A method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch
By constructing a fine-grained attention network model, combining shallow convolutional network and spatial attention network, the problem of inaccurate detection of microstructure of fiber masterbatch is solved, and efficient and accurate fiber masterbatch agglomeration and dispersion ultramicrometers are achieved. Structural testing, suitable for fiber textile industry.
Patent Information
- Application Number
- CN202411046124.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-08-01
AI Technical Summary
The prior art is difficult to effectively characterize the microstructure information of fiber masterbatches, resulting in inaccurate detection of fiber masterbatch agglomeration and dispersion ultra-microstructures, affecting the quality and performance of fiber textile products.
Build and train a fine-grained attention network model, combines shallow convolutional network, weak detector and strong detector, and recognizes the agglomeration structure in the electron microscope image of fiber masterbatch through spatial attention network to realize the agglomeration of fiber masterbatch and Accurate detection of dispersed ultrastructure.
It realizes efficient and fast fiber masterbatch agglomeration and dispersed ultrastructure detection, avoids the artificial feature design method in traditional methods, can obtain multi-dimensional key information from the microscopic scale, and improves the accuracy and generalization ability of detection. Suitable for the actual chemical fiber industry.
Smart Images

Figure CN118587706B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of fiber masterbatch production and intelligent identification, and relates to a method for detecting the agglomeration and dispersion ultrastructure of fiber masterbatch. Background Art
[0002] At present, the development of the chemical fiber industry is facing an urgent need for intelligent and high-end upgrades, and at the same time requires the production model to shift to more intelligent and specialized market segment customization.
[0003] Fiber masterbatch is essentially the raw material in the textile manufacturing process, and its components include carrier resin, pigment, dye, functional powder and additives. When high-concentration fiber masterbatch (or high-concentration dispersion) is in a liquid state, due to the molecular force between particles, the particles in the high-concentration fiber masterbatch are close to each other, forming agglomerates or aggregations, and agglomeration effect occurs. The agglomeration effect is mainly caused by van der Waals force, charge interaction, and repulsive shielding effect. The agglomeration effect of fiber masterbatch not only directly affects the color uniformity, particle dispersion and thermal stability of fiber masterbatch, but also further relates to the strength, appearance and durability of fiber textile products. In the actual production process, this agglomeration effect of high-concentration fiber masterbatch is inevitable, and it can be dissociated by the appropriate use of dispersants during the processing process. Therefore, the evaluation of the dispersion of fiber masterbatch in chemical fibers is a key technology. The existing evaluation technology is mainly achieved by evaluating the filtration effect (dropfiltration, DF value), which is related to the processing process of fiber masterbatch, but a single DF value cannot fully and accurately evaluate the quality of fiber masterbatch, and it is out of touch with actual use.
[0004] Image recognition technology based on artificial intelligence is an important technical means for fiber structure prediction, stability analysis, and material micro-property identification. It has achieved breakthrough research results in the fields of textile image classification, chemical fiber process optimization, and fiber generation online monitoring. These methods mainly use deep neural networks to learn the distribution of data features such as texture, brightness, and color in different pixel areas, and use a small number of fully connected layers to achieve recognition or prediction through feature self-learning of multi-layer perceptron networks. However, the performance of the above methods depends to a large extent on high-quality and massive training samples, cumbersome parameter adjustment, and complex network structure design, and there are still certain doubts about feature generalization ability, computational complexity, and network interpretability. Especially in complex feature processing tasks such as microscopic information and high-resolution fiber masterbatch ultrastructure, existing technical methods cannot learn the effective features of the data ultrastructure, resulting in the final algorithm performance being unstable or even significantly reduced.
[0005] Patent CN220482243U relates to a masterbatch sorting device, which is designed with an image sensor, a signal processor and a light source lamp. The image sensor transmits the signal to the signal processor to sort masterbatches of different colors. The main chute, the first auxiliary chute and the second auxiliary chute are designed to recognize and distinguish two colors. The drainage trough, the plug plate and the slot are designed to facilitate disassembly and installation, and to facilitate centralized collection. This patent solves the problem of distinguishing and identifying fiber masterbatches of two different colors from the perspective of hardware and signal processing, but this method has the problems of high artificial false detection and low recognition efficiency in the actual masterbatch recognition process.
[0006] Patent CN117095005A relates to a plastic masterbatch quality inspection method and system based on machine vision. The method is based on plastic masterbatch samples and adopts multi-sensor synchronous acquisition technology to simultaneously acquire visible light, infrared, and ultraviolet spectral images to generate a multimodal raw image data set. In this patent, the multi-sensor synchronous acquisition technology integrates visible light, infrared, and ultraviolet images to expand the detection dimension. The convolutional neural network and multi-source image fusion algorithm improve image quality and details, create high-definition comprehensive images, and construct a three-dimensional model by point cloud data processing to intuitively display the morphology of plastic masterbatch. The deep convolutional neural network improves the accuracy of defect detection, reduces false alarms, and improves image quality and detection accuracy. However, this method still uses traditional image acquisition equipment, and its resolution is only at the pixel level. It is difficult to intuitively judge the agglomeration and dispersion quality of the prepared masterbatch product at a microscopic scale.
[0007] In the field of artificial intelligence, the Attention Network is a key technology in deep learning and is widely used in natural language processing and computer vision. Its core is to enable the model to selectively focus on different parts of the input data by calculating the attention weights, thereby capturing key information. However, the current attention network needs to store all attention weight matrices, which will significantly increase memory consumption, especially when processing long sequences. At the same time, the attention network is more sensitive to noise in the training data, which affects the stability and generalization ability of the model.
[0008] Therefore, it is of great practical value and significance to develop a method for detecting the ultrastructure of fiber masterbatch agglomeration and dispersion to solve the problem that the existing technology is difficult to effectively characterize the microstructure information of fiber masterbatch. Summary of the invention
[0009] The purpose of the present invention is to solve the problems existing in the prior art and provide a method for detecting the ultra-microstructure of agglomeration and dispersion of fiber masterbatch.
[0010] To achieve the above object, the technical solution adopted by the present invention is as follows:
[0011] A method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch, after constructing and training a fine-grained attention network model, an electron microscope image of the fiber masterbatch to be identified is input into the trained fine-grained attention network model, and the trained fine-grained attention network model outputs the fiber masterbatch agglomeration identification result, that is, the agglomeration structure existing in the electron microscope image of the fiber masterbatch is detected;
[0012] The fine-grained attention network model includes shallow convolutional networks, weak detectors, and strong detectors;
[0013] The shallow convolutional network is used to obtain shallow feature information. The shallow convolutional network includes two convolutional layers, two downsampling layers and one nonlinear layer.
[0014] The weak detector includes a deep convolutional network and an attention network; the deep convolutional network includes two convolutional layers, two downsampling layers and three fully connected layers. The deep convolutional network can calculate the convolutional features based on the input shallow feature information, and the convolutional features are 64×64×128 convolutional features; the attention network includes a global average downsampling 3×3 pooling layer, a nonlinear activation function and a feature recalibration weight W. The attention network can obtain the local salient area based on the input shallow feature information;
[0015] The strong detector includes a spatial attention network; the spatial attention network includes two 3×3 pooling layers (for generating a two-dimensional feature map), a nonlinear activation function and a spatial transformation coefficient α. The spatial transformation coefficient α is a parameter learned inside the spatial attention network, and its size is determined by the type of transformation learned. The present invention adopts affine transformation, and α is a 6-dimensional vector, which is specifically formed by the attention module in the form of , where the weight , , represents the hidden layer state of the network, Represents a multilayer perceptron network (structured as Figure 1 As shown), v represents the extracted feature set, It represents the features that need to be paid attention to, i represents the number of features, t represents the number of hidden layers, j represents the number of weights, and m represents the number of features.
[0016] As the preferred technical solution:
[0017] In the above-mentioned method for detecting ultrastructure of aggregation and dispersion of fiber masterbatch, in the shallow convolution network, the first convolution layer adopts a convolution kernel of 7×7 pixels, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolution layer adopts a convolution kernel of 5×5 pixels, the number of output feature maps is 64, and the convolution step is 2 pixels; both downsampling layers adopt a sampling size of 2×2 pixels and a sampling step of 2 pixels; the nonlinear layer adopts a softmax function.
[0018] In the method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch as described above, in the deep convolutional network, the first convolutional layer adopts a convolution kernel of 11×11 pixels, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolutional layer adopts a convolution kernel of 5×5 pixels, the number of output feature maps is 64, and the convolution step is 2 pixels; the two downsampling layers both adopt a sampling size of 2×2 pixels and a sampling step of 2 pixels; the three fully connected layers contain 5184, 5184 and 2592 neurons respectively; "first" and "second" are numbered according to the order in which the data arrive.
[0019] In the above-mentioned method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch, the nonlinear activation function is a ReLU function.
[0020] In the method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch as described above, the spatial attention network is a spatially weighted Fisher convolutional network.
[0021] In the above-mentioned method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch, the weak detector and the strong detector in the fine-grained attention network model need to be trained separately. The training process of the fine-grained attention network model is as follows:
[0022] (1) Let the number of iterations g = 1;
[0023] (2) Randomly select w training samples, and then use a shallow convolutional network to obtain the shallow feature information of each of the w training samples;
[0024] (3) Input the shallow feature information obtained in step (2) into the weak detector of the fine-grained attention network model. The weak detector learns the input training samples and outputs the encoding features. The learning process is the process of adjusting the connection weights between neurons in the weak detector of the fine-grained attention network model.
[0025] (4) The encoding features output by the weak detector are input into the strong detector of the fine-grained attention network model. The strong detector further learns the local agglomeration structural feature information of the fiber masterbatch electron microscope image from the encoding features. The learning process is the process of adjusting the connection weights between neurons in the strong detector of the fine-grained attention network model.
[0026] (5) Setting a threshold to filter out target boxes with confidence levels higher than the threshold. The target boxes are used to identify the locations of the aggregate structures in the electron microscope image.
[0027] (6) Determine whether the iteration termination condition is met, that is, the number of iterations g is equal to the maximum number of iterations g max , or the percentage of the number of electron microscope images correctly recognized by the fine-grained attention network model to the total number of input electron microscope images is greater than 90%. If so, the training ends. Otherwise, let g = g + 1 and return to step (2).
[0028] (7) Input the verification sample into the fine-grained attention network model trained in step (6) for verification.
[0029] In the process of training the constructed fine-grained attention network model, a key problem is that the detector with the best recognition effect may only notice the images that are easy to detect in the training samples, but cannot find their corresponding positive samples from other complex images. To this end, regularization loss is added during the training process to mine positive samples in the training samples.
[0030] Studies have shown that spatial attention networks can help capture features such as tiny details and textures. This structure is very valuable. It can not only help us remove the interference caused by the background of electron microscope data in the identification of fiber masterbatch agglomeration and dispersion electron microscope ultrastructure, but also accurately locate the agglomeration structure again by learning tiny details, which is conducive to accurately identifying the fiber masterbatch agglomeration structure in electron microscope data.
[0031] In the above-mentioned method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch, the ratio of training samples to verification samples is 50~80:10~30.
[0032] In the method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch as described above, the cross entropy loss function is used as the convergence evaluation index during the training process of the fine-grained attention network model.
[0033] The above-mentioned method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch has a threshold value of 0.5.
[0034] As described above, a method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch has a maximum number of iterations g max Not less than 400.
[0035] Invention mechanism:
[0036] In the process of collecting electron microscope images of fiber masterbatch, it can be found that a lot of agglomeration information in the electron microscope image is hidden in the local area. Therefore, in order to effectively mine the hidden agglomeration information in the ultrastructure of fiber masterbatch, the present invention constructs a fine-grained attention network model, designs two-level attention modules (i.e., weak detectors and strong detectors) of feature graphs, and realizes effective attention to information of different scales by transferring and fusing the feature graphs of the two levels layer by layer, and accurately identifies the agglomeration structure of fiber masterbatch in electron microscope data, which solves the problem that the agglomeration structure of electron microscope images is difficult to manually mark one by one in the prior art, and the agglomeration structure with a smaller particle size is easily ignored, resulting in inaccurate detection.
[0037] Among them, fine-grained refers to the detailed and precise processing of data and tasks. The advantages of fine-grained methods are their high precision and detail, which can capture the subtle differences in the data and improve the performance of the model. For example, under fiber electron microscope data, fine-grained methods can better capture the tiny details and structures under the ultrastructure, and better identify the aggregation and dispersion states of the ultrastructure.
[0038] When the human visual system faces such a complex image, the eyes first perform an initial scan to obtain the overall scene information. After discovering potential areas of interest such as prominent textures, colors or edges, when scanning again, the attention will be focused on the previously determined area of interest, and the microscopic details therein, such as fine-grained size, shape and texture, will be deeply perceived. This hierarchical and staged attention mechanism enables humans to more effectively process and understand the microscopic information in high-resolution images, thereby achieving fine perception and recognition of complex images. Therefore, we consider using the staged local attention mechanism of the biological visual system when processing complex data, and using the advantages of different layers of the convolutional network being sensitive to different specific areas to design a method for detecting the ultrastructure of fiber masterbatch agglomeration and dispersion. The present invention borrows the framework of the traditional attention module, and modifies and improves it on this basis to make it suitable for the case of electron microscope ultramicroscopic images.
[0039] Beneficial effects:
[0040] (1) The present invention provides an efficient and rapid method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatches. It does not require the traditional artificial feature design method, but constructs and trains a fine-grained attention network model, thereby avoiding the problem of agglomeration evaluation being out of touch with reality during the preparation of fiber masterbatches. It can obtain multi-dimensional key information on the ultrastructure of agglomeration and dispersion of fiber masterbatches at a microscopic scale, and complete the accurate detection of agglomeration and dispersion structures in a self-learning manner.
[0041] (2) The present invention provides an efficient and rapid method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatches. The method can effectively detect the ultrastructure of agglomeration and dispersion of fiber masterbatches through the self-learning feature information characterization capability. The result has good generalization ability and has important practical significance for the preparation of masterbatches in the actual chemical fiber industry. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 is a basic structural diagram of the multilayer perceptron network model of the present invention;
[0043] Figure 2 It is the basic structure diagram of the spatial attention network of the present invention;
[0044] Figure 3 The theoretical structure diagram of a method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch according to the present invention;
[0045] Figure 4 Schematic diagram of the detection process of the weak detector of the present invention;
[0046] Figure 5 It is a schematic diagram of the detection process of the strong detector of the present invention;
[0047] Figure 6 This is the original electron microscope image of the fiber masterbatch of the present invention;
[0048] Figure 7 The figure is the electron microscope image detection result of the fiber masterbatch of the present invention. DETAILED DESCRIPTION
[0049] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.
[0050] A method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch, the steps are as follows:
[0051] (1) Construct a fine-grained attention network model;
[0052] The fine-grained attention network model includes shallow convolutional networks, weak detectors, and strong detectors;
[0053] The shallow convolutional network consists of two convolutional layers, two downsampling layers, and a nonlinear layer;
[0054] In the shallow convolutional network, the first convolutional layer of the two convolutional layers uses a 7×7 pixel convolution kernel, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolutional layer uses a 5×5 pixel convolution kernel, the number of output feature maps is 64, and the convolution step is 2 pixels;
[0055] Both downsampling layers use a sampling size of 2 × 2 pixels and a sampling step of 2 pixels;
[0056] The nonlinear layer uses the softmax function;
[0057] The weak detector consists of a deep convolutional network and an attention network;
[0058] The deep convolutional network consists of two convolutional layers, two downsampling layers, and three fully connected layers;
[0059] In the deep convolutional network, the first convolutional layer of the two convolutional layers uses a convolution kernel of 11×11 pixels, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolutional layer uses a convolution kernel of 5×5 pixels, the number of output feature maps is 64, and the convolution step is 2 pixels;
[0060] Both downsampling layers use a sampling size of 2 × 2 pixels and a sampling step of 2 pixels;
[0061] The three fully connected layers contain 5184, 5184, and 2592 neurons respectively;
[0062] The attention network includes a global average downsampling 3×3 pooling layer, a nonlinear activation function and a feature recalibration weight W, and the nonlinear activation function is the ReLU function;
[0063] The strong detector includes a spatial attention network;
[0064] The spatial attention network is a spatially weighted Fisher convolutional network, and its basic structure is shown in the figure below. Figure 2 As shown, it includes two 3×3 pooling layers, a nonlinear activation function and a spatial transformation coefficient α, and the nonlinear activation function is the ReLU function;
[0065] (2) Training a fine-grained attention network model;
[0066] (2.1) Let the number of iterations g = 1;
[0067] (2.2) Randomly select w training samples, and then use a shallow convolutional network to obtain the shallow feature information of each of the w training samples;
[0068] (2.3) The shallow feature information obtained is input into the weak detector of the fine-grained attention network model. The weak detector learns the input training samples and outputs the encoded features.
[0069] (2.4) The encoding features output by the weak detector are input into the strong detector of the fine-grained attention network model. The strong detector further learns the local agglomeration structural feature information of the fiber masterbatch electron microscope image from the encoding features.
[0070] (2.5) Setting a threshold to filter out target boxes with confidence values higher than the threshold, which are used to frame the location of the agglomerated structure in the electron microscope image; wherein the threshold is 0.5;
[0071] (2.6) Determine whether the iteration termination condition is met, that is, the number of iterations g is equal to the maximum number of iterations g max , or the percentage of electron microscope images correctly recognized by the fine-grained attention network model to the total number of electron microscope images input is greater than 90%. If so, the training ends. Otherwise, let g=g+1 and return to step (2.2); where the maximum number of iterations g max Not less than 400;
[0072] (2.7) In the process of steps (2.1) to (2.6), each iteration uses the cross entropy loss function as the convergence evaluation indicator. As the number of training iterations increases, the cost function (loss) is continuously minimized so that it gradually converges to 0;
[0073] (2.8) Input the verification sample into the fine-grained attention network model trained in step (2.7) for verification.
[0074] (3) The electron microscope image of the fiber masterbatch to be identified is input into the trained fine-grained attention network model, and the trained fine-grained attention network model outputs the fiber masterbatch agglomeration identification result.
[0075] The above-mentioned fiber masterbatch agglomeration and dispersion ultrastructure detection method of the present invention is specifically described below:
[0076] A method for detecting the ultrastructure of fiber masterbatch aggregation and dispersion, the theoretical structure of which is as follows Figure 3 As shown, the specific steps are as follows:
[0077] (1) Fiber masterbatch electron microscopy database;
[0078] From the electron microscope image of the original fiber masterbatch (with a resolution of 24k×24k pixels, such as Figure 6 150 image segments of size 512×512 are randomly selected as samples, of which 90 are training samples, 40 are validation samples, and the remaining 20 are detection samples;
[0079] (2) Construct a fine-grained attention network model;
[0080] The fine-grained attention network model includes shallow convolutional networks, weak detectors, and strong detectors;
[0081] The shallow convolutional network is used to obtain shallow feature information, including two convolutional layers, two downsampling layers and one nonlinear layer;
[0082] In the shallow convolutional network, the first convolutional layer of the two convolutional layers uses a 7×7 pixel convolution kernel, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolutional layer uses a 5×5 pixel convolution kernel, the number of output feature maps is 64, and the convolution step is 2 pixels;
[0083] Both downsampling layers use a sampling size of 2 × 2 pixels and a sampling step of 2 pixels;
[0084] The nonlinear layer uses the softmax function;
[0085] The weak detector consists of a deep convolutional network and an attention network;
[0086] The deep convolutional network consists of two convolutional layers, two downsampling layers, and three fully connected layers;
[0087] In the deep convolutional network, the first convolutional layer of the two convolutional layers uses a convolution kernel of 11×11 pixels, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolutional layer uses a convolution kernel of 5×5 pixels, the number of output feature maps is 64, and the convolution step is 2 pixels;
[0088] Both downsampling layers use a sampling size of 2 × 2 pixels and a sampling step of 2 pixels;
[0089] The three fully connected layers contain 5184, 5184, and 2592 neurons respectively;
[0090] The attention network includes a global average downsampling 3×3 pooling layer, a nonlinear activation function (ReLU function), and a feature recalibration weight W;
[0091] The strong detector includes a spatially weighted Fisher convolutional network;
[0092] The spatial weighted Fisher convolutional network consists of two 3×3 pooling layers, a nonlinear activation function (ReLU function) and a spatial transformation coefficient α;
[0093] (3) Training a fine-grained attention network model;
[0094] (3.1) Let the number of iterations g = 1;
[0095] (3.2) If Figure 3 As shown in the figure, 9 training samples are randomly selected from 90 training samples, and then a shallow convolutional network is used to obtain shallow feature information in each training sample, and then the training samples are divided into clustered samples and non-clustered samples according to the obtained information;
[0096] (3.3) If Figure 4 As shown in the figure, the shallow feature information, clustered samples and non-clustered samples are input into the weak detector of the fine-grained attention network model. The weak detector learns the input training samples and outputs the encoding features. Specifically, the deep convolutional network obtains the convolutional features according to the input shallow feature information, and the attention network obtains the local salient area according to the input shallow feature information. Among them, the convolutional features are 64×64×128 convolutional features.
[0097] (3.4) If Figure 5 As shown, the encoding features output by the weak detector are input into the strong detector of the fine-grained attention network model. The strong detector further learns the local agglomeration structural feature information of the fiber masterbatch electron microscope image from the encoding features. Specifically, the strong detector first performs a dot product operation on the convolutional features obtained in step (3.3) and the local salient area, and then the spatial weighted Fisher convolutional network continues to calculate multiple candidate boxes based on the results of the dot product operation, and calculates the confidence of each candidate box respectively.
[0098] (3.5) A threshold of 0.5 is set, and candidate boxes with confidence levels higher than the threshold are retained and determined as target boxes to frame the location of the agglomerated structure in the electron microscope image, and then the target box is output to confirm the location of the agglomerated structure;
[0099] (3.6) When the number of iterations g equals the maximum number of iterations 400, the iteration termination condition is met and the training ends;
[0100] In the process of steps (3.1) to (3.6), each iteration uses the cross entropy loss function as the convergence evaluation indicator. As the number of training iterations increases, the cost function (loss) is continuously minimized so that it gradually converges to 0;
[0101] (3.7) Verify the trained fine-grained attention network model;
[0102] The 40 validation samples were input into the trained fine-grained attention network model for validation. Finally, the accuracy of the trained fine-grained attention network model in detecting the 40 validation samples was 95%.
[0103] (4) Detect the trained fine-grained attention network model;
[0104] The 20 test samples are input into the trained fine-grained attention network model, and then the fine-grained attention network model outputs the prediction results. The prediction results are compared with the true values of the actual samples (that is, the positions where the experts mark the clusters of the 20 test samples). If the prediction results are consistent with the true values, they are correct results, otherwise they are wrong detections.
[0105] The accuracy of the fine-grained attention network model after final statistical verification for 20 test samples was 90%. During the test process, the recognition speed of each test sample was 0.8s, indicating that the trained fine-grained attention network model has the ability to quickly recognize the agglomeration and dispersion ultrastructure of fiber masterbatch. The final test results are as follows: Figure 7 shown.
[0106] By comparing the results of the fine-grained attention network model trained by the present invention with the actual results, it can be seen that the number and position of each correctly detected sample can be accurately identified, indicating that the present invention can use the semi-supervised learning method of artificial intelligence to characterize feature information through self-learning, and effectively identify the agglomerated and dispersed ultrastructures of fiber masterbatches quickly, accurately and with high sensitivity. The results have good generalization ability and are of great practical significance for the actual preparation of industrial fiber masterbatches.
Claims
1. A method for detecting the ultrastructure of fiber masterbatch agglomeration and dispersion, characterized in that: After constructing and training the fine-grained attention network model, the electron microscope image of the fiber masterbatch to be identified is input into the trained fine-grained attention network model, and the trained fine-grained attention network model outputs the fiber masterbatch agglomeration identification result; The fine-grained attention network model includes shallow convolutional networks, weak detectors, and strong detectors; The shallow convolutional network consists of two convolutional layers, two downsampling layers, and a nonlinear layer; The weak detector includes a deep convolutional network and an attention network; the deep convolutional network includes two convolutional layers, two downsampling layers and three fully connected layers; the attention network includes a global average downsampling 3×3 pooling layer, a nonlinear activation function and a feature recalibration weight W; The strong detector includes a spatial attention network; the spatial attention network consists of two 3×3 pooling layers, a non-linear activation function, and a spatial transformation coefficient α.
2. A method for detecting ultrastructure of fiber masterbatch agglomeration and dispersion according to claim 1, characterized in that: In the shallow convolutional network, the first convolution layer uses a 7×7 pixel convolution kernel, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolution layer uses a 5×5 pixel convolution kernel, the number of output feature maps is 64, and the convolution step is 2 pixels; both downsampling layers use a sampling size of 2×2 pixels and a sampling step of 2 pixels; the nonlinear layer uses the softmax function.
3. The method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 1, characterized in that: In the deep convolutional network, the first convolution layer uses a convolution kernel of 11×11 pixels, the number of output feature maps is 32, and the convolution step is 3 pixels; the second convolution layer uses a convolution kernel of 5×5 pixels, the number of output feature maps is 64, and the convolution step is 2 pixels; the two downsampling layers both use a sampling size of 2×2 pixels and a sampling step of 2 pixels; the three fully connected layers contain 5184, 5184, and 2592 neurons, respectively.
4. The method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 1, characterized in that: The nonlinear activation function is the ReLU function.
5. The method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 1, characterized in that: The spatial attention network is a spatially weighted Fisher convolutional network.
6. The method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 1, characterized in that: The training process of the fine-grained attention network model is as follows: (1) Let the number of iterations be g =1; (2) Randomly select from training samples w Then a shallow convolutional network is used to obtain w Shallow feature information of each training sample; (3) The shallow feature information obtained in step (2) is input into the weak detector of the fine-grained attention network model. The weak detector learns the input training samples and outputs the encoded features. (4) The encoding features output by the weak detector are input into the strong detector of the fine-grained attention network model. The strong detector further learns the local agglomeration structural feature information of the fiber masterbatch electron microscope image from the encoding features. (5) Setting a threshold to filter out target boxes with confidence levels higher than the threshold. The target boxes are used to identify the locations of the aggregate structures in the electron microscope image. (6) Determine whether the iteration termination condition is met, that is, the number of iterations g Equal to the maximum number of iterations g max , or the percentage of electron microscope images correctly recognized by the fine-grained attention network model to the total number of electron microscope images input is greater than 90%, if so, the training ends, otherwise g = g +1, return to step (2); (7) Input the verification sample into the fine-grained attention network model trained in step (6) for verification.
7. A method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 6, characterized in that: The ratio of training samples to validation samples is 5~8:1~3.
8. A method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 6, characterized in that: During the training process of the fine-grained attention network model, the cross entropy loss function is used as the convergence evaluation indicator.
9. The method for detecting the ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 6, characterized in that: The threshold is 0.
5.
10. A method for detecting ultrastructure of agglomeration and dispersion of fiber masterbatch according to claim 6, characterized in that: Maximum number of iterations g max Not less than 400.
Citation Information
Patent Citations
Plastic master batch quality inspection method and system based on machine vision
CN117095005A
Color master batch sorting equipment
CN220482243U
Method and device for extracting text area from image, equipment and storage medium
CN111275034A
Semi-supervised image classification method based on generative adversarial network local aggregation coding
CN111832650A
Artificial neural network model and electronic device including the same
CN112149793A
