Intelligent identification method for fiber master batch agglomeration structure based on cross-scale invariant information

Abstract

The invention belongs to the crossing field of computer vision and material micro-analysis, and relates to a cross-scale invariant information-based intelligent identification method for a fiber master batch agglomeration structure, which comprises the following steps of: constructing and training an end-to-end detection frame for extracting cross-scale invariant information, and inputting a to-be-identified fiber color master batch electron microscope image into the end-to-end detection frame; outputting an identification result of the agglomeration structure; the detection framework comprises a convolutional network with selectively enhanced edge information, a self-attention encoder, a self-attention decoder, a scale classifier and a full-connection network. According to the method, an artificial intelligence supervised learning mode is utilized, accurate and robust effective recognition is carried out on cross-scale fiber master batch aggregation by learning invariant information of aggregation in a cross-scale ultrastructure, and the method has important significance on actual industrial fiber color master batch preparation.

Description

TECHNICAL FIELD

[0001] The application belongs to the cross field of computer vision and material microanalysis, and relates to an intelligent identification method for fiber master batch agglomerate structure based on cross-scale invariant information. BACKGROUND

[0002] As a basic raw material for textile manufacturing, fiber master batch is mainly composed of carrier resin, pigment, dye, functional powder and various additives. In a liquid environment, high-concentration master batch (also known as high-concentration dispersion system) has intermolecular forces between particles, which makes particles close to each other and form agglomerate or lump structure, which is defined as agglomeration effect. The fundamental reason for this effect involves van der Waals interaction, electrostatic charge effect and repulsive force shielding mechanism and other physical and chemical factors. It is worth noting that this particle aggregation state is reversible in subsequent processing links, so accurate evaluation of the dispersion characteristics of functional materials in fiber master batch and fiber becomes a core technology point. In the preparation process of fiber master batch, high-resolution scanning electron microscope (SEM) imaging technology is widely used for microstructure research. This technology can clearly present the morphological characteristics and structural details of the master batch, including particle size distribution, geometric shape, surface roughness and internal aggregation morphology and other key information. Although electron microscopy imaging can provide rich information at a fine scale, certain important microscopic features, especially texture edges, are often overlooked or misjudged; moreover, in the micrograph of the master batch, the agglomerate structure to be identified and the texture of the dispersed fibers in the background are similar and very complex, and it is difficult to distinguish them by conventional image recognition methods, resulting in a decrease in recognition accuracy.

[0003] In recent years, the technology system based on artificial intelligence has achieved significant breakthroughs in the application fields of textile image recognition, chemical fiber process optimization, online monitoring of fiber production, etc. The core of these methods is to use deep neural network architecture to learn the performance rules of texture patterns, light intensity and color distribution and other features in different pixel regions, and finally complete the classification or prediction task through a few fully connected layers by means of multi-layer feature self-learning mechanism. However, the performance of the existing technology is heavily dependent on large-scale high-quality training data sets, complex network architecture design and tedious hyperparameter optimization process. There are still significant limitations in feature generalization ability, model interpretability and computational efficiency. Especially in the analysis task of complex features such as microinformation and high-resolution fiber master batch ultrastructure, traditional algorithms are difficult to effectively extract the ultrastructure features in the data, resulting in unstable algorithm performance and even significant performance degradation.

[0004] The patent application CN119295430A discloses a color master batch color detection method and device, computer equipment and storage medium. First, the environment data to be detected is obtained, and the ambient light intensity and ambient light direction are obtained from the environment data to be detected. Then, the color of the color master batch to be detected is obtained, the corresponding detection environment requirement data is obtained according to the color of the color master batch to be detected, and the detection environment requirement data is compared with the ambient light intensity and ambient light direction to obtain a light comparison result. The environment data to be detected is adjusted according to the light comparison result, the color master batch detection image is obtained, the color master batch detection image is subjected to spectral detection, and a color difference detection result is obtained. The color difference detection result is analyzed, the abnormal color difference position is obtained according to the corresponding analysis result, and the abnormal color difference position is marked. Although the application can realize automatic abnormal detection, the final analysis and abnormal marking still depend on the final judgment of artificial, which may introduce subjective error and cannot completely realize full-process unmanned quality inspection.

