Sample uniformity detection method based on convolutional neural network
By using a sample uniformity detection method based on convolutional neural networks, the subjectivity and inconsistency issues in the evaluation of silicon deposition uniformity in silicon-carbon anode materials are solved. This method enables automatic identification and quantitative analysis of sample particles, improving evaluation efficiency and accuracy, especially in the case of particle stacking, where particles can be accurately distinguished.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LIYANG TIANMU PILOT BATTERY MATERIAL TECH CO LTD
- Filing Date
- 2024-12-17
- Publication Date
- 2026-06-19
AI Technical Summary
In the existing technology, the evaluation of silicon deposition uniformity in silicon-carbon anode materials mainly relies on manual observation and experience judgment, which is time-consuming, labor-intensive and highly subjective, making it difficult to guarantee the accuracy and consistency of the evaluation results. Traditional image processing technology has difficulty in accurately distinguishing particles in complex backgrounds, especially when particles are stacked.
A sample uniformity detection method based on convolutional neural networks is adopted. The backscattered image of the sample is preprocessed, a dataset is established and divided into training, testing and validation sets, and the DeepLabV3+ semantic segmentation model and CNN algorithm are used for model training to realize intelligent identification of sample particles and gray value extraction, and calculate particle distribution and gray value uniformity.
It improves the accuracy of sample particle identification, reduces the need for manual observation and experience-based judgment, provides quantitative and objective evaluation results, avoids subjectivity and inconsistency, can accurately distinguish particles in the case of particle stacking, and realizes the scientific analysis of the component distribution of binary materials.
Smart Images
Figure CN122244628A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of material image processing technology, and in particular to a sample uniformity detection method based on convolutional neural networks. Background Technology
[0002] In existing technologies, the assessment of silicon deposition uniformity in silicon-carbon anode materials mainly relies on manual observation and experience. This method is not only time-consuming and labor-intensive but also highly subjective, making it difficult to guarantee the accuracy and consistency of the assessment results. Furthermore, traditional image processing techniques often fail to achieve ideal results when handling particle recognition tasks in complex backgrounds, especially when particles are stacked, making it difficult to accurately distinguish between them. Therefore, developing a new method that can automatically and accurately identify and analyze silicon-carbon particles is particularly important. Summary of the Invention
[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing a sample uniformity detection method based on convolutional neural networks. This detection method can achieve intelligent identification of image particles, extraction and analysis of effective gray values in backscattered images, and can quantitatively, objectively and scientifically evaluate the distribution of each component in binary materials.
[0004] To achieve the above objectives, in a first aspect, the present invention provides a sample uniformity detection method based on a convolutional neural network, the sample uniformity detection method comprising:
[0005] The backscattered image of the sample is acquired and preprocessed to obtain an image after grayscale mean correction.
[0006] A dataset is established based on the image after grayscale mean correction, and the dataset is divided into a training set, a test set, and a validation set;
[0007] The training set is input into the convolutional neural network for model training until the segmentation accuracy of the model meets the accuracy requirements, thus obtaining a trained model.
[0008] The test set is input into the trained model, and the trained model is used to predict the number and distribution of various types of sample particles in the test set, thereby detecting the distribution of various types of particles in the sample particles and the uniformity of gray-scale distribution of various types of particles.
[0009] Preferably, before acquiring the sample backscatter image, the method further includes:
[0010] Select sample materials with a preset particle size distribution for argon ion polishing;
[0011] The field emission scanning electron microscope is adjusted to a preset magnification, and a certain number of backscattered images of the same size are acquired. Each backscattered image includes a preset number of sample particles.
[0012] More preferably, the preset magnification is 1000x or 2000x, the certain quantity is 10-20 images, and the preset quantity is 100-200 pieces.
[0013] Preferably, the step of establishing a dataset based on the image after grayscale mean correction specifically includes:
[0014] Based on the image after grayscale mean correction, the sample particles are labeled according to their geometric shape, particle size, or grayscale distribution characteristics to establish a dataset.
[0015] Preferably, before inputting the training set into the convolutional neural network for model training, the method further includes:
[0016] The images in the dataset are scaled so that the size of the images is 768 pixels × 768 pixels.
[0017] Preferably, the step of inputting the training set into a convolutional neural network for model training until the segmentation accuracy of the model meets the precision requirements, thereby obtaining a trained model, specifically includes:
[0018] The training set is input into a convolutional neural network. The convolutional neural network pre-trained with the DeepLabV3+ semantic segmentation model is used as the basic architecture to build an initial deep learning network model for image segmentation and to predict the pixel mask image of the sample particle category.
