3.1. Feature Extraction Methods and Image Segmentation of Zhuang Brocade Patterns

It systematically collects and classifies representative patterns of Zhuang brocade, using computer image recognition and extraction methods. This can achieve fast and accurate extraction of Zhuang brocade patterns on the ground of their features, laying a preliminary research foundation for subsequent automatic generation and application of patterns. The pattern feature extraction process is shown in Fig. 1.

As shown in Fig. 1, when extracting pattern features, the first step is to clarify the target object of pattern feature extraction, which usually involves systematic collection and classification of Zhuang brocade patterns to ensure the comprehensiveness and pertinence of the research. The steps of constructing extraction algorithms involve designing and developing algorithms that can accurately identify and quantify pattern features, including but not limited to color, texture, and organizational structure recognition. After the algorithm construction is completed, experiments are conducted to verify the effectiveness of the algorithm. The extraction algorithm will analyze the main colors and their distribution in the Zhuang brocade pattern, quantify the color space, and provide a color basis for subsequent design. Pattern feature extraction focuses on identifying and reproducing specific graphic elements of Zhuang brocade patterns. Organizational structure feature extraction focuses on the layout and combination of patterns, analyzing how patterns are organized and arranged in space. Through the collaborative work of three branches, comprehensively extract the key features of Zhuang brocade patterns. In the process of feature extraction, a dataset containing color, pattern, and organizational structure features will be gradually established as the basis for algorithm development and optimization. The color extraction process is shown in Fig. 2.

In Fig. 2, during the color extraction process, samples of any color were selected as examples, and the K-means algorithm was utilized for displaying the hexadecimal encoding and proportion of the main colors in the Zhuang brocade image in an intuitive pie chart. Using the K-means method to binarize the selected image, the main color tone is encoded in hexadecimal and matched with the color closest to the query result in the color dictionary to generate a scale map. This study uses the relative total variance model and Grab cut interactive image segmentation technology to address the impact of sample image blurring and complex textures on the extraction results in traditional pattern extraction. The color extraction process of the pattern is shown in Fig. 3.

To reduce the impact of the texture of the fabric on pattern information processing, a relative total variation model is utilized for achieving smooth extraction of fabric texture. The relative total variation formula for any point in image smoothing processing is shown in formula (1).

In formula (1), $\varepsilon$ is a fixed constant with a value greater than 0, the function is to ensure that the values are always meaningful and to avoid calculation errors; $\Phi(\cdot)$ represents the contribution weight function; $\Psi(\cdot)$ representing texture feature functions; $p$ represents the point of analysis. Its function is for ensuring that the value is always meaningful to avoid calculation errors. The creation of even total and intrinsic variations in the $x$ and $y$ directions at any point $p$ in the graph is shown in formula (2).

In formula (2), $R(p)$ is the region centered on $p$, and $q$ is any point in the variational region $R(p)$. $x$, $y$ is the partial differential of pixel $q$ in the $x$ and $y$ directions, and $h_{p,q}$ is the weight function defined in space. Its expression is shown in formula (3).

In formula (3), $p_x$ and $p_y$ respectively represent the horizontal and vertical coordinates of point $p$, while $\sigma$ controls the spatial scale of the window. The RTV algorithm model is shown in formula (4).

In formula (4), $S$ represents the input Zhuang brocade pattern, $f$ represents the extracted structure of the image, and $\lambda$ is a fixed regularization parameter, that is, the smoothness coefficient. The above model indicates that the degree of image smoothing processed by the relative total variation algorithm mainly depends on two parameters: the smoothness coefficient $\lambda$ and the spatial scale parameter $\sigma$. The algorithm used in image segmentation processing is the GrabCut algorithm, which requires users to manually label the areas to be segmented with rectangles. The connection between Zhuang brocade patterns and patterns, as well as between patterns and tissues, is not high. The GrabCut algorithm is used for selective local segmentation of Zhuang brocade patterns. Its operation first defines a rectangle $T$ containing the target in the image, with the outer region of the rectangle as the background region $T_B$ and the inner region of the rectangle as the foreground region $T_F \neq \emptyset$ and $T_U = T_B$; When the initialization label $a_n = 0$ of a pixel point $n$ in $T_B$ is considered as a background pixel, and when the initialization label $a_n = 1$ of a pixel $n$ in $T_U$ is considered as a possible target pixel; It uses the K-means clustering algorithm to cluster the foreground region $T_F$ and background region $T_B$ into $K$ classes, obtaining $K$ Gaussian mixture models (GMM); Using a Gaussian mixture model to model the background and foreground, two GMM parameters $\theta$ and DE with initial values $(\pi, \mu, \sigma)$ are obtained, where $\pi$ is the weight, $\mu$ is the mean vector, and $\sigma$ is the covariance matrix. The GMM component of each pixel in foreground $T_U$ is the RGB value of the target pixel $n$, and the $k_n$-th Gaussian component of pixel $n$, as shown in formula (5).

