Figure 1 | Fashion Modeling Framework Diagram

This project utilizes synthetic biology to promote sustainable fashion. We have developed a computer vision system that integrates YOLOv8 and MediaPipe to estimate the area of clothing in real time and quantify the water-saving differences between bacterial cellulose and cotton. We also use fractal geometry to generate digital patterns for textile designs, promoting digital and low-water-consumption fashion innovation.

Research shows that garment production—from raw material to finishing—consumes vast amounts of water, yet consumers rarely perceive this hidden usage, contributing to waste and pollution. To address this, we propose a recognition-based solution: using human skeleton measurement to estimate fabric area from body dimensions, with a fixed-size project logo as a portable, recognizable calibration reference that also serves as a brand marker.

Logo Recognition

We compared ORB feature matching and YOLOv8 for logo recognition, and selected YOLOv8 for deployment based on performance across scenarios.

ORB Approaches

ORB logo recognition combines FAST keypoint detection with BRIEF descriptors to capture image features for object matching and localization. FAST identifies keypoints by comparing pixel brightness around a central pixel.

Figure 2 | FAST technology

ORB locates the logo by matching keypoints between the target and camera images using RANSAC and homography, working well at close range. However, at long distances, sparse features and unstable matches reduce recognition reliability, limiting practical stability and robustness.

Figure 3 | ORB Results

Although optimization could be achieved by enhancing the pixel precision of the camera, in practical application, we prioritize algorithms that exhibit strong robustness and low hardware dependency. Consequently, this approach was not ultimately adopted.

Training YOLOv8

We adopted the lightweight, anchor-free YOLOv8 model for its high accuracy, efficiency, and suitability for real-time edge-device deployment.

Figure 4 | YOLO Model Training Process Diagram

For specific training examples, please refer to the following pdf file:

Comparison and display of recognition effects

The trained YOLOv8 model showed robust recognition at both close and long ranges, outperforming ORB with stable, accurate detection boxes.

Figure 5 | YOLO Results

Considering performance, robustness, and compactness, YOLOv8 outperforms ORB, achieving high-accuracy recognition with small custom datasets and supporting mobile or embedded deployment. It is the chosen solution for this project's logo recognition module.

Water Saving Data

Cotton is one of the most widely used textile fibres, but its cultivation and processing are highly water-intensive, requiring around 8,980 L per kilogram of fibre. In comparison, data from published studies indicate that producing 1 kg of bacterial cellulose (including culture, purification, and post-treatment) consumes only about 102 L of water. This corresponds to just 1.1% of the water footprint of cotton, highlighting the significant sustainability advantage of bacterial cellulose as a potential low-carbon alternative for textile applications.

Skeleton Determination

In order to accurately measure the size of clothes using the camera, this system adopts MediaPipe Pose for human pose estimation. This model is based on a lightweight convolutional neural network architecture and realizes real-time and high-precision detection of human key points.

Building upon this, we have constructed a multi-dimensional, high-precision clothing size calculation model based on the spatial topological relationships of human keypoints and scaling factors, combined with anthropometric principles from clothing engineering.

• Shoulder width is measured as the horizontal distance between keypoints 11 and 12, with 20% looseness compensation. Confidence filtering (≥0.5) and nasal tip (keypoint 0) tilt correction improve robustness.

• Chest circumference: Based on ergonomics and hip circumference estimation, the thoracohip ratio is set in combination with gender differences (0.85-0.95 for females and 1.0-1.1 for males), and the final size is obtained by superimposing the bulk coefficient.

• Waist circumference: Based on the statistical rule of the waist-to-chest ratio (0.65-0.75 for females and 0.70-0.80 for males), the chest circumference is calculated and loose compensation is added.

• Thigh circumference: Taking knee width as the proxy index, and mapping it to 20 cm above the knee in combination with the human body proportion relationship, the actual thigh circumference is converted. The calculation formula is as follows:

  $$\text{Actual thigh circumference}=\text{Knee point width}\times1.3~\pm2~cm$$

Firstly, according to the research in clothing engineering, we have designed a matching principle, as shown in the following table:

Table 1 | Principles for Matching Clothing Sizes

Size	Function	Degree of influence
Chest circumference	The primary factor for the fit of a top	65-75%
Waist circumference	Decide whether the trousers can be worn	55-65%
Shoulder width	Affect the comfort of the top	25-35%
Thigh circumference	Affect the freedom of movement of the trousers	35-45%

If a certain size of the user falls within the standard size range, the calculation formula of the matching degree dim score is as follows: $$\text{dim score}=\max\{0,1-\frac{|\text{user value}-\text{mid range}|}{\text{range width}/2}\}$$ When the size exceeds the range of the international standard scale, the distance penalty function is used to calculate the matching degree dist score: $$\text{dist score}=\max\{0,1-\frac{\text{distance}}{\text{range width}}\}$$ After integrating the matching scores of each size, the system conducts a quality assessment of the recommended sizes based on the following table:

Table 2 | Matching Quality Assessment Criteria

Grade	Conditions	User suggestions
Perfect	All key dimensions are perfectly matched	Strongly recommended
Good	The core dimensions are matched, and the deviation of the secondary dimensions is less than two centimeters	It is recommended to pay attention to the secondary dimensions
Fair	The core dimensions do not match, but the matching degree is greater than 0.6	Caution is advised. It is recommended to check the detailed dimensions
Poor	The matching degree is less than 0.6	Not recommended. It is suggested to choose other sizes

To visually demonstrate the environmental advantages of bacterial cellulose, we estimated the weight based on the size of the clothing and, based on the water consumption coefficient of the material, compared and calculated the specific water-saving amount of bacterial cellulose relative to cotton fibers.

Users can view real-time detection results and size suggestions through the intuitive interface (Figures 1 and 2, switch by pressing "d"), make reliable clothing purchasing decisions, and understand the water-saving effect of bacterial cellulose.

Figure 6 | Screenshot of Real-Time Detection

In Figure 1, the quality of the upper garment is detected as "good", while the trousers are judged as "poor" because no nodes were detected. The detection effect is accurate. Figure 2 shows the intelligent matching results, providing users with clear and reliable references for purchasing clothes.

Figure 7 | Detailed Size Recommendations

Above picture interface shows the final output of the system's intelligent matching of the multi-dimensional size calculation results with the international scale. It includes the matching score of each size, quality grade and specific size comparison, providing users with clear and reliable references for purchasing clothes.

Textile pattern design holds an important position in modern fashion textile design. Traditional textile pattern design relies on human brain conception to complete, which is time-consuming and has a single form. Fractals, as a new type of graphic generation method, have emerged as The Times require.

Fractal diagrams, based on mathematical theories, are programmed and designed by adjusting iterative functions, variation rules, and color parameters, etc., to achieve the visualization of program codes and add new fashion elements to textile pattern design. We utilized the implementation method of fractal graphs on the complex plane in Python. By changing the iterative function and its parameters, we generated different fractal graphs, and then used Photoshop for design, applying the fractal graphs to the design of fashion textile patterns.

Generation Principle of Complex Fractals

The visualization generation of fractal images mainly relies on the escape time algorithm. In complex dynamical fractal systems, the most representative iterative form is $z_{n+1}=f(z_n)$ ,$f$ is usually a polynomial or an analytical function (such as $f(z)=z^m+c$), $c$ is a complex parameter. For a given $c$, the point set generated through this iteration is called the Julia set; When $z_0=0$ is fixed and all sets that make the sequence non-divergent are examined, the Mandelbrot set is obtained.

