Research on T-shirt-style design based on Kansei image using back-propagation neural networks

2.1 Clothing pattern attribute category quantification method

Parametric patterning in clothing involves defining mathematical expressions by constraining the geometric and dimensional relationships of the figure, thus creating a mathematical model that interrelates the dimensions of geometric objects [21,22]. In this research, we propose a method for quantifying the attribute categories of patterns based on the theory of parametric pattern making. Starting from the garment body prototype, a program is written in the MatLab software platform to draw the clothing patterns using the coordinate positions of key points and the dimensional calculation formulae based on the inherent structural relationship of the garment. The key design points are identified according to the garment’s key structure, and a three-layer hierarchy of key elements, element attributes, and attribute categories is established. The final quantification process involves assigning element attributes to the key design points in the pattern and subdividing them into attribute category parameters.

2.2 Back-propagation neural network algorithm

The back-propagation neural network is a data processing system that simulates the structure of neural networks in the brain. It can generally be explained using a multi-layer perceptron model and is often used to solve complex problems [23]. Back-propagation networks are suitable for learning multi-layer neural networks, where the network connection weights are stably distributed within a fixed range through repeated learning of a large number of training samples. Its topology is divided into an input layer, a hidden layer, and an output layer. In the Kansei image study of women’s T-shirt styles, the normalised data of the attribute categories and the image evaluation values are used as training samples for supervised learning. After multiple rounds of training, the error oscillation converges to convergence, the research objective is considered achieved, and the model construction is completed.

2.3 Radar chart evaluation method

The radar chart evaluation method, also known as the spider chart evaluation method, determines the main evaluation indicators of the study object according to the actual needs, obtains the evaluation indicator data, and performs a dimensionless process [24]. This method allows for the characterisation of multiple indicators on a two-dimensional plane and visualisation of the strengths and weaknesses of each evaluation object in that indicator. An n-dimensional coordinate system is established using a point as the origin. Each coordinate axis is divided into m segments, and the specific evaluation values of each dimension are labelled and connected on the axes to form an n-dimensional polygon. The area and perimeter of the polygon are used as parameters for a comprehensive evaluation. The user questionnaires were used to rate the ratings, and the radar chart was used to visually characterise the distribution of the rated objects in terms of each Kansei image attribute [25].

3.1 Sample collection

Sample collection was conducted through two methods: product image gathering and computer-aided simulation. Beginning with clothing styles and external features, the samples were acquired from both online platforms and physical stores. We conducted online research by reviewing and recording data from various mainstream online shopping platforms, specialised clothing retail platforms, and electronic resources. T-shirt products that ranked among the top 25 in sales and met our criteria were collected. Details, including brand, style, website, and other pertinent information, were documented and encoded and subsequently incorporated into our sample database. Simultaneously, offline surveys were conducted at three prominent shopping malls in Gusu District, Suzhou. Here, we physically recorded information about the products, encompassing brand, style, and addresses. After encoding, these data were also integrated into our sample database. Following screening for similarity and clarity, we initially selected and collected 279 images of women’s T-shirts. After conducting cluster analysis and comparisons, we retained 137 T-shirt samples with distinctive features. Based on the information in our sample database, we purchased the selected 137 samples for further research. To eliminate potential interference from external factors such as brand logos, fabric, patterns, and colours during the experimentation process, we uniformly pre-processed the initially selected samples. This involved using virtual clothing modelling software to simulate and virtually showcase the product samples.

Virtual garment modelling software is an effective tool for visualising clothing designs. It allows for the quick construction of body models, virtual sewing of fabric panels, and realistic fabric simulation, all with high levels of accuracy and efficiency. Currently, mature virtual clothing design tools include the US Gerber company’s V-stitcher, Germany’s Assyst system, China Lingdi company’s Style 3D, and South Korea’s CLO Virtual Fashion company’s CLO 3D. In this study, we used the V4.7.369 version of the Style 3D apparel digital modelling platform to model samples by measuring their physical dimensions. During the modelling process of the T-shirt samples, common basic cotton knitted fabric was selected. The size of the virtual model was based on the Chinese standard GB/T 1335.2-2008 “Women’s clothing size” for a female model with a height of 165 cm and a size of 88 A. The 21 T-shirt styles with more distinctive image and style differences among the 137 samples were mainly displayed. The samples were modelled, worn on the same virtual model, and paired with consistent basic-style trousers. The results of the pre-processed modelling samples are shown in Figure 1.