It trains and learns GMM parameters for the given image data Z.

The segmentation is performed using the maximum minimum retention algorithm to obtain the minimum energy as shown in formula (7).

It optimizes the GMM model and segmentation results until the energy $E$ reaches a convergence state, thereby outputting high-quality images ^{[18, 19]}. After image preprocessing and segmentation, to objectively evaluate the accuracy of pattern segmentation, it is necessary to calculate the pixel accuracy $PA$, as shown in formula (8).

In formula (8), $p_{ij}$ serves as the total pixels with real pixel category $i$ predicted as category $j$, and $p_{ii}$ serves as the total pixels with real pixel category $i$ predicted as category $i$. Image binarization is a fundamental technique in image processing that preserves sufficient feature information. It is mainly on the ground of the grayscale characteristics of the image, dividing it into two parts: background and foreground, in order to achieve more precise image processing results. The main core idea is to find the maximum grayscale level $k$, and then divide the image into black and white colors that are greater than the threshold and less than the threshold. The specific algorithm principle is to treat grayscale image $F$ as a $M \times N$ matrix, where the pixel value is $(0, 255)$. The probability of a pixel having a grayscale of $i$ is shown in formula (9).

Among them, the value of $\sum_{i=0}^{255} p_i$ is shown in formula (10).

The segmentation threshold for foreground and background is denoted as $k$, and the image is divided into $C_A$ and $C_B$ on the ground of the threshold. The calculation method for the global mean $m_G$ of the image is shown in formula (11).

In formula (11), $p_A$ and $p_B$ are the classification probabilities of images $C_A$ and $C_B$, with grayscale means $m_A$ and $m_B$, respectively. The cumulative mean of grayscale level $K$ is $m$, and the expression of variance is shown in formula (12).

It brings formula (11) into formula (12), as shown in formula (13).

It deforms formula (13) as shown in formula (14).

After obtaining the maximum threshold, the image is segmented by binarization, as shown in formula (15).

3.2. Zhuang Brocade Pattern Design Algorithm on the Ground of Relative Coordinates and Regional Content Replacement

This study investigates the Zhuang brocade pattern design algorithm on the ground of graphic feature elements. It takes the Zhuang brocade pattern as the research object, establishes a Zhuang brocade pattern design method and process on the ground of relative coordinates and regional content replacement by encoding feature materials. This can achieve the exchange of pattern and organizational form target area content. It matches the color scheme of Zhuang brocade, generates multiple types of Zhuang brocade, and evaluates their style similarity. The research process of pattern design is shown in Fig. 4.

As shown in Fig. 4, the pattern encoding process is a key step in achieving automatic generation and customized design of Zhuang brocade patterns. The process is mainly divided into three independent sub processes, including color sample coding, pattern sample coding, and tissue sample coding. Color sample coding focuses on the colors and their proportions used in Zhuang brocade patterns. Establish a color sample library by quantifying and classifying colors. The library will contain the main colors and their corresponding hexadecimal values and proportions that appear in the Zhuang brocade pattern, providing reference for the color matching of the pattern. Pattern sample coding is the analysis and coding of specific graphic elements in Zhuang brocade patterns. Including extracting features such as shape, lines, and texture of patterns, and converting them into editable and reconfigurable data formats. Organizing sample coding involves the spatial arrangement and organizational structure of patterns. During the process, it is necessary to analyze how patterns are laid out in Zhuang brocade, including organizational forms such as repetition, symmetry, and rhythm, and encode them into data structures that can be processed by algorithms. The algorithm framework for replacing relative coordinates and regional content is shown in Fig. 5.

