WO2022061922A1 - 材料微结构的分析方法和装置 - Google Patents

材料微结构的分析方法和装置 Download PDF

Info

Publication number
WO2022061922A1
WO2022061922A1 PCT/CN2020/118536 CN2020118536W WO2022061922A1 WO 2022061922 A1 WO2022061922 A1 WO 2022061922A1 CN 2020118536 W CN2020118536 W CN 2020118536W WO 2022061922 A1 WO2022061922 A1 WO 2022061922A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
region
grain
grain boundaries
analyzing
Prior art date
Application number
PCT/CN2020/118536
Other languages
English (en)
French (fr)
Inventor
亓欣波
李长鹏
陈国锋
Original Assignee
西门子股份公司
西门子（中国）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 西门子股份公司, 西门子（中国）有限公司 filed Critical 西门子股份公司
Priority to CN202080105092.3A priority Critical patent/CN116648723A/zh
Priority to PCT/CN2020/118536 priority patent/WO2022061922A1/zh
Publication of WO2022061922A1 publication Critical patent/WO2022061922A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection

Definitions

  • the present invention relates to the field of deep learning, and in particular, to an analysis method, device, computing device, computer-readable storage medium and program product of material microstructure.
  • Microstructure plays an important role in the influence of material properties. Therefore, materials experts and engineers always want to develop a qualified microstructure to increase the mechanical or physical properties and performance of materials.
  • microstructural information For example, grain size and orientation are microstructures that are critical to the mechanical properties of metallic and ceramic materials, with most materials having many polygrains. Therefore, how to quickly and accurately count the number and orientation of grains is the primary condition for adjusting the microstructure appropriately.
  • Another solution is to use conventional image processing to classify grain boundaries. For example, to study the classification and extraction of grain boundaries of ceramic materials, threshold segmentation is used to convert gray images to binary images, and then a series of binary image-based algorithms are used to extract grain boundaries. Among them, the algorithms include morphological operations and geometric filtering. However, the most important factor is that in the algorithm, the threshold is difficult to determine, so the accuracy of the algorithm is not high.
  • the analysis of material microstructure in the prior art either relies on manual processing and classification, or uses traditional image processing or unsupervised machine learning algorithms for classification.
  • the former requires significant time and processing efficiency is low, and the latter algorithm is not accurate. High or insufficient generalization ability.
  • the present disclosure proposes a solution for material microstructure analysis based on a supervised machine learning algorithm.
  • the first embodiment of the present disclosure proposes a method for analyzing the microstructure of a material.
  • the method includes the following steps: S1 , acquiring a grain image of the material, and performing ROI extraction on the above image to obtain a first region of the image, and obtain the gray value information of the first area; S2, mark a plurality of crystal grains and their grain boundaries in the first area respectively, and perform machine learning training according to the features extracted from the first area and the marked features; S3, Image processing is performed based on the results of the machine learning training, and skeletonization processing and region closure processing are respectively performed on the grain boundaries in the image after performing the image processing.
  • a second embodiment of the present disclosure proposes an analysis device for material microstructure, the device includes: an image acquisition unit configured to acquire a grain image of a material, and perform ROI extraction on the image to obtain a first image of the image. a region, and obtains the gray value information of the first region; the machine learning unit is configured to label a plurality of crystal grains and their grain boundaries in the first region respectively, according to the features and labels extracted from the first region The feature performs machine learning training; the first image processing unit is configured to perform image processing based on the result of the machine learning training, and respectively perform skeletonization processing and region closure processing on the grain boundaries in the image after performing the image processing.
  • a third embodiment of the present disclosure provides a computing device comprising: a processor; and a memory for storing computer-executable instructions that, when executed, cause the processor to The method described in the first embodiment is performed.
  • a fourth embodiment of the present disclosure proposes a computer-readable storage medium having computer-executable instructions stored thereon for executing the steps described in the first embodiment. method described.
  • a fifth embodiment of the present disclosure proposes a computer program product tangibly stored on a computer-readable storage medium and comprising computer-executable instructions that, when executed, cause At least one processor executes the method described in the first embodiment.
  • grain boundaries can be determined more accurately than traditional image processing methods through a supervised machine learning algorithm.
  • the machine learning algorithm has a high generalization ability and can be applied to similar or For different types of grain images, since there is no need to manually set the adjustment parameters, the algorithm has strong robustness, and the fast training speed enables the algorithm to meet the needs of practical industrial applications.
  • Figure 1 shows a flow chart of an exemplary analysis method for the microstructure of a material.
  • FIG. 2 shows an exemplary material grain image according to an embodiment of the present disclosure.
  • FIG. 3 shows an exemplary image after ROI extraction according to an embodiment of the present disclosure.
  • FIG. 4 shows an exemplary die labeling image according to an embodiment of the present disclosure.
  • FIG. 5 illustrates an exemplary machine learning processed image according to an embodiment of the present disclosure.
  • FIG. 6 shows an exemplary denoised image according to an embodiment of the present disclosure.
  • FIG. 7 shows an exemplary skeletonized image according to an embodiment of the present disclosure.
  • FIG. 8 shows an exemplary region closure image according to an embodiment of the present disclosure.
  • FIG. 9 shows an image after ellipse fitting processing according to an embodiment of the present disclosure.
  • Figure 10 shows another exemplary material grain image.
  • FIG. 11 shows an exemplary analysis device for the microstructure of a material according to an embodiment of the present disclosure.
  • FIG. 12 illustrates an exemplary computing device for analyzing material microstructure, according to embodiments of the present disclosure.
  • the terms “including”, “including” and similar terms are open-ended terms, ie, “including/including but not limited to,” meaning that other content may also be included.
  • the term “based on” is “based at least in part on.”
  • the term “one embodiment” means “at least one embodiment”; the term “another embodiment” means “at least one additional embodiment” and so on.
  • FIG. 1 shows a flowchart of an exemplary analysis method 100 of material microstructure in accordance with an embodiment of the present disclosure.
  • Method 100 may be implemented by example computing device 300 in FIG. 12 .
  • the method 100 is described below with reference to FIGS. 2-9 , wherein FIGS. 2-9 show various images obtained during the processing of the method 100 according to an embodiment of the present disclosure.
  • the method 100 begins at step 101 .
  • step 101 a grain image of the material is obtained, ROI extraction is performed on the above image, to obtain a first area of the image, and a gray value of the first area is obtained.
  • a grain image of the material may be produced by means of a scanning electron microscope (SEM) or a focused ion beam apparatus or obtained from a storage device associated therewith.
  • SEM scanning electron microscope
  • this step preprocessing of the original image is achieved.
  • this step 101 may also include ruler identification to identify the actual size represented by a unit distance in the image, which may be used to determine the size of various objects in the image (eg, grains, noise, etc.).
  • FIG. 2 shows an exemplary material grain image obtained by step 101 .
  • ROI extraction may be performed by zooming in on the image, searching for or determining a target region, and segmenting the target region from the original image to obtain an exemplary image after ROI extraction of the embodiment of the present disclosure as shown in FIG. 3 .
  • the gray value of the first region can be determined, that is, the color value of each pixel can be used as gray value information.
  • step 102 a plurality of grains and their grain boundaries in the first region are marked respectively, and machine learning training is performed according to the features extracted from the first region and the marked features.
  • the recognition accuracy and generalization ability of the algorithm are improved by annotating multiple grains and their grain boundaries.
  • Annotation can be done visually, for example, by displaying the image on an interactive graphical interface, and using specific mouse events (eg, moving and clicking on the displayed image) to achieve image annotation at the location of the annotation.
  • grain boundaries and grains can be annotated by zooming in on the image to the pixel level for more precise annotation.
  • FIG. 4 shows an exemplary grain labeling image schematically depicting grain boundaries and grain labels, such as grain boundaries labelled in a particular color (illustrated shallow on the boundaries), in accordance with embodiments of the present disclosure. gray callout lines) and grains (white callout lines inside the boundary shown).
