[go: up one dir, main page]

CN111915630B - A Grain Segmentation Algorithm for Ceramic Materials Driven by Joint Data and Model - Google Patents

A Grain Segmentation Algorithm for Ceramic Materials Driven by Joint Data and Model Download PDF

Info

Publication number
CN111915630B
CN111915630B CN202010836869.5A CN202010836869A CN111915630B CN 111915630 B CN111915630 B CN 111915630B CN 202010836869 A CN202010836869 A CN 202010836869A CN 111915630 B CN111915630 B CN 111915630B
Authority
CN
China
Prior art keywords
image
segmentation
result
rcf
ceramic material
Prior art date
Legal status (The legal status 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 status listed.)
Active
Application number
CN202010836869.5A
Other languages
Chinese (zh)
Other versions
CN111915630A (en
Inventor
雷涛
李云彤
加小红
周文政
袁启斌
王成兵
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shaanxi University of Science and Technology
Original Assignee
Shaanxi University of Science and Technology
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 Shaanxi University of Science and Technology filed Critical Shaanxi University of Science and Technology
Priority to CN202010836869.5A priority Critical patent/CN111915630B/en
Publication of CN111915630A publication Critical patent/CN111915630A/en
Application granted granted Critical
Publication of CN111915630B publication Critical patent/CN111915630B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/12—Edge-based segmentation
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G06T5/20—Image enhancement or restoration using local operators
    • G06T5/30—Erosion or dilatation, e.g. thinning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/13—Edge detection
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/60—Analysis of geometric attributes
    • G06T7/62—Analysis of geometric attributes of area, perimeter, diameter or volume
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10056—Microscopic image
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20036—Morphological image processing

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Geometry (AREA)
  • Image Analysis (AREA)

Abstract

本发明公开了一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，首先通过卷积神经网络输出图像轮廓；其次利用鲁棒分水岭变换实现图像中晶粒的预分割，解决了相关技术分水岭算法存在的过分割以及分割区域个数与轮廓精度难以平衡的问题；最后根据卷积神经网络输出的图像轮廓对预分割结果进行优化，本发明利用鲁棒分水岭变换联合卷积神经网络来构建分割模型，能够实现陶瓷材料晶粒尺寸的自动测量，而且具有较高的计算效率和准确性，在基于扫描电镜图像的陶瓷材料分析领域具有广泛的应用前景。

Figure 202010836869

The invention discloses a ceramic material grain segmentation algorithm based on joint drive of data and model, firstly output image outline through convolutional neural network; secondly utilize robust watershed transform to realize pre-segmentation of grain in image, solve related technology watershed The problem of over-segmentation in the algorithm and the difficulty in balancing the number of segmented regions and contour accuracy; finally, the pre-segmentation results are optimized according to the image contour output by the convolutional neural network. The model can realize the automatic measurement of the grain size of ceramic materials, and has high calculation efficiency and accuracy, and has broad application prospects in the field of ceramic material analysis based on scanning electron microscope images.

