Material growth analysis method based on image features

Abstract

The invention discloses a material growth analysis method based on image features. The method comprises the following steps: acquiring the growth condition of a nano-film material under a microscope, and selecting a later frame for pixel labeling; performing pixel classification and aggregation on the key frames according to the labels by using a tree model, and obtaining simplified position and shape information of a correction point, a crystal domain and a liquid growth center; acquiring a video jitter correction matrix by analyzing the center point coordinate of the correction point, and applying the matrix to the domain edge and the liquid growth nucleus; and matching the coordinates of each liquid growth center to obtain growth trajectories of different paths, filtering to relieve partial abnormal and error value interference, and drawing a liquid growth center trajectory and a domain mask pattern. According to the method, the growth condition of the nano-film material in the liquid growth process can be captured in real time, and traversal is provided for subsequent analysis.

Description

Technical Field

[0001] The invention belongs to the field of nano film material growth result analysis, and in particular relates to a material growth analysis method based on image features. Background Art

[0002] Nanofilm materials have broad application prospects in the fields of electronics, optics, catalysis, energy storage and sensors. Especially in electronics, nanofilms are widely used in the manufacture of high-performance semiconductor devices, memory and sensors, promoting the development of information technology and intelligent hardware. Therefore, how to efficiently and controllably prepare high-quality nanofilm materials is a key research topic in materials science.

[0003] The preparation method of using liquid cores as growth centers has attracted attention as a special thin film preparation technology. In this process, the material first forms tiny liquid cores in a liquid state, and then these liquid cores are transformed into crystal domains on a solid substrate and finally grow into nanofilms. Compared with traditional gas or solid phase growth methods, liquid core growth has significant advantages in growth rate. The formation of liquid cores makes the growth process more efficient and improves the uniformity and stability of the film.

[0004] The introduction of machine learning and image processing has significantly improved the monitoring and optimization capabilities of the film growth process. Through methods based on image feature extraction, the dynamic changes in the liquid core growth process can be captured in real time, key data can be extracted and the film growth can be analyzed. The processing and analysis of these image data, combined with machine learning algorithms, can predict the growth trend of the film and optimize the growth parameters, thereby improving the preparation efficiency and film quality. Summary of the invention

[0005] The problem to be solved by the present invention is to provide a material growth analysis method based on image features, which is used to capture the dynamic changes of nano-thin film materials in the liquid growth process in real time, extract key data, and analyze the growth of the film.

[0006] The present invention adopts the following technical solution: a material growth analysis method based on image features, comprising the following steps:

[0007] S1, obtaining the liquid growth of nanofilm materials under a microscope, and selecting later frames for pixel annotation;

[0008] S2. Using a tree model to classify and aggregate the pixels of key frames according to pixel annotation, construct a decision tree model, and obtain simplified position and shape information of inert impurities, crystal domains, and liquid growth centers;

[0009] S3, taking the geometric center point of the inert impurity as the correction point and the geometric center point of the liquid growth center as the liquid growth core, obtaining the video jitter correction matrix by analyzing the coordinates of the correction points, and applying it to the crystal domain edge and the liquid growth core to obtain the corrected crystal domain edge information and the liquid growth core coordinates;

[0010] S4, matching the coordinates of each liquid growth core to obtain growth trajectories of different paths;

[0011] S5. Filter each path obtained in S4 to alleviate the interference of abnormalities and erroneous values, and draw the liquid growth center trajectory and the crystal domain mask map according to the crystal domain edge and liquid growth core in S3.

[0012] Preferably, in step S1, one early growth frame and one mid growth frame are selected as key frames, and the inert impurities, crystal domains, liquid growth centers, and background used for correction are pixel-annotated to obtain the annotated data set Piexl_Matrix (Piexl-Matrix = [R1, R2, R3, R4]).

