Nano powder material dispersity intelligent detection method and system based on deep learning
By combining an improved YOLOv8 model with a dynamic threshold algorithm, high-precision detection of carbon nanotube dispersion was achieved, solving the subjectivity and efficiency problems of traditional methods and meeting the real-time detection needs of industrial production.
Patent Information
- Application Number
- CN202510890985.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-30
- Publication Date
- 2025-11-21
AI Technical Summary
Existing technologies are insufficient to achieve high-precision dispersion detection of carbon nanotubes in liquid media. Traditional methods suffer from problems such as strong subjectivity, missing local detection areas, and low efficiency, failing to meet the real-time detection requirements of industrial production.
An improved YOLOv8 model combined with channel attention mechanism was used for image segmentation. A multi-dimensional dispersion evaluation system was constructed by combining it with a dynamic threshold algorithm. Images were acquired through an optical microscopy imaging system, and contrast enhancement and noise suppression were performed. The area ratio of the agglomerated region, the maximum equivalent diameter, and the distribution uniformity index were calculated. The threshold was dynamically adjusted to adapt to the concentration fluctuations of different batches of dispersion.
It achieves high-precision detection of carbon nanotube dispersion, reduces the false judgment rate, shortens the detection time, adapts to industrial online monitoring, improves detection accuracy by 11.2%, reduces the false judgment rate to 2.3%, and the detection time for a single sample is less than 3 seconds.
Smart Images
Figure CN120997828A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of nanomaterial detection, and particularly relates to an intelligent detection method and system for dispersion degree of nanometer powder material based on deep learning, which is particularly suitable for realizing quantitative evaluation of dispersion state of carbon nanotubes through image segmentation technology. BACKGROUND
[0002] Carbon nanotubes are widely used in composite materials, new energy and other fields due to their excellent mechanical, electrical and thermal properties. However, the dispersion degree of carbon nanotubes in liquid medium directly affects their performance, and the agglomeration phenomenon will lead to material interface failure, discontinuous conductive network and other problems. The current mainstream detection methods include ultraviolet-visible spectroscopy and manual microscopic observation method, but both have significant defects.
[0003] The ultraviolet absorption method can only reflect the overall dispersion concentration through absorbance, and cannot detect the spatial distribution characteristics of the local agglomeration area. The manual microscopic observation method relies on the experience of the operator and has subjective judgment deviation, and is low in efficiency, which is difficult to meet the real-time detection needs of industrialized production. The traditional image processing algorithm (such as threshold segmentation, edge detection) has poor adaptability to low-contrast microscopic images, and is easily disturbed by noise, which leads to misjudgment.
[0004] In recent years, deep learning technology has been gradually applied in the field of material characterization, but the existing scheme still has the following problems: the recall rate of the YOLO series model is insufficient in the detection of micro targets, and the boundary segmentation accuracy of nanoscale agglomerates is low; the static threshold judgment method cannot adapt to the concentration fluctuation of different batches of dispersion liquids, which leads to the disconnection between the qualified standard and the actual situation.
[0005] Therefore, it is urgent to develop an intelligent detection method that combines adaptive image segmentation and dynamic evaluation index. SUMMARY
[0006] Based on the above problems, the purpose of the present application is to provide an intelligent detection method and system for dispersion degree of nanometer powder material based on deep learning, and to apply it to the detection of dispersion degree of carbon nanotubes, to realize high-precision agglomeration area segmentation through an improved YOLOv8 model, and to construct a multi-dimensional dispersion evaluation system combined with a dynamic threshold algorithm, to solve the problems of strong subjectivity and local detection deficiency of traditional methods.
[0007] To achieve the above purpose, the present application provides an image recognition detection method for dispersion state of nanometer powder material, wherein the nanometer powder material can be carbon nanotubes or other nanometer particle shaped nanometer powder material, comprising the following steps:
[0008] (1) obtaining a digital image of the nanometer powder material dispersion liquid through an optical microscopic imaging system;
[0009] (2) performing a pre-processing operation of contrast enhancement and noise suppression on the digital image;
[0010] (3) identifying the agglomeration area using a deep learning-based image segmentation model, the model comprising a channel attention mechanism;
[0011] (4) calculating the area proportion of the agglomeration area, the maximum equivalent diameter, and the distribution uniformity index;
[0012] (5) determining the dispersion level based on a dynamic threshold algorithm, the threshold being automatically adjusted according to historical detection data.