[0005] Therefore, it is of great significance to develop an intelligent method for automatically identifying the agglomeration and dispersion of fiber master batch ultra-microstructure with cross-scale generalization to improve the identification efficiency and accuracy. SUMMARY

[0006] The purpose of the present application is to solve the technical problems existing in the prior art, and to provide an intelligent identification method for the agglomeration structure of fiber master batch based on cross-scale invariant information.

[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] An intelligent identification method for the agglomeration structure of fiber master batch based on cross-scale invariant information, characterized in that it comprises the following steps:

[0009] (a) Constructing an end-to-end detection framework, inputting the micrograph of the fiber master batch to be identified into the framework, and outputting the identification result of the agglomeration structure;

[0010] (b) The detection framework comprises the following sub-steps:

[0011] b.1) Edge information selective enhancement convolution network processing: processing the input image information through the edge information selective enhancement convolution network to obtain global edge enhancement features;

[0012] b.2) Spatial attention mechanism feature selection: spatial attention mechanism processing is performed on the multi-scale feature map subjected to edge enhancement to emphasize the valuable edge information region and weaken the irrelevant region;

[0013] b.3) Channel attention mechanism feature selection: the feature map processed by the spatial attention mechanism is processed by the channel attention mechanism, the weighted selection of the multi-channel features is realized, and the important channel information is highlighted;

[0014] b.4) Self-attention encoder encoding: the processed features are input into the self-attention encoder, and the advanced semantic features of the target and the background are fully decoupled through the multi-head attention mechanism;

[0015] b.5) Self-attention decoder decoding and full connection network output: the encoded features are input into the self-attention decoder, the prediction set of the target is output through the object query method, and the category and position information results are output through the full connection network;

[0016] (c) The features obtained by the detection framework are made to have cross-scale invariance through the adversarial training method;

[0017] (d) The detection framework is trained in a supervised learning manner, and the trained detection framework is used for intelligent identification of the fibrogranulosa aggregation structure.

[0018] The edge information selective enhancement convolutional network obtains enhanced features through multi-scale pooling operation and edge enhancement operation, and the enhanced features are spliced to form a stacked feature map.

[0019] The specific processing steps of the global edge enhancement module include: (a) adaptively average-pooling the input feature map to multiple different scales; (b) processing each scale feature map through the global edge enhancement module respectively; (c) splicing the enhanced features of multiple scales in the channel dimension to form multi-scale edge enhanced features;

[0020] The specific processing steps of the global edge enhancement module include: performing average pooling operation on the input scale feature map to obtain low-frequency information of the feature map; performing element-wise subtraction operation on the original scale feature map and the low-frequency information to obtain high-frequency information; performing element-wise addition operation on the high-frequency information and the original scale feature map to obtain the edge enhanced feature map.

[0021] The spatial attention mechanism includes: performing average pooling and maximum pooling operations on the spliced multi-scale feature map respectively; splicing the pooling results and generating a spatial attention weight map through a convolution layer; multiplying the spatial attention weight map and the spliced multi-scale feature map to focus on the spatial region containing important edge information.

[0022] The channel attention mechanism comprises: performing channel-level global average pooling on the feature map after spatial attention processing to calculate the channel feature mean; subtracting the channel feature mean of the feature map from the feature map to obtain a high-frequency component; and performing element-wise multiplication of the high-frequency component and the feature map to selectively enhance the edge texture channel related to the structure.

[0023] The pooling kernel size of the channel-level global average pooling is equal to the spatial dimension of the input feature map, and the weight range of the element-wise multiplication operation is [0.1, 2.0].

[0024] The output channel number of the convolutional network is between 32 and 128, the convolutional layer stride is 1 or 2, and the number of cascading times of the edge enhancement module and the attention module is 1 to 4 times; the feature of the input image after selective enhancement of edge information is used as the input of the subsequent encoding process.

