Macroscopic imaging-oriented multi-feature fusion visual evaluation method and system for carbon nanotubes

Abstract

This invention discloses a multi-feature fusion visual evaluation method and system for carbon nanotubes in macroscopic imaging, belonging to the field of nanomaterial performance testing technology. The method uses the central hole as an anchor point to locate and generate a solid preliminary region. Then, it extracts the outer contamination ring using geometrical corrosion difference, while simultaneously using grayscale content adaptively to extract the real sample region. This gradually eliminates the triple interference of background mixing, edge contamination, and residual abnormal pixels, ensuring that the final quality score is based on the objectively cleaned real sample region, significantly improving the accuracy and anti-interference capability of macroscopic quality evaluation. Furthermore, based on the aforementioned three parameters, multiple macroscopic visual features are calculated and mapped to individual scores. A weighted fusion is then used to calculate the comprehensive quality score and perform grade sorting, thereby enabling the numbering, quality scoring, and grade sorting of multiple independent carbon nanotube samples arranged on a wafer, facilitating automated sorting on the production line.

Description

Technical Field

[0001] This invention relates to a multi-feature fusion visual evaluation method and system for carbon nanotubes for macroscopic imaging, belonging to the field of nanomaterial performance testing technology. Background Technology

[0002] Carbon nanotubes (CNTs) have become key materials for advanced composite materials, new energy batteries, and semiconductor devices due to their excellent mechanical, electrical, and thermal properties. In their industrial-scale preparation, macroscopic yield is a core indicator for measuring process stability, controlling production costs, and evaluating batch quality.

[0003] Traditional quality assessment relies on microscopic characterization methods such as scanning electron microscopy (SEM) and transmission electron microscopy (TEM), which have the following drawbacks: (1) insufficient sample representativeness, only micrometer-level areas can be observed each time, making it difficult to reflect the overall quality of macroscopic samples above the centimeter level; (2) long sample preparation and detection cycles and high costs; (3) unable to meet the urgent need for rapid, non-destructive, online macroscopic yield assessment and sorting of large batches of samples in the production site. Therefore, there is an urgent need for a method and system that is geared towards macroscopic imaging scenarios and can perform automated quality assessment and grade sorting of batches of carbon nanotube samples on wafers.

[0004] In the prior art, patent CN117788843A discloses a carbon nanotube image processing method based on a neural network algorithm. This scheme uses a neural network to perform pixel-level enhancement processing on carbon nanotube images. Its core technology lies in performing general image optimization operations such as denoising and contrast enhancement on the input image through a trained network model, and outputting an image with improved visual effects. This scheme has the following shortcomings: (1) It focuses on pixel-level optimization in the image preprocessing stage, solving the problem of "how to make carbon nanotube images clearer", rather than the problem of "how to evaluate the quality of carbon nanotubes from the image"; (2) It does not involve how to extract visual features related to macroscopic yield from the processed image, such as the overall blackness of the sample (reflecting the amount of carbon nanotube deposition), coverage (reflecting the filling integrity), spatial uniformity (reflecting the deposition consistency), etc.; (3) It does not construct a quantitative model or hierarchical decision system that associates the above visual features with yield evaluation, and cannot directly output the quality score and grade of each sample based on the image.

[0005] Patent CN120997211A discloses a quantitative method for characterizing the dispersion of CNTs in resin-based composite materials. This method employs the following steps: Gamma enhancement of a microscopic two-dimensional distribution image, Sobel edge extraction, binarization, connected component identification, dividing the target region into rectangular micro-element grids, calculating the CNT area proportion of each micro-element, and using the variance and standard deviation of the area proportion vector as quantitative indicators of dispersion. Therefore, the essence of this scheme is "single-step segmentation followed by uniform grid statistics," that is, first dividing the image into CNT foreground and resin background through binarization, and then statistically analyzing the area proportion distribution within each micro-element on a uniformly divided grid. This method is difficult to apply to carbon nanotube quality detection based on macroscopic imaging, and has the following key shortcomings: (1) It is geared towards microscopic imaging scenarios (SEM / high-resolution microscopic images). There is a discernible grayscale difference between the target carbon nanotubes (CNTs) and the background (resin). Sobel edge detection plus single-step binarization is sufficient to complete the separation. However, in the macroscopic imaging scenarios that carbon nanotubes (CNTs) are geared towards, the wafer substrate and the CNTs sample are both dark colors, and their grayscale histograms overlap significantly. Simple binarization cannot accurately separate the sample area from the wafer background. (1) Forced use will result in a large number of background pixels being mixed into the sample or sample pixels being mistakenly deleted; (2) It uses a uniform grid division to calculate a single statistic (variance / standard deviation), without performing any structured regional semantic extraction on the image. Similarly, in macroscopic imaging, there are differences in spatial structure within each sample disk. Since its center is the CNT body (dark area) and the edge is a contamination transition zone (bright area) caused by wafer boundary effect and light attenuation, the two contribute completely differently to the "area ratio" statistics: edge contamination pixels should not be included in the CNT coverage area. Therefore, the uniform grid division of the above method cannot distinguish between these two semantic regions, resulting in the statistical results being severely interfered with by edge contamination; (3) Its output is a global dispersion variance value, which can only reflect "whether the CNT distribution in the whole image is uniform". It does not have the ability to number, score and sort multiple independent samples arranged on the wafer one by one, and cannot be directly used for automated sorting on the production line.