  • the features of the first region include gray value information, grain boundary information, and grain shape information
  • the labeling features include labeling positions and labeling gray values.
  • labeling features can determine where grain boundaries or grains are labelled.
  • the profile of the grains/grain boundaries can be determined from the features of the first region.
  • step 102 may include the following steps: selecting a plurality of grains, marking the grain boundaries of the plurality of grains in the first region with a first color, and marking the plurality of grains in the first region with a second color grains to generate grain boundary labels and grain labels; perform denoising on the first region based on the Gaussian blur algorithm, then perform boundary extraction on the grains in the first region based on the Sobel filtering algorithm, and perform the boundary extraction based on the gradient structure tensor algorithm.
  • the grains in the first area perform shape judgment to complete the extraction of the features of the first area; extract the coordinate positions and gray values of the grain boundary label and the grain label to obtain the labeling feature; generate the first area of A corresponding list of features and said annotated features.
  • the grain boundaries and grains are marked with different colors (as shown in Figure 3), so that it is easy to obtain from the image where the grain boundaries or grains are marked, so as to obtain the marking features (marking position, marking grayscale value).
  • the Gaussian blur algorithm can achieve smooth filtering of the image and is suitable for removing Gaussian noise.
  • boundary extraction is performed on the crystal grains in the first region based on the Sobel filtering algorithm to obtain boundary information (edge information) about the crystal grains.
  • boundary information edge information
  • the image can be filtered based on the Sobel operator to obtain gradient information in vertical and horizontal directions as boundary information.
  • shape judgment may be performed on the grains in the first region based on a gradient structure tensor algorithm to obtain shape information about the grains, such as shape structure (for example, flat regions of the image may be distinguished by the structure tensor) , corner regions, etc. to provide shape information), the structure tensor can better reflect the local feature direction of the image than the gradient.
  • the gradient structure tensor algorithm can determine the structure tensor based on the image or determine the gradient structure tensor based on the gradient information of the image (for example, the gradient information in the vertical and horizontal directions obtained by filtering the image through the Sobel operator) to Get shape information.
  • the gradient information of the image for example, the gradient information in the vertical and horizontal directions obtained by filtering the image through the Sobel operator.
  • Gaussian Blur, Sobel Filter, and Gradient Structure Tensor Algorithms may be performed multiple times (eg, quadratic, triple, quadratic, etc.) on the image to achieve better denoising and better targeting of the first region.
  • Feature extraction may be performed multiple times (eg, quadratic, triple, quadratic, etc.) on the image to achieve better denoising and better targeting of the first region.
  • step 102 may further include the following steps: according to the information corresponding to the marked grains and grain boundaries in the corresponding list, setting a training sample set, and training the training sample set through a machine learning algorithm to generate a prediction model ; According to the information corresponding to the unmarked grains and grain boundaries in the corresponding list, use the prediction model to predict the unmarked grains and grain boundaries.
  • training can be performed, for example, by selecting at least some or all of the labeled data as a training sample set, and the labeled data can be obtained from the corresponding list of information (gray value information, Boundary information, shape information, labeling position information, labeling gray value information) to select.
  • the labeled gray value information (that is, the information indicating whether it is a grain boundary or a crystal grain) can be used as a target variable, and gray value information, boundary information, shape information, etc. can be used as feature variables to train the training sample set to generate prediction model.
  • the prediction model can be used to predict the unlabeled grains and grain boundaries, for example, by inputting the information corresponding to the unlabeled grains in the corresponding list into the prediction model, the characteristic variables and the target variables are predicted.
  • the choice can be the same as above.
  • the machine learning algorithm includes a random forest algorithm.
  • a random forest algorithm can be adopted as the above-mentioned machine learning algorithm.
  • the random forest algorithm is an ensemble learning method based on decision trees.
  • the random forest classification model consists of multiple decision trees. When the sample to be classified enters the random forest, the multiple decision trees are used for classification. The decision tree selects the category with the most times as the final classification result.
  • the random forest algorithm obtains K training sets by performing random sampling with replacement on the original training sample set, namely training set 1, training set 2, ..., training set K; randomly select each training set A number of features are generated to generate K classification models, namely classification model 1, classification model 2, ..., classification model K; the optimal classification is determined by voting on the K classification models.
  • n_estimators the number of decision tree models included in the random forest model
  • max_depth the maximum depth of the decision tree model
  • max_features The maximum number of features selected when building a decision tree
  • min_samples_leaf the minimum number of samples of leaf nodes
  • min_samples_split The minimum number of samples that the current node allows for splitting
  • hyperparameters can be set to meet the actual application requirements and accuracy requirements.
  • optimization of hyperparameters can be achieved by grid search or random search without manually setting tuning parameters.
  • step 103 the method proceeds to step 103 to perform image processing through the results of the machine learning training, and perform skeletonization processing and region closure processing on the grain boundaries in the image after performing the image processing, respectively.
  • the images can be post-processed based on the results of the machine learning training to more clearly identify the grains.
  • step 103 may include the steps of: generating an image with contour lines based on the result of machine learning training and the features of the first region; performing denoising processing on the image with contour lines based on a connected domain algorithm; The skeletonization process is performed on the grain boundaries in the denoised image to generate a skeletonized image; the region closure process is performed on the grain boundaries in the skeletonized image based on the marked watershed algorithm.
  • an image with contour lines may be generated based on prediction results obtained by performing machine learning training (eg, predicted grain boundaries or grains) and features of the first region (eg, gray value, boundary, shape information, etc.). Images with contour lines can be binarized to highlight the contours.
  • FIG. 5 illustrates an exemplary machine learning processed image showing grain/grain boundary contour lines according to embodiments of the present disclosure. It can be seen that the image in Figure 5 still includes some noise, such as small black dots located in the inner regions of the grain boundaries. Therefore, next, a denoising process can be performed on the image with contour lines based on a connected domain algorithm.
  • Performing connected region analysis based on the connected domain algorithm can find and label each connected region in the image, and can judge whether the connected region is a noise point based on the circularity of the connected domain.
  • the connected regions are removed as noise.
  • FIG. 6 shows a denoised image according to an embodiment of the present disclosure. It can be seen that after denoising, the noise in the connected region is significantly eliminated, so that the grains can be more accurately identified.
  • a skeletonization process may be further performed on the grain boundaries in the denoised image to generate a skeletonized image.
  • skeletonization can be performed by refining the outline of the grain boundary by reducing the pixel width of the outline (eg, to a single pixel), thereby culling branches or burrs that do not align with the backbone.
  • FIG. 7 shows an exemplary skeletonized image according to an embodiment of the present disclosure. It can be seen that after skeletonization, branches and burrs are significantly eliminated, allowing for more accurate grain identification. However, as shown in FIG. 7 , because the boundaries of some grains are not completely closed, it may cause errors in the identification of the grains. For example, some connected areas may be 1 die, and some connected areas may be 2 die.
  • a region closure process can be further performed on the grain boundaries in the skeletonized image based on a marked watershed algorithm.
  • the Watershed algorithm is an image segmentation algorithm based on the analysis of geographic morphology, imitating geographic structures (such as mountains, ravines, and basins) to classify different objects.
  • the common watershed algorithm is to find the dividing line based on the results of grayscale and distance transformation, which can easily lead to over-segmentation.