[0025] The self-attention encoder adopts a multi-head attention mechanism, the number of heads is 4 to 16, and the output dimension of each layer is 128 to 512; each feature position in the self-attention encoder structure is added with a sine-cosine position encoding and is input in the form of a flattened tensor, the target and background features are decoupled through the attention mechanism; the self-attention decoder introduces a fixed number of object queries, the number of decoder layers is 1 to 8, and the output dimension is 64 to 512; the decoding output is input into a fully connected network containing 2 to 5 layers, and the activation function is ReLU, and finally the class and position information of each detection target is output.

[0026] The specific process of the adversarial training comprises:

[0027] (a) inputting the edge enhancement feature into the gradient reversal layer;

[0028] (b) during the forward propagation process, the gradient reversal layer acts as a transparent layer and does not process the input feature;

[0029] (c) during the backward propagation process, the gradient reversal layer multiplies the loss gradient by a negative coefficient α, wherein the value range of α is [0.1, 2.0];

[0030] (d) inputting the output of the gradient reversal layer into the scale classifier for scale discrimination;

[0031] (e) calculating the joint loss function of the scale classification loss and the main task detection loss for end-to-end optimization.

[0032] The scale classifier is a single-layer full connection network, the activation function is ReLU, the scale classification result is output, and the cross entropy loss function is used to calculate the classification loss; the joint loss function is a weighted combination of the main task detection loss and the adversarial loss, and the weight ratio range is [0.7, 0.3]: [0.9, 0.1]; the learning rate range of the adversarial training is [0.0001, 0.01], and the optimizer uses Adam or SGD; the training set is obtained by cutting and data enhancement on the original microscopic image in the size interval of [128*128, 512*512], including one or more of image horizontal flip, random rotation, brightness disturbance and noise disturbance.

[0033] The method outputs the agglomerated structure region with a rectangular boundary box, and each boundary box predicts four vertex coordinates, and the coordinate range is within the width and height of the original image.

[0034] The number of samples used is 50 to 1000, the observation scale is 300 nanometers to 500 microns, and unified training and prediction are performed on an end-to-end detection framework.

[0035] The beneficial effects of the present application are: the fiber master batch agglomerated structure intelligent identification method based on cross-scale invariant information can capture the detail texture edge feature and cross-scale invariant feature of the fiber agglomerated structure, and complete the accurate identification of the agglomerated and dispersed structure in a self-learning manner; the influence of the reduction of the agglomerated identification accuracy in the fiber master batch preparation process due to the lack of scale generalization of the existing design method is avoided. The attention mechanism in the present application can dynamically and intelligently determine which scale and which region of the edge information should be enhanced and which should be inhibited, so that the edge enhancement technology has better target perception ability through this selectivity, and can selectively sharpen the agglomerated structure rather than the fiber structure, greatly improving the discrimination accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 It is a theoretical structure diagram of the cross-scale invariant information fiber master batch agglomerated structure intelligent identification method of the present application;

[0037] Figure 2 It is an edge selective enhancer structure diagram in the edge selective enhancement convolutional network in the present application;

[0038] Figure 3 It is a fiber master batch electron microscope image;

[0039] Figure 4 It is an effect diagram output after identification in the control group 3; Figure 3

[0040] Figure 5 It is an effect diagram output after identification in the control group 3; Figure 3 ​The effect picture output after the recognition. DETAILED DESCRIPTION

[0041] By constructing and training an end-to-end detection framework for extracting cross-scale invariant information, the fiber master batch electron microscope image to be recognized is input into the end-to-end detection framework, and the recognition result of the agglomerated structure is output; the detection framework includes a convolutional network with selective enhancement of edge information, a self-attention encoder, a self-attention decoder, a scale classifier, and a fully connected network. The present application uses the supervised learning method of artificial intelligence, learns the invariant information in the agglomeration of cross-scale ultrastructure, accurately and robustly identifies the cross-scale fiber master batch agglomeration, and has important significance for the actual industrial fiber master batch preparation.

[0042] The fiber master batch suitable for the detection method of the present application can be composed of the following components:

[0043] Polymer material: as the main carrier, including but not limited to polyethylene terephthalate (PET), polypropylene (PP), polyethylene (PE), polyamide (PA), etc. thermoplastic resin. For example, PET resin (44.98wt%) is used as the carrier in the patent embodiment.