Figure 202010836869

Description

Ceramic material grain segmentation algorithm based on data and model combined driving
Technical Field
The invention belongs to the technical field of image processing, and particularly relates to a ceramic material grain segmentation algorithm based on data and model combined driving.
Background
Ceramic is a polycrystalline material and researchers typically use scanning electron microscopes (Scanning Electron Microscope, SEM) to scan the ceramic sample to estimate the physical properties of the ceramic material sample by analyzing the grain size distribution in the image. The scanning electron microscope works in the principle that electron beams are irradiated on the surface of a sample through high voltage, electrons interact with materials on the surface of the sample to generate electric signals, and imaging results are displayed after the electric signals are received and processed. SEM images of ceramic materials consist of material regions (i.e. grains) and interstices between the grains (i.e. grain boundaries), which are easily broken down by high voltage electricity during imaging, since ceramics are an insulating material, and do not have conductive properties. To avoid this, it is necessary to control the grain size as small as possible, i.e. there are more grain boundaries within the same size area, so that high voltage electricity is conducted away from the grain boundaries, protecting the ceramic sample from breakdown. However, the grain size directly determines the performance of the ceramic material, so that the distribution of the grain size in the SEM image needs to be counted, so that the distribution relationship between the experimental condition and the grain size is conveniently obtained, and the corresponding relationship between the experimental condition and the material performance is indirectly constructed. Currently, grain analysis in SEM images mainly relies on manual means, and measurement results have obvious limitations: firstly, the number of grains in one image is numerous, the manual statistics is time-consuming and labor-consuming, the measurement difficulty is high, and the efficiency is low; secondly, the grain size and shape are irregular, and the manual measurement is easily influenced by subjective factors, so that the result is inaccurate. Therefore, how to improve the working efficiency and the accuracy of the grain size measurement is still a challenging task.
In order to obtain the grain size distribution of the ceramic material, SEM images of the ceramic material are first analyzed, and the images have the following characteristics: the image has rich edge information but missing texture information, the grain size in the image is uneven and the shape is irregular, and the contrast of the image is lower. Based on the characteristics, the image segmentation technology is utilized to realize the segmentation of the crystal grains, and then the size distribution of the crystal grains is counted, so that the method is a feasible method. The commonly used image segmentation methods are classified into an unsupervised image segmentation method and a supervised image segmentation method.
The unsupervised image segmentation method includes a pixel-based image segmentation method, a contour-based image segmentation method, and a region-based image segmentation method. In the unsupervised image segmentation algorithm, image segmentation based on pixel classification and image segmentation based on region information are both dependent on the texture characteristics of the image, and the texture information in the SEM image of the ceramic material is missing, so that the two methods are difficult to realize effective image segmentation. In view of this abundance of image edge information, we will employ contour-based image segmentation strategies. Among such methods, watershed is one of the most popular algorithms. However, the traditional watershed transformation is sensitive to noise, and the segmentation result has the problem of excessive segmentation, so in order to solve the problem, students put forward a morphological reconstruction algorithm, correct the gradient image through corrosion and expansion operation, and then perform watershed transformation on the corrected gradient image, thereby achieving the purpose of overcoming the problem of excessive segmentation. The morphological reconstruction algorithm (Adaptive Morphological Reconstruction, AMR) of the related art can obtain a good reconstruction effect by performing a combined morphological open-close operation on the gradient image, while keeping the large target unsmoothly, and effectively filtering out the small target. However, the reconstruction result of the algorithm is greatly influenced by the structural element parameter r of gradient reconstruction, if the value of r is smaller, the segmentation result still has over-segmentation, and if the value of r is larger, the contour precision of the segmentation result is lower due to the smoothing of the structural element on the gradient.
With the continuous rapid development of deep learning, image semantic segmentation is increasingly focused by students. Unlike traditional image segmentation, image semantic segmentation essentially classifies pixels of an image, assigning each pixel in an input image a semantic class to yield a pixelized dense classification. Aiming at the contour prediction of the image, the related technology adopts a rich convolution feature network (Richer Convolutional Features, RCF), the loss function of the rich convolution feature network is calculated in each convolution layer, and the feature information of all layers is fused to obtain the final feature, wherein the deep feature can output more contour information, and the shallow feature can supplement details for the deep feature. However, since SEM images have very high resolution, the existing image segmentation algorithm consumes a long time for segmenting SEM images, and SEM image acquisition equipment is high, it is difficult to form a massive data set like a conventional image, and thus it is difficult to directly realize end-to-end target segmentation by using a deep convolutional neural network. Further investigation is needed on how to obtain accurate segmentation results of SEM images of ceramic materials quickly.
Disclosure of Invention
In order to solve the problems in the prior art, the invention provides a ceramic material grain segmentation algorithm based on data and model combined driving, which can realize automatic measurement of the size of ceramic material grains and has higher calculation efficiency and accuracy.
In order to achieve the above object, the technical scheme adopted by the invention comprises the following steps:
1) Inputting a scanning electron microscope image f, and outputting a corresponding gradient image g by using an RCF network rcf ；
2) Obtaining a gradient image g corresponding to the scanning electron microscope image f by utilizing a structure edge algorithm;
3) Pre-segmentation is carried out on the gradient image g by utilizing the robust watershed transformation to obtain a pre-segmentation result f 0 ；