[0013] Preferably, in step S2, a tree model is used to extract the features of R1, R2, R3, and R4 from Piexl_Matrix and construct a classifier. The tree model is divided based on the pixel position and color channel features in the training set, and the Gini score is calculated:

[0014]

[0015] Among them, R i Indicates the data source for calculation (inert impurities R1, crystal domains R2, liquid growth centers R3, background R4) Gini represents the Gini score calculation formula, X j represents the feature used for segmentation, and t represents the segmentation threshold in the decision tree model.

[0016] Recursively update the selection of leaf nodes to find the optimal partitioning features and partitioning points. Specifically, by minimizing the partitioning Gini score of the training set, construct leaf nodes and finally form a decision tree:

[0017]

[0018] in, t * Indicates the best partitioning features and optimal partitioning points, Gini split Represents the partitioning feature X corresponding to each feature calculated according to the partitioning threshold j The Gini score of .

[0019] Finally, the pixel-scale division of the four regions of inert impurities, crystal domains, growing liquid nuclei and background is achieved.

[0020] Furthermore, the Canny operator is used to extract the inert impurity edge Edge_r ([f, N0, 2], only one impurity point is selected for correction, f is the number of selected key frames, Ni is the number of edge fitting points), the crystal domain edge Edge_s ([f, n1, N1, 2], n1 is the number of crystal domains, and if the crystal domain disappears, the value -1 is used to fill the gap to ensure that the matrix is ​​full of elements), and the liquid growth core Edge_l ([f, n2, N2, 2], n2 is the number of liquid growth cores, and if the growth core disappears, the value -1 is used to fill the gap to ensure that the matrix is ​​full of elements).

[0021] Preferably, in step S3, according to the zeroth dimension of the inert impurity edge Edge-r, f edge information is extracted. Since the area of ​​the inert impurity is small, the center coordinates are obtained by taking the average value of i (i=0, 1, ..., N0-1) contour coordinates as the impurity center center_i, and all center_i are stored in the center point matrix Center_Matrix in sequence order, that is:

[0022] Center_Matrix=[center_i|i∈{0,1,…,N0-1}]

[0023] Similarly, considering the small size of the liquid growth core, the geometric center of the liquid growth center is used as the liquid growth core. For each of the n2 growth centers, the weighted average of the edge points (the number is N2) is taken as the coordinates of the liquid growth core, and finally the liquid growth core matrix Center_liquid is obtained, with a shape of [f, n2, 2].

[0024] Furthermore, by comparing the changes in the values ​​of the correction impurity center point matrix Center_Matrix on a time scale, the image jitter correction matrix Stable_Matrix of the selected key frame can be obtained, specifically:

[0025] Stabled_Matrix=[0,Center_Matrix[i+1]-Center_Matrix[i]∣i∈{0,1,…,N0-2}]

[0026] Next, Stable_Matrix is ​​applied to Edge_s and Edge_l to obtain the corrected domain edge information and the corrected liquid growth core matrix Center_liquid′:

[0027] Edge s ′ =Edge_s-Stable_Matrix

[0028] Center_liquid′=Center_liquid-Stabled_Matrix

[0029] Preferably, in step S4, the corrected liquid growth core matrix Center_liquid' is examined, and frames are extracted from the original video at a certain interval in time to convert the video into multiple images as key frames, and f is used to represent the number of selected key frames, and n2 represents the number of liquid growth cores in each frame.

[0030] The practical operation is as follows: 1. The interval between key frames is not large (no more than 2s); 2. Multiple growth nuclei are formed by the fission of a main nucleus of a growth center. The main function of the liquid growth nucleus is to promote the accelerated growth around the nucleus and the liquid nucleus keeps moving until it disappears.

[0031] For the case of f>1, the nearest neighbor matching points are found in the previous frame for all points in n2 whose values ​​are not (-1, -1), allowing multiple points to match the same point in the previous frame (for example, Center_liquid'[i+1, j], Center_liquid'[i+1, k] are simultaneously matched to Center_liquid'[i, m], i, i+1, j, k are all within the index range), and the distance threshold is set at the same time.

[0032] If the Euclidean distance between the point in the current frame and the point in the previous frame that matches it is greater than the threshold, the matching is terminated and this path will not be matched subsequently. To ensure consistency with the length of other paths, the filling point is selected as (-1, -1).

[0033] If the longest path contains L points, since there are no subsequent points to be matched or the distance threshold is exceeded, all matching is terminated and all matching results are stored in Paths with a shape of [M, L, 2], where M is the total number of paths screened out and L is the number of liquid growth centers contained in the longest path.

[0034] Preferably, in step 5, for the path matrix Paths, each element of the zeroth dimension represents a path, which is expressed as: (path_0, path_1, ... path_M-1).

[0035] When the point in the path is not (-1, -1), the path is filtered using a Gaussian kernel:

[0036]

[0037] Among them, x is the horizontal or vertical coordinate, k represents the Gaussian kernel size, σ is the smoothing weight used to control the smoothing strength, i represents the current position number, and j selects the number of surrounding points of the current position i for smoothing.

[0038] Furthermore, according to the non-1 coordinates in the crystal domain edge Edge_s, a white mask is drawn, and the background is drawn in black; according to the Paths' obtained after filtering and each path therein, a path map is drawn on the mask.

[0039] The technical solution of the present invention also provides: an electronic device, comprising:

[0040] one or more processors;

[0041] a storage device having one or more programs stored thereon;

[0042] When the one or more programs are executed by the one or more processors, the one or more processors implement any of the above-mentioned material growth analysis methods based on image features.

[0043] The technical solution of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in any of the above-mentioned material growth analysis methods based on image features are implemented.

[0044] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0045] 1. Liquid growth is a novel growth form in the field of material growth. The analysis method of the present invention can effectively extract the information contained in this novel growth mode, which is convenient for researchers to analyze the growth process and growth mechanism to promote the progress of this growth method;

[0046] 2. The analysis method of the present invention adopts machine learning to improve the efficiency of information collection, making it possible to process data in high throughput and large batches. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the material growth analysis method based on image features of the present invention;

[0048] Figure 2 This is a schematic diagram comparing the mask result after pixel-level fitting extracted by the tree model in the present invention with the original image;

[0049] Figure 3 A crystal domain mask and a liquid growth path diagram drawn for an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the application is further elaborated in detail below in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in the field on this embodiment belong to the protection scope of the present invention. At the same time, for the step numbering in the embodiment of the present invention, it is only set for the convenience of explanation, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0051] The present invention focuses on the liquid growth form of nano-thin film materials, integrates the technical assistance of machine learning and computer vision, and realizes efficient growth information extraction and presentation for subsequent analysis.

[0052] In one embodiment of the present invention, the liquid core growth information of nano-thin film materials is extracted based on image features. Taking MoS2 as an example, the liquid growth path capture method of nano-thin film materials is as follows: Figure 1 As shown, the following steps are included:

[0053] S1, obtain the growth of MoS2 under the microscope and select the later frames for pixel annotation;

[0054] S2, using a tree model to classify and aggregate pixels of key frames according to annotations, and obtain simplified position and shape information of calibration points, crystal domains, and liquid growth centers;

[0055] S3, obtaining a video jitter correction matrix by analyzing the center point coordinates of the correction points and applying it to the crystal domain edge and liquid growth core;

[0056] S4, matching the coordinates of each liquid growth center to obtain the growth trajectories of different paths, and then filtering to alleviate some abnormalities and error value interference;

[0057] S5. Draw the liquid growth center trajectory and the crystal domain mask map according to the crystal domain edge in S3 and the liquid growth core in S4.

[0058] Specifically, in step S1, the key frame interval is set to 5 frames for a total of 18 key frames, and the 3rd and 17th frames are selected as key frames as annotation frames to perform pixel annotation of inert impurities, crystal domains, and growing liquid nuclei and save them as an annotation data set.

[0059] Specifically, in step S2, a decision tree model is constructed using the Gini score as a partitioning parameter, and the maximum depth is set to 4 to ensure generalization. The decision tree model is trained on the labeled data set obtained in S1.

[0060] Save the model and load other data to be analyzed to obtain pixel classification maps of inert impurities, crystal domains, and liquid growth centers.

[0061] The Canny operator is used to fit the pixels into a closed graph, and the inert impurity edge information Edge_r ([18, 151, 2], the crystal domain edge information Edge_s ([18, 1, 104, 2]), and the liquid growth core Edge_l ([18, 2, 189, 2], in this embodiment, the number of growth cores is split from 1 to 2) are obtained. The schematic diagram after fitting is shown as follows: Figure 2 .

[0062] Specifically, in step S3, 18 edge information are extracted according to the zeroth dimension of the inert impurity edge Edge_r. Since the area of ​​the inert impurity is small, the average value of i (i = 0, 1, ..., 17) contour coordinates is taken to obtain the center coordinate, which is regarded as the impurity center center_i, and all center_i are stored in the center point matrix Center_Matrix in sequence order.

[0063] Taking into account the small size of the liquid growth core, the geometric center of the liquid growth center is used as the liquid growth core. For each of the two growth cores, the weighted average of the edge points (the maximum value is 189) is taken as the growth center of the liquid growth core. Finally, the liquid core growth core matrix Center_liquid is obtained, with a shape of [18, 2, 2]. After correction by Center_Matrix, Center_liquid' is obtained, and the shape remains unchanged.

[0064] Specifically, in step S4, the key frame interval is 0.25s, the number of paths in Center_liquid' is 2, the distance threshold is set to 250 pixels, and point matching is performed. The path is split at 2 points and stored as Paths = [path0, path1], where the length of path0 is 8 points and the length of path1 is 18 points. That is, the liquid growth core of the first path disappears at the 8th key frame, while the second path is maintained until the last key frame.

[0065] Specifically, in step S5, a Gaussian kernel with k=2 and σ=1 is selected for path0 and path1 to filter, and the crystal domain mask path is drawn, such as Figure 3 shown.

[0066] It can be seen that the method of this embodiment can more effectively extract the information contained in the growth mode of liquid growth, which is convenient for researchers to study the growth process and growth mechanism to promote the progress of this growth method.

[0067] In an embodiment of the present invention, an electronic device is also provided, including: one or more processors; a storage device on which one or more programs are stored; when the one or more programs are executed by the one or more processors, the one or more processors implement the material growth analysis method based on image features described in any of the above embodiments.

[0068] In an embodiment of the present invention, a computer-readable storage medium is further provided, on which a computer program is stored. When the program is executed by a processor, the steps in any one of the material growth analysis methods based on image features in the above embodiments are implemented.

[0069] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A material growth analysis method based on image features, characterized in that: The steps include: S1, obtaining the liquid growth of nanofilm materials under a microscope, and selecting later frames for pixel annotation; S2. Using a tree model to classify and aggregate the pixels of key frames according to pixel annotation, construct a decision tree model, and obtain simplified position and shape information of inert impurities, crystal domains, and liquid growth centers; S3, taking the geometric center point of the inert impurity as the correction point and the geometric center point of the liquid growth center as the liquid growth core, obtaining the video jitter correction matrix by analyzing the coordinates of the correction points, and applying it to the crystal domain edge and the liquid growth core to obtain the corrected crystal domain edge information and the liquid growth core coordinates; S4, matching the coordinates of each liquid growth core to obtain growth trajectories of different paths; S5. Filter each path obtained in S4 to alleviate the interference of abnormalities and erroneous values, and draw the liquid growth center trajectory and the crystal domain mask map according to the crystal domain edge and liquid growth core in S3.

