CN-122023753-A - Intelligent recognition method for multi-domain feature fusion fiber master batch aggregation structure

CN122023753ACN 122023753 ACN122023753 ACN 122023753ACN-122023753-A

Abstract

The invention relates to the field of computer vision and material microscopic analysis intersection, in particular to an intelligent recognition method of a fiber master batch aggregation structure based on multi-domain feature fusion, which comprises the steps of constructing and training a multi-feature fusion network model, inputting a microstructure image to be detected into the trained multi-feature fusion network model, and outputting a microstructure image with a rectangular boundary frame, wherein the rectangular boundary frame region corresponds to the aggregation structure region of the fiber master batch; compared with the traditional manual analysis, the invention can effectively improve the analysis efficiency and accuracy, and has wide application prospect in the fields of material science, manufacturing industry and scientific research.

Inventors

Kui Bing
REN TIANFEI
WANG HUAPING
XU YIMING

Assignees

苏州宝丽迪材料科技股份有限公司

Dates

Publication Date: 20260512
Application Date: 20250523

Claims (10)

1. The intelligent fiber master batch aggregation structure identification method based on multi-domain feature fusion is characterized by comprising the following steps of: step 1, obtaining microstructure images of fiber master batches, wherein the microstructure images are acquired by an electron microscope or an optical microscope; Step 2, preprocessing the microstructure image and labeling a region with an agglomeration structure; step 3, extracting high-frequency texture features from the preprocessed image, and extracting low-frequency morphological features; Step 4, inputting the high-frequency texture features and the low-frequency morphological features into a fully-connected network, and constructing a model by taking the position with the aggregation structure area as an output value; and 5, inputting the image to be detected into a model, and identifying the agglomeration structure.
2. The intelligent recognition method of the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 1 is characterized in that in the step 3, the step of extracting high-frequency texture features comprises the steps of (a) constructing a neural network of an encoder-decoder structure, capturing a multi-scale feature map of an input image through downsampling operation of an encoder, (b) performing discrete cosine transform on each channel of the multi-scale feature map, converting the discrete cosine transform from a spatial domain to a frequency domain, selecting high-frequency components based on energy distribution, and (c) calculating response weights of the high-frequency components, performing weighted fusion on channels of an original feature map, and generating a high-frequency enhanced texture feature map so as to strengthen detail texture information in the image.
3. The intelligent recognition method of the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 2, wherein the encoder-decoder structure is a U-Net architecture and comprises symmetrical encoding paths and decoding paths, the encoding paths are gradually downsampled through a convolution layer, the channel number of a feature map is increased, the decoding paths are gradually upsampled through a transposed convolution layer, and the fusion features are connected with corresponding levels of the encoding paths through hops.
4. The intelligent recognition method for the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 3 is characterized in that an encoding path is composed of a plurality of layers of downsampling modules, each layer is subjected to feature extraction by adopting convolution kernels with fixed sizes, the number of filters is increased step by step along with the deepening of the layers, a decoding path is composed of a plurality of layers of upsampling modules, the resolution of each layer is recovered step by transposed convolution, the number of filters is reduced step by step along with the deepening of the layers, the output of each layer of an encoder is sequentially subjected to standardization processing and nonlinear activation, feature fusion is achieved by jump connection with the corresponding layers of the decoding path, and the filling mode of all convolution operations is symmetrical filling, so that the continuity of the size of a feature map is ensured.
5. The intelligent recognition method for the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 2 is characterized in that in the step (b), the Discrete Cosine Transform (DCT) is specifically implemented by (b 1) dividing a multi-scale feature map into a plurality of groups according to channels, wherein each group comprises a plurality of channels, (b 2) independently performing DCT on each group of channels to extract frequency domain components, and (b 3) selecting the first k high-frequency component groups according to the distribution proportion of frequency domain energy, and calculating the response weight of each component.
6. The intelligent recognition method for the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 1 is characterized in that the selection mode of the high-frequency components is that frequency domain components of each channel are ordered from high to low according to frequency, k components farthest from a frequency domain origin are selected, the response weight is normalized and calculated through the energy duty ratio of each component, the number k of channel groups is 10-20, the number of channels in each group is equally distributed, the high-frequency enhanced texture feature map is generated by multiplying the weighted frequency domain response with an original feature map obtained by an encoder element by element, and then the decoder gradually restores the spatial resolution.
7. The intelligent recognition method for the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 1 is characterized in that the extraction of low-frequency morphological features in the step 3 comprises the following steps of (a) carrying out discrete wavelet transformation on a multi-scale feature map output by an encoder to separate low-frequency components, (b) generating a space weight map by using a gating convolution through an attention mechanism to enhance the response strength of a key contour region in the low-frequency components, and (c) inputting the optimized low-frequency components into a feature optimization module to output the low-frequency feature map representing the macroscopic morphology of the fiber master batch.
8. The intelligent recognition method of the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 7, wherein in the step (a), the input of the Discrete Wavelet Transform (DWT) is a feature map output by a certain level of an encoder, the separation of the low-frequency components is realized through multi-resolution analysis, and profile information representing the overall shape of the fiber master batch is reserved.
9. The intelligent recognition method for the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 7 is characterized in that in the step (b), the attention mechanism generates a space weight graph through a gating convolution operation and multiplies the space weight graph by the low-frequency component element by element, and dynamically focuses on an aggregation boundary or a shape mutation region, the feature optimization module in the step (c) comprises a residual error connection structure, and the expression capability of the low-frequency feature is enhanced through a multi-layer convolution operation, so that information loss is avoided.
10. The intelligent recognition method of the fiber master batch aggregation structure based on multi-domain feature fusion according to claim 1, wherein the preprocessing in the step 2 comprises normalization processing and data enhancement operation, and the data enhancement comprises horizontal/vertical overturning, rotation, contrast adjustment, gaussian noise injection and sharpening processing.