  • a marker-based watershed algorithm is used to avoid over-segmentation and to identify grains more accurately.
  • the process of the marked watershed algorithm can include: converting the input image to a grayscale image, generating markers (each marked point is equivalent to a water injection point in the watershed) through distance transformation, and then using the watershed algorithm to obtain the image segmentation result, And visualize the image segmentation results (for example, showing adjacent regions in different colors).
  • FIG. 8 shows an exemplary region closure image according to an embodiment of the present disclosure. It can be seen that by performing the region closing process, the separation of different grains in the same connected region can be achieved, so that the grains can be more accurately identified.
  • the method 100 may further include the step of: counting the number of grains in the image.
  • a computer can easily be used to count the number of grains in the image in FIG. 8 .
  • the method 100 may further include the step of: performing an ellipse fitting process on the grain boundaries in the image after performing the region closure process.
  • the semi-axis is b, and ellipse fitting is performed by finding the largest circumscribed ellipse of each closed region.
  • the orientation of the crystal grains represented by each region can be determined through the semimajor axis, semiaxis minor axis, or the angle between the fitted ellipse and the horizontal plane, so that the orientation of the crystal grains in the image can be counted.
  • FIG. 9 shows an image after ellipse fitting processing according to an embodiment of the present disclosure.
  • grain size and orientation are microstructures that are critical to the mechanical properties of metallic and ceramic materials, and the relevant properties of materials can be more favorably determined by statistical grain orientation.
  • the above-mentioned method 200 adopts a supervised machine learning algorithm, it has a high generalization ability.
  • the above-mentioned prediction model obtained by machine learning training based on the image in FIG. 2 can be directly applied to the other method shown in FIG. 10 .
  • grain boundaries can be determined more accurately than traditional image processing methods through a supervised machine learning algorithm, which has a high generalization ability and can be applied to similar or different types of grains Since there is no need to manually set the adjustment parameters, the algorithm has strong robustness, and the fast training speed enables the algorithm to meet the needs of practical industrial applications.
  • FIG. 11 shows a block diagram of an exemplary analysis apparatus 200 for material microstructure in accordance with embodiments of the present disclosure.
  • the apparatus 200 includes an image acquisition unit 201 , a machine learning unit 202 and a first image processing unit 203 .
  • the image acquisition unit 201 is configured to acquire a grain image of the material, perform ROI extraction on the above image to acquire a first region of the image, and acquire gray value information of the first region.
  • the machine learning unit 202 is configured to respectively label a plurality of crystal grains and their grain boundaries in the first region, and perform machine learning training according to the features extracted from the first region and the labeled features.
  • the first image processing unit 203 performs image processing based on the result of the machine learning training, and performs skeletonization processing and region closure processing respectively on the grain boundaries in the image after performing the image processing.
  • the apparatus 200 may further include a number counting unit configured to count the number of die in the image.
  • the features of the first region include gray value information, grain boundary information, and grain shape information
  • the labeling features include labeling positions and labeling gray values.
  • the machine learning unit 202 may be further configured to: select a plurality of grains, mark the grain boundaries of the plurality of grains in the first region with a first color, respectively mark all the grains in the first region with a second color to generate grain boundary labels and grain labels; perform denoising on the first region based on a Gaussian blurring algorithm, and then perform boundary extraction on the grains in the first region based on the Sobel filtering algorithm;
  • the quantitative algorithm performs shape judgment on the grains in the first area, so as to complete the feature extraction of the first area; extracts the coordinate position and gray value of the grain boundary label and the grain label to obtain the labeling feature; generates the first A corresponding list of features of a region and said labeled features.
  • the machine learning unit 202 may be further configured to: set a training sample set according to the information corresponding to the marked grains and grain boundaries in the correspondence list, and perform training on the training sample set through a machine learning algorithm to generate a prediction model; according to the information corresponding to the unlabeled grains and grain boundaries in the corresponding list, use the prediction model to predict the unlabeled grains and grain boundaries.
  • the machine learning algorithm may include a random forest algorithm.
  • the first image processing unit 103 may be further configured to: generate an image with contour lines based on the result of the machine learning training and the characteristics of the first region; and perform decompression on the image with contour lines based on a connected domain algorithm.
  • Noise processing perform skeletonization processing on the grain boundaries in the denoised image to generate a skeletonized image; perform region closure processing on the grain boundaries in the skeletonized image based on the marked watershed algorithm.
  • the apparatus 200 may further include a second image processing unit, and the second image processing unit may be configured to perform an ellipse fitting process on the grain boundaries in the image after performing the region closure process.
  • the apparatus 200 may further include an orientation statistic unit configured to count grain orientations in the image based on ellipse properties in the image after performing the ellipse fitting process.
  • Computing device 300 includes processor 301 and memory 302 coupled with processor 301 .
  • the memory 302 is used to store computer-executable instructions that, when executed, cause the processor 301 to perform the methods in the above embodiments (eg, any one or more steps of the aforementioned method 100 ).
  • the above-described method can be implemented by a computer-readable storage medium.
  • the computer-readable storage medium carries computer-readable program instructions for carrying out various embodiments of the present disclosure.
  • a computer-readable storage medium may be a tangible device that can hold and store instructions for use by the instruction execution device.
  • the computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing.
  • Non-exhaustive list of computer readable storage media include: portable computer disks, hard disks, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM) or flash memory), static random access memory (SRAM), portable compact disk read only memory (CD-ROM), digital versatile disk (DVD), memory sticks, floppy disks, mechanically coded devices, such as printers with instructions stored thereon Hole cards or raised structures in grooves, and any suitable combination of the above.
  • RAM random access memory
  • ROM read only memory
  • EPROM erasable programmable read only memory
  • flash memory static random access memory
  • SRAM static random access memory
  • CD-ROM compact disk read only memory
  • DVD digital versatile disk
  • memory sticks floppy disks
  • mechanically coded devices such as printers with instructions stored thereon Hole cards or raised structures in grooves, and any suitable combination of the above.
  • Computer-readable storage media are not to be construed as transient signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (eg, light pulses through fiber optic cables), or through electrical wires transmitted electrical signals.
  • the present disclosure presents a computer-readable storage medium having computer-executable instructions stored thereon for performing various implementations of the present disclosure method in the example.
  • the present disclosure presents a computer program product tangibly stored on a computer-readable storage medium and comprising computer-executable instructions that, when executed, cause At least one processor executes the methods in various embodiments of the present disclosure.
  • the various example embodiments of the present disclosure may be implemented in hardware or special purpose circuits, software, firmware, logic, or any combination thereof. Certain aspects may be implemented in hardware, while other aspects may be implemented in firmware or software that may be executed by a controller, microprocessor or other computing device. While aspects of the embodiments of the present disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it is to be understood that the blocks, apparatus, systems, techniques, or methods described herein may be taken as non-limiting Examples are implemented in hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controllers or other computing devices, or some combination thereof.
  • Computer-readable program instructions or computer program products for executing various embodiments of the present disclosure can also be stored in the cloud, and when invoked, the user can access the data stored in the cloud for execution through the mobile Internet, fixed network or other network.
  • the computer-readable program instructions of an embodiment of the present disclosure implement the technical solutions disclosed in accordance with various embodiments of the present disclosure.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