[0044] Fillers are used to enhance material performance or reduce cost, including inorganic particles (such as calcium carbonate, silica, talc, carbon black) or functional powders (such as conductive carbon fiber, glass fiber). Carbon black (20wt%) and calcium carbonate (35wt%) are used as fillers in the patent.

[0045] Additives are used to improve processing performance or stability, including: antioxidants (such as thiol compounds, 0.01wt%), lubricants (such as stearic acid, zinc stearate, 0.01wt%), dispersants (such as silane coupling agents), flame retardants (such as aluminum hydroxide, phosphorus compounds), coloring agents (such as organic pigments or dyes).

[0046] The preparation of the fiber masterbatch needs to ensure uniform dispersion of the components and the formation of detectable microstructure through the following processes: premixing and drying: the polymer matrix, fillers and additives are premixed in proportion, and the moisture is removed by vacuum drying. Melt blending: using a twin-screw extruder at high temperature (such as PET processing temperature 260-280℃), the fillers and additives are uniformly distributed in the polymer matrix. Extrusion granulation: the blend is made into masterbatch by underwater cutting or hot cutting process, and the particle size is usually 2-4mm. Post-processing: the masterbatch is cooled, sieved and surface treated (such as spraying anti-blocking agent) to avoid secondary agglomeration during storage. The masterbatch needs to form agglomeration structure (such as carbon black aggregation, calcium carbonate particle adhesion) during processing due to charge accumulation or component interaction, and the agglomeration size is in the micron to sub-micron level, which can be clearly imaged (resolution ≥512×512) by scanning electron microscopy (SEM). The surface of the masterbatch needs to have sufficient electrical conductivity (such as gold spraying treatment) to ensure the quality of SEM imaging, and there is a clear texture or morphological difference between the agglomeration area and the matrix, which facilitates the extraction and identification of multi-domain feature fusion algorithm.

[0047] The microscopic images of the masterbatch obtained in this patent can be collected by electron microscope or optical microscope; and the image resolution is preferably ≥512×512 pixels to ensure sufficient extraction of texture and morphological features by the algorithm. For magnification, SEM photos are preferably adjusted in the range of 500×-50,000×, and the specific selection is based on the agglomeration size (such as 1,000× for micron-level agglomeration and 10,000× for sub-micron-level agglomeration). Optical microscope (OM) photos are preferably usually between 100×-1,000×, and need to be combined with high numerical aperture (NA) objective lens to improve clarity. Before constructing the training set and the validation set, the agglomeration area in the image needs to be manually annotated by experts in the field of materials science or chemistry, and the annotation format is a rectangular bounding box (coordinate format such as [x_min, y_min, x_max, y_max]). The annotation needs to cover all visible agglomeration structures, including partially occluded or edge blurred areas, to ensure the comprehensiveness and consistency of the annotation.

[0048] In the following examples, the masterbatch samples used are from Suzhou Baolide New Materials Technology Co., Ltd. The total number of fiber masterbatch electron microscope images is 70.

[0049] The intelligent recognition method of fiber masterbatch agglomeration structure based on cross-scale invariant information is as follows:

[0050] (1) Construct an end-to-end detection framework;

[0051] The workflow of the end-to-end detection framework is as shown in Figure 2 , and is as follows:

[0052] (1.1) The fiber master batch electron microscope image is subjected to edge information selective enhancement convolution network to obtain enhanced picture features A;

[0053] The fiber master batch electron microscope image is processed by an edge information selective enhancement convolution network to obtain enhanced picture features A, and the steps are as follows:

[0054] (a) The fiber master batch electron microscope image is subjected to a convolution layer, the input dimension is 1, the output dimension is 64, the convolution kernel size is 3, and the step is 2 to obtain feature A1;