[0019] Based on the pixel mask image, the loss function is calculated using the backpropagation algorithm, and the training hyperparameters are adjusted using the Adam optimizer until the loss function converges or reaches a predetermined training threshold, thus obtaining the pre-trained model.
[0020] The validation set is input into the pre-trained model to evaluate the segmentation accuracy of the model until the segmentation accuracy meets the preset requirements, thus obtaining a trained model.
[0021] Preferably, the accuracy evaluation parameters include precision and recall.
[0022] Preferably, the tuning of training hyperparameters using the Adam optimizer specifically includes:
[0023] The initial learning rate was 0.001, and the learning rate was reduced to 50% every 10 epochs, for a total of 90 epochs of training.
[0024] In a second aspect, the present invention provides a computer-readable storage medium including a program or instructions that, when run on a computer, implement the sample uniformity detection method as described in any of the first aspects above.
[0025] In a second aspect, the present invention provides a computer system including a memory and one or more processors communicatively connected to the memory;
[0026] The memory stores instructions that can be executed by the one or more processors to enable the one or more processors to implement the sample uniformity detection method as described in any of the first aspects above.
[0027] The sample uniformity detection method based on convolutional neural networks provided in this invention, by employing deep learning technology, particularly the semantic segmentation model DeepLabV3+ and CNN algorithms, can effectively improve the recognition accuracy of sample particles. Especially in the case of stacked sample particles, it can accurately distinguish particles using ResNet or U-Net as the backbone network, solving the problem that traditional image processing techniques are difficult to handle in complex backgrounds and stacked particles. In addition, it realizes automatic identification and classification of sample particles, reducing the need for manual observation and experience-based judgment, which not only improves the evaluation efficiency but also avoids the subjectivity and inconsistency caused by human factors. By performing grayscale value analysis and feature extraction on sample particles in backscattered images, the distribution of each component in binary materials can be quantitatively evaluated. That is, the particle distribution and grayscale distribution uniformity in the material can be evaluated by calculating the grayscale mean and standard deviation of each particle, thereby providing more objective and scientific analysis results. Attached Figure Description
[0028] Figure 1 This is a flowchart of a sample uniformity detection method based on a convolutional neural network provided in an embodiment of the present invention. Detailed Implementation
[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.
[0030] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0031] The sample uniformity detection method based on convolutional neural networks provided in this invention applies image processing and analysis technology to the field of materials analysis, changing the current situation that mainly relies on manual observation and experience judgment, and realizing the quantitative analysis of each component of binary materials.
[0032] Figure 1 The flowchart below shows the sample uniformity detection method based on convolutional neural networks provided in the embodiments of the present invention. Figure 1 The technical solution of the present invention will be described with reference to specific embodiments.
[0033] This invention provides a sample uniformity detection method based on a convolutional neural network, which mainly includes the following: Figure 1 The steps shown are as follows:
[0034] Step 110: Obtain the backscattered image of the sample and perform preprocessing to obtain the image after grayscale mean correction;
[0035] Specifically, backscattered images can be obtained by randomly acquiring sample materials using a field emission scanning electron microscope, and they reflect the morphology and composition information of the sample surface.
[0036] The preprocessing specifically involves converting the format of the backscattered image and normalizing its grayscale values to ensure that the average grayscale value of the backscattered image is a preset grayscale value. The preset grayscale value can be 90. The format conversion transforms the color information of each pixel in the backscattered image into a single grayscale value, i.e., converting it to an 8-bit grayscale image. This simplifies the image processing and ensures that images captured under different shooting conditions maintain consistent grayscale distribution characteristics.
[0037] Before performing step 110, the detection method also includes:
[0038] First, sample materials with a preset particle size distribution are selected and subjected to argon ion polishing.
[0039] Specifically, the preset particle size distribution can be D50 = 3μm-12μm. Argon ion polishing can remove contaminants and oxide layers from the sample material surface, and also improve the surface roughness of the sample material, providing a prerequisite for obtaining high-quality backscattered images.
[0040] Next, the field emission scanning electron microscope is adjusted to a preset magnification, and a certain number of backscattered images of the same size are acquired. Each backscattered image includes a preset number of sample particles.
[0041] Specifically, the preset magnification can be 1000x or 2000x. The quantity is 10-20 images, and the preset quantity is 100-200 samples, meaning that each backscattered image contains 100-200 micrometer-sized sample particles.