4) Pre-segmentation result f by using contour optimization method based on morphology 0 Optimizing to obtain an optimized result f 1 ；
5) Gradient image g using watershed transformation rcf Dividing to obtain a result f rcf According to result f rcf For the optimized result f 1 Optimizing to obtain a final result f 2 ；
6) According to the final result f 2 The grain area distribution was counted and the grain size was calculated.
Further, setting an image gray level variance threshold eta in the step 2), calculating an image gray level variance var, and outputting a corresponding gradient image g by using a structural edge algorithm if the image gray level variance var is smaller than or equal to the image gray level variance threshold eta; if the image gray value variance var is larger than the image gray value variance threshold eta, preprocessing the scanning electron microscope image f to homogenize the image gray value, and then outputting a corresponding gradient image g by utilizing a structure edge algorithm.
Further, the image gray variance threshold η=1.5x10 3 。
Further, the preprocessing in the step 2) adopts a multi-scale Retinex algorithm, and the model is as follows:
Figure BDA0002640021430000031
R n (x,y)＝logI n (x,y)-log[F n (x,y)*I n (x,y)]
Figure BDA0002640021430000041
wherein R is MSR Is a reflected image output by a multi-scale Retinex algorithm; n=3, N is 1, 2, 3, respectively representing three scales of low, medium and high; w (w) n Represents the weight coefficient, w 1 ＝w 2 ＝w 3 ＝1/3；R n (x, y) represents a multi-scale reflected image, I n (x, y) represents an experimental image, wherein x represents an abscissa and y represents an ordinate; f (F) n (x, y) is a gaussian surround function, wherein the value of K meets the value of ≡ ≡F n (x,y)dxdy＝1；c n Represents a gaussian surround space constant, where c 1 ＝15，c 2 ＝80，c 3 ＝200。
Further, the pre-dividing in the step 3) includes:
3.1 Through a) a process of
Figure BDA0002640021430000042
Carrying out gradient reconstruction on the gradient image g;
wherein,,
Figure BDA0002640021430000043
representing reconstruction of a pre-segmentation result f from a gradient image g 0 Is a combined morphological closing operation, R represents morphological reconstruction, R φ Represents morphological closing operations, ε represents morphological erosion operations, and δ represents morphological dilation operations; />
Figure BDA0002640021430000044
Representing reconstruction of a pre-segmentation result f from a gradient image g 0 Morphological dilation operation of h E N + And meet the following
Figure BDA0002640021430000045
Figure BDA0002640021430000046
Representing reconstruction of a pre-segmentation result f from a gradient image g 0 Morphological erosion operations of h.epsilon.N + And satisfy->
Figure BDA0002640021430000047
m represents the scale of the largest structural element, the multiscale structural element satisfying the relationship +.>
Figure BDA0002640021430000048
3.2 If the reconstruction result contains multiple small regions, then by
Figure BDA0002640021430000049
Merging the multiple small regions of the reconstruction result in step 3.1), wherein I r Representing the combined small area image; i represents the image before merging the small areas; k is a structural element parameter; b k Representing the combined structural element dimensions.
Further, the contour optimization method based on morphology in the step 4) includes:
4.1 For the pre-segmentation result f) 0 Labeling each region in the image, ensuring that each label in the image can only cover one region;
4.2 Performing morphological open operation on each tag region from the pre-segmentation result f 0 Subtracting the result of the open operation from the image of (a):
Figure BDA00026400214300000410
wherein b l The first label area is indicated as such, l=1.. L is; f (f) m Representing the pre-segmentation result f 0 Removing an open operation part from the image of (a);
4.3 F) to f m Reassigning to adjacent regions to make the labels of each reassigned region different and reassigning the labels to obtain an optimized result f 1 。
Further, the step 5) uses the RCF network loss function to optimize the result f 1 Optimizing, wherein the RCF network loss function is as follows:
Figure BDA0002640021430000051
wherein θ represents all parameters learned in the structure;
Figure BDA0002640021430000052
representing an activation value from the k layer; />
Figure BDA0002640021430000053
Representing an activation value from the fusion layer; i represents the number of pixels in image I; k represents the number of layers.
Further, in the step 6), the grain size is calculated by approximating the grain to a circle.
Compared with the prior art, the method outputs the image contour through the convolutional neural network; secondly, the pre-segmentation of crystal grains in the image is realized by utilizing the robust watershed transformation, and the problem that the number of segmented areas and the contour precision are difficult to balance in the watershed algorithm of the related technology is solved; and finally, optimizing the pre-segmentation result according to the image contour output by the convolutional neural network. The invention realizes the rapid pre-segmentation of the crystal grains by utilizing the robust watershed transformation, thereby not only avoiding the problem of excessive segmentation of the traditional watershed algorithm, but also solving the problem that the number of segmented areas and the contour precision are difficult to balance, and being capable of obtaining better ceramic material SEM image pre-segmentation results compared with the super-pixel algorithm of the related technology; the main contour of the image is obtained by utilizing the convolutional neural network, and the pre-segmentation result of the robust watershed is combined, so that the optimization of the segmentation contour line is completed, the contour precision of the robust watershed transformation is improved, the problem of region erroneous segmentation caused by the fact that the convolutional neural network is sensitive to the image texture is solved, on one hand, more accurate region positioning is realized by utilizing the robust watershed transformation, and on the other hand, more accurate contour is obtained by utilizing the fusion of the low-layer and high-layer features of the image, and the advantage complementation of two segmentation algorithms is realized. The invention utilizes the robust watershed transformation and the rich convolution characteristic network (Richer Convolutional Features jointing Robust Watershed Transform, RCF-RWT) to construct the segmentation model, can realize the automatic measurement of the grain size of the ceramic material, has higher calculation efficiency and accuracy, and has wide application prospect in the field of ceramic material analysis based on scanning electron microscope images.