[0006] In summary, existing technologies for carbon nanotube samples in macroscopic imaging, batch sampling, and anti-interference evaluation scenarios have the following unresolved challenges: (1) Gray-scale overlap between the wafer substrate and CNTs samples makes it impossible to accurately separate sample regions using conventional single-step thresholding methods; (2) Contamination transition zones exist at the sample edges in macroscopic imaging, which will seriously interfere with the accuracy of quality evaluation after being mixed into statistics; (3) The technology does not have the capability to number, score, and sort multiple independent carbon nanotube samples arranged on a wafer one by one, and cannot be directly used for automated sorting on production lines. Summary of the Invention

[0007] To address the problems of grayscale mixing between the wafer substrate and CNTs samples, contamination transition zones at sample edges, and severe interference with statistical analysis of carbon nanotube samples in macroscopic imaging, batch sampling, and anti-interference evaluation scenarios, as well as the lack of quality scoring and grading capabilities, this invention provides a multi-feature fusion visual evaluation method and system for carbon nanotubes in macroscopic imaging. The technical solution is as follows: In a first aspect, this application provides a visual evaluation method for carbon nanotubes using multi-feature fusion for macroscopic imaging, the method comprising: Step S1: Obtain a top view image of a sample containing multiple carbon nanotubes mounted on a wafer under standard illumination conditions, and perform grayscale processing on the top view image. Step S2: Locate and label each sample from the grayscale top view image, using the center hole as the anchor point, to determine the solid preliminary sample area corresponding to each sample. ; Step S3: For each preliminary sample area Geometric corrosion difference was used to extract the peripheral contaminated area located on its periphery. ; Step S4: Based on the preliminary sample area The grayscale content within is used to adaptively extract the real sample region. (Area covered by the carbon nanotube body); Step S5: Based on the preliminary sample area , edge pollution area Real sample area For each sample, multiple macroscopic visual features are calculated; Step S6: Map multiple macroscopic visual feature values ​​to individual scores, calculate the comprehensive quality score through weighted fusion, and perform grade sorting.

[0008] Furthermore, step 2 includes: Step S21: Sample localization, threshold segmentation is performed on the grayscale image of the carbon nanotube sample to extract candidate central pores. Then, DBSCAN density clustering is used to divide the image into R×C grid sub-regions, each sub-region corresponding to one sample, and assigned numbers i=1,2,…,N (N=R×C) from top to bottom and from left to right. Step S22: Preliminary sample region extraction, Gaussian filtering is applied to the sub-regions, and threshold segmentation is then performed. The obtained segmentation thresholds are binarized, and morphological opening and closing operations are performed to obtain the final sample region. , and then from Extract all connected components from the anchor point. The nearest and sufficiently large connected region is used to extract its outermost contour and fill it with solid fill, thus obtaining the preliminary sample region. .

[0009] Furthermore, in step S21, the extraction of candidate center holes... Then, DBSCAN density clustering is used to perform clustering on candidate central holes. Calculate the area of ​​each connected component in the given data. (Number of pixels) and perimeter (Number of pixels), connected components that meet the threshold condition are retained. The centroid coordinates of the retained connected components are calculated to obtain... spatial anchor points DBSCAN density clustering is performed on all anchor points to divide the image into R×C grid sub-regions. Each sub-region corresponds to one sample and is assigned numbers i=1,2,…,N (N=R×C) from top to bottom and from left to right.

[0010] Furthermore, in step S21, the preservation of connected components that satisfy the threshold condition includes: area. satisfy and roundness The connected components are preserved, where , It is an area threshold. Roundness threshold.