[0013] Further, the image segmentation model in step (3) is based on the YOLO architecture, and a channel attention module is embedded in the backbone network, the channel attention module comprising a plurality of global pooling layers and a plurality of fully connected layers.
[0014] Further, the calculation method of the distribution uniformity index in step (4) is to divide the image into MxN grids, and calculate the coefficient of variation of the agglomeration area proportion of each grid, wherein M and N are positive integers greater than 5.
[0015] Further, the dynamic threshold algorithm in step (5) performs the following operations: when the consecutive K batches of detection qualification rates exceed the first preset value, the area proportion and equivalent diameter thresholds are tightened in proportion; when the qualification rate is lower than the second preset value, the initial threshold is restored; wherein K is a positive integer of 5-20, the first preset value is 90-98%, and the second preset value is 80-90%.
[0016] The present application also provides a detection system for implementing the above method, comprising:
[0017] an image acquisition module (100) comprising an optical microscope and a digital camera;
[0018] an image processing module (200) carrying a deep learning processor supporting a channel attention mechanism;
[0019] a human-computer interaction module (300) equipped with a display device and a threshold adjustment interface.
[0020] Compared with the prior art, the present application has the following advantages:
[0021] 1. Improved detection accuracy: the mAP@0.5 of the improved YOLOv8 model reaches 92.7%, which is 11.2% higher than the baseline model.
[0022] 2. Dynamic adaptability: the threshold is adjusted through historical data feedback, and the misjudgment rate is reduced to 2.3%.
[0023] 3. Industrial applicability: single sample detection time < 3 seconds, supporting online real-time monitoring. BRIEF DESCRIPTION OF DRAWINGS
[0024] Figure 1 Structure diagram of image recognition detection system for dispersion state of carbon nanotubes used in each embodiment of the present application.
[0025] Figure 2 Flowchart of image recognition detection method for dispersion state of carbon nanotubes of each embodiment of the present application.
[0026] Figure 3 Structure diagram of improved YOLOv8 model in the detection system of the present application, in which the position and connection mode of the channel attention module are marked. DETAILED DESCRIPTION
[0027] The present application will be further described below in conjunction with specific embodiments. It should be understood that the following embodiments are only used to illustrate the present application, but not to limit the scope of the present application.
[0028] The following embodiments of the present application take the detection of the dispersion state of carbon nanotubes as an example to specifically introduce the detection system and the detection method, but the following does not limit the application of the detection system and the detection method of the present application in the dispersion state detection of other nanometer powder materials.
[0029] As shown in Figure 1 The image recognition detection system provided by the present application includes an image acquisition module 100, an image processing module 200 and a man-machine interaction module 300.
[0030] The image acquisition module 100 includes an optical microscope and a digital camera. In the following embodiments of the present application, the optical microscope adopts an Olympus BX53M optical microscope; the digital camera adopts a CMOS digital camera, which has a frame rate of 30 fps and a pixel size of 2.2 μm, and the microscope is equipped with a ring light source to provide uniform illumination.
[0031] The image processing module 200 includes a deep learning processor supporting channel attention mechanism. In the following embodiments of the present application, the deep learning processor includes a computing unit of an NVIDIA Jetson AGX Xavier processor, which includes a deep learning-based image segmentation model, a built-in TensorRT acceleration engine, which is used to perform image preprocessing, deep learning model inference and dispersion quantization analysis. The deep learning-based image segmentation model adopts an improved YOLOv8 model loaded with pre-training, and a channel attention module is embedded in the backbone network, which includes a global pooling layer and two fully connected layers; as Figure 3The improved YOLOv8 model structure diagram of each embodiment of the application is shown in the figure, and the embedding position and connection mode of the channel attention module in the backbone network of the YOLOv8 model structure are indicated.