Further, the method sets the image gray level variance threshold eta, calculates the image gray level variance var, preprocesses the image with the image gray level variance var larger than the image gray level variance threshold eta, adopts a multi-scale Retinex algorithm for preprocessing, solves the problem that the image gray level is uneven due to reflection and other reasons on the surface of the material, avoids the problem that the error segmentation is easy to be caused by the excessive brightness of partial areas in the image, ensures that the whole gray level of the preprocessed image is relatively even, and is favorable for the accurate segmentation of grains by a subsequent algorithm.
Drawings
FIG. 1 is a general flow chart of an embodiment of the present invention;
FIG. 2 is a comparison of artwork and pre-processed images, wherein (a), (b), (c) and (d) in the first row are artwork, respectively, and (e), (f), (g) and (h) in the second row are corresponding pre-processed images, respectively;
FIG. 3 is a graph comparing the morphology watershed algorithm (MGR-WT) without small regions removed, the Robust Watershed Transform (RWT), and the manual labeling results, wherein (a) is the morphology watershed algorithm segmentation result without small regions removed, (b) is the RWT segmentation result, and (c) is the RWT segmentation result compared to the manual labeling result;
FIG. 4 is a graph comparing results before and after using a morphology-based contour optimization method, wherein (a) is a segmentation result of an untreated double line, (b) is a segmentation result of a removed double line, and (c) is a comparison result of an untreated double line and a removed double line;
FIG. 5 is a graph of results comparison of pre-and post-optimization and manual labeling incorporating convolutional neural network-based image profiles, wherein (a) is post-optimization result, (b) is pre-and post-optimization result comparison, and (c) is pre-and post-optimization result comparison with manual labeling;
FIG. 6 is a graph showing a comparison of segmentation results of a first set of four non-gold plated images using the method of the present invention and five other methods, respectively, wherein (a) is a SLIC method, (b) is an LSC method, (c) is an MR-WT method, (d) is an AMR-WT method, (e) is an RCF-MR-WT, and (f) is an RCF-RWT method of the present invention;
FIG. 7 is a graph showing the comparison of the segmentation results of the second set of four gold-plated images using the method of the present invention and the other five methods, respectively, wherein (a) is the SLIC method, (b) is the LSC method, (c) is the MR-WT method, (d) is the AMR-WT method, (e) is the RCF-MR-WT, and (f) is the RCF-RWT method of the present invention.
Detailed Description
The present invention will be further illustrated by the following description, taken in conjunction with the accompanying drawings and specific embodiments, and it will be apparent that the embodiments described are some, but not all, of the embodiments of the present application. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are within the scope of the present disclosure.
Because the grain size of the ceramic material is one of important indexes for measuring physical properties of the material, the grain size is measured by a manual method at present, and the grain shape is irregular and different in size, the manual method has low measurement efficiency and large error, so that the embodiment provides a ceramic material grain segmentation algorithm based on data and model combined driving, relates to the technologies of watershed transformation, morphological processing, convolutional neural network and the like, can be applied to segmentation of ceramic material scanning electron microscope images and grain size calculation, and lays a foundation for the subsequent study of the relation between experimental conditions of the ceramic material and material performance.
The method solves the problem of uneven image gray values caused by the problems of material surface reflection and the like through a multi-scale Retinex algorithm; secondly, pre-segmenting crystal grains in the image by utilizing robust watershed transformation; and finally, optimizing the pre-segmentation result according to the image contour output by the convolutional neural network.
Specifically, referring to fig. 1, the method comprises the following steps:
1) Inputting a scanning electron microscope image f, and outputting a corresponding gradient image g by using an RCF network rcf ；
2) Obtaining a gradient image g corresponding to the scanning electron microscope image f by utilizing a structure edge algorithm;
3) Pre-segmentation is carried out on the gradient image g by utilizing the robust watershed transformation to obtain a pre-segmentation result f 0 ；
4) Pre-segmentation result f by using contour optimization method based on morphology 0 Optimizing to obtain an optimized result f 1 ；
5) Gradient image g using watershed transformation rcf Dividing to obtain a result f rcf According to result f rcf For the optimized result f 1 Optimizing to obtain a final result f 2 ；
6) According to the final result f 2 The grain area distribution was counted and the grain size was calculated.
Specifically, in step 2), setting an image gray level variance threshold η, calculating an image gray level variance var, and if the image gray level variance var is smaller than or equal to the image gray level variance threshold η, outputting a corresponding gradient image g by using a structural edge algorithm; if the image gray value variance var is larger than the image gray value variance threshold eta, preprocessing the scanning electron microscope image f to homogenize the image gray value, and then outputting a corresponding gradient image g by utilizing an RCF network structure edge algorithm. Preferably, an image gray variance threshold η=1.5x10 is set 3 。
Specifically, the model of the multiscale Retinex algorithm (Multi Scale Retinex, MSR) employed for preprocessing is:
Figure BDA0002640021430000081
R n (x,y)＝logI n (x,y)-log[F n (x,y)*I n (x,y)]
Figure BDA0002640021430000082
wherein R is MSR Is a reflected image output by a multi-scale Retinex algorithm; n=3, N is 1, 2, 3, respectively representing three scales of low, medium and high; w (w) n Represents the weight coefficient, w 1 ＝w 2 ＝w 3 ＝1/3；R n (x, y) represents a multi-scale reflected image, I n (x, y) represents an experimental image, wherein x represents an abscissa and y represents an ordinate; f (F) n (x, y) is a gaussian surround function, wherein the value of K meets the value of ≡ ≡F n (x,y)dxdy＝1；c n Represents a gaussian surround space constant, where c 1 ＝15，c 2 ＝80，c 3 ＝200。
Specifically, the pre-segmentation in step 3) includes:
3.1 Through a) a process of
Figure BDA0002640021430000083
Carrying out gradient reconstruction on the gradient image g;
wherein,,
Figure BDA0002640021430000084
representing reconstruction of a pre-segmentation result f from a gradient image g 0 Is a combined morphological closing operation, R represents morphological reconstruction, R φ Represents morphological closing operations, ε represents morphological erosion operations, and δ represents morphological dilation operations; />
Figure BDA0002640021430000085