[0042] Step 120: Establish a dataset based on the image after grayscale mean correction, and divide the dataset into a training set, a test set, and a validation set;
[0043] Specifically, based on the image after grayscale mean correction, the sample particles are labeled according to their geometric shape features, and / or particle size, and / or grayscale distribution features, resulting in labeled image data. This labeled image data is then low-pass filtered to create a mask image, where the foreground represents the sample particles (different particles can be represented by different colors), and the background represents the interparticle gaps (which can be represented by black). The image is then converted to numpy.array format and encoded using One-Hot encoding, transforming the label image data into a numerical form suitable for machine learning algorithms—the encoded label data—thus establishing the dataset. The geometric shape features of the sample particles can specifically include spherical, irregular, and polygonal shapes. The particle size can specifically be within the following ranges: <1μm, [1μm-3μm), [3μm-5μm), [5μm-7μm), [7μm-9μm), [9μm-12μm), 12μm. The grayscale distribution characteristics of sample particles can be categorized as high grayscale (grayscale value 150–255), medium grayscale (grayscale value 80–150), and low grayscale (grayscale value 10–80). Generally, there is a certain positive proportionality between atomic number and image grayscale value. For example, when the sample particles are silicon-carbon particles, particles with high grayscale values are considered to have high silicon content, and particles with low grayscale values are considered to have high carbon content.
[0044] In this application, the training set, test set, and validation set can be randomly partitioned. Each image in the dataset has a first label. The first label can be spherical, 12μm, or high grayscale. Furthermore, within the same category, different label values can be defined for different subcategories; for example, a label value of 100 can be defined for spherical images, a label value of 200 for irregular images, and so on.
[0045] Step 130: Input the training set into the convolutional neural network to train the model until the segmentation accuracy of the model meets the accuracy requirements, and obtain the trained model.
[0046] To facilitate the training of the Convolutional Neural Network (CNN) model, the detection method further includes the following steps before performing step 130:
[0047] The images in the dataset are scaled so that each image is 768 pixels × 768 pixels.
[0048] The specific training process is as follows:
[0049] First, the training set is input into a convolutional neural network. The DeepLabV3+ semantic segmentation model is used as the basic architecture to build an initial deep learning network model for image segmentation and to predict the pixel mask image of the sample particle category.
[0050] Specifically, the pre-trained convolutional neural network can be ResNet or U-Net. ResNet can be used for classification and feature extraction, while U-Net excels at fine pixel-level segmentation and is suitable for handling the task of separating stacked particles.
[0051] The structure of the model includes:
[0052] Input layer: Receives images with a size of 768 pixels × 768 pixels;
[0053] Encoder: Using ResNet or U-Net as the backbone network, it is used to extract multi-scale features (the network's tasks include extracting the edges, geometric features, particle size features, and gray-scale distribution differences of sample particles, thereby automatically distinguishing different types of particles).
[0054] Decoder: Restores the original resolution through upsampling to generate pixel-level segmentation results;
[0055] Output layer: Pixel mask image for predicting particle type.
[0056] Specifically, the particle category is consistent with the category used for annotation. The pixel mask image has a second label. The second label can be understood as a virtual label corresponding to the first label.
[0057] Then, based on the pixel mask image, the loss function is calculated using the backpropagation algorithm, and the training hyperparameters are adjusted using the Adam optimizer until the loss function converges or reaches a predetermined threshold of training iterations, thus obtaining the pre-trained model.
[0058] Specifically, the loss function in this application can be the cross-entropy loss function. The cross-entropy loss function measures the difference between the model prediction and the true label. In this application, it calculates the difference between the first label and the second label. The value of the difference can be in the range of 0-1, preferably 0.05.
[0059] Specifically, the Adam optimizer has an initial learning rate of 0.001, which is reduced to 50% every 10 epochs, for a total of 90 epochs.
[0060] Secondly, the validation set is input into the pre-trained model to evaluate the segmentation accuracy of the model until the segmentation accuracy meets the preset requirements, thus obtaining a trained model.
[0061] Specifically, the preset requirements are defined as the segmentation effect of stacked particles and the accuracy of label value matching.
[0062] Segmentation performance is defined as 95% of the particles in the image being selected and segmented one by one. Accuracy is defined as an evaluation parameter, and its calculation includes precision and recall. Precision and recall can be directly called from the model. The formula for calculating accuracy is as follows:
[0063]
[0064] Example, not constraint: R is not less than 95%.
