality reduction mapping of the data set and a quantitative characterization of the microstructure.

[0078] S4. Physical interpretation and quantitative analysis of material microstructure characteristics: The first two UMAP components after UMAP dimensionality reduction in step S3, namely UMAP-1 and UMAP-2, are used as quantitative characterization parameters of the complex microstructure characteristics of the material, and the physical interpretability analysis of the dimensionality reduction results is carried out.

[0079] A total of 1000 images of 40 kinds of alloy microstructures were characterized by global dimensionality reduction. Figure 7 The dimensionality reduction results of various microstructures of a nickel-based single crystal superalloy in the global UMAP topological space are given, where the horizontal and vertical coordinates of the schematic diagram are UMAP-1 and UMAP-2 components respectively. It can be found that similar microstructure morphologies are clustered in local areas of the UMAP space: area A is a positive needle-like morphology; area B is an oblique bamboo-raft-like morphology; the positive cubic morphology and the smaller bamboo-raft-like strengthening phase morphology are similar, and these morphologies are clustered in area C; positive bamboo-raft-like morphologies of different sizes are also clustered in the lower left part of the UMAP space.

[0080] In terms of morphological evolution, Figure 7 As shown, the positive cubic morphology is distributed in the upper right area of ​​the C region. As the UMAP-1 and UMAP-2 components decrease, the vertical matrix phase gradually disappears and presents a slight bamboo raft-like morphology. When the UMAP-1 component decreases to around -7, as the size of the bamboo raft-like morphology further increases, the UMAP-1 component begins to increase, and this process is also accompanied by a small increase in the UMAP-2 component. In addition, whether it is a positive bamboo raft-like morphology, a positive needle-like morphology, or an oblique bamboo raft-like morphology, as the UMAP-2 component decreases, the tissue size shows an increasing trend. In the microstructural state studied in the embodiment, the UMAP-1 component and the UMAP-2 component jointly control the evolution of the microstructural morphology and size, and the UMAP-2 component is closely related to the size evolution.

[0081] Further, all microscopic images of the alloy in a specific positive bamboo-raft morphology were characterized by local dimensionality reduction. Figure 8 The dimensionality reduction results of the specific microstructure morphology of a nickel-based single crystal high-temperature alloy in the local UMAP topological space are given, where the horizontal and vertical coordinates of the schematic diagram are the UMAP-1 and UMAP-2 components, respectively. Compared with the global UMAP space, the morphology distribution in the local space is more uniform. The size of the positive bamboo-raft morphology is related to both components of the local UMAP space, and is more closely related to the UMAP-1 component. With the increase of the UMAP-1 component, the slightly bamboo-raft-like vertical matrix phase completely disappears, forming a completely horizontally connected bamboo-raft-like morphology. With the decrease of the UMAP-2 component, the width of the matrix phase increases significantly, and the severe bamboo-raft-like morphology forms a cluster distributed in the lower right corner of the local space.

[0082] According to the above process, professionals in this field can judge the microstructure morphology, size and evolution of nickel-based single crystal high-temperature alloys based on the values ​​of UMAP-1 and UMAP-2 components. In summary, the technology of coupling image enhancement and machine learning has successfully achieved a comprehensive quantitative characterization of the complex microstructural characteristics of the material.

[0083] The above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. The above embodiment of the present invention can also be modified in various ways. That is, all simple, equivalent changes and modifications made according to the claims and the description of the present invention application fall within the protection scope of the claims of the present invention.