[0055] (b) (b.1) A1 is adaptively averaged to four size scales of 3*3, 6*6, 9*9 and 12*12, and for each scale, a global edge enhancement module is used; this step is used to simulate observing objects from different perspectives, and the small scale of 3*3 can capture very macro and global structure information, and the large scale of 12*12 can retain more local and detailed information, by analyzing the four scales at the same time, the model can understand different sizes of agglomerate structures and their edge features, and avoid missing detection due to different target sizes; (b.2) The global edge enhancement module performs average pooling on the input to obtain the low-frequency information A1', average pooling smooths out the details of the feature map and retains the overall and gently changing regions, then A1 and A1' are subtracted element by element to obtain the high-frequency information A1'', this step uses the original clear image to subtract the blurred version to obtain the details ignored in the blurring process, including some edge, texture and noise data information, as high-frequency information; then A1'' and A1 are added element by element to obtain the global edge enhanced feature A g ; this step adds the extracted high-frequency information back to the original feature map, so that the edge signal is significantly enhanced; (b.3) A g of multiple scales obtained in steps (b.1) and (b.2) are spliced to obtain A' g , and input to an edge information selection module. The edge information selection module includes spatial attention and channel attention mechanisms. The spatial attention mechanism performs average pooling and maximum pooling operations on A' g , splices the pooling results, and generates a general feature through convolution. Then, the channel is separated through deep convolution, and the general feature is expanded along the channel dimension, multiplied element by element with the deep convolution result, and added with a residual connection to obtain the spatial selection feature A'' g This process helps the network focus on important areas containing valuable edge information and weakens the influence of irrelevant background areas. The channel attention mechanism calculates the feature mean through channel-level global average pooling, subtracts the mean from A'' g to obtain a high-frequency component. Then the high-frequency component and A'' gElement-wise multiplication is performed, and the final enhanced feature A2 is obtained through residual connection. This operation selectively emphasizes edge texture information and suppresses irrelevant texture interference, making the network more focused on texture features related to the detection task. By stacking A g Stacking in the channel dimension forms a stacked feature map containing all enhanced edge information from macro to micro. This module uses spatial attention and channel attention mechanisms to selectively calculate edge information, ultimately achieving the effect of edge enhancement only on target information. For the channel attention mechanism, different channels corresponding to different topography or texture features have different effects, so weights are assigned. Similarly, features appearing in different spatial positions also have different effects and will be assigned weights.

[0056] (c) Repeat operation (b) 4 times to obtain the edge information selectively enhanced feature A2. In this feature map, the edge profile of the target agglomerate structure is significantly sharpened and highlighted, while the edge information of the background or other irrelevant objects is suppressed by the attention mechanism.

[0057] (1.2) Input the enhanced picture feature into the self-attention encoder, capture global information through the self-attention mechanism, decouple the target feature and the background feature, and obtain higher-level semantic features B. The self-attention mechanism can associate any point in the graph with other points and determine how strong the association is, allowing the model to focus on enhancing the foreground or region of interest in the image, and thus the model can better distinguish which features truly belong to the "agglomerate structure" and separate them from other ordinary backgrounds composed of topography.

[0058] The steps for processing the fibrous master granule electron microscope image to obtain higher-level semantic features B are as follows:

[0059] (d) Flatten the A2 feature map into a tensor, and superimpose the sine and cosine position encodings for each location's feature to obtain feature A';

[0060] (e) Input A' into the self-attention encoder structure, use the general multi-head attention mechanism to calculate the relevance of each location, and obtain higher-level semantic features B. The number of encoder layers is 1, and the output dimension is 256.

[0061] (1.3) Input the output B of the encoder into the self-attention decoder, which accepts a fixed number of object queries and the output of the encoder, and generates a prediction set of the target through the decoding process. Use a fully connected network to process the output of the decoder to obtain the final output of the class information and position information.

[0062] The steps for processing the fibrous master granule electron microscope image to obtain higher-level semantic features B and output the prediction results are as follows:

[0063] (f) Feature B is input into a self-attention decoder structure, and a fixed number of object query mechanisms are introduced. The total number of layers of the decoder is 6, and the output dimension is 300;

[0064] (g) The decoding output is sent to a fully connected network, and the final output is the predicted position information and category information. The fully connected layer contains three layers, the activation function is ReLU, and each output contains five parameters, which are the position information of the predicted target box in the figure (top left, top right, bottom left, bottom right) and the category information.

[0065] (1.4) Only in the training process, the detection framework inputs feature A into the scale classifier, and through the adversarial training method, the feature obtained by the above framework has cross-scale invariance; the feature A extracted through the adversarial training method can not only recognize the agglomeration structure well, but also eliminate the scale information of the original image; because different experimental conditions and experimenters do not necessarily obtain the same size of microstructure photos when taking the microstructure photos of the master batch, therefore, after this step, whether the input electron microscope image is high magnification or low magnification, the model can extract stable and consistent agglomeration essential features, thereby realizing robust cross-scale recognition.

[0066] The steps of processing the fiber master batch electron microscope image to obtain cross-scale invariant information are as follows:

[0067] (h) Feature A is input into the gradient reversal layer. During training, the gradient reversal layer is equivalent to a transparent layer during the forward propagation process, and does not process A. When the loss gradient is back propagated, the gradient reversal layer will multiply the gradient by a negative parameter -a;

[0068] (i) The output of the gradient reversal layer is accepted, and a fully connected network is used to classify the features. The classification loss is calculated using the cross-entropy loss. The fully connected network has only one layer, the activation function is ReLU, and the output contains one parameter, which is the scale classification result.

[0069] (2) Train the end-to-end detection framework;

[0070] (2.1) Construct the training set: first, randomly extract image segments with a size of 256x256 from the original fiber master batch electron microscope image, perform horizontal flipping and rotation operations on the image, and then use manual labeling to label the agglomeration structure region in the image segment under the guidance of a fiber master batch expert;

[0071] (2.2) Let the iteration number g = 1;

[0072] (2.3) Randomly select a training sample from the training set and input it into the end-to-end detection framework, which outputs the fiber master batch electron microscope image with a rectangular bounding box;

[0073] (2.4) judge whether the iteration end condition is met, i.e. the iteration number g is equal to the maximum iteration number, if yes, the training ends, otherwise, return to step (III) and let g = g + 1;

[0074] (2.5) input the test sample into the end-to-end detection framework trained in step (2.4) for verification.

[0075] (3) input the fiber master batch electron microscope image (as shown in Figure 3 ) into the trained end-to-end detection framework, and output the fiber master batch electron microscope image (as shown in Figure 4 ) with a rectangular boundary box, and the rectangular boundary box area is the agglomeration structure area.

[0076] (3.1) detect the trained end-to-end detection framework:

[0077] Input 70 cross-scale pictures, the enlarged observation range is between 300 nm and 500 μm, containing thousands of cross-scale samples into the trained end-to-end detection framework, then the end-to-end detection framework outputs the prediction result, and compares the prediction result with the true value of the actual sample (i.e. the expert marks the position of the agglomeration of the detection sample), the prediction result consistent with the true value is the correct result, otherwise it is an error detection. Finally, the accuracy rate of the trained end-to-end detection framework for detecting samples in 70 pictures is 95%, the total number of samples is 1053, the number of missed detection (undetected target) is 22, and the number of false detection (mistaking background as target) is 30. Comparing the results identified by the trained end-to-end detection framework of the present application with the true results, it can be seen that the present application can learn the cross-scale invariant feature information representation ability by using the supervised learning method of artificial intelligence, and quickly, accurately and high-sensitivity identify the agglomeration and dispersion ultra-microstructure of fiber color master batch, and the results have very good cross-scale generalization ability, which has important practical significance for actual industrial fiber color master batch preparation.

[0078] Control group 1

[0079] This control group is used to verify the role of channel attention in screening edge feature types. The algorithm flow is modified as follows: in the edge information selection module of step (b.3), only the spatial attention mechanism is retained, and the channel attention calculation step is completely removed; the complete flow of the spatial attention mechanism is: average pooling and maximum pooling operations are performed on A g′ , the pooled results are spliced, a general feature is generated through convolution, channel separation is performed through deep convolution, and the general feature is expanded along the channel dimension, multiplied element by element with the deep convolution result and added with a residual connection to obtain the spatial selection feature A" g ; no longer execute "calculate the feature mean value through channel-level global average pooling, A" gSubtract the mean value to obtain the high-frequency component, then multiply the high-frequency component with A g The operation of element-by-element multiplication is performed; A g is directly output as the final enhanced feature A2, that is, A2=A g . In addition, the two groups use the same multi-scale edge feature extraction structure, Transformer encoder structure, classifier, and the same training data set, hyperparameter setting, and performance evaluation index. The recognition accuracy in the control group is 68%, which means that a real cluster structure edge in the target region and a noise edge with similar intensity are given the same high weight, and because it lacks the ability to distinguish feature channels, its accuracy is lower than that of the algorithm with double screening.

[0080] Control group 2

[0081] This control group is used to verify the role of spatial attention in screening edge spatial positions. The algorithm flow is modified as follows: in the edge information selection module of step (b.3), only the channel attention mechanism is retained, and the spatial attention calculation step is completely removed; "A g′ The average pooling and maximum pooling operations are performed, the pooled results are spliced, and the general features are generated through convolution, followed by channel separation through deep convolution and other spatial attention related operations; the multi-scale spliced features A g′ are directly input into the channel attention mechanism, that is, A g =A g ; the feature mean value is calculated through channel-level global average pooling, A g′ is subtracted from the mean value to obtain the high-frequency component, then the high-frequency component is multiplied with A g′ , and the final enhanced feature A2 is obtained through residual connection. In addition, the two groups use the same multi-scale edge feature extraction structure, Transformer encoder structure, classifier, and the same training data set, hyperparameter setting, and performance evaluation index. The recognition accuracy in the control group is 73%, here because it can enhance the sharpening feature channel, but this enhancement will be applied to every corner of the image, so that some positions related to background noise will also be enhanced, which is equivalent to increasing the interference to the recognition process, therefore, its accuracy will also be lower than that of the algorithm with double screening.

[0082] From control group 1 and control group 2, it can be seen that channel attention and spatial attention are not simply functionally superimposed in this algorithm, but a combination with synergistic effect, both of which are indispensable for achieving high-precision target edge enhancement, and together constitute the improvement of model recognition accuracy.

[0083] Control group 3

[0084] This control group is used to verify the role of making the features obtained by the framework have cross-scale invariance through adversarial training means. It is completely consistent with the original method in network structure, training data, hyperparameter setting, loss function except for the adversarial part. The difference is that the adversarial training module is removed, and the scale classifier and gradient reversal layer are removed from the detection framework. During training, the features A generated by the edge information selective enhancement convolutional network will be directly and only input into the self-attention encoder, and will no longer be sent to any adversarial branch for scale classification; the total loss function of the model total loss = target detection loss. In the training process of the control group, the only goal of the model optimization is to accurately identify and locate the aggregation structure as much as possible, without being guided to eliminate the scale information in the features. The recognition accuracy in the control group is 82%, and the test results can be seen that the adversarial training mechanism can make the model more effectively identify the aggregation structure, without being affected by the observation size, has good generalization ability, and can stably and accurately identify the target in electron microscope images of different magnification.

Claims

1. A method for intelligent identification of fiber master pellet agglomerate structure based on cross-scale invariant information, characterized in that, The method comprises the following steps: (a) constructing an end-to-end detection framework, inputting a micrograph of a fiber master batch to be identified into the framework, and outputting an identification result of the agglomerated structure; (b) the detection framework comprises the following sub-steps: b.1) edge information selective enhancement convolution network processing: processing the input image information through an edge information selective enhancement convolution network to obtain globally enhanced edge features; b.2) spatial attention mechanism feature selection: performing spatial attention mechanism processing on the edge-enhanced multi-scale feature map to emphasize the region containing valuable edge information and weaken the irrelevant region; b.3) channel attention mechanism feature selection: performing channel attention mechanism processing on the feature map processed by the spatial attention mechanism to realize the weighted selection of multi-channel features and highlight important channel information; b.4) self-attention encoder coding: inputting the processed feature into a self-attention encoder to obtain high-level semantic features with the target and background sufficiently decoupled through a multi-head attention mechanism; b.5) self-attention decoder decoding and full connection network output: inputting the encoded features into a self-attention decoder to output a prediction set of the target through an object query method, and then outputting the class and position information results through a full connection network; (c) making the features obtained by the detection framework have cross-scale invariance through adversarial training means; (d) training the detection framework in a supervised learning manner, and using the trained detection framework to intelligently identify the agglomerated structure of the fiber master batch.

