CN-121999904-A - Textile fiber component analysis system and method based on machine learning

CN121999904ACN 121999904 ACN121999904 ACN 121999904ACN-121999904-A

Abstract

The invention relates to the technical field of detection data processing, in particular to a textile fiber component analysis system and a method based on machine learning, wherein the directional gradient attention module provided by the embodiment has strong directional bias before generating a candidate region proposal, so that a region proposal network can be more easily positioned to the geometric center of the region proposal network according to the extending direction of fibers, thereby generating a more compact and more accurate candidate frame and laying a foundation for the subsequent disentangled adhesion fibers. By focusing attention on several main directions associated with the fibers, the directional gradient attention module naturally suppresses edge response from random directional background noise, decontaminates features, and improves model robustness.

Inventors

ZHOU SHAOQIANG
YAN GANGHUA
ZHOU YUHANG
WANG XIANGXIANG
WANG XIAOQIONG
DING YOUCHAO
DONG SHAOWEI
LI BING

Assignees

南京海关工业产品检测中心
广州海关技术中心
深圳市菲雀兰博科技研究中心有限公司

Dates

Publication Date: 20260508
Application Date: 20260408

Claims (10)

1. A machine learning based textile fiber composition analysis system, comprising: the image acquisition and pretreatment module is used for realizing image acquisition and data pretreatment of the textile fiber sample; the fiber area segmentation module is used for separating each fiber in the pretreated textile fiber sample image from the background thereof to form an independent and complete fiber instance mask; The fiber region segmentation module comprises a main network, a region proposal network and a three-task head, wherein the main network comprises a basic main network, a feature pyramid network and a directional gradient attention module, and the directional gradient attention module is embedded in the back of an output layer of the feature pyramid network and in front of the region proposal network and is used for carrying out directional enhancement processing on a multi-scale feature pyramid generated by the feature pyramid network, and then transmitting an enhanced feature map to the region proposal network for candidate region proposal.
2. The machine learning based textile fiber composition analysis system of claim 1, wherein, The region proposal network generates 9 anchor frames with different scales and length-width ratios at each position of each direction enhanced output characteristic diagram based on a sliding window mechanism, screens out candidate regions containing fiber targets through two classification and frame regression, and combines the candidate regions with overlapping degree larger than a preset threshold value through a non-maximum suppression algorithm to finally output 200-300 high-quality candidate regions.
3. The machine learning based textile fiber composition analysis system of claim 2, wherein, The three-task head comprises a classification head, a frame regression head and a mask head, wherein the classification head classifies the characteristics of the candidate areas through a full connection layer, predicts that the candidate areas belong to fiber images or background images, the frame regression head finely adjusts the coordinates of the candidate areas through the full connection layer to enable the coordinates to be more fit with the actual boundaries of fibers, and outputs a corrected frame, and the mask head is used for outputting an example mask of each fiber target.
4. The machine learning based textile fiber composition analysis system of claim 1, wherein, The image acquisition and preprocessing module comprises a standardized image acquisition interface and an image enhancement processing unit, wherein the standardized image acquisition interface is a standardized hardware interface for a system and is connected with an optical microscope, a digital microscope or a scanning electron microscope, the image enhancement processing unit adopts a non-local mean value filtering algorithm, gaussian noise and spiced salt noise are removed, meanwhile, edge details of textile fibers are reserved, and meanwhile, an adaptive histogram equalization algorithm is applied to improve image contrast.
5. The machine learning based textile fiber composition analysis system of claim 1, wherein, The feature extraction and representation module automatically calculates geometric parameter features of each fiber image to form a basic feature vector, wherein the geometric parameter features comprise dimension features such as length, width/diameter, length-width ratio and curvature, shape features such as circularity, rectangularity, eccentricity and Fourier descriptors, and texture features such as roughness and directionality of fiber surfaces extracted by using a gray level co-occurrence matrix method.
6. A machine learning based textile fiber composition analysis method, wherein the method is run on the machine learning based textile fiber composition analysis system of any of claims 1-5, the method comprising: Image acquisition and data pretreatment of textile fiber samples; Separating each fiber in the pre-processed textile fiber sample image from its background to form an independent, complete fiber instance mask; performing feature extraction operation on the segmented fiber image; And realizing fiber category classification through a machine learning classification and recognition model.
7. The machine learning based textile fiber composition analysis method of claim 6, wherein, In the step of separating each fiber in the pretreated textile fiber sample image from the background thereof to form an independent and complete fiber instance mask, the multi-scale feature pyramid generated by a feature pyramid network is subjected to directional enhancement treatment through a directional gradient attention module, and the enhanced feature image is transferred to the area proposal network for candidate area proposal, wherein the specific steps are as follows: S1, carrying out directional gradient map calculation on a multi-scale feature pyramid generated by the feature pyramid network to obtain four-directional initial gradient response maps; S2, performing feature stitching and attention learning operation on the initial gradient response graphs in the four directions to generate direction attention weights; And S3, carrying out weighted fusion on the multi-scale feature pyramid generated by the feature pyramid network according to the direction attention weight, so as to realize feature enhancement of direction perception.
8. The machine learning based textile fiber composition analysis method of claim 7, wherein, The S1 specifically comprises the steps of carrying out average pooling on a multi-scale feature pyramid generated by the feature pyramid network to obtain a two-dimensional feature map, carrying out convolution operation on the two-dimensional feature map by using Scharr operators K0, K45, K90 and K135 in four directions respectively, and outputting initial gradient response maps in four directions 。
9. The machine learning based textile fiber composition analysis method of claim 8, wherein, The S2 specifically comprises the steps of splicing the initial gradient response graphs in the four directions along the channel dimension to form a temporary characteristic tensor Then, inputting the temporary characteristic tensor into a lightweight convolution layer, and then connecting with a Softmax activation function to output attention mechanics learning weight, wherein the number of output channels of the convolution layer is 4, and the expression is as follows: ; in the formula, To note the mechanical learning weights, W c and b c are the weights and offsets of the convolutional layers.
10. The machine learning based textile fiber composition analysis method of claim 9, wherein, The S3 specifically comprises the steps of multiplying a multi-scale feature pyramid generated by the feature pyramid network with the corresponding attention weight element by element to obtain four gradient feature images with enhanced directions, adding the gradient feature images with enhanced directions to obtain an aggregate feature image fused with direction information, and finally, carrying out residual connection on the aggregate feature image fused with the direction information and the multi-scale feature pyramid generated by the feature pyramid network after carrying out channel adjustment on the aggregate feature image through a 1X 1 convolution layer to obtain a final direction enhanced output feature image.