Claims

1. A quantitative characterization method for material microstructure by coupling image enhancement and machine learning, characterized in that: The following steps are involved: S1: preparing an observation sample of the target alloy material, and obtaining a microstructure image set of the observation sample by a microscopic observation method; S2: Image enhancement, multiple noise reduction and binarization processing of microscopic tissue image sets based on Gaussian distribution adaptive threshold segmentation, mask processing and adaptive histogram equalization; The target alloy material is a two-phase metal alloy material, and step S2 specifically includes the following steps: S2.1: A Gaussian filter is used to process the noise in the microstructure image set, and the Otsu method is used to perform preliminary binarization processing on the denoised microstructure image set. The black pixel value of 1 in the black-and-white binary image represents the I phase of the target alloy material, and the white pixel value of 0 represents the II phase of the target alloy material; S2.2: The microstructure image set binarized by the Otsu method is used as a mask for mask processing, and is applied to the microstructure image set after Gaussian noise reduction, so that the phase boundary area in the material microstructure is more prominent, and a mask image set is obtained; S2.3: using an adaptive histogram equalization method to process the mask image set, so that the grayscale histogram is nonlinearly mapped from the foreground interval to the entire grayscale range, thereby enhancing the contrast between different phases in the target alloy material, and obtaining a balanced mask image set; S2.4: The grayscale histogram of the mask image set after equalization is a bimodal distribution. The second-order Gaussian distribution is used to fit the grayscale histogram, and the grayscale corresponding to the left peak value is mainly corresponding to phase II. S2.5: The grayscale corresponding to the left peak of the second-order Gaussian distribution in step S2.4 is used as the threshold, and the pixels in the image that are less than the threshold are set to 0 and the pixels that are greater than the threshold are set to 1 using a fixed threshold function, thereby achieving the final binarization processing of the equalized mask image set; S2.6: Based on the opening operation denoising algorithm, the discrete points near phase I are expanded and eroded by pixels, and the boundary of phase I after binarization is smoothed. The connected domain denoising method is used to eliminate the discrete foreground points scattered in phase II and the discrete pixel blocks that affect the image quality. Finally, a microstructure image set is obtained after image enhancement, multiple denoising and binarization processing. S3: using a deep convolutional neural network algorithm to extract the features of the microstructure image set processed in step S2 to a high-dimensional latent space, obtaining high-dimensional latent space features, and then using a consistent manifold dimensionality reduction algorithm to reduce the dimension and quantitatively characterize the high-dimensional latent space features, that is, to obtain a reduced-dimensional quantitative characterization of the microstructure of the target alloy material; S3 specifically includes the following steps: S3.1: Select a deep convolutional neural network algorithm, remove the classification layer, retain the convolution layer and pooling layer, and add a linear fully connected layer to average pool and flatten the output of the deep convolutional neural network algorithm; S3.2: Based on the microscopic tissue image set obtained in step S2 after image enhancement, multiple noise reduction and binarization processing, select an image cropping algorithm, crop the microscopic tissue image set to the input dimension of the deep convolutional neural network algorithm selected in step S3.1, and expand the cropped microscopic tissue image set to 3 channels to form the deep convolutional neural network algorithm input image set; S3.3: Input the deep convolutional neural network algorithm input image set in step S3.2 into the deep convolutional neural network algorithm one by one, extract the microscopic tissue image features and map them to the high-dimensional latent space, and obtain the high-dimensional latent space features of the microscopic tissue image set; S3.4: Setting the upper limit and selected lower limit parameters for the number of local connections of data points in the consistent manifold dimensionality reduction algorithm; S3.5: inputting the high-dimensional latent space features of the microscopic tissue image set obtained in step S3.3 into the consistent manifold dimensionality reduction algorithm model, constructing a weighted neighborhood graph and a connected neighborhood graph, thereby learning the manifold structure of the microscopic tissue image set in the high-dimensional latent space; S3.6: Project the approximate manifold structure of the high-dimensional latent space to the low-dimensional space, convert the variable distance of the approximate manifold structure into the standard Euclidean distance of the low-dimensional global coordinate system, take the distance invariance as the optimization goal, and use the cross entropy as the loss function. By adjusting the upper limit and the lower limit of the number of local connections of the data points in the consistent manifold dimensionality reduction algorithm, find the optimal low-dimensional popular representation, and finally obtain the dimensionality reduction quantitative representation of the microscopic tissue image set; S4: Conduct physical interpretability analysis of dimensionality reduction quantitative characterization of the target alloy material microstructure, and evaluate the effect of dimensionality reduction quantitative indicators on the quantitative characterization of the target alloy material microstructure.

2. The material microstructure quantitative characterization method coupled with image enhancement and machine learning according to claim 1, characterized in that: In the step S1, sample blocks are cut from target alloy materials with different service histories or pre-treated, and are prepared into microscopic observation samples after mechanical sample preparation or fine slicing; images of different microstructural states of the observation samples are captured using a microscopic observation method to form a microstructural image set including various microstructural features of the target alloy materials.

3. The material microstructure quantitative characterization method coupled with image enhancement and machine learning according to claim 1, characterized in that: In the step S1, the microstructure image set includes microstructure images of different microstructure morphologies and sizes of the target alloy material.

4. The material microstructure quantitative characterization method coupled with image enhancement and machine learning according to claim 1, characterized in that: In the step S4, the consistent manifold component after dimensionality reduction by the consistent manifold dimensionality reduction algorithm in step S3 is used as a quantitative characterization parameter of the microstructure characteristics of the target alloy material, and a physical interpretability analysis of the dimensionality reduction results is carried out; global consistent manifold topological space mapping and local consistent manifold topological space mapping are performed on the microstructure image set to evaluate the characterization effect of the dimensionality reduction quantitative index, thereby achieving a comprehensive quantitative characterization of the microstructure of the target alloy material.

5. The material microstructure quantitative characterization method coupled with image enhancement and machine learning according to claim 1, characterized in that: The step S1 uses an optical microscope, a scanning electron microscope or a transmission electron microscope to perform microscopic observation of the sample.

Images