一种材料微结构的分析方法和装置，材料微结构的分析方法包括：获取材料的晶粒图像，对图像执行ROI提取，以获得图像的第一区域，并获取第一区域的灰度值信息（101）；分别标注第一区域中多个晶粒及其晶界，根据对第一区域提取的特征和标注特征执行机器学习训练（102）；通过机器学习训练的结果来执行图像处理，并对执行图像处理后的图像中的晶界分别执行骨架化处理和区域闭合处理（103）。比传统的图像处理方法更精确地确定晶界，且具有较高的泛化能力，可以应用于相似或不同类型的晶粒图像，无需手动设置调整参数，具有较强的健壮性，快速的训练速度可以满足实际工业应用的需求。

Description

材料微结构的分析方法和装置 技术领域
本发明涉及深度学习领域，尤其涉及材料微结构的分析方法、装置、计算设备、计算机可读存储介质和程序产品。
背景技术
在材料特性的影响中，微结构扮演了重要的角色。因此，材料专家和工程师总是想要发展一个合格的微结构，以增加材料的机械或者物理特性和表现。
为了评估材料微结构的表现，重要步骤是精确测量微结构信息。例如，晶粒尺寸和取向是对金属和陶瓷材料的机械特性至关重要的微结构，大部分材料具有许许多多晶粒。因此，如何快速精确地统计晶粒数量和取向则是相应地适当调节微结构的首要条件。
为了解决这个问题，现有技术其中一个方案是人工一个接一个地手绘晶粒边界。这种方案最精确，但是明显非常难处理和花时间。对于那种具有上百个晶粒的图像，一个经验丰富的专家会花费半小时来完成工作。因此，如果有许多图像需要人工分析和处理，人工劳动就非常巨大。
另一种方案是利用传统图像处理来对晶粒界限分类。例如，研究分类并提取陶瓷材料的晶粒界限，其利用了阈值分割来转换灰色图像为二进制图像，然后用一系列的基于二进制图像算法来提取晶粒界限。其中，算法包括形态运算和几何滤波。然而，最重要的因素是在算法中，阈值很难确定，因此算法的精确性并不高。
此外，还有一种方案是利用无监督的机器学习算法来提取叶图片的脉络，其中，机器学习算法包括k均值聚类等。这种算法具有非常好的分类结果，然而，由于算法对图像分类和品质较敏感，因此其在其他图片上的泛化能力并不好。
发明内容
现有技术中对材料微结构的分析要么依赖于人工处理分类，要么利用传统图像处理或无监督的机器学习算法来进行分类，前者需要花费显著的时间且处理效率低下，后者算法精确度不高或泛化能力不足。为了解决现有技术中的各种问题，本公开提出了基于有监督的机器学习算法的材料微结构分析的解决方案。
本公开的第一实施例提出了一种材料微结构的分析方法，该方法包括以下步骤：S1，获取材料的晶粒图像，对上述图像执行ROI提取，以获得所述图像的第一区域，并获取第一区域的灰度值信息；S2，分别标注所述第一区域中多个晶粒及其晶界，根据对所述第一区域提取的特征和标注特征执行机器学习训练；S3，通过机器学习训练的结果来执行图像处理，并对执行所述图像处理后的图像中的晶界分别执行骨架化处理和区域闭合处理。
本公开的第二实施例提出了一种材料微结构的分析装置，该装置包括：图像获取单元，被配置为获取材料的晶粒图像，对上述图像执行ROI提取，以获得所述图像的第一区域，并获取第一区域的灰度值信息；机器学习单元，被配置为分别标注所述第一区域中多个晶粒及其晶界，根据对所述第一区域提取的特征和标注特征执行机器学习训练；第一图像处理单元，被配置为通过机器学习训练的结果来执行图像处理，并对执行所述图像处理后的图像中的晶界分别执行骨架化处理和区域闭合处理。
本公开的第三实施例提供了一种计算设备，所述计算设备包括：处理器；以及存储器，其用于存储计算机可执行指令，当所述计算机可执行指令被执行时使得所述处理器执行第一实施例中所述的方法。
本公开的第四实施例提出了一种计算机可读存储介质，所述计算机可读存储介质具有存储在其上的计算机可执行指令，所述计算机可执行指令用于执行第一实施例中所述的方法。
本公开的第五实施例提出了一种计算机程序产品，所述计算机程序产品被有形地存储在计算机可读存储介质上，并且包括计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行第一实施例中所述的方法。
根据本公开的实施例的分析方法和装置，通过有监督的机器学习算法可以比传统的图像处理方法更精确地确定晶界，该机器学习算法具有较高的泛化能力，可以应用于相似或不同类型的晶粒图像，由于无需手动设置调整参数，因此该算法具有较强的健壮性，此外快速的训练速度使得该算法可以满足实际工业应用的需求。
附图说明
结合附图并参考以下详细说明，本公开的各实施例的特征、优点及其他方面将变得更加明显，在此以示例性而非限制性的方式示出了本公开的若干实施例，在附图中：
图1示出了材料微结构的示例性分析方法的流程图。
图2示出了根据本公开的实施例的示例性材料晶粒图像。
图3示出了根据本公开的实施例的提取ROI之后的示例性图像。
图4示出了根据本公开的实施例的示例性晶粒标注图像。
图5示出了根据本公开的实施例的示例性机器学习处理后的图像。
图6示出了根据本公开的实施例的示例性去噪后图像。
图7示出了根据本公开的实施例的示例性骨架化图像。
图8示出了根据本公开的实施例的示例性区域闭合图像。
图9示出了根据本公开的实施例的椭圆拟合处理后的图像。
图10示出了另一种示例性材料晶粒图像。
图11示出了根据本公开的实施例的材料微结构的示例性分析装置。
图12示出了根据本公开的实施例的用于分析材料微结构的示例性计算设备。
具体实施方式
以下参考附图详细描述本公开的各个示例性实施例。虽然以下所描述的示例性方法、装置包括在其它组件当中的硬件上执行的软件和/或固件，但是应当注意，这些示例仅仅是说明性的，而不应看作是限制性的。例如，考虑在硬件中独占地、在软件中独占地、或在硬件和软件的任何组合中可以实施任何或所有硬件、软件和固件组件。因此，虽然以下已经描述了示例性的方 法和装置，但是本领域的技术人员应容易理解，所提供的示例并不用于限制用于实现这些方法和装置的方式。
此外，附图中的流程图和框图示出了根据本公开的各个实施例的方法和系统的可能实现的体系架构、功能和操作。应当注意，方框中所标注的功能也可以按照不同于附图中所标注的顺序发生。例如，两个接连地表示的方框实际上可以基本并行地执行，或者它们有时也可以按照相反的顺序执行，这取决于所涉及的功能。同样应当注意的是，流程图和/或框图中的每个方框、以及流程图和/或框图中的方框的组合，可以使用执行规定的功能或操作的专用的基于硬件的系统来实现，或者可以使用专用硬件与计算机指令的组合来实现。
本文所使用的术语“包括”、“包含”及类似术语是开放性的术语，即“包括/包含但不限于”，表示还可以包括其它内容。术语“基于”是“至少部分地基于”。术语“一个实施例”表示“至少一个实施例”；术语“另一实施例”表示“至少一个另外的实施例”等等。
图1示出了根据本公开的实施例的材料微结构的示例性分析方法100的流程图。方法100可以由图12中的示例性计算设备300实现。下面结合图2-图9来描述方法100，其中，图2-图9示出了根据本公开的实施例的方法100的处理过程得到的各种图像。
首先，方法100开始于步骤101。在步骤101中，获取材料的晶粒图像，对上述图像执行ROI提取，以获得图像的第一区域，并获取第一区域的灰度值。例如，可以通过扫描电子显微镜(SEM)或者聚焦离子束设备之类的装置来产生材料的晶粒图像或从与其相关联的存储装置来获取材料的晶粒图像。在该步骤，实现对原始图像的预处理。此外，该步骤101还可以包括标尺识别，以识别图像中单位距离表示的实际尺寸，从而可以用于确定图像中各种物体的尺寸(例如，晶粒，噪点等)。图2示出了通过步骤101获取的示例性材料晶粒图像。例如，可以通过放大图像，搜索或确定目标区域，从原始图像中分割出该目标区域来执行ROI提取，以得到如图3所示的本公开的实施例的提取ROI之后的示例性图像。通过读取每个像素可以确定第一区域的灰度值，即每个像素的颜色值作为灰度值信息。