Description

Intelligent recognition method for multi-domain feature fusion fiber master batch aggregation structure Technical Field The invention belongs to the field of intersection of computer vision and material microscopic analysis, and particularly relates to an intelligent recognition method for a fiber master batch aggregation structure based on multi-domain feature fusion. Background The fiber master batch is used as a core raw material for textile manufacture, and the components of the fiber master batch mainly comprise carrier resin, pigment, dye, functional powder and auxiliary agent. Under the microscopic scale, the fiber master batch is easy to adsorb water molecules, dust or charged particles in the air due to higher surface energy, so that charges are continuously accumulated. This phenomenon exacerbates the interactions of electrostatic forces, van der Waals forces, liquid bridge forces, etc. between the fiber masterbatches, causing them to exhibit a significant tendency to agglomerate in air. The formation of the agglomerated structure not only causes uneven distribution of the colorant in the carrier resin and causes the chromatic aberration problem of the fiber product, but also reduces the plasticity of the material in subsequent processing, and seriously affects the mechanical property of the fiber and the quality stability of the finished product. At present, the assessment of the agglomeration effect of the fiber master batch mainly depends on the traditional mechanical test method (such as a tensile test and a bending test), and the agglomeration degree is estimated by indirectly reflecting the change of the material performance. However, the method has the limitations that firstly experimental conditions (such as temperature and humidity and loading rate) have larger influence on the result, the testing precision is difficult to ensure, and secondly, the mechanical test cannot intuitively capture the morphology, the distribution and the dynamic evolution process of the agglomeration structure under the microscopic scale, so that the guiding effect on process optimization is limited. In recent years, deep learning technology has made breakthrough progress in the field of image processing, and the performance of tasks such as image classification, target detection and the like is remarkably improved by replacing traditional manual feature extraction through end-to-end feature learning. However, existing deep learning methods still have problems such as insufficient detail capture and insufficient robustness in the face of high resolution, complex ultrastructural microstructure images. In the prior art, patent application CN119223817A 'method for detecting EVA high-concentration color master batch based on multispectral imaging' proposes a method for detecting color master batch based on multispectral imaging, color difference and defect identification are realized through multispectral image analysis, and an intelligent feedback control system is utilized to regulate a production line. Although the method provides a visual result to assist manual quality inspection, the core of the method still depends on manual final judgment, subjective error introduction risk exists, a special algorithm is not designed for a microscopic agglomeration structure of the fiber master batch, and full-process unmanned accurate identification cannot be realized. Therefore, an intelligent method capable of automatically and accurately identifying the agglomerate structure of the fiber master batch is developed, the limitations of the traditional mechanical test and the existing image processing technology are broken through, and the method has important industrial application value for improving the quality control efficiency of the fiber master batch preparation process and reducing the production cost. Disclosure of Invention The invention aims to overcome the defects that the identification of the agglomerate structure of the fiber master batch depends on manual judgment and the traditional deep learning model is insufficient in capturing microcosmic details in the prior art, and provides an intelligent identification method of the agglomerate structure of the fiber master batch based on multi-domain feature fusion, so that full-process automation and high-precision detection are realized, and the quality control efficiency of an industrial preparation process is improved. In order to achieve the above purpose, the present invention adopts the following technical scheme: an intelligent recognition method for a fiber master batch aggregation structure based on multi-domain feature fusion comprises the following steps: step 1, obtaining microstructure images of fiber master batches, wherein the microstructure images are acquired by an electron microscope or an optical microscope; Step 2, preprocessing the microstructure image and labeling a region with an agglomeration structure; step 3, extracting high-frequency tex