Representing reconstruction of a pre-segmentation result f from a gradient image g 0 Morphological dilation operation of h E N + And meet the following
Figure BDA0002640021430000086
Figure BDA0002640021430000087
Represented by a gradient mapReconstructing the pre-segmentation result f from the image g 0 Morphological erosion operations of h.epsilon.N + And satisfy->
Figure BDA0002640021430000088
m represents the scale of the largest structural element, the multiscale structural element satisfying the relationship +.>
Figure BDA0002640021430000089
3.2 If the reconstruction result contains multiple small regions, then by
Figure BDA0002640021430000091
Merging the multiple small regions of the reconstruction result in step 3.1), wherein I r Representing the combined small area image; i represents the image before merging the small areas; k is a structural element parameter, and by changing a small region in the k value combined image, the larger the k value is, the more the combined region is; b k Representing the combined structural element dimensions.
Specifically, the pre-segmentation result f is solved 0 The morphology-based contour optimization method in the step 4) comprises the following steps:
4.1 For the pre-segmentation result f) 0 Labeling each region in the image of (a) and ensuring that the labels of each region are different;
4.2 Performing morphological open operation on each tag region from the pre-segmentation result f 0 Subtracting the result of the open operation from the image of (a):
Figure BDA0002640021430000092
wherein b l The first label area is indicated as such, l=1.. L is; f (f) m Representing the pre-segmentation result f 0 Removing an open operation part from the image of (a);
4.3 F) to f m Reassigning to adjacent regions to make the labels of each reassigned region different and reassigning the labels to obtain an optimized result f 1 。
Specifically, the optimized result f is obtained in step 5) by using the RCF network loss function 1 Optimizing and improving the divisionThe cutting precision, RCF network loss function is:
Figure BDA0002640021430000093
where θ represents all parameters learned in the structure,
Figure BDA0002640021430000094
representing an activation value from the k layer; />
Figure BDA0002640021430000095
Representing an activation value from the fusion layer; i represents the number of pixels in image I; k represents the number of layers.
Specifically, the grain size is calculated in step 6) by approximating the grains to circles.
The embodiment of the invention utilizes a robust watershed transformation and convolution rich characteristic network (Richer Convolutional Features jointing Robust Watershed Transform, RCF-RWT) to construct a segmentation model, and firstly solves the problem of uneven gray scale caused by reflection of the surface of a material through image preprocessing; secondly, the pre-segmentation of grains in the image is realized by utilizing the robust watershed transformation, and the problems that the traditional watershed algorithm is excessive and the number of segmented areas and the contour precision are difficult to balance are solved; and finally, optimizing the pre-segmentation result according to the image contour output by the convolutional neural network. Compared with the main stream image segmentation algorithm, the method realizes more accurate region positioning by utilizing the robust watershed transformation on one hand, and acquires more accurate contours by utilizing the fusion of the low-layer and high-layer features of the image on the other hand. The rapid pre-segmentation of the crystal grains is realized by utilizing the robust watershed transformation, so that the problem of excessive segmentation of the traditional watershed algorithm is avoided, the problem that the number of segmented areas and the contour accuracy are difficult to balance is solved, and better ceramic material SEM image pre-segmentation results can be obtained compared with the mainstream super-pixel algorithm; the main contour of the image is obtained by using the convolutional neural network, and the pre-segmentation result of the robust watershed is combined, so that the optimization of the segmentation contour line is completed, the contour precision of the robust watershed transformation is improved, the problem of region error segmentation caused by the fact that the convolutional neural network is sensitive to the image texture is solved, and the advantage complementation of two segmentation algorithms is realized. The embodiment of the invention can realize automatic measurement of the grain size of the ceramic material, has higher calculation efficiency and accuracy, provides objective and accurate data for ceramic material attribute analysis, and has wide application prospect in the field of ceramic material analysis based on scanning electron microscope images.
The effect of the embodiment of the invention is verified by the following experiment:
in order to verify the segmentation effect of the present invention on the SEM image, two sets of ceramic material images with a resolution of 1024×885 are selected as test data, the first set is an untreated ceramic SEM image, i.e. an undeplated ceramic SEM image, and the second set is a ceramic SEM image with the illumination influence removed after gold plating, and related hardware is configured as follows: training of the RCF network was performed on Intel Core i9 9900X@3.5GHZ 128GB RAM, double NVIDIA GeForce RTX 2080Ti GPU workstations, with an operating environment of Win 10 and a programming environment of PyTorrch1.2.
Referring to fig. 2, comparing the original image with the image after preprocessing, it can be seen from fig. 2 that the gray value and the overall difference of the partial area of the image without preprocessing are large, and especially the partial area of a single crystal grain is too bright, which is easy to cause erroneous segmentation. The whole gray value of the preprocessed image is uniform, which is favorable for the accurate segmentation of the crystal grains by the subsequent algorithm.
The gradient image is pre-segmented by adopting a morphology watershed algorithm (MGR-WT) without removing small areas, a Robust Watershed Transformation (RWT) and manual marking, and the RWT has the advantages of solving the problem of difficult selection of the size of the MGR-WT structural element, ensuring that the precision of the contour is not influenced by the change of a parameter k, and ensuring that the parameter change only influences the number of final segmented areas, but not the final contour precision.
The pre-segmentation result is optimized by using a contour optimization method based on morphology, and referring to fig. 4, it can be seen from the figure that the method adopted in the embodiment can solve the problem of contour double lines, and does not change the shape of the segmented contour. In order to further improve the precision of the segmentation contour, avoid the segmentation result from excessively depending on the gradient obtained by SE, introduce the image contour prediction model based on convolutional neural network, the result is shown in figure 5, purple is the optimized result in the figure, green is the pre-optimized result, and yellow is the manual marking result, and it can be seen from the figure that after the RCF network model is introduced, the segmentation precision is improved to some extent.