接着，方法100行进到步骤102。在步骤102中，分别标注第一区域中 多个晶粒及其晶界，根据对第一区域提取的特征和标注特征执行机器学习训练。不同于现有的无监督的机器学习方法，在该步骤中通过对多个晶粒及其晶界进行标注来提高算法的识别精确度和泛化能力。可以通过可视化方式来进行标注，例如，通过在可交互的图形界面上显示图像，并且使用特定的鼠标事件(例如，在显示的图像上移动和点击)来在标注位置实现图像标注。在一个示例中，可以通过放大图像至像素级水平来标注晶界和晶粒，以更精确地进行标注。图4示出了根据本公开的实施例的示例性晶粒标注图像，其示意性地描绘了晶界和晶粒的标注，例如以特定颜色标注的晶界(图示的位于边界上的浅灰色标注线)和晶粒(图示的位于边界内部的白色标注线)。
在一些实施例中，第一区域的特征包括灰度值信息、晶粒边界信息、晶粒形状信息，标注特征包括标注位置、标注灰度值。例如，通过标注特征可以确定在哪些位置标注了晶界还是晶粒。例如，通过第一区域的特征可以确定晶粒/晶界的轮廓。
在一些实施例中，步骤102可以包括如下步骤：选取多个晶粒，分别通过第一颜色标注第一区域中多个晶粒的晶界，通过第二颜色标注第一区域中所述多个晶粒，以生成晶界标签和晶粒标签；基于高斯模糊算法对第一区域执行去噪，接着分别基于Sobel滤波算法对第一区域中的晶粒执行边界提取以及基于梯度结构张量算法对第一区域中的晶粒执行形状判断，从而完成对第一区域的特征的提取；提取晶界标签和晶粒标签的坐标位置以及灰度值，以获得标注特征；生成所述第一区域的特征和所述标注特征的对应列表。
在该步骤中，通过不同颜色来标注晶界和晶粒(如图3所示)，使得可以容易从图像中获取在哪些位置标注了晶界还是晶粒，从而获得标注特征(标注位置、标注灰度值)。数字图像在数字化和传输过程中经常受到成像设备与外部环境噪声干扰等影响而包含噪声，在该步骤中通过高斯模糊算法可以实现图像的平滑滤波，并适用于去除高斯噪声。
在该步骤中，基于Sobel滤波算法对第一区域中的晶粒执行边界提取，以获取关于晶粒的边界信息(边缘信息)。例如，可以基于Sobel算子对图像进行滤波以获得垂直和水平方向上的梯度信息以作为边界信息。在该步骤中，可以基于梯度结构张量算法对第一区域中的晶粒执行形状判断，以获取关于晶粒的形状信息，例如形状结构(例如，可以通过结构张量来区分图像的平 坦区域、角点区域等以提供形状信息)，结构张量比梯度能够更好地反映图像的局部特征方向。例如，梯度结构张量算法可以基于图像来确定结构张量或基于图像的梯度信息(例如，通过Sobel算子对图像进行滤波获得的垂直和水平方向上的梯度信息)来确定梯度结构张量以获取形状信息。通过执行边界提取和形状判断，可以获得第一区域的特征。
在一些实施例中，可以对图像重复多次(例如，二次、三次、四次等)执行高斯模糊算法、Sobel滤波算法和梯度结构张量算法来实现更好的去噪和对第一区域特征的提取。
在一些实施例中，步骤102还可以包括如下步骤：根据对应列表中与已标注晶粒及晶界相对应的信息，设置训练样本集，通过机器学习算法对训练样本集进行训练以生成预测模型；根据所述对应列表中与未标注晶粒及晶界相对应的信息，使用所述预测模型对未标注晶粒及晶界进行预测。
在该步骤中，可以例如通过选择已标注数据的至少一些或全部作为训练样本集来进行训练，标注数据可以从对应列表中与已标注晶粒及晶界相对应的信息(灰度值信息、边界信息、形状信息、标注位置信息、标注灰度值信息)进行选择。例如，可以将标注灰度值信息(即表示是晶界还是晶粒的信息)作为目标变量，将灰度值信息、边界信息、形状信息等作为特征变量，来对训练样本集进行训练以生成预测模型。在生成预测模型后，可以使用预测模型对未标注晶粒及晶界进行预测，例如通过将对应列表中与未标注晶粒相对应的信息输入到预测模型来进行预测，特征变量和目标变量的选择可以同上。
在一些实施例中，机器学习算法包括随机森林算法。例如，可以采用随机森林算法作为上述的机器学习算法。随机森林算法是一种基于决策树的集成(ensemble)学习方法，随机森林分类模型由多棵决策树组成，当待分类样本进入随机森林时，由该多棵决策树进行分类，最后选取被所有决策树选择次数最多的类别作为最终的分类结果。具体地，随机森林算法通过对原始训练样本集进行有放回的随机抽样，来获得K个训练集，即训练集1、训练集2、……、训练集K；对每个训练集随机选择若干个特征，从而生成K个分类模型，即分类模型1、分类模型2、……、分类模型K；通过K个分类模型进行投票确定最优分类。
例如，可以设定关于随机森林算法的以下超参数：
n_estimators：随机森林模型中包含决策树模型的个数；
max_depth：决策树模型的最大深度；
max_features：用于构建决策树时选取的最大特征数量；
min_samples_leaf：叶子节点最少样本数；
min_samples_split：当前节点允许分裂的最小样本数；
criterion：节点分裂依据。
可以基于算法代码执行速度，设定合适的超参数以满足实际应用需求和精确度要求。或者，可以通过网格搜索或随机搜索来实现超参数的优化，而无需手动设置调整参数。
通过使用作为集成学习的随机森林算法来对未标注的晶界和晶粒进行预测，可以实现较稳定的预测结果和较强的泛化能力。
接着，方法行进到步骤103，通过机器学习训练的结果来执行图像处理，并对执行图像处理后的图像中的晶界分别执行骨架化处理和区域闭合处理。在该步骤中，可以基于机器学习训练的结果来对图像进行后处理，以便更清楚地识别晶粒。
在一些实施例中，步骤103可以包括如下步骤：基于机器学习训练的结果和所述第一区域的特征，生成具有轮廓线的图像；基于连通域算法对具有轮廓线的图像执行去噪处理；对执行去噪处理后的图像中的晶界执行骨架化处理以生成骨架化图像；基于带标记的分水岭算法对骨架化图像中的晶界执行区域闭合处理。
在该步骤中，基于执行机器学习训练获得的预测结果(例如，预测晶界或晶粒)和第一区域的特征(如灰度值、边界、形状信息等)可以生成具有轮廓线的图像。具有轮廓线的图像可以被二值化处理以凸显轮廓。图5示出了根据本公开的实施例的示例性机器学习处理后的图像，其示出了晶粒/晶界的轮廓线。可以看出，图5的图像中仍然包括一些噪点，例如位于晶粒边界内部区域的小黑点。因此，接着，可以基于连通域算法对具有轮廓线的图像执行去噪处理。基于连通域算法执行连通区域分析可以将图像中的各个连通区域找出并标记，并且可以基于连通域圆度来判断连通区域是否为噪点，例如，可以将具有小的连通域直径且接近圆形的连通区域作为噪点进行去除。 图6示出了根据本公开的实施例的去噪后图像。可以看出，经过去噪后，显著消除了连通区域内的噪点，从而可以更准确地识别晶粒。
在该步骤中，可以进一步对执行去噪处理后的图像中的晶界执行骨架化处理以生成骨架化图像。例如，可以通过减少轮廓的像素宽度(例如，减少至单像素)细化晶界的轮廓来执行骨架化处理，从而剔除与主干不一致的分支或毛刺。图7示出了根据本公开的实施例的示例性骨架化图像。可以看出，经过骨架化处理后，显著消除了分支和毛刺，从而可以更准确地识别晶粒。然而，如图7所示，由于某些晶粒的边界并未完全闭合，可能对晶粒的识别造成错误。例如，有的连通区域可能是1个晶粒，而有的连通区域可能是2个晶粒。
