Evaluation Method for Appearance Uniformity of Polyester Filament Heather Gray Fabric Based on Convolution Feature Fusion

Abstract

This invention belongs to the technical field of evaluation methods, and relates to a method for evaluating the appearance uniformity of polyester filament heather gray fabric using convolutional feature fusion. The method includes sequentially processing images of the polyester filament heather gray fabric... I The input is fed into a pre-trained VGG16 network. Feature maps from the first to third convolutional layers are extracted. Structural feature maps of each feature map are calculated. The structural feature maps are scaled and multiplied point-by-point to obtain a fused structural feature map. Based on the fused structural feature map, an appearance uniformity evaluation index is calculated. U The U-value is used to determine the uniformity of the appearance. This invention achieves effective fusion of multi-scale texture information of polyester filament heather gray fabric, improves the accuracy of uniformity evaluation, reflects local changes in fabric texture, and avoids the subjectivity of manual evaluation.

Description

Technical Field

[0001] This invention belongs to the technical field of evaluation methods, and relates to an evaluation method for the appearance uniformity of polyester filament heather gray fabric based on convolution feature fusion. Background Technology

[0002] Polyester filament heather fabric is typically made by laminating two different colors or materials of filament together in a web pattern to create a heather composite yarn before weaving. This fabric is widely used in the apparel industry due to its flexible and efficient production method, as well as its unique appearance and performance. For heather fabric, the uniformity of the heather effect is one of the important indicators of its quality. Fabric with good uniformity not only has a better visual effect but also ensures stability during subsequent processing and use.

[0003] Currently, most methods for evaluating the uniformity of polyester filament heather fabric rely on manual experience, resulting in high subjectivity and low efficiency. The paper ("Research on Strip Quality Detection Based on Image Processing" [D]. Jiangnan University, 2024. DOI:10.27169 / d.cnki.gwqgu.2024.002547.) uses the coefficient of variation (a statistic used to measure the dispersion of data, defined as the ratio of standard deviation to mean) to calculate the uniformity of the cotton web, considering the differences in light transmittance on carding machines. However, this method depends on the image grayscale range and is sensitive to noise, making it unsuitable for directly evaluating the appearance uniformity of polyester filament heather fabric.

[0004] Therefore, developing an objective, accurate, and efficient method for evaluating the appearance uniformity of polyester filament heather gray fabric is of significant practical importance. Summary of the Invention

[0005] The purpose of this invention is to solve the problems in the prior art and to propose a method for evaluating the appearance uniformity of polyester filament heather gray fabric by convolution feature fusion.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for evaluating the appearance uniformity of polyester filament heather gray fabric based on convolutional feature fusion includes the following steps:

[0008] (1) Feature map extraction;

[0009] Image I of polyester filament heather fabric was input into a pre-trained VGG16 network, and feature maps of the first to third convolutional layers were extracted, denoted as F1, F2, and F3, respectively. The pre-training process of the VGG16 network is described in the reference (Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014.).

[0010] (2) Calculation of structural feature diagram;

[0011] Calculate the structural feature maps of F1, F2, and F3, denoted as S1, S2, and S3 respectively. The calculation formulas are as follows:

[0012] ;

[0013] In the formula, j = 1, 2, 3; Representing F j The structural feature map of the i-th channel; M represents F j The total number of central channels;

[0014] (3) Feature fusion;

[0015] Obtain the structural feature map S0 of image I, and scale S2 and S3 to the same size as S1. Then, multiply S0, S1, the scaled S2, and the scaled S3 point by point to obtain the fused structural feature map. ;

[0016] (4) Calculation of appearance uniformity evaluation index;

[0017] The formula for calculating the appearance uniformity evaluation index U is as follows:

[0018] ;

[0019] In the formula, N represents The total number of pixels; represent The structural feature value of the k-th pixel;

[0020] (5) Evaluation of appearance uniformity;

[0021] Determine whether U≤4.1 is true. If it is true, the appearance uniformity is good; otherwise, the appearance uniformity is poor.

[0022] The VGG16 network achieves efficient extraction of multi-scale features through multiple stacking operations of small convolutional kernels. This unique structural design endows it with powerful feature filtering capabilities: on the one hand, it can effectively suppress uniformly distributed conventional features; on the other hand, it emphasizes local irregular details.