In order to further demonstrate the superiority of the embodiments of the present invention, the RCF-RWT method of the embodiments of the present invention is compared with the main stream image segmentation algorithms SLIC, LSC, MR-WT, AMR-WT and RCF-MR-WT algorithms, the comparison results of the algorithms are shown in fig. 6 and 7, and it can be seen from the results that the segmentation results of SLICs are uniform in size and shape, the ceramic grains are different in size and irregular in shape, and SLICs and LSCs are sensitive to gray values, so that the super-pixel algorithm is not suitable for experimental images. Compared with the MR-WT, the AMR-WT solves the problem of over-segmentation, but the segmentation result has the problem of double lines, the RCF-MR-WT utilizes the deep features of the network to perform region positioning, the shallow features supplement details, the segmentation edge accuracy is superior to the AMR-WT, but the segmentation result of the algorithm is over-serious. The RCF-RWT algorithm provided by the invention has the advantages of accurate grain segmentation, double-line problem solving and best segmentation effect. In FIG. 7, the gray value variation range of the whole image after gold plating is reduced compared with that of FIG. 6, the segmentation accuracy of MR-WT, AMR-WT and RCF-RWT is improved, and the RCF-RWT still shows the best result.
To further compare the performance of different segmentation algorithms, the embodiment of the present invention uses four algorithm indexes to test the segmentation results, namely, overlap ratio (CV), variation information (Variation of Information, VI), global consistency error (Global Consistency Error, GCE) and boundary displacement error (Boundary Displacement Error, BDE). Wherein the larger the CV value, the better the segmentation result; the smaller the values of VI, GCE and BDE, the better the segmentation result. Table 1 shows the mean experimental index values of the first group of non-gold plated images, and Table 2 shows the mean experimental index values of the second group of gold plated images.
CV↑ VI↓ GCE↓ BDE↓
SLIC 0.3547 3.0524 0.4396 10.1678
LSC 0.3455 2.8820 0.3563 7.5911
MR-WT 0.4680 2.3887 0.1364 5.0346
AMR-WT 0.8287 1.1280 0.1122 1.6261
RCF-MR-WT 0.6636 1.4952 0.0955 3.5651
RCF-RWT 0.8697 0.8710 0.0763 1.6262
TABLE 1
CV↑ VI↓ GCE↓ BDE↓
SLIC 0.3279 3.0962 0.4070 11.3350
LSC 0.3347 2.8418 0.3265 8.0651
MR-WT 0.7979 1.2031 0.1033 2.0565
AMR-WT 0.8757 0.9909 0.1110 1.2623
RCF-MR-WT 0.5771 1.7691 0.0895 4.8813
RCF-RWT 0.9217 0.6699 0.0628 1.0201
TABLE 2
As can be seen from fig. 6, the SLIC division result is not accurate, and although the LSC division result is higher than the grain boundary coincidence ratio, the algorithm is greatly affected by the pixel gray value, and a closed region is not formed at the grain edge, so that CV indexes of the two algorithms are low. The MR-WT constitutes a closed region compared to the LSC, so the CV value of the algorithm is greater than that of the LSC. However, MR-WT over-segmentation is severe, so that CV indicators are greatly improved when AMR-WT overcomes over-segmentation problems. The segmentation result of the RCF-MR-WT is also very severe, so the CV value of the algorithm is lower than that of the AMR-WT, but the accuracy of the segmentation edges of the RCF-MR-WT is high, which is manifested in that the change information and errors of the algorithm are smaller than those of the AMR-WT. Compared with other algorithms, the RCF-RWT obtains the highest experimental index. Analysis of the experimental data of table 2, the experimental indicators of SLIC and LSC after removal of the light interference were similar to those in table 1, and these algorithms all showed lower segmentation accuracy. The RCF-MR-WT utilizes the deep and shallow features, the algorithm result is stable, and the segmentation result is not affected by illumination. The experimental indexes of MR-WT, AMR-WT and RCF-RWT are improved compared with those of Table 1, and the segmentation results of the three algorithms in FIG. 7 are better than those of FIG. 6. The analysis results show that the visual effect of the segmentation result diagram is consistent with the visual data of the experimental index, the RCF-RWT segmentation effect is best, and the segmentation accuracy is highest.
After the image segmentation is completed, the grain size can be calculated. In the image segmentation process, the grains are already separated, and the number of grains and the grain size distribution of the grains in one image can be obtained. And selecting grains with uniform sizes in each image, calculating the average value of the sizes, wherein the calculation method can approximate the grains to be round, and indirectly calculating the diameters of the grains by obtaining the areas of the grains, namely the sizes of the grains. Wherein, the manual measurement mode selects 5 crystal grains with uniform shape and size for each image, and the crystal grain size is measured and averaged. Since the artificial measurement results are greatly subjectively influenced by the measurer, 5 measurers are selected to respectively measure the experimental images, the obtained measurement data are shown in table 3, the measured values are averaged after the maximum value and the minimum value are removed, so that the influence of subjective factors on the measurement is weakened, and the finally obtained artificial measurement results are shown in table 4.
Measurer 1 Measurer 2 Measurer 3 Measurer 4 Measurer 5
1 94.55 89.17 93.39 94.22 88.51
2 90.92 100.33 105.38 91.48 99.91
3 107.50 100.91 102.09 96.49 89.91
4 101.61 89.91 92.08 94.42 93.38
5 108.31 103.88 95.16 102.45 93.52
6 112.51 108.21 112.34 109.70 107.84
7 101.85 104.13 102.80 94.40 89.73
TABLE 3 Table 3
Figure BDA0002640021430000141
TABLE 4 Table 4
As can be seen from tables 3 and 4, the error of manual measurement is large, and the measurement is greatly influenced by the subjective effect, and the measurement is time-consuming and labor-consuming. Compared with a comparison algorithm, the RCF-RWT provided has the advantages that the obtained grain size is closer to a real result, and the embodiment of the invention is further verified.
Therefore, the invention utilizes the robust watershed transformation and the convolution neural network to construct the segmentation model, can realize the automatic measurement of the grain size of the ceramic material, has higher calculation efficiency and accuracy, provides objective and accurate data for the attribute analysis of the ceramic material, lays a foundation for the subsequent research of the relation between the experimental condition of the ceramic material and the material performance, and has wide application prospect in the field of the analysis of the ceramic material based on the scanning electron microscope image.