[0032] The human-computer interaction module 300 includes a display device and a threshold adjustment interface. In the following embodiments of the application, the human-computer interaction module 300 includes a touch display screen and a control panel, providing a real-time segmentation visualization interface, a threshold parameter adjustment slide bar, and a PDF format detection report generation function.
[0033] Based on the above system, the following embodiments of the application take the image recognition detection of the dispersion state of carbon nanotubes as an example to specifically introduce the image recognition intelligent detection method of the dispersion state of nanomaterials, and the detection process of each embodiment is as shown in Figure 2 The specific implementation is as follows.
[0034] Embodiment 1: Detection of aqueous carbon nanotube dispersion
[0035] Step 1, sample preparation and image acquisition, as shown in Figure 2 Step 1 in the foregoing, is obtained by an optical microscope, specifically: take an aqueous dispersion containing 0.5wt% multi-walled carbon nanotubes, after ultrasonic dispersion treatment for 30 minutes, take 100μL of the sample and place it on a glass slide, then use a cover glass to seal the sample and place it on the microscope stage. Set the microscope magnification to 400x, adjust the focusing and illumination parameters, and collect a 2048x1536 resolution RGB digital image through a CMOS camera, and save it as a PNG format.
[0036] Step 2, image preprocessing, as shown in Figure 2 Step 2 in the foregoing, mainly performs contrast enhancement and noise suppression, specifically: apply the CLAHE (Contrast Limited Adaptive Histogram Equalization) algorithm to the collected original image to enhance the image contrast, wherein the contrast limiting parameter is set to 3.0 and the block size is 8x8 pixels. Then, non-local mean filtering with a parameter of h=10 is used to eliminate shot noise and preserve the details of the agglomeration structure.
[0037] Step 3, deep learning model segmentation, as shown in Figure 2Step 3 primarily involves improving the YOLOv8 model and its embedded channel attention mechanism for cluster region identification and segmentation. Specifically, this involves loading a pre-trained improved YOLOv8 model. This model embeds the CBAM (Convolutional Block Attention Module) attention mechanism into the backbone network. The channel attention module includes two branches: global average pooling and global max pooling, generating channel weights through two fully connected layers (dimensionality reduction ratio of 16). The model uses the Focal Loss function (parameters γ = 2.0, α = 0.25) instead of the standard cross-entropy loss to alleviate the sample imbalance problem. The pre-processed image is input into the model, and a cluster region segmentation mask is output with a confidence threshold of 0.25.
[0038] Step 4, as follows Figure 2 In step 4, the three indicators of carbon nanotube dispersion are quantified, specifically: the dispersion quantification analysis is based on the segmentation mask to calculate the following three quantitative indicators: (1) area ratio (R area ): Calculate the ratio of the number of pixels with a value of 1 in the mask to the total number of pixels, and the result is 3.8%; (2) Maximum equivalent diameter (D max ): Connectivity analysis was used to extract the connected region with the largest area, and its equivalent circle diameter was calculated to be 16.7 μm; (3) Uniformity index (U index The image was uniformly divided into an 8×8 grid, and the coefficient of variation of the proportion of clustered area in the 64 sub-regions was calculated, with a result of 12.3%.
[0039] Step 5, as follows Figure 2 Step 5, dynamic threshold determination, is mainly based on historical data and dynamically adjusts the threshold. Specifically, the initial system threshold is set to: R area ≤5%, D max ≤20μm, U index ≤15%. After applying the dynamic threshold algorithm, all three indicators of this sample meet the requirements, and it is classified as "well dispersed". The system records this result and stores it in the historical database. If the pass rate of 10 consecutive batches of samples reaches or exceeds 95%, the system will automatically tighten the threshold to R. area ≤3%, D max ≤15μm; if the pass rate drops below 85% in subsequent tests, the initial threshold setting will be restored and an alert signal will be triggered.
[0040] Example 2: Detection of carbon nanotube dispersions in organic solvents
[0041] Step 1: Sample preparation and image acquisition Take a single-walled carbon nanotube dispersion liquid containing 0.3wt% in N-methyl pyrrolidone (NMP), after ultrasonic treatment for 45 minutes, prepare a microscope sample. Set the microscope magnification to 600x, and collect an RGB image with a resolution of 2048x1536.