2. The material growth analysis method based on image features according to claim 1, characterized in that: The pixel labeling method in step S1 is as follows: Select one early growth frame and one middle growth frame as key frames, annotate the inert impurity R1, crystal domain R2, liquid growth center R3, and background R4, and save them as the annotated data set Piexl-Matrix = [R1, R2, R3, R4].

3. The material growth analysis method based on image features according to claim 2, characterized in that: The decision tree model is constructed in step S2 as follows: Use the tree model to extract the features of R1, R2, R3, and R4 from Piexl_Matrix and build a classifier. Perform division based on the pixel position and color channel features in the training set, calculate the Gini score as the recursive indicator of the decision tree, find the optimal division features and division points, recursively update the leaf nodes, and form a decision tree model.

4. The material growth analysis method based on image features according to claim 3, characterized in that: Step S2 divides the pixel scale of the four areas of inert impurities, crystal domains, liquid growth centers, and background based on the trained decision tree model, uses the Canny operator to fit the pixels into a closed graph, and extracts the inert impurity edge Edge_r, the crystal domain edge Edge_s, and the liquid growth core Edge_l.

5. The material growth analysis method based on image features according to claim 4, characterized in that: Step S3, the method is as follows: S3.

1. Extract edge information according to the zeroth dimension of the inert impurity edge Edge-r, take the average value based on the contour coordinates to obtain the center coordinates, obtain the impurity center, and store all impurity centers in the impurity center point matrix Center_Matrix in sequence: S3.2, use the geometric center of the liquid growth center as the liquid growth core, for each liquid growth center, take the edge points for weighted average, get the coordinates of the liquid growth core, and construct the liquid growth core matrix Center_liquid; S3.3, compare the change of the value of the impurity center point matrix Center_Matrix on the time scale, and obtain the image jitter correction matrix Stable_Matrix of the key frame; S3.

4. Apply Stable_Matrix to Edge_s and Edge_l to obtain the corrected crystal domain edge information Edge_s′ and the corrected liquid growth core matrix Center_liquid′.

6. The material growth analysis method based on image features according to claim 5, characterized in that: Step S4: before matching the coordinates of each liquid growth center, based on the corrected liquid growth core matrix Center_liquid′, extract frames at a certain interval from the original video to convert the video into multiple images as key frames. The number of key frames is f, the time interval is no more than 2s, and the number of liquid growth cores in each frame is n2; A plurality of liquid growth cores are formed by the splitting of a liquid growth center main core, the position of the liquid growth center main core remains relatively unchanged, the liquid growth core is used to promote the growth acceleration around the core, and the liquid core is always in motion until it disappears.

7. The material growth analysis method based on image features according to claim 6, characterized in that: Step S4 matches the coordinates of each liquid growth center in the following way: For all points in n2 whose values ​​are not (-1, -1), find the nearest neighbor matching point in the previous frame, allowing multiple liquid growth cores to match the same point in the previous frame at the same time, and set the distance threshold to limit the matching range; If the Euclidean distance between the point in the current frame and the point in the previous frame that matches it is greater than the threshold, the matching is terminated. To ensure the consistency of the path length, the empty point is filled with (-1, -1); If the longest path contains L points, since there are no subsequent points that need to be matched or the distance threshold is exceeded, all matching is terminated and all matching results are stored in the path matrix Paths with a shape of [M, L, 2], where M is the number of screened paths and L is the number of liquid growth centers contained in the longest path.

8. The material growth analysis method based on image features according to claim 7, characterized in that: Step S5 filters each path in the following way: For the path matrix Paths, each element of the zeroth dimension represents a path. When a point in the path is not (-1, -1), the path is subjected to Gaussian kernel filtering to reduce the interference of outliers.

9. An electronic device, characterized in that: include: one or more processors; a storage device having one or more programs stored thereon; When the one or more programs are executed by the one or more processors, the one or more processors implement the material growth analysis method based on image features as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the program is executed by a processor, the steps in the material growth analysis method based on image features described in any one of claims 1 to 8 are implemented.

Images