Claims (7)

1.一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，其特征在于，包括以下步骤：1. A ceramic material grain segmentation algorithm based on data and model joint drive, is characterized in that, comprises the following steps: 1)输入扫描电镜图像f，利用RCF网络输出对应的梯度图像grcf；1) Input the scanning electron microscope image f, and use the RCF network to output the corresponding gradient image g rcf ; 2)利用结构边缘算法获得扫描电镜图像f对应的梯度图像g；2) Using the structural edge algorithm to obtain the gradient image g corresponding to the scanning electron microscope image f; 3)利用鲁棒分水岭变换对梯度图像g进行预分割，得到预分割结果f0；3) Pre-segment the gradient image g by using the robust watershed transform, and obtain the pre-segmentation result f 0 ; 所述步骤3)中预分割包括：Said step 3) pre-segmentation includes: 3.1)通过
Figure FDA0004236861480000011
对梯度图像g进行梯度重建；3.1) pass
Figure FDA0004236861480000011
Perform gradient reconstruction on the gradient image g; 其中，
Figure FDA0004236861480000012
表示由梯度图像g重建预分割结果f0的组合形态学闭运算，R表示形态学重建，Rφ表示形态学闭运算，ε表示形态学腐蚀运算，δ表示形态学膨胀运算；
Figure FDA0004236861480000013
表示由梯度图像g重建预分割结果f0的形态学膨胀运算，h∈N+，且满足
Figure FDA0004236861480000014
表示由梯度图像g重建预分割结果f0的形态学腐蚀运算，h∈N+，且满足/>
Figure FDA0004236861480000015
m表示最大结构元素的尺度，多尺度结构元素满足关系/>
Figure FDA0004236861480000016
in,
Figure FDA0004236861480000012
Represents the combined morphological closing operation of the pre-segmentation result f 0 reconstructed from the gradient image g, R represents the morphological reconstruction, R φ represents the morphological closing operation, ε represents the morphological erosion operation, and δ represents the morphological expansion operation;
Figure FDA0004236861480000013
Represents the morphological expansion operation of the pre-segmentation result f 0 reconstructed from the gradient image g, h∈N + , and satisfies
Figure FDA0004236861480000014
Represents the morphological erosion operation of reconstructing the pre-segmentation result f 0 from the gradient image g, h∈N + , and satisfies />
Figure FDA0004236861480000015
m represents the scale of the largest structural element, and multi-scale structural elements satisfy the relation />
Figure FDA0004236861480000016
 3.2)若重建结果包含多个小区域，则通过
Figure FDA0004236861480000017
对步骤3.1)中重建结果的多个小区域进行合并，其中，Ir表示合并小区域后的图像；I表示合并小区域前的图像；k是结构元素参数；bk表示合并的结构元素尺度；3.2) If the reconstruction result contains multiple small areas, pass
Figure FDA0004236861480000017
Merge multiple small regions of the reconstruction results in step 3.1), where Ir represents the image after merging the small regions; I represents the image before merging the small regions; k is the structural element parameter; b k represents the combined structural element scale ; 4)利用基于形态学的轮廓优化方法对预分割结果f0进行优化，得到优化后结果f1；4) Using the contour optimization method based on morphology to optimize the pre-segmentation result f 0 to obtain the optimized result f 1 ; 5)利用分水岭变换对梯度图像grcf进行分割，得到结果frcf，根据结果frcf对优化后结果f1进行优化，得到最终结果f2；5) Using the watershed transform to segment the gradient image g rcf to obtain the result f rcf , optimize the optimized result f 1 according to the result f rcf to obtain the final result f 2 ; 6)根据最终结果f2，统计晶粒面积分布，并计算晶粒尺寸。6) According to the final result f 2 , count the grain area distribution and calculate the grain size. 2.根据权利要求1所述的一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，其特征在于，所述步骤2)中设定图像灰度方差阈值η，计算图像灰度值方差var，若图像灰度值方差var小于等于图像灰度方差阈值η，则利用结构边缘算法输出对应的梯度图像g；若图像灰度值方差var大于图像灰度方差阈值η，则先对扫描电镜图像f进行预处理使图像灰度值均匀化，再利用结构边缘算法输出对应的梯度图像g。2. a kind of ceramic material grain segmentation algorithm based on data and model joint drive according to claim 1, it is characterized in that, described step 2) in setting image grayscale variance threshold n, calculate image grayscale value variance var, if the image gray value variance var is less than or equal to the image gray value variance threshold η, then use the structural edge algorithm to output the corresponding gradient image g; if the image gray value variance var is greater than the image gray value variance threshold η, first scan the The image f is preprocessed to make the gray value of the image uniform, and then the corresponding gradient image g is output by using the structural edge algorithm. 3.根据权利要求2所述的一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，其特征在于，所述图像灰度方差阈值η＝1.5×103。3. A ceramic material grain segmentation algorithm based on data and model joint drive according to claim 2, characterized in that the image gray level variance threshold η=1.5×10 3 . 4.根据权利要求2所述的一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，其特征在于，所述步骤2)中预处理采用多尺度Retinex算法，模型为：4. a kind of ceramic material grain segmentation algorithm based on the joint drive of data and model according to claim 2, is characterized in that, described step 2) in preprocessing adopts multi-scale Retinex algorithm, and model is:
Figure FDA0004236861480000021
Figure FDA0004236861480000021
 Rn(x,y)＝logIn(x,y)-log[Fn(x,y)*In(x,y)]R n (x,y)=logI n (x,y)-log[F n (x,y)*I n (x,y)]
Figure FDA0004236861480000022
Figure FDA0004236861480000022
 其中，RMSR是多尺度Retinex算法输出的反射图像；N＝3，n取1、2、3，分别表示低、中、高三个尺度；wn表示权重系数，w1＝w2＝w3＝1/3；Rn(x,y)表示多尺度的反射图像，In(x,y)表示实验图像，其中，x表示横坐标，y表示纵坐标；Fn(x,y)是高斯环绕函数，其中，K的取值满足∫∫Fn(x,y)dxdy＝1；cn表示高斯环绕空间常数，其中，c1＝15，c2＝80，c3＝200。Among them, R MSR is the reflection image output by the multi-scale Retinex algorithm; N=3, n takes 1, 2, and 3, representing the three scales of low, medium, and high respectively; w n represents the weight coefficient, w 1 =w 2 =w 3 =1/3; R n (x, y) represents the multi-scale reflection image, I n (x, y) represents the experimental image, where x represents the abscissa, y represents the ordinate; F n (x, y) is Gaussian surround function, where the value of K satisfies ∫∫F n (x,y)dxdy=1; c n represents the Gaussian surround space constant, where c 1 =15, c 2 =80, c 3 =200. 5.根据权利要求1所述的一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，其特征在于，所述步骤4)中基于形态学的轮廓优化方法包括：5. a kind of ceramic material grain segmentation algorithm based on data and model joint drive according to claim 1, is characterized in that, described step 4) in the profile optimization method based on morphology comprises: 4.1)对预分割结果f0的图像中的每个区域进行标签化，确保图像中每个标签只能覆盖一个区域；4.1) Label each region in the image of the pre-segmentation result f 0 to ensure that each label in the image can only cover one region; 4.2)对每个标签区域执行形态学开运算，从预分割结果f0的图像中减去开运算结果：
Figure FDA0004236861480000023
其中，bl表示第l个标签区域，l＝1,...,L；fm表示预分割结果f0的图像中去掉开运算部分；4.2) Perform a morphological opening operation on each labeled region, and subtract the result of the opening operation from the image of the pre-segmentation result f 0 :
Figure FDA0004236861480000023
Wherein, b 1 represents the lth label region, l=1,...,L; f m represents the image of the pre-segmentation result f 0 and removes the opening operation part; 4.3)将fm重新分配给相邻区域，使重新分配后的每个区域的标签不同，并重新标号，得到优化后结果f1。4.3) Reassign f m to adjacent regions, make the labels of each region different after reassignment, and relabel, and obtain the optimized result f 1 . 6.根据权利要求1所述的一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，其特征在于，所述步骤5)中利用RCF网络损失函数对优化后结果f1进行优化，RCF网络损失函数为：6. A kind of ceramic material grain segmentation algorithm based on data and model joint drive according to claim 1, it is characterized in that, in described step 5), utilize RCF network loss function to optimize result f 1 after optimization, RCF The network loss function is:
Figure FDA0004236861480000031
Figure FDA0004236861480000031
 其中，θ表示在结构中学习到的所有参数；
Figure FDA0004236861480000032
表示来自k层的激活值；/>
Figure FDA0004236861480000033
表示来自融合层的激活值；|I|表示图像I中的像素数；K表示层数。where θ represents all the parameters learned in the structure;
Figure FDA0004236861480000032
Indicates the activation value from layer k; />
Figure FDA0004236861480000033
Denotes the activation value from the fusion layer; |I| denotes the number of pixels in image I; K denotes the number of layers. 7.根据权利要求1所述的一种基于数据与模型联合驱动的陶瓷材料晶粒分割算法，其特征在于，所述步骤6)中将晶粒近似为圆计算晶粒尺寸。7. A kind of ceramic material grain segmentation algorithm based on joint driving of data and model according to claim 1, characterized in that, in said step 6), the grain is approximated as a circle to calculate the grain size.
CN202010836869.5A 2020-08-19 2020-08-19 A Grain Segmentation Algorithm for Ceramic Materials Driven by Joint Data and Model Active CN111915630B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010836869.5A CN111915630B (en) 2020-08-19 2020-08-19 A Grain Segmentation Algorithm for Ceramic Materials Driven by Joint Data and Model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010836869.5A CN111915630B (en) 2020-08-19 2020-08-19 A Grain Segmentation Algorithm for Ceramic Materials Driven by Joint Data and Model