在该步骤中，可以进一步地基于带标记的分水岭算法对骨架化图像中的晶界执行区域闭合处理。分水岭(Watershed)算法是基于地理形态的分析的图像分割算法，模仿地理结构(比如山川、沟壑，盆地)来实现对不同物体的分类。常见的分水岭算法是基于灰度与距离变换结果寻找分割线，很容易导致过度分割。这里采用了基于标记(marker)的分水岭算法来避免过度分割，从而更准确地识别晶粒。例如，带标记的分水岭算法的过程可以包括：将输入的图像转换为灰度图像，通过距离变换生成标记(每个标记点相当于分水岭中的注水点)，然后使用分水岭算法得到图像分割结果，并对图像分割结果进行可视化(例如，以不同的颜色显示相邻的区域)。图8示出了根据本公开的实施例的示例性区域闭合图像。可以看出，通过执行区域闭合处理，可以实现同一连通区域内不同晶粒的分离，从而更准确地识别晶粒。
在一些实施例中，在步骤103之后，方法100还可以包括如下步骤：统计图像中的晶粒数量。例如，可以容易地利用计算机来统计图8中图像中的晶粒的数量。
在一些实施例中，在步骤103之后，方法100还可以包括如下步骤：对执行区域闭合处理后的图像中的晶界执行椭圆拟合处理。该步骤之后，方法100还可以包括：基于执行椭圆拟合处理后的图像中的椭圆特性，统计图像中的晶粒取向。例如，可以设定设椭圆拟合方程：f(x,y)＝(x-h)2/a2+(y-k)2/b2-1，即椭圆中心为(h,k)，长半轴为a，短半轴为b，通过寻找每个闭合区域的最大外接椭圆来进行椭圆拟合。例如，可以通过所拟合的椭圆的长半轴、 短半轴、或其与水平面的夹角来确定每个区域表示的晶粒的取向，从而统计图像中的晶粒取向。图9示出了根据本公开的实施例的椭圆拟合处理后的图像。如前所述，晶粒尺寸和取向是对金属和陶瓷材料的机械特性至关重要的微结构，通过统计晶粒取向可以更有利地确定材料的相关特性。
此外，由于上述方法200采用了有监督的机器学习算法，有着较高的泛化能力，例如上述基于图2的图像进行机器学习训练得到的预测模型，可以直接适用于如图10所示的另一种示例性材料晶粒图像，而无需重新进行样本训练。
根据本公开的实施例，通过有监督的机器学习算法可以比传统的图像处理方法更精确地确定晶界，该机器学习算法具有较高的泛化能力，可以应用于相似或不同类型的晶粒图像，由于无需手动设置调整参数，因此该算法具有较强的健壮性，此外快速的训练速度使得该算法可以满足实际工业应用的需求。
图11示出了根据本公开的实施例的材料微结构的示例性分析装置200的框图。装置200包括图像获取单元201、机器学习单元202和第一图像处理单元203。
图像获取单元201被配置为获取材料的晶粒图像，对上述图像执行ROI提取，以获得图像的第一区域，并获取第一区域的灰度值信息。
机器学习单元202被配置为分别标注第一区域中多个晶粒及其晶界，根据对第一区域提取的特征和标注特征执行机器学习训练。
第一图像处理单元203为通过机器学习训练的结果来执行图像处理，并对执行图像处理后的图像中的晶界分别执行骨架化处理和区域闭合处理。
在一些实施例中，装置200还可以包括数量统计单元，被配置为统计图像中的晶粒数量。
在一些实施例中，第一区域的特征包括灰度值信息、晶粒边界信息、晶粒形状信息，标注特征包括标注位置、标注灰度值。
在一些实施例中，机器学习单元202还可以被配置为：选取多个晶粒，分别通过第一颜色标注第一区域中多个晶粒的晶界，通过第二颜色标注第一区域中所述多个晶粒，以生成晶界标签和晶粒标签；基于高斯模糊算法对第一区域执行去噪，接着分别基于Sobel滤波算法对第一区域中的晶粒执行边 界提取以及基于梯度结构张量算法对第一区域中的晶粒执行形状判断，从而完成对第一区域的特征的提取；提取晶界标签和晶粒标签的坐标位置以及灰度值，以获得标注特征；生成所述第一区域的特征和所述标注特征的对应列表。
在一些实施例中，机器学习单元202还可以被配置为：根据所述对应列表中与已标注晶粒及晶界相对应的信息，设置训练样本集，通过机器学习算法对训练样本集进行训练以生成预测模型；根据对应列表中与未标注晶粒及晶界相对应的信息，使用预测模型对未标注晶粒及晶界进行预测。
在一些实施例中，机器学习算法可以包括随机森林算法。
在一些实施例中，第一图像处理单元103还可以被配置为：基于机器学习训练的结果和第一区域的特征，生成具有轮廓线的图像；基于连通域算法对具有轮廓线的图像执行去噪处理；对执行去噪处理后的图像中的晶界执行骨架化处理以生成骨架化图像；基于带标记的分水岭算法对骨架化图像中的晶界执行区域闭合处理。
在一些实施例中，装置200还可以包括第二图像处理单元，第二图像处理单元可以被配置为对执行区域闭合处理后的图像中的晶界执行椭圆拟合处理。
在一些实施例中，装置200还可以包括取向统计单元，取向统计单元被配置为基于执行椭圆拟合处理后的图像中的椭圆特性，统计图像中的晶粒取向。
图12示出了根据本公开的实施例的用于分析材料微结构的示例性计算设备300的框图。计算设备300包括处理器301和与处理器301耦合的存储器302。存储器302用于存储计算机可执行指令，当计算机可执行指令被执行时使得处理器301执行以上实施例中的方法(例如，前述的方法100中的任何一个或多个步骤)。
此外，替代地，上述方法能够通过计算机可读存储介质来实现。计算机可读存储介质上载有用于执行本公开的各个实施例的计算机可读程序指令。计算机可读存储介质可以是可以保持和存储由指令执行设备使用的指令的有形设备。计算机可读存储介质例如可以是但不限于电存储设备、磁存储设备、光存储设备、电磁存储设备、半导体存储设备或者上述的任意合适的组 合。计算机可读存储介质的更具体的例子(非穷举的列表)包括：便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦式可编程只读存储器(EPROM或闪存)、静态随机存取存储器(SRAM)、便携式压缩盘只读存储器(CD-ROM)、数字多功能盘(DVD)、记忆棒、软盘、机械编码设备、例如其上存储有指令的打孔卡或凹槽内凸起结构、以及上述的任意合适的组合。这里所使用的计算机可读存储介质不被解释为瞬时信号本身，诸如无线电波或者其他自由传播的电磁波、通过波导或其他传输媒介传播的电磁波(例如，通过光纤电缆的光脉冲)、或者通过电线传输的电信号。
因此，在另一个实施例中，本公开提出了一种计算机可读存储介质，该计算机可读存储介质具有存储在其上的计算机可执行指令，计算机可执行指令用于执行本公开的各个实施例中的方法。
在另一个实施例中，本公开提出了一种计算机程序产品，该计算机程序产品被有形地存储在计算机可读存储介质上，并且包括计算机可执行指令，该计算机可执行指令在被执行时使至少一个处理器执行本公开的各个实施例中的方法。
一般而言，本公开的各个示例实施例可以在硬件或专用电路、软件、固件、逻辑，或其任何组合中实施。某些方面可以在硬件中实施，而其他方面可以在可以由控制器、微处理器或其他计算设备执行的固件或软件中实施。当本公开的实施例的各方面被图示或描述为框图、流程图或使用某些其他图形表示时，将理解此处描述的方框、装置、系统、技术或方法可以作为非限制性的示例在硬件、软件、固件、专用电路或逻辑、通用硬件或控制器或其他计算设备，或其某些组合中实施。
用于执行本公开的各个实施例的计算机可读程序指令或者计算机程序产品也能够存储在云端，在需要调用时，用户能够通过移动互联网、固网或者其他网络访问存储在云端上的用于执行本公开的一个实施例的计算机可读程序指令，从而实施依据本公开的各个实施例所公开的技术方案。
虽然已经参考若干具体实施例描述了本公开的实施例，但是应当理解，本公开的实施例并不限于所公开的具体实施例。本公开的实施例旨在涵盖在所附权利要求的精神和范围内所包括的各种修改和等同布置。权利要求的范 围符合最宽泛的解释，从而包含所有这样的修改及等同结构和功能。