In Fig. 5, it can be seen that the Zhuang brocade pattern reconstruction algorithm generates corresponding mapping files by binary encoding each element. On this basis, the received pattern genes are decoded to obtain the corresponding pattern elements, and their localization is performed. This can make it consistent with the center coordinates of the pattern to be reconstructed, and use the position displacement method specified in the image to obtain the reconstructed pattern, and color it to obtain the final effect. In the algorithm framework, its running steps include: generating element encoding mapping files; Decoding; Pattern reconstruction and pattern coloring. Pattern reconstruction mainly involves reorganizing the extracted pattern patterns and organizational forms, using various combination methods to generate a novel Zhuang brocade pattern in batches ^[20]. The center point (CP) of the pattern is aligned with the CP of the organizational form to be embedded, thus completing the recombination of the pattern and organizational form. The process of pattern position matching is shown in Fig. 6.

Fig. 6 shows that in Fig. 6(a), when the pasted area does not exist or the coordinates are (0,0), it indicates that the top left corner (TLC) of the pattern coincides with the TLC of the prohibited image, and the pattern is pasted to the TLC of the organizational form; In Fig. 6(b), if the coordinates of the CPs of the organizational pattern area and the pasted area exist, then when the pasted area is the same as it, the upper left corner of the pattern is located at the CP of the pasted area, and the pattern and organizational attributes overlap; In Fig. 6(c), if the pattern image wants to be offset upwards, the CP of the pattern image will coincide with the CP of the pasted area. When using research methods for Zhuang brocade pattern design in practical scenarios, in the initial stage of technical operation, requirement analysis is first conducted to clarify design goals and expected effects. Using computer image recognition technology to extract features from Zhuang brocade patterns. Identification and classification of colors, patterns, and organizational structures. Encode the extracted features to form color sample encoding, pattern sample encoding, and tissue sample encoding. Develop pattern design algorithms based on encoded data, including the development and optimization of algorithms for pattern recombination, color matching, and style similarity evaluation. And through computer programs, multiple design schemes are automatically output to achieve the automatic generation of Zhuang brocade patterns. The automatically generated patterns are evaluated, including style similarity, color coordination, and pattern innovation. Select the solution that best meets the design objectives for further refinement and improvement. After user customization and final adjustment by the designer, the output of the Zhuang brocade pattern meets the requirements of modern packaging design.

4.1. Analysis of the Application Effect of Zhuang Brocade Pattern Design Algorithm on the Ground of Relative Coordinates and Regional Content Replacement

For verifying the effect of pattern reconstruction, the study first decodes the selected patterns, calls the decode function, and passes a parameter code = “1000110100100100010”. The code is separated and searched for the corresponding code value in the imported mapping file code_map. json, and returns the color_group, organization structure, pattern A, and pattern B. Subsequently, pattern reconstruction and pattern coloring are performed. The color function is selected, and the reconstructed image obtained from the first two steps is combined with the color input. The resulting reconstructed pattern coloring scheme is shown in Fig. 7.

In Fig. 7, through case analysis, it is confirmed that using relative coordinates and regional content substitution methods with Zhuang brocade as the main feature element can effectively achieve automatic generation of Zhuang brocade patterns. On this basis, utilizing the color, pattern, organization and other characteristics of Zhuang brocade patterns, rapid generation of Zhuang brocade patterns can be achieved. We compared the reconstructed Zhuang brocade patterns with the collected samples and found that there were stylistic differences between them. On this basis, on the ground of user similarity evaluation, the style similarity of Zhuang brocade patterns is evaluated for testing the effectiveness of this method. The similarity evaluation of Zhuang brocade pattern styles was conducted using 30 users as subjects to evaluate the styles of 10 patterns. This survey consists of 30 people. There are three evaluation methods, namely “similar”, “dissimilar”, and “neutral”.