Publications (2)

Publication Number Publication Date
CN111915630A CN111915630A (en) 2020-11-10
CN111915630B true CN111915630B (en) 2023-06-23

Family

ID=73278295

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010836869.5A Active CN111915630B (en) 2020-08-19 2020-08-19 A Grain Segmentation Algorithm for Ceramic Materials Driven by Joint Data and Model

Country Status (1)

Country Link
CN (1) CN111915630B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112465838B (en) * 2020-12-10 2023-01-31 桂林理工大学 Ceramic grain image segmentation method, system, storage medium and computer equipment
CN114596295B (en) * 2022-03-15 2024-10-25 中南大学 Method and device for extracting tortuosity of lithium ion cathode material SEM (scanning electron microscope) graph

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
FR2705005A1 (en) * 1993-05-05 1994-11-10 Philips Laboratoire Electroniq Device for segmenting images composed of textures from digital signals representing the said images
JP2006292544A (en) * 2005-04-11 2006-10-26 Univ Of Fukui Crystal structure inspection support device and support method for metal materials, etc.
CN102947727A (en) * 2010-06-15 2013-02-27 皇家飞利浦电子股份有限公司 Automated crystal identification achieved via modifiable templates
CN106599793A (en) * 2016-11-21 2017-04-26 江苏大学 Marked watershed segmentation-based steel grain boundary automatic extraction method
CN108226159A (en) * 2017-12-29 2018-06-29 钢铁研究总院 The full filed quantitative statistics distribution characterizing method of precipitated phase particle in metal material
CN109961451A (en) * 2019-03-22 2019-07-02 西北工业大学 A Grain Structure Segmentation Method Based on Edge Information
CN109993773A (en) * 2019-03-28 2019-07-09 北京科技大学 A multi-target tracking method and device for a series of cross-sectional images

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
FR2705005A1 (en) * 1993-05-05 1994-11-10 Philips Laboratoire Electroniq Device for segmenting images composed of textures from digital signals representing the said images
JP2006292544A (en) * 2005-04-11 2006-10-26 Univ Of Fukui Crystal structure inspection support device and support method for metal materials, etc.
CN102947727A (en) * 2010-06-15 2013-02-27 皇家飞利浦电子股份有限公司 Automated crystal identification achieved via modifiable templates
CN106599793A (en) * 2016-11-21 2017-04-26 江苏大学 Marked watershed segmentation-based steel grain boundary automatic extraction method
CN108226159A (en) * 2017-12-29 2018-06-29 钢铁研究总院 The full filed quantitative statistics distribution characterizing method of precipitated phase particle in metal material
CN109961451A (en) * 2019-03-22 2019-07-02 西北工业大学 A Grain Structure Segmentation Method Based on Edge Information
CN109993773A (en) * 2019-03-28 2019-07-09 北京科技大学 A multi-target tracking method and device for a series of cross-sectional images

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于图像处理技术的多晶体材料金相图像分割算法研究;王宝珠;李芳影;刘翠响;胡洋;;铸造(第11期);全文 *
数据与模型联合驱动的陶瓷材料晶粒分割;雷涛;李云彤;周文政;袁启斌;王成兵;张小红;自动化学报;第48卷(第004期);全文 *

Also Published As

Publication number Publication date
CN111915630A (en) 2020-11-10

Similar Documents

Publication Publication Date Title
CN113435460B (en) A recognition method for bright crystal granular limestone images
CN107256558B (en) Unsupervised automatic cervical cell image segmentation method and system
CN109658381B (en) Method for detecting copper surface defects of flexible IC packaging substrate based on super-pixels
CN117315289B (en) Aeroengine blade contour edge detection method based on image processing
CN113592845A (en) Defect detection method and device for battery coating and storage medium
CN114723704B (en) Textile quality evaluation method based on image processing
CN102494976B (en) Method for automatic measurement and morphological classification statistic of ultra-fine grain steel grains
CN109636772A (en) The defect inspection method on the irregular shape intermetallic composite coating surface based on deep learning
WO2018107939A1 (en) Edge completeness-based optimal identification method for image segmentation
Li et al. Application of multi-scale feature fusion and deep learning in detection of steel strip surface defect
CN100545865C (en) An Automatic Segmentation Method for Optimizing the Initial Segmentation Boundary of Image
AU2015392660A1 (en) Stepwise-refinement pavement crack detection method
CN115937143A (en) Fabric defect detection method
CN103543394A (en) A method for extracting quantitative parameters of high-voltage electrical equipment discharge ultraviolet imaging
CN111915630B (en) A Grain Segmentation Algorithm for Ceramic Materials Driven by Joint Data and Model
CN111192273A (en) A digital shot peening coverage measurement method based on computer vision technology
CN116664565A (en) Hidden crack detection method and system for photovoltaic solar cell
CN110910367A (en) A method for evaluating the quality of cell culture in bioreactors
CN119850605A (en) Visual detection method and system for welding quality of heat exchanger
CN115908362A (en) Method for detecting wear resistance of skateboard wheel
CN110047079A (en) A kind of optimum segmentation scale selection method based on objects similarity
CN114152211A (en) Fracturing propping agent roundness measuring method based on microscopic image processing
CN119991670B (en) Visual detection method for titanium alloy bar based on image segmentation
CN118470006B (en) Novel heat exchange tube quality visual detection method and system
CN117994271B (en) Human brain image segmentation method based on watershed algorithm

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant