Abstract:

Objective From the perspective of intelligent pattern making, the closest pattern in the pattern library is matched by the identification of the garment style drawings and new patterns can be developed based on that pattern. This method of pattern making makes maximum use of existing pattern information and simplifies the structure drawing process of the pattern. In order to achieve similarity matching from garment style drawings to patterns, a pattern matching technique based on garment structural feature recognition is proposed.

Method The implementation of this technique consists of two main parts. The first is category label design, where certain structural features in the flat style drawing are defined and multi-label categories are set according to the definition according to the knowledge of women's trouser structural drawing. Then the multi-label categories were transformed into single-label multi-categories, and the link between the flat style diagram, structural features and the pattern was established. Finally, examining the correlation between the labels, the final labels for the experiment were designed. The second part is the model validation. In this part of work, the apparel dataset was established, which took the women's trousers flat style drawing as the sample. Then the AlexNet network was improved in the experiment. These changes mainly include simplifying the network structure and adding batch normalization operations after each convolutional layer.

Results In the process of label design, 18 categories of women's trousers were set through data visualization analysis and the study of correlations between labels. One of the results of the model testing shows the model converges faster after adding the batch normalization after the convolution layer (Fig. 10 and Tab. 4). The recognition accuracy is higher, with the highest recognition accuracy achieved by adding four layers of batch normalization. Second, it can be seen that the final training accuracy of the model tends to be stable around 99% (Fig. 8). The accuracy of the improved model on the validation set is 83.4%, and the recall and F₁ values are 0.834 and 0.835, respectively. Third, the accuracy of the improved model is 6.7% higher than that of the original AlexNet model and 6% and 3.6% higher than that of the ResNet18 and VGG11 models, respectively (Fig. 11, Tab. 5). The number of model parameters is also less than that of the original model.

Conclusion In this paper, the identification of structural features and matching of similar patterns is achieved by defining 18 structural features in the silhouette, waist position, waist shape, waist process and hip circumference, and using them as similarity representations of the pattern, constructing a model using the garment style drawings as input. The experimental research has certain value and significance to the intelligent pattern making of garments. In the process of label design, through the visual analysis of data and the study of the correlation between labels, invalid labels can be eliminated and label categories can be set reasonably, which can improve the problem of uneven data distribution in label categories and improve the utilization rate of data; In the process of model validation, by improving the structure of the original AlexNet model and introducing batch normalization operation in the convolution layer, the convergence speed of the model can be speed up and the recognition accuracy of the model be increased.

Key words: garment structure characteristic, pattern matching, multi-label classification, garment dataset, AlexNet neural network, garment pattern making