Figure 1

Representative sample modelling diagram.

3.2 Analysis of key elements

3.2.1 Parametric pattern making

Popular general drafting CAD software includes AutoCAD developed by Autodesk in the United States, MatLab developed by Mathworks, and Lectra developed in France. MatLab is primarily used for scientific calculation, data visualisation, and interactive programming. It features a powerful image processing toolkit that allows designers to handle complex curves in paper samples more easily and solves issues such as the inability of traditional automatic code pushing to adjust each dimension independently [26]. For parametric mapping in this study, we selected the MatLab software platform to draw clothing patterns and determined the dimensional formulae for the function programs based on the intrinsic structural relationships.

Taking the front piece and sleeve piece of a women’s T-shirt as an example, the design method of the women’s clothing style and pattern series in the reference book for clothing students in Chinese higher education institutions is referred to, and the intrinsic calculation formula and parametric drawing method of the paper pattern structure are proposed. The key variables for the front piece of a women’s T-shirt are the bust, collar, shoulder width, and garment length, while the key variables for the sleeve piece are the sleeve cage and sleeve length. The standardised front piece and sleeve piece structure design is shown in Figure 2. The front piece of the garment is designed using the cage depth line as the horizontal axis and the front centre line as the vertical axis to establish a co-ordinate system with the intersection of the two lines as the origin of the front piece. For the sleeve piece, the horizontal axis is the drop line and the vertical axis is the sleeve centre line. The sizing formula is determined based on the structural relationships inherent in the garment, as shown in Table 1. The coordinate points are determined according to how the key variables and other secondary variables are plotted in the clothing pattern, as shown in Table 2.

Figure 2

Structure of the front piece and sleeve piece of the women’s standardised T-shirt.

Table 1

Women’s standardised T-shirt front piece and sleeve piece structure size relationship

Location of the structure	Structural lines	Size calculation formula	Location of the structure	Structural lines	Size calculation formula
Length of clothing	f 1 f 20	L	Side seam hem rise	f 9 f 24	a 4
Armhole depth	f 7 f 10	B/6 + 7	Side seam bottom hem retraction	f 9 f 14	a 5
Front neck drop	f 1 f 2	N/5 + a 1	Waist retraction amount	f 21 f 22	a 6
Front neck across	f 1 f 3	N/5 + a 2	Front sleeve diagonal	g 2 g 4	AH/2
Side shoulder	f 4 f 13	S/2	Back sleeve diagonal	g 4 g 7	AH/2 + 1
Protruding front shoulder line	f 4 f 18	2	Sleeve cap height	g 1 g 4	AH/3 − b 1
Front drop shoulder line	f 12 f 18	B/18	Sleeve length	g 4 g 13	SL
Chest width	f 7 f 17	B/6 + 3	Sleeve side seam bottom hem retraction	g 8 g 10	b 2
Width of front	f 6 f 7	B/4 + 2.5	Sleeve side seam hem rise	g 10 g 12	b 3
Rise of front midsole swing	f 8 f 20	a 3	Rise of the middle hem of the sleeve	g 11 g 13	b 4

Note: B is the bust, N is the neck, S is the shoulder breadth, L is the length of garment, AH is the arm hole, and a ₁–a ₆ and b ₁–b ₄ are the variables (set based on empirical).