[0011] Furthermore, step 3 includes: using elliptical structural elements. (Both the major and minor axes are) For the preliminary sample area Morphological etching was performed, followed by extraction of annular candidate contamination zones. ,exist Traversing the area, the width of the edge contamination zone is estimated based on the grayscale profile inside the initial sample region boundary, and then mapped to the corrosion scale. Take the final edge pollution area as ,in, Represents the morphological opening operation. It is a small-scale elliptical structural element.

[0012] Furthermore, step 4 includes: in the preliminary sample area Dark pixels (carbon nanotubes appear dark in the image) are extracted as candidate regions. For candidate regions Perform morphological closing and opening operations to obtain ,from Extract all connected components, and retain only the connected component with the largest area as the real sample region. .

[0013] Furthermore, in step 5, multiple macroscopic visual features include blackness features, coverage features, spatial consistency features, and pollution level features.

[0014] Furthermore, in step 6, each macroscopic visual feature is normalized and mapped, and the scores of the four features are weighted and summed to obtain a comprehensive quality score. Based on the grading rules, the corresponding grade evaluation is obtained.

[0015] Secondly, this application provides a multi-feature fusion visual evaluation system for carbon nanotubes for macroscopic imaging. Based on the aforementioned multi-feature fusion visual evaluation method for carbon nanotubes for macroscopic imaging, the system includes: A standard imaging unit is used to acquire a top-view image of multiple carbon nanotube samples carried on a wafer under standard illumination conditions. The image processing and control unit is communicatively connected to the standard imaging unit. It has a built-in processor and memory, with a computer program stored in the memory. When the processor executes this program, it implements all steps of the anti-interference carbon nanotube multi-feature fusion visual evaluation method for macroscopic imaging and outputs a comprehensive quality score. With sorting instructions; The sorting execution unit is communicatively connected to the image processing and control unit, receives sorting instructions, and performs quality sorting operations on the samples.

[0016] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned carbon nanotube multi-feature fusion visual evaluation method.

[0017] Fourthly, this application provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the aforementioned carbon nanotube multi-feature fusion visual evaluation method.

[0018] The beneficial effects of this invention are: By processing the top-view image of the grayscale carbon nanotube sample, a solid preliminary region is first generated using the central pore as an anchor point. Then, the peripheral contamination ring is extracted using geometric corrosion difference. Simultaneously, it adaptively extracts the real sample region using grayscale content. By gradually eliminating the triple interference of background intrusion, edge contamination, and residual abnormal pixels, the final quality score is based on the real sample area after objective purification. This significantly improves the accuracy and anti-interference ability of macroscopic quality assessment; at the same time, based on the aforementioned three parameters (preliminary sample area)... , edge pollution area Real sample area The process calculates multiple macroscopic visual features, maps these macroscopic visual feature values ​​to individual scores, calculates a comprehensive quality score through weighted fusion, and performs grade sorting. This enables the numbering, quality scoring, and grade sorting of multiple independent carbon nanotube samples arranged on a wafer, facilitating automated sorting on the production line. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of a hardware structure implementation according to an embodiment of the present invention; Figure 2 This is the preliminary sample area provided in Embodiment 2 of the present invention. The result image shows a total of 20 solid white disks; Figure 3 This invention provides an edge contamination area in Embodiment 2. The resulting image shows that the contamination zone is located only on the periphery of each disk; Figure 4 This is the second embodiment of the present invention, which provides a real sample area. Resulting image; Figure 5 This invention provides a comprehensive quality score in Embodiment 2. Sorted bar chart. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0022] See also Figure 1 As shown, the hardware facilities used in this invention for acquiring image information are as follows: a sealed imaging chamber (101) is constructed. The top of the imaging chamber (101) is equipped with a uniform illumination source (102) and an industrial camera (103). A gray background plate is laid at the bottom of the imaging chamber. The wafer (104) carrying the carbon nanotube sample is placed at a designated position on the background plate. The uniform illumination source (102) ensures that the sample surface is uniformly illuminated without shadows. The industrial camera (103) acquires a top view image of the sample under standard illumination conditions, providing a consistent and reliable image source for subsequent analysis.

[0023] Example 1 This embodiment provides a multi-feature fusion visual evaluation method for carbon nanotubes for macroscopic imaging, the method comprising: Step S1: Obtain a top-view image of a sample containing multiple carbon nanotubes mounted on a wafer, acquired under standard illumination conditions. Let the top-view image, after grayscale processing, be denoted as... , The smaller the value, the darker the pixel. The image width is denoted as . Altitude is recorded as .

