Abstract:

Objective The generation of garment patterns has long been an important research focus in the field of garment product development and garment CAD. Addressing the issue of poor pattern accuracy due to the lack of consideration of garment-specific knowledge during the conversion of 3-D garments into 2-D patterns, which results in patterns that cannot be directly applied, this paper proposes an automatic method for high-precision 3-D garment pattern generation based on the combination of deep learning and expert knowledge.

Method This paper adopted a deep learning-based approach, improving the model by integrating garment pattern requirements and expert knowledge into the NeuralTailor hybrid framework. As the first step, cubic and quartic Bezier curves, as well as right-angle constraints, were added to enhance the garment pattern dataset generator, producing a professional high-precision pattern and 3-D garment model dataset, solving the previous issue of accurately representing complex curves in garment patterns. Then, an edge length loss function was introduced in the training loss of the NeuralTailor framework. Combined with expert knowledge of garment structural design, a fuzzy mathematical model was used to assess garment fit, adjusting the corresponding pattern arcs and optimizing edge details of the generated patterns. This made the improved model capable of automatically generating more precise patterns that better met industrial application requirements. Finally, physical simulations and real-world scanned 3-D garment models were used for case validation.

Results The improved model was evaluated through comparative experiments. Quantitative analysis of the evaluation metrics for different models showed that the pattern shape error of this model was reduced by 0.69 cm compared to the pre-improvement model, with the pattern shape error being less than 2 cm, which falls within the acceptable bust tolerance range for garment production. Translation and rotation errors were also reduced, and the accuracy of the number of pattern edges increased to 100%, indicating improvements in pattern similarity and the prediction of pattern position information. Validation was performed using the 3-D garments from the dataset, 3-D models simulated with other software, and 3-D models from the real-world scanned public dataset MGN. The experimental results indicate that there were significant discrepancies between the patterns generated by the original NeuralTailor model and the actual garment patterns, such as the neckline and sleeve shapes highlighted by black boxes, large differences in seam length, missing pattern pieces, and the inability to segment the placket for closed garments. The method proposed in this paper was shown to be able to accurately predict the pattern shapes of the 3-D garments in the dataset. Although errors might occur around the placket and the accuracy of dart prediction needs improvement for physically simulated and real scanned garments, the garment patterns exhibited good accuracy, capable of predicting higher-precision patterns with pattern curves.

Conclusion The research reported in this paper improves the deep learning model by incorporating garment drafting standards and expert knowledge, using the enhanced NeuralTailor framework to generate garment patterns, followed by professional optimization based on expert knowledge of garment structure. Experimental results from physical simulations and actual 3-D garment scans demonstrate that this method can predict standardized garment patterns. The improved model provides a new professional pattern generation method for virtual try-on and garment design and manufacturing. In the future, robustness studies should be conducted on fabric properties, body shapes, or posture changes, and more extensive garment structure expert knowledge can be integrated for the generation of specialized patterns.

Key words: 3-D garment, pattern generation, expert knowledge, deep learning, garment digital modeling