2. The method of claim 1, wherein, The edge information selective enhancement convolution network obtains enhanced features through multi-scale pooling operations and edge enhancement operations, and the enhanced features are spliced to form a stacked feature map.

3. The method of claim 2, wherein, The specific processing steps of the global edge enhancement module include: (a) adaptively average pooling the input feature map to multiple different scales; (b) processing each scale feature map through a global edge enhancement module; (c) splicing the enhanced features of multiple scales in the channel dimension to form multi-scale edge enhanced features. The specific processing steps of the global edge enhancement module include: performing average pooling operation on the input scale feature map to obtain low-frequency information of the feature map; performing element-wise subtraction operation on the original scale feature map and the low-frequency information to obtain high-frequency information; performing element-wise addition operation on the high-frequency information and the original scale feature map to obtain the edge-enhanced feature map.

4. The method of claim 3, wherein, The spatial attention mechanism includes: performing average pooling and maximum pooling operations on the spliced multi-scale feature map respectively; splicing the pooling results and generating a spatial attention weight map through a convolution layer; multiplying the spatial attention weight map and the spliced multi-scale feature map to focus on the spatial region containing important edge information.

5. The method of claim 1, wherein, The channel attention mechanism includes: performing channel-level global average pooling on the feature map processed by the spatial attention to calculate the channel feature mean; subtracting the channel feature mean from the feature map to obtain a high-frequency component; and multiplying the high-frequency component and the feature map element by element to selectively enhance the edge texture channel related to the agglomerated structure. The pooling kernel size of the channel-level global average pooling is equal to the spatial dimension of the input feature map, and the weight range of the element-wise multiplication operation is [0.1, 2.0].

6. The method of claim 2, wherein, The output channel number of the convolutional network is between 32 and 128, the convolutional layer step size is 1 or 2, and the number of cascaded edge enhancement modules and attention modules is 1 to 4 times; the image features input are selectively enhanced by the edge information, and the enhanced features are used as the input of the subsequent encoding process; The self-attention encoder adopts a multi-head attention mechanism, the number of heads is 4 to 16, and the output dimension of each layer is 128 to 512; each feature position in the self-attention encoder structure is added with a positive and negative sine position encoding, and is input in the form of a flattened tensor; the target and background features are decoupled through the attention mechanism; The self-attention decoder introduces a fixed number of object queries, the number of decoder layers is 1 to 8, and the output dimension is 64 to 512; the decoding output is input into a fully connected network containing 2 to 5 layers, and the activation function is ReLU; finally, the class and position information of each detection target is output.

7. The method of claim 1, wherein, The specific process of the adversarial training includes: (a) inputting the edge-enhanced features into the gradient reversal layer; (b) during the forward propagation process, the gradient reversal layer acts as a transparent layer and does not process the input features; (c) during the backward propagation process, the gradient reversal layer multiplies the loss gradient by a negative coefficient α, wherein the value of α ranges from 0.1 to 2.0; (d) the output of the gradient reversal layer is input into the scale classifier for scale discrimination; (e) the joint loss function of the scale classification loss and the main task detection loss is calculated for end-to-end optimization.

8. The method of claim 7, wherein, The scale classifier is a single-layer fully connected network with a ReLU activation function, which outputs the scale classification result and calculates the classification loss using a cross-entropy loss function; the joint loss function is a weighted combination of the main task detection loss and the adversarial loss, and the weight ratio ranges from 0.7 to 0.3: 0.9 to 0.1; the learning rate of the adversarial training ranges from 0.0001 to 0.01, and the optimizer uses Adam or SGD; the training set is obtained by cropping and data augmentation on the original microscopic images with a size range of [128x128, 512x512], including one or more of image horizontal flipping, random rotation, brightness disturbance, and noise disturbance.

9. The method of claim 1, wherein, The output agglomerate structure region is labeled with a rectangular bounding box, and four vertex coordinates are predicted for each bounding box, with a coordinate range within the original image width and height.

10. The method according to any one of claims 1 to 9, characterized in that, The number of samples used is 50 to 1000, the observation scale is 300 nanometers to 500 microns, and the unified training and prediction are performed on the end-to-end detection framework.

Images