[0065] Step 140: Input the test set into the trained model and use the trained model to predict the number and distribution of various types of sample particles in the test set, thereby detecting the distribution of various types of particles in the sample particles and the uniformity of gray-scale distribution of various types of particles.
[0066] Specifically, the "classes" in this step refer to the categories labeled when the dataset was created. After predicting the number and distribution of sample particles of each class in the test set, the mean and standard deviation of grayscale for each particle can be calculated using ResNet or U-Net, or the area of silicon-carbon particles with various grayscale distribution characteristics can be statistically analyzed.
[0067] The sample uniformity detection method based on convolutional neural networks provided in this invention, by employing deep learning technology, particularly the semantic segmentation model DeepLabV3+ and CNN algorithms, can effectively improve the recognition accuracy of sample particles. Especially in the case of stacked sample particles, it can accurately distinguish particles using ResNet or U-Net as the backbone network, solving the problem that traditional image processing techniques are difficult to handle in complex backgrounds and stacked particles. In addition, it realizes automatic identification and classification of sample particles, reducing the need for manual observation and experience-based judgment, which not only improves the evaluation efficiency but also avoids the subjectivity and inconsistency caused by human factors. By performing grayscale value analysis and feature extraction on sample particles in backscattered images, the distribution of each component in binary materials can be quantitatively evaluated. That is, the particle distribution and grayscale distribution uniformity in the material can be evaluated by calculating the grayscale mean and standard deviation of each particle, thereby providing more objective and scientific analysis results.
[0068] In summary, this invention uses image processing, intelligent recognition, and grayscale image analysis technology to quantitatively determine the deposition uniformity of sample materials and the distribution of each component in binary materials. It has the advantages of high recognition accuracy, high degree of automation, strong quantitative analysis capability, and wide application range.
[0069] This invention also provides a computer-readable storage medium, including a program or instructions, which, when run on a computer, implement the sample uniformity detection method as described in any of the preceding embodiments.
[0070] This invention also provides a computer system, including a memory and one or more processors communicatively connected to the memory;
[0071] The memory stores instructions that can be executed by the one or more processors to enable the one or more processors to implement the sample uniformity detection method as described above.
[0072] The sample uniformity detection method based on convolutional neural networks provided by this invention can be applied to the detection of binary materials in energy storage devices such as supercapacitors, lithium-ion batteries, sodium-ion batteries, and dye-sensitized batteries, such as graphene composite materials.
[0073] To better understand the technical solution provided by the present invention, the following uses several specific examples to illustrate the specific process of detecting silicon-carbon particulate materials using the method provided in the above embodiments of the present invention.
[0074] Example 1
[0075] The first step is to select silicon-carbon particles with a preset particle size distribution D50 = 9 μm and perform argon ion polishing.
[0076] The second step is to adjust the field emission scanning electron microscope to 1000x, acquire 10 backscattered images of the same size, and save them. Each backscattered image contains approximately 200-300 silicon-carbon particles.
[0077] The third step is to convert the format of the backscattered image and normalize its grayscale value so that the average grayscale value of the backscattered image is 90.
[0078] The fourth step involves labeling the image after grayscale mean correction according to the geometric characteristics of the silicon carbon particles (spherical, irregular, polygonal), resulting in labeled image data. This labeled image data is then used to create a mask image through low-pass filtering. The foreground represents the silicon carbon particles, with different particles represented by different colors. The background represents the gaps between particles, which can be represented by black. The image is then converted to NumPy array format and encoded using One-Hot encoding, transforming the labeled image data into a numerical form suitable for machine learning algorithms—the encoded label data—thus establishing the dataset. The dataset is then divided into training, testing, and validation sets in a 4:4:2 ratio.
[0079] The fifth step is to scale the images in the dataset so that each image is 768 pixels × 768 pixels.
[0080] The sixth step involves inputting the images from the training set into a pre-trained ResNet convolutional neural network for feature extraction and classification, primarily focusing on extracting the edge and shape features of the silicon carbide particles. First, modify the test image path and set the saved model parameter path in `evaute.py`, then run `evaute.py`. The semantic segmentation model DeepLabV3+ will predict and output pixel mask images of the silicon carbide particles.
[0081] The seventh step involves calculating the loss function using the backpropagation algorithm and adjusting the training hyperparameters using the Adam optimizer until the loss function converges or reaches a predetermined threshold for the number of training iterations, thus obtaining the pre-trained model.