Claims (21)

  1. 材料微结构的分析方法，其中，包括如下步骤：
    S1，获取材料的晶粒图像，对上述图像执行ROI提取，以获得所述图像的第一区域，并获取第一区域的灰度值信息；
    S2，分别标注所述第一区域中多个晶粒及其晶界，根据对所述第一区域提取的特征和标注特征执行机器学习训练；
    S3，通过机器学习训练的结果来执行图像处理，并对执行所述图像处理后的图像中的晶界分别执行骨架化处理和区域闭合处理。
  2. 根据权利要求1所述的材料微结构的分析方法，其特征在于，所述步骤S3之后还包括如下步骤：统计所述图像中的晶粒数量。
  3. 根据权利要求1所述的材料微结构的分析方法，其特征在于，所述第一区域的特征包括灰度值信息、晶粒边界信息、晶粒形状信息，所述标注特征包括标注位置、标注灰度值。
  4. 根据权利要求3所述的材料微结构的分析方法，其特征在于，所述步骤S2还包括如下步骤：
    S21，选取多个晶粒，分别通过第一颜色标注所述第一区域中所述多个晶粒的晶界，通过第二颜色标注所述第一区域中所述多个晶粒，以生成晶界标签和晶粒标签；
    S22，基于高斯模糊算法对所述第一区域执行去噪，接着分别基于Sobel滤波算法对所述第一区域中的晶粒执行边界提取以及基于梯度结构张量算法对所述第一区域中的晶粒执行形状判断，从而完成对第一区域的特征的提取；
    S23，提取晶界标签和晶粒标签的坐标位置以及灰度值，以获得标注特征；
    S24，生成所述第一区域的特征和所述标注特征的对应列表；
  5. 根据权利要求4所述的材料微结构的分析方法，其特征在于，所述步骤S2还包括如下步骤：
    S25.根据所述对应列表中与已标注晶粒及晶界相对应的信息，设置训练样本集，通过机器学习算法对训练样本集进行训练以生成预测模型；
    S26.根据所述对应列表中与未标注晶粒及晶界相对应的信息，使用所述预测模型对未标注晶粒及晶界进行预测。
  6. 根据权利要求4所述的材料微结构的分析方法，其特征在于，所述机器学习算法包括随机森林算法。
  7. 根据权利要求1所述的材料微结构的分析方法，其特征在于，所述步骤S3还包括如下步骤：
    S31，基于所述机器学习训练的结果和所述第一区域的特征，生成具有轮廓线的图像；
    S32，基于连通域算法对具有轮廓线的图像执行去噪处理；
    S33，对执行去噪处理后的图像中的晶界执行骨架化处理以生成骨架化图像；
    S34，基于带标记的分水岭算法对骨架化图像中的晶界执行区域闭合处理。
  8. 根据权利要求1所述的材料微结构的分析方法，其特征在于，所述步骤S3之后还包括如下步骤：
    S4.对执行区域闭合处理后的图像中的晶界执行椭圆拟合处理。
  9. 根据权利要求8所述的材料微结构的分析方法，其特征在于，所述步骤S4之后还包括如下步骤：基于执行椭圆拟合处理后的图像中的椭圆特性，统计图像中的晶粒取向。
  10. 材料微结构的分析装置，其中，所述分析装置包括：
    图像获取单元，被配置为获取材料的晶粒图像，对上述图像执行ROI提取，以获得所述图像的第一区域，并获取第一区域的灰度值信息；
    机器学习单元，被配置为分别标注所述第一区域中多个晶粒及其晶界，根据对所述第一区域提取的特征和标注特征执行机器学习训练；
    第一图像处理单元，被配置为通过机器学习训练的结果来执行图像处理，并对执行所述图像处理后的图像中的晶界分别执行骨架化处理和区域闭合处理。
  11. 根据权利要求10所述的材料微结构的分析装置，其特征在于，所述分析装置还包括数量统计单元，被配置为统计所述图像中的晶粒数量。
  12. 根据权利要求10所述的材料微结构的分析装置，其特征在于，所 述第一区域的特征包括灰度值信息、晶粒边界信息、晶粒形状信息，所述标注特征包括标注位置、标注灰度值。
  13. 根据权利要求12所述的材料微结构的分析装置，其特征在于，所述机器学习单元还被配置为：
    选取多个晶粒，分别通过第一颜色标注所述第一区域中所述多个晶粒的晶界，通过第二颜色标注所述第一区域中所述多个晶粒，以生成晶界标签和晶粒标签；
    基于高斯模糊算法对所述第一区域执行去噪，接着分别基于Sobel滤波算法对所述第一区域中的晶粒执行边界提取以及基于梯度结构张量算法对所述第一区域中的晶粒执行形状判断，从而完成对第一区域的特征的提取；
    提取晶界标签和晶粒标签的坐标位置以及灰度值，以获得标注特征；
    生成所述第一区域的特征和所述标注特征的对应列表。
  14. 根据权利要求13所述的材料微结构的分析装置，其特征在于，所述机器学习单元还被配置为：
    根据所述对应列表中与已标注晶粒及晶界相对应的信息，设置训练样本集，通过机器学习算法对训练样本集进行训练以生成预测模型；
    根据所述对应列表中与未标注晶粒及晶界相对应的信息，使用所述预测模型对未标注晶粒及晶界进行预测。
  15. 根据权利要求14所述的材料微结构的分析装置，其特征在于，所述机器学习算法包括随机森林算法。
  16. 根据权利要求10所述的材料微结构的分析装置，其特征在于，所述第一图像处理单元还被配置为：
    基于所述机器学习训练的结果和所述第一区域的特征，生成具有轮廓线的图像；
    基于连通域算法对具有轮廓线的图像执行去噪处理；
    对执行去噪处理后的图像中的晶界执行骨架化处理以生成骨架化图像；
    基于带标记的分水岭算法对骨架化图像中的晶界执行区域闭合处理。
  17. 根据权利要求10所述的材料微结构的分析装置，其特征在于，所述分析装置还包括第二图像处理单元，被配置为对执行区域闭合处理后的图像中的晶界执行椭圆拟合处理。
  18. 根据权利要求17所述的材料微结构的分析装置，其特征在于，所述分析装置还包括取向统计单元，被配置为基于执行椭圆拟合处理后的图像中的椭圆特性，统计图像中的晶粒取向。
  19. 计算设备，所述计算机备包括：
    处理器；以及
    存储器，其用于存储计算机可执行指令，当所述计算机可执行指令被执行时使得所述处理器执行根据权利要求1-9中任一项所述的方法。
  20. 计算机可读存储介质，所述计算机可读存储介质具有存储在其上的计算机可执行指令，所述计算机可执行指令用于执行根据权利要求1-9中任一项所述的方法。
  21. 计算机程序产品，所述计算机程序产品被有形地存储在计算机可读存储介质上，并且包括计算机可执行指令，所述计算机可执行指令在被执行时使至少一个处理器执行根据权利要求1-9中任一项所述的方法。
PCT/CN2020/118536 2020-09-28 2020-09-28 材料微结构的分析方法和装置 WO2022061922A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202080105092.3A CN116648723A (zh) 2020-09-28 2020-09-28 材料微结构的分析方法和装置
PCT/CN2020/118536 WO2022061922A1 (zh) 2020-09-28 2020-09-28 材料微结构的分析方法和装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2020/118536 WO2022061922A1 (zh) 2020-09-28 2020-09-28 材料微结构的分析方法和装置