[0024] Step S2: From the grayscale top view image The individual carbon nanotube samples were located and labeled, with the central pore serving as the anchor point for positioning, thus determining the corresponding solid preliminary sample area for each sample. .

[0025] This step specifically includes: Step S21: Sample localization. Due to the presence of bright pores in the central region of the carbon nanotube sample (the central pore of the carbon nanotube), the grayscale top-view image is used. Threshold segmentation is performed to extract candidate center holes: , in, High brightness threshold (this value can be selected) ), for the central hole candidate Calculate the area of ​​each connected component in the given data. (Number of pixels) and perimeter (Number of pixels).

[0026] The connected components that satisfy the threshold condition are retained, that is, those that satisfy the threshold condition are retained. and roundness Connected components (in this embodiment) , , ), including roundness ,in Representing a perfect circle. Calculate the centroid coordinates for each preserved connected component, obtaining... Spatial anchor points: , in, For all anchor points DBSCAN density clustering is performed on the coordinates: , in, for directional neighborhood radius, This represents the minimum number of samples required to form a cluster. The number of valid clusters in the sample is the number of rows. Anchor points within each cluster The mean of the coordinates is used as the row center. ( ), each row according to Sort in ascending order. Then sort the... Inline anchors DBSCAN clustering of coordinates: , in, , For adjacent anchor points within each row Set of coordinate differences Scaling factor ( The number of effective clusters in each row is equal to the number of columns. If the number of valid column clusters obtained from all rows is consistent, then this consistent number of valid clusters is determined as the column number C. The columns are then... Coordinate mean as column center ( Row and column boundaries are the arithmetic midpoints of adjacent centers: , Among them, the boundary , , , Therefore, the image is divided into Each grid sub-region corresponds to one sample, and the sub-regions are numbered from top to bottom and left to right. ( ).

[0027] Step S22: Preliminary sample region extraction, for the first For each sample, a grid sub-region is given, and its grayscale image is... The size is ( For height, (Width in pixels). First, perform Gaussian filtering: , in, Gaussian filter kernel (kernel size in this embodiment) Pixels, standard deviation It is automatically determined by OpenCV. ), This represents a two-dimensional convolution operation. In the filtered image... The Otsu thresholding process is performed, and the optimal threshold is obtained by maximizing the inter-class variance. , in, And it is grayscale Pixel ratio, And it is grayscale pixel ratio, And it is the mean of the dark class. And it is the mean of the bright class. This is a normalized grayscale histogram. Segmentation threshold. ( (For offset). Binarization: , Then binarize the obtained Perform morphological closing operation (structuring element) ,size The elliptical kernel fills the tiny internal gaps, and then an opening operation (structuring element) is performed. ,size To remove isolated noise, see the following expression: , in, This indicates the closing operation (expansion followed by erosion). This indicates the opening operation (erosion followed by dilation). From... Extract all connected components from , and let the set of connected components be . Select a distance from the aforementioned spatial anchor point. The nearest connected component with a sufficiently large area: , in, For connected components The number of pixels, Minimum area threshold (in this embodiment) ), for The coordinates of the centroid, The coordinates of the spatial anchor point of the sample in step S21. Represents the Euclidean distance. For connected components. Extract the outermost contour and fill it with solid fill to obtain the preliminary sample area. : , in, This indicates that only the outermost contour is extracted (RETR_EXTERNAL mode). This indicates that the entire area inside the outline is filled with white, and a morphological closing operation (ellipse kernel) is applied after filling. By smoothing the boundaries, the internal voids caused by the central hole are eliminated, resulting in a completely solid disc-shaped region. .

[0028] In this step, center hole detection combined with row and column DBSCAN density clustering is used for sample localization. This method is independent of preset sample quantity and arrangement rules, and can adaptively detect and identify sample arrays with arbitrary row and column arrangements. Simultaneously, an outer contour filling method is used to eliminate internal voids caused by center holes, ensuring... The solid disk is designed to address the grayscale aliasing issue between the wafer and the CNTs sample.

[0029] Step S3: For each preliminary sample area Geometric corrosion difference was used to extract the peripheral contaminated area located on its periphery. Specifically, it adopts elliptical structural elements. (Both the major and minor axes are) For the preliminary sample area Morphological erosion is performed using the following expression: , in, This represents the grayscale morphological erosion operation, which transforms the structuring element... In the preliminary sample area Slide up when the structuring element Completely contained within the initial sample area Pixels corresponding to the origin of the structuring element are retained if necessary; otherwise, they are deleted. The annular candidate contamination band is the difference between the eroded and uneroded sets, as shown in the following expression: , Furthermore, the width of the edge contamination zone is estimated based on the grayscale distribution inside the preliminary sample area boundary, and the corrosion scale is adaptively determined accordingly. Let the average grayscale of each layer extending inwards from the boundary constitute the boundary grayscale profile. Then, the width of the edge contamination zone is determined based on the grayscale transition relationship between the boundary bright band and the internal main dark area. And map it to corrosion scale Therefore, take

[0030] The final edge of the contaminated area, in which, Represents the morphological opening operation. It is a small-scale elliptical structural element used to remove local burrs and small noise.