[0082] The eighth step is to input the validation set into the pre-trained model and evaluate the segmentation accuracy of the model until the segmentation accuracy reaches 95.7%, thus obtaining the trained model.
[0083] The ninth step involves inputting the test set into the trained model, using the trained model to predict the number and distribution of silicon-carbon particles of various geometric shapes in the test set, and calculating the gray mean and standard deviation of each particle using ResNet, thereby detecting the distribution of silicon-carbon particles of various geometric shapes and the uniformity of gray distribution of particles of various geometric shapes.
[0084] Experimental results show that the average gray value of silicon-carbon particles with three geometric shapes is: spherical ≈ irregular shape > polygonal shape. This indicates that the amount of silicon deposited by porous carbon with different geometric shapes is different. The analysis is that the pore volume or activated pore structure of porous carbon particles with different geometric shapes are different, which leads to the difference in the amount of silicon adsorbed by porous carbon in the same batch.
[0085] The standard deviation of grayscale values for silicon-carbon particles of three geometric shapes was: polygonal > irregular > spherical, indicating that spherical particles exhibited better deposition uniformity while polygonal particles showed poorer uniformity. This analysis provides both qualitative and quantitative insights for optimizing the deposition uniformity of silicon-carbon materials.
[0086] Example 2
[0087] Steps one through three are the same as in Example 1.
[0088] The fourth step involves labeling the image after grayscale mean correction according to the particle size of silicon carbon: <1μm, [1μm-3μm), [3μm-5μm), [5μm-7μm), [7μm-9μm), [9μm-12μm), 12μm. This results in labeled image data. The labeled image data is then low-pass filtered to create a mask image. The foreground represents the silicon carbon particles (different particles can be represented by different colors), and the background represents the gaps between particles (which can be represented by black). The image is then converted to numpy.array format and encoded using One-Hot encoding, transforming the labeled image data into a numerical form suitable for machine learning algorithms—the encoded label data—thus establishing the dataset. The dataset is then divided into training, testing, and validation sets in a 4:4:2 ratio.
[0089] The fifth step is to scale the images in the dataset so that each image is 768 pixels × 768 pixels.
[0090] The sixth step involves inputting the images from the training set into a pre-trained ResNet convolutional neural network for feature extraction and classification, primarily focusing on extracting the edge and shape features of the silicon carbide particles. First, modify the test image path and set the saved model parameter path in `evaute.py`, then run `evaute.py`. The semantic segmentation model DeepLabV3+ will predict and output pixel mask images of the silicon carbide particles.
[0091] The seventh step involves calculating the loss function using the backpropagation algorithm and adjusting the training hyperparameters using the Adam optimizer until the loss function converges or reaches a predetermined threshold for the number of training iterations, thus obtaining the pre-trained model.
[0092] The eighth step is to input the validation set into the pre-trained model and evaluate the segmentation accuracy of the model until the segmentation accuracy reaches 95.5%, thus obtaining the trained model.
[0093] The ninth step involves inputting the test set into the trained model, using the trained model to predict the number and distribution of silicon-carbon particles of various sizes in the test set, and calculating the mean and standard deviation of grayscale for each particle using ResNet, thereby detecting the particle distribution of various sizes in silicon-carbon particles and the uniformity of grayscale distribution of particles of various sizes.
[0094] Experimental results show that the gray standard deviation is the highest for particles with a diameter <1μm, indicating that the deposition uniformity of porous carbon with a diameter <1μm is poor.
[0095] Example 3
[0096] Steps one through three are the same as in Example 1.
[0097] The fourth step involves labeling the image after grayscale mean correction according to the grayscale distribution characteristics of silicon carbon particles: high grayscale, low grayscale, and medium grayscale. This results in labeled image data. The labeled image data is then used to create a mask image through low-pass filtering. The foreground represents the silicon carbon particles, with different particles represented by different colors. The background represents the gaps between particles and can be represented by black. The image is then converted to NumPy array format and encoded using One-Hot encoding, transforming the labeled image data into a numerical form suitable for machine learning algorithms—the encoded label data—thus establishing the dataset. The dataset is then divided into training, testing, and validation sets in a 4:4:2 ratio.
[0098] The fifth step is to scale the images in the dataset so that each image is 768 pixels × 768 pixels.