As shown in Table 1, the data displayed in the generated graphic style similarity evaluation table shows that 10 generated graphics were evaluated for style similarity, of which 9 were similar, accounting for 90%. By evaluating the style similarity of 10 generated patterns, an average level of 75.80% was obtained, indicating that participants with and without knowledge of Zhuang brocade had a higher level of identification with clothing patterns. The use of pattern genes to generate patterns, tissue morphological features, and color schemes effectively inherits the style and characteristics of Zhuang brocade. The convergence and accuracy of the algorithm proposed in this article were compared with the commonly used Bresenham algorithm for image rendering. The comparison results are shown in Fig. 8.

Fig. 8 shows that compared to the Bresenham algorithm, the proposed algorithm consistently maintains significant advantages in accuracy and convergence at the same number of iterations. When the number of iterations reaches 200, the accuracy increases from 0.723 to 0.905, indicating a significant improvement in accuracy as the number of iterations increases. This indicates that the model gradually optimizes during the learning process and can better fit the training data.

4.2. Color extraction results in Zhuang brocade features

By calculating the proportion of each color in the image and creating a visual histogram, the proportion of each recognized color in the image can be clearly displayed. The proportion of color cluster pixels is shown in Table 2.

In Table 2, this image contains five main colors, each with its own name, pixel, and proportion. By viewing the data in the table, you can clearly understand the detailed information of each color, including their distribution and proportion in the image. This table presentation is very intuitive and allows for easy analysis and understanding of color information in images. After extracting colors from the image, the proportions of each color are shown in Fig. 9.

In Fig. 9, the discount represents the standard values obtained from color extraction, with colors with RGB values of # 2b314f accounting for 12.5% of all colors; Colors with RGB values of # 474a5f account for 23.2% of all colors; Colors with RGB values of # 645d69 account for 26.9% of all colors; Colors with RGB values of # 7e6e75 account for 26.1% of all colors; Colors with an RGB value of # 93848c account for 11.3% of all colors. Fig. 9(a) shows the relationship between the size of each color region, with a small difference from the standard value; Fig. 9(b) shows that the color ratio obtained by using other methods for color extraction differs significantly from the standard value. The image segmentation algorithm manually marks the object segmentation area with a rectangle, uses the pattern outside the rectangle as the background, and segments the pattern inside the rectangle. The connectivity between Zhuang brocade patterns and patterns, as well as between patterns and tissues, is poor. It uses the GrabCut algorithm to perform local selection segmentation on four sample patterns A, B, C, and D, and calculates the pixel accuracy and intersection to union ratio of the segmentation results. Then, on the ground of pixel accuracy and intersection to union ratio as evaluation criteria. The outcomes obtained are shown in Fig. 10.

In Fig. 10, Fig. 10(a) shows the pixel accuracy of experimental samples A, B, C, and D obtained by using the GrabCut algorithm to extract pattern features, which are 0.991, 0.986, 0.965, and 0.989, respectively; In the results obtained from feature extraction of pattern patterns using other methods, the pixel accuracy of experimental samples A, B, C, and D were 0.726, 0.668, 0.703, and 0.619, respectively; Fig. 10(b) shows the results obtained by using the GrabCut algorithm to extract pattern features. The intersection to union ratios of experimental samples A, B, C, and D are 0.986, 0.819, 0.974, and 0.968, respectively; In the results obtained from feature extraction of pattern patterns using other methods, the intersection to union ratios of experimental samples A, B, C, and D were 0.736, 0.596, 0.527, and 0.688. The research results indicate that the GrabCut algorithm has good segmentation performance and can be further applied.

Table 3 compares the performance of the method based on relative coordinates and region content replacement proposed in the article with CNN based deep learning and GAN based generative adversarial network methods on multiple quantitative metrics. The pattern generation time of the article method is 0.83 seconds per pattern, which is faster than CNN based deep learning (1.17 seconds per pattern) and GAN based generative adversarial network methods (1.48 seconds per pattern). The color extraction accuracy of the article's method is 98.73%, which is higher than CNN based deep learning (96.53%) and slightly higher than GAN based generative adversarial network method (97.84%). In terms of algorithm convergence, the proposed method achieves convergence within 198 iterations, while CNN based deep learning and GAN based generative adversarial network methods require 243 and 307 iterations, respectively. The memory consumption of the article method is relatively low, only 2.37GB. In comparison, the memory consumption of CNN based deep learning and GAN based generative adversarial network methods are 4.83GB and 5.62GB, respectively.