[0042] Step 2-5: Perform the corresponding operations according to steps 2-5 in Example 1, but in the uniformity index calculation of step 4, divide the image into a 10x10 grid to improve spatial resolution. The results of the three indicators of this sample are: R area = 6.2%, D max = 22.5 μm, U index = 18.7%, all exceeding the initial threshold, the system determines the "poor dispersion" level and gives the treatment suggestion "suggest extending the ultrasonic time or adjusting the amount of dispersant".
[0043] Experimental verification shows that the method of the present application has significant advantages over traditional methods. On 100 test samples, the consistency rate of the method of the present application with the results of artificial expert determination reaches 94.5%, which is 23.7 percentage points higher than the traditional threshold segmentation method; the average detection time of a single sample is 2.8 seconds, meeting the real-time monitoring needs of industrial production; the mAP@0.5 of the improved YOLOv8 model reaches 92.7%, which is 11.2% higher than the baseline model.
Claims
1. A deep learning-based intelligent detection method for dispersion of nano-powder material, characterized in that The detection method is based on image recognition, comprising the following steps: (1) obtaining a digital image of the nanometer powder material dispersion liquid by an optical microscopic imaging system; (2) performing a pretreatment operation of contrast enhancement and noise suppression on the digital image; (3) identifying the agglomeration area using a deep learning-based image segmentation model, the model containing a multi-channel attention mechanism; (4) calculating the agglomeration area proportion, maximum equivalent diameter and distribution uniformity index; (5) determining the dispersion level based on a dynamic threshold algorithm, the threshold being automatically adjusted according to historical detection data.
2. The method of claim 1, wherein: The nanometer powder material is carbon nanotubes.
3. The method of claim 1, wherein The image segmentation model in step (3) is based on the YOLO architecture, with a channel attention module embedded in the backbone network, the multi-channel attention module including multiple global pooling layers and multiple fully connected layers.
4. The method of claim 1, wherein In step (4), the calculation method of the distribution uniformity index is to divide the image into an M×N grid and calculate the coefficient of variation of the agglomeration area proportion of each grid, where M and N are positive integers greater than 5.
5. The method of claim 1, wherein In step (5), the dynamic threshold algorithm performs the following operations: when the consecutive K batches of detection qualification rate exceeds the first preset value, the area proportion and equivalent diameter threshold are tightened by a certain proportion; when the qualification rate is lower than the second preset value, the initial threshold is restored; wherein K is a positive integer from 5 to 20, the first preset value is 90-98%, wherein the first preset value includes 90%, and the second preset value is 80-90%, wherein the second preset value does not include 90%.
6. A detection system for carrying out the method according to any one of claims 1 to 5, characterized in that The detection system comprises: an image acquisition module (100) comprising an optical microscope and a digital camera; an image processing module (200) carrying a deep learning processor supporting a channel attention mechanism; a human-computer interaction module (300) equipped with a display device and a threshold adjustment interface.
7. The detection system of claim 6, wherein: The deep learning processor includes a deep learning-based image segmentation model, which uses a pre-trained improved YOLO series model; the improved YOLO series model refers to embedding a channel attention module in the backbone network of the YOLO series model, and the channel attention module includes multiple global pooling layers and multiple fully connected layers.
8. A deep learning-based intelligent detection system for the dispersion degree of nano-powder materials, characterized in that, The detection system comprises: an image acquisition module comprising an optical microscope and a digital camera; an image processing module carrying a deep learning processor supporting a channel attention mechanism; a human-computer interaction module equipped with a display device and a threshold adjustment interface.
9. The detection system of claim 8, wherein: The deep learning processor includes a deep learning-based image segmentation model for performing image preprocessing, deep learning model inference and dispersion quantification analysis; The deep learning-based image segmentation model uses an improved YOLO series model; The improved YOLO series model refers to embedding a channel attention module in the backbone network of the YOLO series model; The channel attention module includes multiple global pooling layers and multiple fully connected layers.
10. The detection system of claim 9, wherein: The dispersion quantification analysis includes the quantification analysis of the agglomeration area proportion, maximum equivalent diameter and uniformity index.