Publications (1)

Publication Number Publication Date
WO2022061922A1 true WO2022061922A1 (zh) 2022-03-31

Family

ID=80846150

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/118536 WO2022061922A1 (zh) 2020-09-28 2020-09-28 材料微结构的分析方法和装置

Country Status (2)

Country Link
CN (1) CN116648723A (zh)
WO (1) WO2022061922A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117314906A (zh) * 2023-11-28 2023-12-29 无锡九霄科技有限公司 基于阵列分布信息引导的晶圆晶粒检测方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117431394A (zh) * 2023-12-14 2024-01-23 深圳市中航盛世模切机械有限公司 一种调节热处理变形的万向装置及方法

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101949810A (zh) * 2010-08-12 2011-01-19 中国石油天然气集团公司 一种针状铁素体型管线钢组织鉴别和评定方法
CN106023134A (zh) * 2016-04-29 2016-10-12 江苏大学 一种钢材晶粒的自动晶界提取方法
CN106599793A (zh) * 2016-11-21 2017-04-26 江苏大学 一种基于标记分水岭分割的钢材晶粒晶界自动提取方法
CN107274413A (zh) * 2017-05-27 2017-10-20 江苏大学 一种钢材金相组织晶界自动提取方法
CN107358005A (zh) * 2017-07-24 2017-11-17 哈尔滨工业大学 基于真实微观组织结构sem‑ebsd图像的有限元模型建模方法
US20190050628A1 (en) * 2017-08-08 2019-02-14 Uchicago Argonne, Llc Machine learning technique to identify grains in polycrystalline materials samples
CN110414492A (zh) * 2019-08-29 2019-11-05 广东工业大学 一种晶体材料图像识别方法及装置
CN111325706A (zh) * 2018-12-13 2020-06-23 阿里巴巴集团控股有限公司 一种晶界提取、晶界提取模型生成方法及装置

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101949810A (zh) * 2010-08-12 2011-01-19 中国石油天然气集团公司 一种针状铁素体型管线钢组织鉴别和评定方法
CN106023134A (zh) * 2016-04-29 2016-10-12 江苏大学 一种钢材晶粒的自动晶界提取方法
CN106599793A (zh) * 2016-11-21 2017-04-26 江苏大学 一种基于标记分水岭分割的钢材晶粒晶界自动提取方法
CN107274413A (zh) * 2017-05-27 2017-10-20 江苏大学 一种钢材金相组织晶界自动提取方法
CN107358005A (zh) * 2017-07-24 2017-11-17 哈尔滨工业大学 基于真实微观组织结构sem‑ebsd图像的有限元模型建模方法
US20190050628A1 (en) * 2017-08-08 2019-02-14 Uchicago Argonne, Llc Machine learning technique to identify grains in polycrystalline materials samples
CN111325706A (zh) * 2018-12-13 2020-06-23 阿里巴巴集团控股有限公司 一种晶界提取、晶界提取模型生成方法及装置
CN110414492A (zh) * 2019-08-29 2019-11-05 广东工业大学 一种晶体材料图像识别方法及装置

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117314906A (zh) * 2023-11-28 2023-12-29 无锡九霄科技有限公司 基于阵列分布信息引导的晶圆晶粒检测方法
CN117314906B (zh) * 2023-11-28 2024-03-15 无锡九霄科技有限公司 基于阵列分布信息引导的晶圆晶粒检测方法

Also Published As

Publication number Publication date
CN116648723A (zh) 2023-08-25

Similar Documents

Publication Publication Date Title
CN110543837B (zh) 一种基于潜在目标点的可见光机场飞机检测方法
Khan et al. An efficient contour based fine-grained algorithm for multi category object detection
CN108562589B (zh) 一种对磁路材料表面缺陷进行检测的方法
CN109670429B (zh) 一种基于实例分割的监控视频多目标人脸检测方法及系统
CN109767422B (zh) 基于深度学习的管道检测识别方法、存储介质及机器人
CN111145209B (zh) 一种医学图像分割方法、装置、设备及存储介质
CN108038481A (zh) 一种结合最大极值稳定区域和笔画宽度变化的文本定位方法
WO2019232852A1 (zh) 手写字训练样本获取方法、装置、设备及介质
Li et al. Saliency based image segmentation
Laga et al. Image-based plant stornata phenotyping
WO2022061922A1 (zh) 材料微结构的分析方法和装置
CN109598681B (zh) 一种对称唐卡修复后图像的无参考质量评价方法
CN110751619A (zh) 一种绝缘子缺陷检测方法
Xue et al. Curved text detection in blurred/non-blurred video/scene images
CN114565675A (zh) 一种在视觉slam前端去除动态特征点的方法
CN113393454A (zh) 活检组织中病理目标实例分割方法和装置
Zhao et al. Research on detection method for the leakage of underwater pipeline by YOLOv3
CN107704864B (zh) 基于图像对象性语义检测的显著目标检测方法
CN114708645A (zh) 物件辨识装置及物件辨识方法
CN114926635B (zh) 与深度学习方法相结合的多焦图像中目标分割方法
Cheng et al. Power pole detection based on graph cut
Jin et al. A color image segmentation method based on improved K-means clustering algorithm
CN110188601B (zh) 一种基于学习的机场遥感图像检测方法
Rani et al. Object Detection in Natural Scene Images Using Thresholding Techniques
Wu et al. Fast cloud image segmentation with superpixel analysis based convolutional networks

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20954770

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 202080105092.3

Country of ref document: CN

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20954770

Country of ref document: EP

Kind code of ref document: A1