[0023] When faced with a polyester filament heather gray fabric exhibiting poor uniformity in appearance, the localized uneven distribution of features on the fabric surface becomes the focus of the VGG16 network. During the network's forward propagation, different channels and convolutional layers work together to continuously reinforce and highlight these uneven features. As the network deepens, these reinforced features are ultimately reflected in the structural feature map, manifesting as significantly increased values. The uniformity U-value calculated based on the structural feature map also increases accordingly, intuitively reflecting the degree of unevenness in the fabric's appearance. Conversely, if the polyester filament heather gray fabric has good uniformity, the VGG16 network effectively suppresses its uniformly distributed features. This suppression results in relatively small values ​​in the structural feature map, leading to a smaller uniformity U-value, accurately reflecting the uniformity of the fabric's appearance.

[0024] As a preferred technical solution:

[0025] As described above, in the method for evaluating the appearance uniformity of polyester filament heather gray fabric by convolutional feature fusion, the process of obtaining the structural feature map in steps (2) and (3) is as follows: (a) A fixed-size window is slid across the image, and the variance of all pixels in each window is calculated. and mean ratio (a) Use it as the structural feature value of the center pixel of the window; (b) Repeat step (a) until all pixels of the image are traversed to obtain the structural feature map.

[0026] The above-described method for evaluating the appearance uniformity of polyester filament heather gray fabric through convolutional feature fusion uses a window size of 20 pixels × 20 pixels and a step size of 1 pixel during the acquisition of the structural feature map. The sliding method is: first from left to right, then from top to bottom.

[0027] In the above-mentioned method for evaluating the appearance uniformity of polyester filament heather gray fabric by convolution feature fusion, in step (2), the total number of channels in F1, F2 and F3 are 64, 128 and 256 respectively.

[0028] In the above-mentioned method for evaluating the appearance uniformity of polyester filament heather fabric by convolution feature fusion, in step (3), the scaling adopts bilinear interpolation.

[0029] Beneficial effects

[0030] This invention achieves effective fusion of multi-scale texture information of polyester filament heather fabric by extracting feature maps from the first to third convolutional layers of the VGG16 network and calculating their structural feature maps. The feature maps of different convolutional layers contain texture features at different scales. By fusing these features, the texture structure of the fabric can be described more comprehensively, improving the accuracy of uniformity evaluation. Using the ratio of variance to mean as the structural feature value reflects the local variations in the fabric texture, avoiding the subjectivity of manual evaluation. Attached Figure Description

[0031] Figure 1 A schematic diagram of extracting feature maps from the first to third convolutional layers of a VGG16 network;

[0032] Figure 2 This is a structural feature diagram of a fabric with low appearance uniformity.

[0033] Figure 3 This is a structural feature diagram of a fabric with moderate appearance uniformity.

[0034] Figure 4 This is a structural feature diagram of a fabric with high appearance uniformity. Detailed Implementation

[0035] The present invention will be further described below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0036] A method for evaluating the appearance uniformity of polyester filament heather gray fabric based on convolutional feature fusion includes the following steps:

[0037] (1) Feature map extraction;

[0038] Image I of polyester filament heather fabric is input into a pre-trained VGG16 network to extract feature maps from the first to third convolutional layers (e.g., ...). Figure 1 As shown), they are denoted as F1, F2, and F3 respectively; the pre-training process of the VGG16 network is described in the reference (Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition[J]. arXiv preprint arXiv:1409.1556, 2014.).

[0039] (2) Calculation of structural feature diagram;

[0040] Calculate the structural feature maps of F1, F2, and F3, denoted as S1, S2, and S3 respectively. The calculation formulas are as follows:

[0041] ;

[0042] In the formula, j = 1, 2, 3; Representing F j The structural feature map of the i-th channel; M represents F j The total number of middle channels (the total number of middle channels in F1, F2, and F3 are 64, 128, and 256 respectively).

[0043] (3) Feature fusion;

[0044] Obtain the structural feature map S0 of image I. Then, scale S2 and S3 to the same size as S1 using bilinear interpolation. Finally, multiply S0, S1, the scaled S2, and the scaled S3 point-by-point to obtain the fused structural feature map. ;

[0045] In steps (2) and (3), the process of obtaining the structural feature map is as follows: (a) Using a window with a size of 20 pixels × 20 pixels and a step size of 1 pixel, slide it on the image first from left to right and then from top to bottom, and calculate the variance of all pixels in each window. and mean ratio (a) and use it as the structural feature value of the corresponding center pixel of the window; (b) repeat step (a) until all pixels of the image are traversed to obtain the structural feature map;

[0046] (4) Calculation of appearance uniformity evaluation index;

[0047] The formula for calculating the appearance uniformity evaluation index U is as follows:

[0048] ;

[0049] In the formula, N represents The total number of pixels; represent The structural feature value of the k-th pixel;

[0050] (5) Evaluation of appearance uniformity;

[0051] Determine whether U≤4.1 is true. If it is true, the appearance uniformity is good; otherwise, the appearance uniformity is poor.