[0031] In this step, the morphological corrosion difference method is still used to extract the final edge contamination area. The method changes the determination of corrosion scale from a comparison-based selection based on truth value annotation to an adaptive determination based on the gray-scale transition features of the sample boundary. This allows for edge contamination region extraction without the need for truth value annotation during runtime. This method only captures the outer annular band of the sample, and compared to abnormal pixel detection methods based on gray-scale thresholds, it is less susceptible to interference from brightness variations within the sample, effectively addressing the issue of statistical infiltration of edge contamination transition bands.

[0032] Step S4: Based on the preliminary sample area The grayscale content within is used to adaptively extract the real sample region. (i.e., the area covered by the carbon nanotube body).

[0033] The specific operation is as follows: Dark pixels (carbon nanotubes appear dark in the image) are extracted as candidate regions: , in, Let be the grayscale threshold for the dark pixels of the i-th sample. For the initial sample region... We statistically analyze the grayscale values ​​within the range, and let the median be 1. Standard deviation is The grayscale threshold for dark pixels is then set as follows: , Where λ is a preset scaling factor. Then, the candidate region... Perform morphological closing operation (elliptic kernel) ,size (filling tiny gaps caused by local unevenness within the CNT coverage area) and opening operations (elliptic kernel) ,size (to remove isolated noise points), see the following expression: , from Extract all connected components, and retain only the connected component with the largest area as the real sample region. See the following expression: , in, for Connected components in the array. Parameters Determined through a three-dimensional mesh search.

[0034] In this step, a content-adaptive dark pixel extraction method is used to obtain the real sample area. This method determines the coverage area of ​​carbon nanotubes based on the sample's own grayscale distribution, rather than through simple geometric subtraction. This approach can adapt to grayscale differences between different samples and effectively addresses localized inhomogeneities and interference from residual abnormal pixels. Furthermore, the structured region processing chain, from "preliminary localization" to "contamination stripping" and then to "purification and extraction," makes subsequent evaluation results more reliable.

[0035] Step S5: Based on the preliminary sample area , edge pollution area Real sample area For each sample, multiple macroscopic visual features are calculated.

[0036] The aforementioned macroscopic visual features include blackness features, coverage features, spatial consistency features, and pollution level features, specifically: (a) Blackness characteristics Calculate the real sample region Average gray value of inner pixels : , , in, This represents the number of pixels in the actual sample area. (Actual sample area) The denser the internal carbon nanotube (CNT) deposition, the better. The lower, The closer it is to 1.

[0037] (b) Coverage characteristics Calculate the ratio of the actual sample area to the preliminary sample area: , in, The larger the value, the more fully the carbon nanotube material fills the sample disk.

[0038] (c) Spatial consistency characteristics : The actual sample area Divided evenly along the longitudinal and transverse directions into common Sub-regions. For the first... Sub-regions ( ), assuming it is similar to the actual sample area The intersection is (Require Greater than the preset minimum number of pixels Otherwise, the sub-region is skipped), and the blackness within the sub-region is calculated. and coverage : , , in, For the first Sub-regions and preliminary sample areas The intersection, Its pixel count. Calculate the blackness and coverage of each sub-region and compare them with the overall value ( Mean Absolute Deviation (MAD): , in, The set of valid sub-regions that satisfy the pixel count condition. (Require ,otherwise Take directly The mean absolute deviation is passed through a tolerance threshold. and Normalization and truncation to : , in, For blackness deviation tolerance, This is the tolerance for coverage deviation. (or The meaning of ) is: if the average deviation of darkness (or coverage) between sub-regions reaches or exceeds the tolerance threshold, this item is 0; the smaller the deviation, the closer this item is to 1. Spatial consistency is the equal-weighted average of the two items: , in, The closer this value is to 1, the more uniform the distribution of carbon nanotubes inside the sample, meaning the smaller the difference in blackness and coverage between sub-regions.

[0039] (d) Characteristics of pollution level Calculate the edge pollution area Area occupies the initial sample area Area ratio: , in, The larger the value, the more severe the influence of edge contamination on the sample.