Description

Textile fiber component analysis system and method based on machine learning Technical Field The invention relates to the technical field of detection data processing, in particular to a textile fiber component analysis system and method based on machine learning. Background The accurate identification of the fiber components of the textile is a key link of quality control, performance evaluation, trade compliance and authenticity identification. In the prior art, semantic/example segmentation models based on deep learning are commonly adopted, and the models are excellent in a standard data set, but when facing an actual textile fiber microscopic image, boundaries of different fibers are fuzzy and even are completely adhered together in a fiber interweaving dense area to form a fiber bundle, and the existing models are often difficult to perfectly segment the fiber bundle into independent individuals, so that one fiber is identified as a plurality of fibers or a plurality of fibers are identified as one fiber. Disclosure of Invention In order to solve the technical problems, the invention provides a textile fiber component analysis system and a method based on machine learning, which are used for solving the problems in the prior art. The invention provides a textile fiber composition analysis system based on machine learning, which comprises the following components: the image acquisition and pretreatment module is used for realizing image acquisition and data pretreatment of the textile fiber sample; the fiber area segmentation module is used for separating each fiber in the pretreated textile fiber sample image from the background thereof to form an independent and complete fiber instance mask; The fiber region segmentation module comprises a main network, a region proposal network and a three-task head, wherein the main network comprises a basic main network, a characteristic pyramid network and a directional gradient attention module, the directional gradient attention module is embedded in the rear of an output layer of the characteristic pyramid network and the front of the Region Proposal Network (RPN) and is used for carrying out directional enhancement processing on a multi-scale characteristic pyramid generated by the characteristic pyramid network, and then the enhanced characteristic map is transmitted to the region proposal network for candidate region proposal. Preferably, the area proposal network generates 9 anchor frames with different scales and length-width ratios at each position of each direction enhanced output characteristic diagram based on a sliding window mechanism, screens out candidate areas containing fiber targets through two classification and frame regression, and combines the candidate areas with the overlapping degree larger than a preset threshold value through a non-maximum suppression algorithm to finally output 200-300 high-quality candidate areas. The three-task head comprises a classification head, a frame regression head and a mask head, wherein the classification head classifies the characteristics of the candidate areas through a full connection layer, predicts that the candidate areas belong to fiber images or background images, the frame regression head finely adjusts the coordinates of the candidate areas through the full connection layer to enable the coordinates to be more fit with the actual boundaries of fibers, and outputs a corrected frame, and the mask head is used for outputting an example mask of each fiber target. Preferably, the image acquisition and preprocessing module comprises a standardized image acquisition interface and an image enhancement processing unit, wherein the standardized image acquisition interface is a standardized hardware interface for a system and is connected with an optical microscope, a digital microscope or a scanning electron microscope, the image enhancement processing unit adopts a non-local mean value filtering algorithm, gaussian noise and spiced salt noise are removed, meanwhile, edge details of textile fibers are reserved, and meanwhile, an adaptive histogram equalization algorithm is used for improving image contrast. Preferably, the feature extraction and representation module automatically calculates geometric parameter features of each fiber image to form a basic feature vector. Preferably, the geometric parameter features comprise size features, such as length, width/diameter, length-width ratio, curvature, shape features, such as circularity, rectangularity, eccentricity and Fourier descriptor, and texture features, such as roughness and directionality of the fiber surface extracted by using a gray level co-occurrence matrix method. According to another aspect of the present invention, there is provided a machine learning-based textile fiber composition analysis method, operating in a machine learning-based textile fiber composition analysis system as described above, the method comprising: Image acquisition an