[0052] To verify the effectiveness and accuracy of the uniformity evaluation method provided by this invention, experiments were conducted using three samples with different uniformities (e.g., ...). Figure 2 , Figure 3and Figure 4 , Figures 2 to 4 (Appearance uniformity increases sequentially), and the evaluation results are shown in Table 1:

[0053] Table 1 Comparison of appearance uniformity results

[0054]

[0055] Table 1 shows that the uniformity value (U) calculated by the method of this invention is completely consistent with the ranking judged by the human eye: Figure 2 The U-value is the largest (4.34 points), corresponding to a human eye score of 4.5 points (the worst uniformity, ranked 1st). Figure 4 The U-value was the smallest (3.63 points), corresponding to a human eye score of 2.0 points (best uniformity, ranked 3rd). There is a strong positive correlation between the human eye score and the U-value of this invention; that is, the larger the U-value, the worse the uniformity of the human eye's judgment, which aligns with practical physical meaning. Regarding accuracy, traditional human eye evaluations are influenced by subjective factors (such as observation angle and fatigue), while this invention eliminates individual differences through quantitative calculations. For example, Figure 2 and Figure 3 The difference in scores given by the human eye was 1.3 points, while the difference in U-values ​​given by this invention was 0.27. Both reflect a significant difference in uniformity between samples, and the numerical fluctuation of this invention is more stable, proving that this method can objectively and accurately quantify uniformity. In summary, the experimental data show that the method of this invention is highly consistent with the results of human eye evaluation, and improves the accuracy and reliability of uniformity evaluation through quantitative indicators, providing a standardized evaluation tool for the industry.

Claims

1. A method for evaluating the uniformity of the appearance of a polyester filament gray fabric with convolution feature fusion, characterized in that, Comprising the following steps: (1) Feature map extraction; The image I of the polyester filament gray fabric is input into the pre-trained VGG16 network, and the feature maps of the first to third convolutional layers are extracted, respectively denoted as F1, F2, and F3; (2) Structural feature map calculation; The structural feature maps of F1, F2, and F3 are calculated, respectively denoted as S1, S2, and S3, and the calculation formula is as follows: ； wherein j = 1, 2, 3; represents F j the structure feature map of the i-th channel; M represents F j the total number of channels in In step (2), the total number of channels in F1, F2, and F3 is 64, 128, and 256, respectively; (3) Feature fusion; A structure feature map S0 of the image I is obtained, and after S2 and S3 are scaled to the same size as S1, S0, S1, the scaled S2 and the scaled S3 are multiplied point by point to obtain a fused structure feature map ; In step (2) and step (3), the acquisition process of the structure feature map is as follows: (a) a fixed size window is used to slide on the image, and the ratio of the variance and the mean value of all pixel points in each window is calculated as the structure feature value of the center pixel point position of the window; (b) repeat step (a) until all pixel points of the image are traversed, that is, the structure feature map is obtained. and mean value ratio , and the structure feature value of the center pixel point position of the window is obtained; (b) repeat step (a) until all pixel points of the image are traversed, that is, the structure feature map is obtained. (4) Appearance uniformity evaluation index calculation; The formula for calculating the appearance uniformity evaluation index U is as follows: ； In the formula, N represents the total number of the middle pixel points; represents the structural feature value of the kth middle pixel point (5) Appearance uniformity evaluation; Determine whether U≤4.1 is true, if yes, the appearance uniformity is good; otherwise, the appearance uniformity is poor.

2. The method according to claim 1, wherein the method is characterized by, In the process of obtaining the structural feature map, the size of the window is 20 pixels x 20 pixels, the step is 1 pixel, and the sliding mode is: first from left to right, then from top to bottom.

3. The method according to claim 1, wherein the method is characterized by, In step (3), the scaling adopts the bilinear interpolation method.

Images