[0040] In this step, the blackness feature is used to characterize the macroscopic deposition amount and density of carbon nanotubes within the actual sample area. The darker the sample, the more fully the carbon nanotubes are deposited. The coverage feature is used to characterize the effective filling integrity of carbon nanotubes within the sample area. The higher the coverage ratio, the more fully the sample is spread and the fewer the gaps. The spatial consistency feature is used to characterize the uniformity of carbon nanotube distribution at different locations within the sample. The higher the consistency, the smaller the differences in blackness and coverage between sub-regions and the more stable the internal quality. The contamination level feature is used to characterize the degree of interference of edge contamination areas on the evaluation of the main sample. The larger the contamination ratio, the more significant the impact of edge anomalies on the evaluation results.

[0041] Step S6: Map the four macroscopic visual feature values ​​from step S5 to individual scores, calculate the comprehensive quality score through weighted fusion, and perform grade sorting.

[0042] Assume the current batch has a total of For each sample, for each feature Let its set of values ​​be , , The batch-normalized mapping for individual scores is as follows: Positive features ( These correspond to the blackness feature, coverage feature, and spatial consistency feature, respectively; larger values ​​are better. , Reverse features ( (i.e., pollution level characteristic, the smaller the value, the better). , Overall quality score The weighted sum of the four scores: , in, to For the weighting coefficients, satisfying Recommended values ​​in this embodiment (Blackness) (Coverage) (consistency), (Pollution). Reverse characteristics Its characteristics vary depending on the degree of pollution. It increases and decreases, therefore when pollution worsens... It automatically lowers the score, thus deducting points.

[0043] Grading and sorting rules: , in, For the grading threshold ( In this embodiment, we take... , .

[0044] The beneficial effects of this application are as follows: (1) The core processing difficulty of this application lies in the scene where the wafer and CNT sample are both dark colors and their grayscale histograms overlap severely in macroscopic imaging. Therefore, in the initial sample area extraction, the processing method of using the central hole as a spatial anchor point, DBSCAN to automatically identify the sample arrangement, Otsu segmentation of each part, and outer contour filling is adopted to obtain a stable initial sample area. ; (2) To address the issue of confusion between edge contamination and the true CNT coverage area, i.e., the problem that bright rings at the sample edge are easily mixed into the statistics, instead of using simple grayscale anomaly detection, the contamination identification is transformed into a geometric localization problem. The outer ring band is extracted through multi-scale elliptical kernel corrosion difference, thereby obtaining the edge contamination area. ; (3) This application adopts a three-level serial mode of “positioning → purification → extraction” for the extraction of the preliminary sample area, edge contamination area and real sample area of ​​carbon nanotube sample area. There is a progressive dependency between each step, which is aimed at the complex interference environment under macroscopic imaging.

[0045] (4) After obtaining the real sample area, this application extracts four macroscopic visual features: blackness, coverage, spatial consistency and contamination level, and maps them into individual scores and then performs weighted fusion to obtain a comprehensive quality score and implement grade sorting. This can avoid the one-sidedness of evaluation caused by relying on a single gray level index or a single area index, and realize the joint characterization of carbon nanotube sample deposition intensity, coverage integrity, internal uniformity and edge interference degree. It has the beneficial effects of more comprehensive results, stronger interpretability, more intuitive batch comparison and convenient production line automated screening and quality grading.

[0046] Example 2 This embodiment uses a specific example for illustration. The carbon nanotube samples analyzed in this embodiment are derived from chemical vapor deposition (CVD) on a silicon substrate. After being formed, multiple samples with different morphologies, along with their silicon substrates, are regularly and tightly arranged and fixed on a dedicated macroscopic evaluation carrier (silicon wafer), forming a standardized macroscopic sample set. This embodiment contains... Several carbon nanotube samples, arranged as follows OK, List.

[0047] Step S1: Place the wafer carrying the sample in Figure 1 The imaging unit shown captures a top-view image containing all samples and converts it into a grayscale image. Image size in this embodiment Pixels, grayscale range The average value is approximately 196.

[0048] Step S2: Sample positioning and preliminary sample area determination .

[0049] Sub-step S21: Sample localization. Set the high brightness threshold. right Binarization is used to extract candidate bright-colored center holes. Candidate connected components are then processed by area. Pixels and roundness Filtering. Within each column Coordinates clustered by DBSCAN ( , This results in 4 lines; each line contains... Coordinates clustered by DBSCAN ( Depend on The median of the distance is determined adaptively. This results in 5 columns. Row center. for , column center for Based on this division Grid, 20 sample numbers .

[0050] Sub-step S22: Preliminary sample area Extraction. For each grid sub-region, first apply Gaussian filtering ( ,correspond After kernel) Otsu threshold segmentation ( Perform morphological closing operation again. Opening operation Elliptic kernel. Minimum area of ​​connected region. Fill the outer contour with solid fill to eliminate the central hole. Figure 2 The preliminary sample areas of all 20 samples are shown. The results showed that all were solid disks without any internal voids. The accuracy was 0.983.

[0051] Step S3: Edge Contamination Area Extraction. For each preliminary sample region. A grayscale profile is constructed inward along its boundary, and the width of the edge contamination band is estimated based on the grayscale transition relationship between the boundary bright band and the internal main dark area. The median of the bandwidth estimates obtained from all samples is taken to obtain a global edge contamination band width of 5 pixels, which is then mapped to an elliptical erosion kernel size of 11×11. The elliptical kernel is then used to etch the initial sample area. Morphological etching was performed, and the etching result was subtracted from the original preliminary sample area to obtain an annular candidate contamination zone. Then, an opening operation with a 3×3 elliptic kernel size was performed to remove local burrs and small noise points, thus obtaining the edge contamination region. . Figure 3 The results show the edge contamination areas of all 20 samples, with the contamination bands located only in the outermost annular region of each disk.

[0052] Step S4: Real Sample Area Extraction. For each preliminary sample region. In eliminating marginal contamination areas After the influence of the data, the internal grayscale values ​​are statistically analyzed, and the median is set as . Standard deviation is Then, an adaptive grayscale threshold τ_i is used. +0.5 Dark pixel candidate regions are extracted, where dark pixels correspond to the carbon nanotube main body coverage area. Then, based on the scale of the initial sample region, the sizes of the morphological closing and opening kernels are adaptively determined, and closing and opening operations are performed on the candidate regions to fill local micro-gaps and remove isolated noise. Finally, all connected components are extracted, and only the connected component with the largest area is retained as the true sample region. In this embodiment, the average value of the adaptive threshold for each sample is approximately 50.6, the minimum value is 45, and the maximum value is 56. Figure 4 The results show the actual sample area of ​​all 20 samples.

[0053] Step S5: Calculation of four macroscopic visual features. Based on the preliminary sample area. , edge pollution area Real sample area Calculate each sample Distribution of characteristic values ​​for 20 samples: (mean 0.880) (Mean 0.922) (mean 0.798) (Mean 0.054). Emissivity and pollution The small intra-batch variations (ranges of 0.028 and 0.012, respectively) indicate consistent imaging conditions and similar levels of edge contamination. Coverage and consistency The intra-batch variability was relatively large (ranges of 0.049 and 0.363, respectively), which is the main characteristic dimension for distinguishing sample quality.

[0054] Step S6: Quality scoring and grading. Intra-batch normalized mapping is used, with weights... threshold , .like Figure 5 As shown, the horizontal axis represents the sample number. The vertical axis represents the overall quality score. The black solid column is Grade A. The gray diagonally filled column is grade B. The light gray cross-line filled column is grade C ( ). Horizontal solid lines and dashed lines are marked respectively. and The rating distribution is as follows: 4 grades of A ( Samples 13, 18, 20, and 6), 3 Grade B samples ( Samples 1, 12, and 9), 13 Grade C samples ( The highest-rated sample was 13 (). It excels in all four characteristics: , , , The lowest-scoring sample 3 ( Multiple features ranked low. Figure 5 It can be observed that the blackness The highest sample 20 ( Because the coverage score was only 47.8, the overall ranking dropped to 3rd, indicating that a single feature cannot fully reflect the sample quality. This invention achieves a multi-dimensional comprehensive evaluation by weighted fusion of four features, and the evaluation results are accurate.

[0055] Example 3 A multi-feature fusion visual evaluation system for carbon nanotubes for macroscopic imaging, the system comprising: A standard imaging unit is used to acquire a top-view image of multiple carbon nanotube samples carried on a wafer under standard illumination conditions. The image processing and control unit is communicatively connected to the standard imaging unit. It has a built-in processor and memory, with a computer program stored in the memory. When the processor executes this program, it implements all steps of the anti-interference carbon nanotube multi-feature fusion visual evaluation method for macroscopic imaging and outputs a comprehensive quality score. With sorting instructions; The sorting execution unit is communicatively connected to the image processing and control unit, receives sorting instructions, and performs quality sorting operations on the samples.