Table 2

Key points of women’s standardised T-shirt front piece and sleeve piece

Point	X-coordinate	Y-coordinate	Point	X-coordinate	Y-coordinate
f 1	0	B/6 + 7 − 0.5	f 20	0	−L + B/6 + 7 − 0.5
f 2	0	B/6 + 7 − N/5 − a 1	f 21	−(B/4 + 2.5)	−(L/2 − B/12 − 3.5)
f 3	−(N/5 + a 2)	B/6 + 7 − 0.5	f 22	−(B/4 + 2.5 − a 6)	−(L/2 − B/12 − 3.5)
f 4	−S/2	B/6 + 7 − B/18	f 23	0	−(L/2 − B/12 − 3.5)
f 5	−(B/6 + 3)	B/18 + 0.5	f 24	−(B/4 + 2.5)	−(L − B/6 − 7)
f 6	−(B/4 + 2.5)	0	g 1	0	0
f 7	0	0	g 2	5(AH)2/36	0
f 8	0	−(L − B/6 – 7 − a 3)	g 3	5(AH)2/72 − 0.7	AH/6 − 0.7
f 9	−(B/4 + 2.5 − a 4)	−(L − B/6 – 7 − a 5)	g 4	0	AH/3 − b1
f 10	0	B/6 + 7	g 5	−5(AH)2/72 + 0.7	AH/6 − 0.7
f 11	−(N/5 + a 2)	B/6 + 7	g 6	5(AH)2/72	AH/6
f 12	−(B/6 + 3)	B/6 + 7	g 7	−5(AH)2/36 − AH − 1	0
f 13	0	B/6 + 7 − B/18	g 8	−5(AH)2/36 − AH − 1	−(SL − AH/3)
f 14	−(B/4 + 2.5)	−(L − B/6 – 7)	g 9	5(AH)2/36	−(SL − AH/3)
f 15	−(B/12 + 1.5)	−(L − B/6 – 7)	g 10	−5(AH)2/36 − AH − 1 − b 2	−(SL − AH/3 + b 3)
f 16	−(B/12 + 1.5)	0	g 11	0	−(SL − AH/3 − b 4)
f 17	−(B/6 + 3)	0	g 12	5(AH)2/36 − b 2	−(SL − AH/3 + b 3)
f 18	−(B/6 + 3)	B/9 + 7	g 13	0	−(SL − AH/3)
f 19	−(N/5 + a 2)	0	g 14	−5(AH)2/36 − AH − 1 − b 2	−(SL − AH/3)

Note: B is the bust, N is the neck, S is the shoulder breadth, L is the length of garment, AH is the arm hole, and a ₁–a ₆ and b ₁–b ₄ are the variables (set based on empirical).

The straight lines in the pattern were drawn using the plot function on the MatLab software platform, including the centre front line f ₂ f ₈, the shoulder line f ₃ f ₄, the side seam line f ₆ f ₉, the hem line f ₈ f ₉, the sleeve side seam lines g ₇ g ₁₀ and g ₂ g ₁₂, and the sleeve hem lines g ₁₀ g ₁₁ and g ₁₁gg ₁₂. The curves in the pattern were drawn using the Bezier curve model [27,28], including the front collar curve f ₂ f ₁₁, the front sleeve cage line f ₄ f ₆, the front sleeve cage line g ₂ g ₄ in the sleeve, and the back sleeve cage line g ₄ g ₇. Finally, the women’s T-shirt front piece and sleeve piece parametric planar platemaking function program was completed in MatLab. When the program is run, the mssback function in the Command Window is called, three samples are randomly selected as examples, and the corresponding parameter values are entered as shown in Table 3, resulting in the front and sleeve paper samples for the different parameter values shown in Figure 3.

Table 3

Corresponding parameter values

Name	B	N	S	L	AH	SL	a 1	a 2	a 3	a 4	a 5	a 6	b 1	b 2	b 3	b 4
Example 1/cm	81.00	36.00	38.50	55.00	42.00	18.50	0.00	3.00	0.00	0.50	0.00	0.00	0.00	0.50	0.00	0.00
Example 2/cm	86.00	41.00	40.00	40.00	52.00	17.00	0.00	2.50	2.50	0.00	1.50	9.00	30.00	0.50	0.00	0.00
Example 3/cm	90.00	44.00	38.00	52.00	44.80	19.50	0.00	2.00	0.00	1.00	0.50	0.00	1.00	0.50	0.50	0.00

Figure 3

Women’s T-shirt front and sleeve pattern based on parameter values.

3.3 Kansei image acquisition

The 62 T-shirt-style intention words initially collected were analysed and aggregated into 11 words with high frequency of occurrence. After the second round of cluster analysis, five basic Kansei intention words were obtained, which were avant-garde, casual, sporty, classic, and elegant. The Likert scale, the most widely used in survey research, was selected to combine the 5 words with 137 samples to build a 5-level scale, and a questionnaire on the perception of women’s T-shirt style Kansei image was distributed through an online platform. A total of 76 valid questionnaires were returned, and the image evaluation values were obtained using statistical analysis, as shown in Table 6.

3.4 Construction of cognitive model for styling intention

Back-propagation neural network algorithms are trained using MatLab software. The network training results are judged on the basis of the errors in the test and validation sets. The number of iterations of the network training should be moderate, too few iterations of the algorithm are not accurate, and too many may lead to overfitting. The smaller the mean square error value of the training data, the closer the output value is to the target value. The size of the error gradient indicates the size of the change in weights and thresholds. The closer the parameter curve in the training process graph is to the target curve, the better the network is trained [33]. The number of neuron layers in the input layer is 18 and the number of neuron layers in the output layer is 5, which is consistent with the structure of the training sample data. Sigmoid was chosen as the activation function of the hidden layer, linear was chosen as the activation function of the output layer, and the back-propagation method was used to perform the operation. The algorithm path is shown in Figure 4. More than 50 iterations of training were carried out during the training iteration. The network reached the optimum when the 46th training was performed, so the current network parameters were applied. The training results are shown in Figure 5, where the R value indicates the prediction accuracy of the neural network (taking values in the range of 0–1). The model constructed by multiple training (R ≥ 0.81) has a high accuracy, which is in line with the experimental design research needs.

Figure 4

Algorithm path.

Figure 5

Optimal training results.

4.1 Conclusion

The generalisation ability of the neural network can reasonably predict the probability of a change in the same type of sample, and by combining it with the enumeration algorithm, the optimal combination form can be derived by verifying the pattern of all combinations of attribute categories. A function is created from the aforementioned back-propagation neural network model and called in the enumeration algorithm to obtain an optimisation algorithm for the combination of women’s T-shirt attribute categories. The optimal combination of attribute categories under the perception of the user’s target Kansei image is shown in Table 7. The optimal-style design form was modelled and displayed in the Style 3D platform, as shown in Figure 6.

Figure 6

Optimal-style design form under target image.

4.2 Evaluation

The radar chart method is applied to visually characterise the weight of each indicator of the research object, quantify the contribution of each indicator to the overall morphology, and evaluate the distribution of the research object in each Kansei image attribute. Determine five evaluation indicators with the same weight based on emotional intention style. The radar diagram was constructed based on the indicators, and the average of the corresponding research scores was represented by different shaped “dots” on the indicator dimension axes. By connecting the dots on adjacent axes in the diagram, a closed polygon is formed, which represents the overall state of the sample’s Kansei image. The standardised radar chart is confined to a pentagon with a radius of 1, and the indicators are divided into 0.2, 0.4, 0.6, 0.8, and 1. The closer the value of the indicator is to 1, the higher the level of the indicator.

A questionnaire survey was conducted in three large shopping malls in Suzhou, China, in the clothing shopping area. A total of 89 questionnaires were returned, 86 of which were valid, with an effective rate of 96.63%. In the questionnaire, users were invited to rate a range of optimal-style designs for the target image on five main evaluation criteria dimensions. The data collected were statistically analysed and plotted as a radar chart. As can be seen from the results in Figure 7, the five optimal women’s T-shirt-style designs all fit well with the users’ Kansei perceptions. Therefore, the design method proposed in this study can be better applied in product design. Summarising the technology roadmap of this study is shown in Figure 8.

Figure 7

Indicator evaluation results.

Figure 8

Technology roadmap.