Table 4 provides a comparative analysis of the Zhuang brocade pattern design effect based on relative coordinates and regional content replacement, covering six samples and their average values in terms of color extraction accuracy, pattern segmentation accuracy, tissue morphology matching accuracy and overall design efficiency improvement. Percent performance on four key indicators. The highest color extraction accuracy is 0.9942 for Sample C, the lowest is 0.9675 for Sample B, and the average is 0.9819. The highest pattern segmentation accuracy is 0.9924 for Sample A, the lowest is 0.9819 for Sample D, and the average is 0.9886. The highest tissue morphology matching accuracy is 0.9813 for Sample C, the lowest is 0.9685 for Sample D, and the average is 0.9745. The highest overall design efficiency improvement percentage is 35.73% for Sample C, the lowest is 28.67% for Sample B, and the average improvement is 32.02%. The results show that this design method has significant effects in improving the efficiency of Zhuang brocade pattern design and the accuracy of key design links.

Table 5 presents the feature extraction of Zhuang brocade pattern and its design performance index analysis.

In terms of primary color accuracy, Sample E performed best, reaching 0.9912, while Sample D was relatively low, at 0.9651, and the average primary color accuracy was 0.9786. Sample E also had the highest secondary color accuracy, at 0.9727, while Sample B had the lowest secondary color accuracy, at 0.9516, and the average secondary color accuracy was 0.9638. In terms of pattern complexity, measured by the number of points per unit area, Sample B had the most complex pattern, at 63.4 points, while Sample D had the lowest pattern complexity, at 57.6 points, and the average pattern complexity was 60.8 points/unit area. The shortest segmentation time was Sample C, at 1.142 seconds, and the longest was Sample A, at 1.267 seconds, with an average segmentation time of 1.197 seconds. In terms of reconstruction time, Sample B has the shortest reconstruction time of 2.392 seconds, while Sample C has the longest reconstruction time of 2.501 seconds, with an average reconstruction time of 2.434 seconds. The matching error is in pixels. Sample E has the smallest matching error of 1.87 pixels, while Sample B has the largest matching error of 2.31 pixels, with an average matching error of 2.09 pixels. In terms of data transmission efficiency, Sample E has the highest data transmission efficiency of 13.142 MB/s, while Sample B has the lowest data transmission efficiency of 12.543 MB/s, with an average data transmission efficiency of 12.836 MB/s. These data show that different samples have differences in feature extraction and performance indicators of Zhuang brocade pattern design, but overall maintain high accuracy and efficiency.

Table 6 provides the performance distribution analysis of the designed methods based on relative coordinates and region content replacement. In terms of color extraction, the average color deviation is 1.8234, indicating that the accuracy of color extraction is high. The algorithm complexity is $O(n * \log(n))$, which means that the process has good efficiency when processing large-scale data. The performance of pattern segmentation shows that the pattern smoothness change rate is 12.47% and the pixel segmentation overlap rate (IOU) is 0.9312, showing the high accuracy of the segmentation process. The memory usage is 215.43MB and the execution time is 1.154 seconds, indicating that the segmentation process performs well in resource consumption and time efficiency. In the performance of region replacement, the memory usage is 198.35MB, the execution time is 1.987 seconds, and the pattern detail retention rate is 96.32%, indicating that the details of the pattern can be well maintained during the replacement process. The algorithm complexity is $O(n^3)$, suggesting that this step may require more computing resources when processing large-scale data. The performance analysis of pattern reconstruction shows that the average color deviation is 1.5647, the pattern smoothness change rate is 10.83%, the pixel segmentation overlap rate (IOU) is 0.9468, the memory usage is 325.28MB, the execution time is 2.417 seconds, and the pattern detail retention rate is 92.74%. The algorithm complexity is also $O(n * \log(n))$, indicating that the reconstruction process has an advantage in processing efficiency. In terms of overall performance (average), the average color deviation is 1.6941, the pattern smoothness change rate is 11.65%, the pixel segmentation overlap rate (IOU) is 0.9390, the memory usage is 210.93MB, the execution time is 1.574 seconds, and the pattern detail retention rate is 94.53%. These data show that the design method based on relative coordinates and regional content replacement has a balanced and efficient performance in terms of color accuracy, detail retention, resource consumption, and execution time.