[0056] Example 4 This application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned carbon nanotube multi-feature fusion visual evaluation method.

[0057] Example 5 This application provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the aforementioned carbon nanotube multi-feature fusion visual evaluation method.

[0058] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0059] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-feature fusion visual evaluation method for carbon nanotubes for macroscopic imaging, characterized in that, The method includes: Step S1: Obtain a top view image of a sample containing multiple carbon nanotubes mounted on a wafer under standard illumination conditions, and perform grayscale processing on the top view image. Step S2: Locate and label each sample from the grayscale top view image, using the center hole as the anchor point, to determine the solid preliminary sample area corresponding to each sample. ; Step S3: For each preliminary sample area Geometric corrosion difference was used to extract the peripheral contaminated area located on its periphery. ; Step S4: Based on the preliminary sample area The grayscale content within is used to adaptively extract the real sample region. ; Step S5: Based on the preliminary sample area , edge pollution area Real sample area For each sample, multiple macroscopic visual features are calculated; Step S6: Map multiple macroscopic visual feature values ​​to individual scores, calculate the comprehensive quality score through weighted fusion, and perform grade sorting.

2. The method according to claim 1, characterized in that, Step 2 includes: Step S21: Sample localization, threshold segmentation is performed on the grayscale image of the carbon nanotube sample to extract candidate central pores. Then, DBSCAN density clustering is used to divide the image into R×C grid sub-regions, each sub-region corresponding to one sample, and assigned numbers i=1,2,…,N (N=R×C) from top to bottom and from left to right. Step S22: Preliminary sample region extraction, Gaussian filtering is applied to the sub-regions, and then threshold segmentation is performed. The obtained segmentation thresholds are binarized, and morphological opening and closing operations are performed to obtain the final sample region. , and then from Extract all connected components from the anchor point. The nearest and sufficiently large connected region is used to extract its outermost contour and fill it with solid fill, thus obtaining the preliminary sample region. .

3. The method according to claim 2, characterized in that, In step S21, the extraction of candidate central holes Then, DBSCAN density clustering is used to perform clustering on candidate central holes. Calculate the area of ​​each connected component in the given data. and perimeter Connected components that meet the threshold condition are preserved. The centroid coordinates of the preserved connected components are calculated to obtain... spatial anchor points DBSCAN density clustering is performed on all anchor points to divide the image into R×C grid sub-regions. Each sub-region corresponds to one sample and is assigned numbers i=1,2,…,N (N=R×C) from top to bottom and from left to right.

4. The method according to claim 3, characterized in that, In step S21, the connected components that satisfy the threshold condition are preserved, including: area. satisfy and roundness The connected components are preserved, where , It is an area threshold. Roundness threshold.

5. The method according to claim 1, characterized in that, Step 3 includes: using elliptical structural elements. For the preliminary sample area Morphological etching was performed, followed by extraction of annular candidate contamination zones. ,exist Traversing the area, the width of the edge contamination zone is estimated based on the grayscale profile inside the initial sample region boundary, and then mapped to the corrosion scale. The final edge contamination area is: ， in, Represents the morphological opening operation. It is a small-scale elliptical structural element.

6. The method according to claim 1, characterized in that, Step 4 includes: in the preliminary sample area Extract dark pixels from within as candidate regions For candidate regions Perform morphological closing and opening operations to obtain ,from Extract all connected components, and retain only the connected component with the largest area as the real sample region. .

7. The method according to claim 1, characterized in that, In step 5, multiple macroscopic visual features include blackness features, coverage features, spatial consistency features, and pollution level features.

8. The method according to claim 1, characterized in that, In step 6, each macroscopic visual feature is normalized and mapped, and the scores of the four features are weighted and summed to obtain a comprehensive quality score. According to the grading rules, the corresponding grade evaluation is obtained.

9. A multi-feature fusion visual evaluation system for carbon nanotubes for macroscopic imaging, based on the method described in any one of claims 1 to 8, characterized in that, The system includes: A standard imaging unit is used to acquire a top-view image of multiple carbon nanotube samples carried on a wafer under standard illumination conditions. The image processing and control unit is communicatively connected to the standard imaging unit. It has a built-in processor and memory, with a computer program stored in the memory. When the processor executes this program, it implements all steps of the anti-interference carbon nanotube multi-feature fusion visual evaluation method for macroscopic imaging and outputs a comprehensive quality score. With sorting instructions; The sorting execution unit is communicatively connected to the image processing and control unit, receives sorting instructions, and performs quality sorting operations on the samples.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method of any one of claims 1 to 8.

Images