[0099] The sixth step involves inputting the images from the training set into a pre-trained ResNet convolutional neural network for feature extraction and classification, primarily focusing on extracting the edge and shape features of the silicon carbide particles. First, modify the test image path and set the saved model parameter path in `evaute.py`, then run `evaute.py`. The semantic segmentation model DeepLabV3+ will predict and output pixel mask images of the silicon carbide particles.
[0100] The seventh step involves calculating the loss function using the backpropagation algorithm and adjusting the training hyperparameters using the Adam optimizer until the loss function converges or reaches a predetermined threshold for the number of training iterations, thus obtaining the pre-trained model.
[0101] The eighth step is to input the validation set into the pre-trained model and evaluate the segmentation accuracy of the model until the segmentation accuracy reaches 96.3%, thus obtaining the trained model.
[0102] The ninth step involves inputting the test set into the trained model, using the trained model to predict the quantity and distribution of various types of silicon-carbon particles corresponding to the test set, and using ResNet to statistically analyze the area of silicon-carbon particles with various gray-scale distribution characteristics.
[0103] Experimental results show that the area of silicon-carbon particles with different gray scale distribution characteristics varies greatly. Among them, the area ratio of low gray scale (high carbon) particles is 20%, and the area ratio of high gray scale (high silicon) particles is 10%, indicating that the deposition uniformity of silicon-carbon particles is generally poor.
[0104] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0105] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0106] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A sample uniformity detection method based on a convolutional neural network, characterized in that, The sample uniformity detection method includes: The backscattered image of the sample is acquired and preprocessed to obtain an image after grayscale mean correction. A dataset is established based on the image after grayscale mean correction, and the dataset is divided into a training set, a test set, and a validation set; The training set is input into the convolutional neural network for model training until the segmentation accuracy of the model meets the accuracy requirements, thus obtaining a trained model. The test set is input into the trained model, and the trained model is used to predict the number and distribution of various types of sample particles in the test set, thereby detecting the distribution of various types of particles in the sample particles and the uniformity of gray-scale distribution of various types of particles.
2. The sample uniformity detection method according to claim 1, characterized by, Before acquiring the sample backscatter image, the method further includes: Select sample materials with a preset particle size distribution for argon ion polishing; The field emission scanning electron microscope is adjusted to a preset magnification, and a certain number of backscattered images of the same size are acquired. Each backscattered image includes a preset number of sample particles.
3. The sample uniformity detection method according to claim 2, characterized by, The preset magnification is 1000x or 2000x, the certain quantity is 10-20 images, and the preset quantity is 100-200 pieces.
4. The sample uniformity detection method according to claim 1, characterized by, The step of establishing a dataset based on the image after grayscale mean correction specifically includes: Based on the image after grayscale mean correction, the sample particles are labeled according to their geometric shape characteristics, and / or particle size, and / or grayscale distribution characteristics, thereby establishing a dataset.
5. The sample uniformity detection method according to claim 1, characterized by, Before inputting the training set into the convolutional neural network for model training, the method further includes: The images in the dataset are scaled so that the size of the images is 768 pixels × 768 pixels.
6. The sample uniformity detection method according to claim 1, characterized by, The step of inputting the training set into a convolutional neural network for model training until the segmentation accuracy of the model meets the precision requirements, thereby obtaining a trained model, specifically includes: The training set is input into a convolutional neural network. The convolutional neural network pre-trained with the DeepLabV3+ semantic segmentation model is used as the basic architecture to build an initial deep learning network model for image segmentation and to predict the pixel mask image of the sample particle category. Based on the pixel mask image, the loss function is calculated using the backpropagation algorithm, and the training hyperparameters are adjusted using the Adam optimizer until the loss function converges or reaches a predetermined training threshold, thus obtaining the pre-trained model. The validation set is input into the pre-trained model to evaluate the segmentation accuracy of the model until the segmentation accuracy meets the preset requirements, thus obtaining a trained model.
7. The sample uniformity detection method according to claim 1, characterized by, The accuracy evaluation parameters include precision and recall.
8. The sample uniformity detection method according to claim 1, characterized by, The tuning of training hyperparameters using the Adam optimizer specifically includes: The initial learning rate was 0.001, and the learning rate was reduced to 50% every 10 epochs, for a total of 90 epochs of training.
9. A computer-readable storage medium, characterized in that, Includes a program or instructions that, when run on a computer, implement the sample uniformity detection method as described in any one of claims 1 to 8.
10. A computer system, characterized by Includes a memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors, which, when executed by the one or more processors, enable the one or more processors to implement the sample uniformity detection method as described in any one of claims 1 to 8.