To convert iterative operations on the complex plane into specific calculations of the real and imaginary parts, we need to expand the function over the complex number field, for example: $$z^3=(x+iy)^3=(x^3-3xy^2)+(3x^2y-y^3)$$ $$z^4=(x+iy)^4=(x^4-6x^2y^2+y^4)+(4x^3y-4xy^3)$$ $$\sin z=\frac12\sin x(e^{-y}+e^y)-\frac{1}{2}\cos x(e^{-y}-e^y)$$ $$\cos z=\frac{1}{2}\cos x(e^{-y}+e^y)+\frac12\sin x(e^{-y}-e^y)i$$

In addition, a modulus threshold $R$ needs to be set. For each grid point, no more than $N_\text{max}$ iterations should be performed, and the number of iteration steps maintained before reaching the modulus threshold $R$ should be recorded, that is, the "escape time" : $$ img(x,y) = \sum_{n=0}^{N_{\text{max}} - 1} 1_{\{\,|Z_n(x,y)| < R\,\}} $$ Directly using the number of integer iterations as the coloring index will lead to a significant striping effect. To enhance the continuity and aesthetic effect of the image, we have introduced the smooth coloring method. For power function mapping $f(z)=z^m$, the smooth iteration index can be expressed as: $$v=n+1-\frac{\log(\log|z_n|)}{\log m}$$This formula, through logarithmic scale correction, extends the number of integer iterations to a real number, thereby eliminating color bands and generating continuous color transitions.

Generation and Application of Fractal Graphs

Based on the above mathematical theories, we have carried out the generation and application practice of fractal graphs relying on the design of this project. We use black, blue, purple and light flesh color as the basic colors of the fractal graph. In the Python software, we define the fractal generation function and batch generation function, randomly select the fractal type (m-set or j-set) and transformation function and perform iterative calculations to batch generate a variety of fractal patterns to form a set. The following figure is a component diagram generated by the code running:

Figure 8 | Fractal Graphs Generated in Batches

It can be seen that the generated graphics have a rather unique style, and they contain fine structures and novel elements, which can bring a fresh visual experience to people. Therefore, fractal diagrams can be used as new pattern materials in textile pattern design.

By performing matting, copying, sorting and splicing on the above patterns, a large number of innovative fractal patterns can be designed. We selected some high-quality fractal patterns from them, deepened the design through permutation and combination, and conducted effect simulation with the help of image processing software Photoshop and Doubao AI. Eventually, these patterns were applied to clothing and home textile products. The physical presentation effect is shown in the figure:

Figure 9 | The Simulation Application of Fractal Diagrams on Fabrics

Based on the generation principle of fractal graphs on the complex plane, by running the program code multiple times, more creative fractal images can be continuously produced. This process can greatly enrich the material library of fashion textile patterns and provide sufficient inspiration for more subsequent fashion designs. From the perspective of practical application, these fractal patterns perform extremely well on textile objects, with both the visual texture and compatibility meeting expectations. The entire process is in a digital mode, featuring high precision, short cycle and good effect. In conclusion, fractal diagrams not only provide a brand-new tool for textile pattern design, but also help modern textile production models break through traditional thinking, demonstrating broad application prospects in the fashion field.

• [1]Kim J-W, Choi J-Y, Ha E-J, Choi J-H. Human Pose Estimation Using MediaPipe Pose and Optimization Method Based on a Humanoid Model. Applied sciences, 2023, 13(4): 2700.
• [2]Chapagain A K, et al. The Water Footprint of Cotton Consumption: An Assessment of the Impact of Worldwide Consumption of Cotton Products on the Water Resources in the Cotton Producing Countries. Ecological Economics, 2006, 60(1): 186-203.
• [3]Li Z, Liu F, Yang W, et al. A Survey of Convolutional Neural Networks: Analysis, Applications, and Prospects. IEEE Transactions on Neural Networks and Learning Systems, 2021, 33(12): 6999-7019.
• [4]Aglave P, Kolkure V S. Implementation of High Performance Feature Extraction Method Using Oriented Fast and Rotated Brief Algorithm. International Journal of Research in Engineering and Technology, 2022, 4(2): 394-397.
• [5]Wang S, Yang X. Generation of Fractal Diagrams on Complex Planes and Their Application in Textiles. Chinese Silk, 2017, 54(08): 56-61.
• [6]Zhang R, Yu S, Qiu H. Iterative Function System and CAD Application in Textile Industry. Mechanical and Electrical Engineering, 2008, (09): 89-91.