Abstract:

Objective In order to address the low efficiency in conventional clothing pattern making methods and the limitation of existing AIGC-based generation methods which only output images without binding process dimension information, this study aims to develop a method for automatically generating digital clothing patterns with both visual accuracy and production usability, so as to connect the design and subsequent production links (such as grading, marker making and cutting) in the clothing manufacturing process and provide a feasible solution for the engineering application of intelligent garment pattern making.

Method A two-stage pattern generation framework integrating YOLOv8 key-point recognition and parametric dimension constraints was proposed. First, based on the YOLOv8-pose model, optimizations were made according to the characteristics of T-shirt patterns: three groups of anchor boxes (small anchor box 12×12, medium anchor box 24×24, large anchor box 48×48) suitable for the scale of pattern were generated by re-clustering to replace the default anchor boxes. On the basis of the original bounding box loss (CIoU) + key-point loss (MSE), a line consistency constraint loss was added to force key-points to fit the structural lines of the pattern, so as to realize the accurate positioning of the contour and internal structural key-points of the clothing back sheet pattern. Then, a size-driven generation module was adopted to convert discrete key-points into structured patterns. Parametric basic templates were constructed based on clothing pattern making principles, and the coordinates of template vertices were dynamically calculated by input key size variables (such as garment length, half bust, shoulder width, collar width, sleeve length, cuff width). Skeleton alignment and key-point mapping were carried out (including bilateral filtering for denoising, Otsu threshold method for converting to binary images, morphological thinning for extracting skeleton, KD-tree nearest-neighbor search for adsorbing key-points to skeleton lines) to eliminate image noise and recognition deviation. Adjacency matrices were built to quantify topological relationships through connection validity judgment, subgraph division and adjacency matrix generation, so as to convert discrete key-points into structured connections. Finally, smooth contours were generated using quadratic Bezier curves, and a dual-slider interactive mechanism was introduced to realize the dynamic adjustment of size parameters and curve shapes.

Results The experimental results showed that the optimized key-point detection model achieved a key-point localization accuracy (KLA) of 96.8%, which was significantly higher than the original YOLOv8-pose model with an mAP50 of 0.93. The generated T-shirt patterns had a structural similarity index (SSIM) of 0.85 and a key-point connection accuracy (CA) of 96.2%, which effectively ensured the accuracy of the contour and the rationality of the topological relationship. In terms of efficiency, the average time consumption of the entire processing flow was only 1.2 s per pattern (based on NVIDIA RTX 4060 GPU), which was much faster than the conventional manual CAD pattern making that usually takes tens of minutes per pattern. Compared with GAN-based methods (SSIM=0.76) and diffusion-based methods (SSIM=0.82) that only output static images, the output of this method was a parametric and editable digital pattern that could be directly connected to subsequent production links such as grading, marker making and cutting. In addition, through the developed dual-slider interactive interface, users could dynamically adjust size parameters and curve control points, and the system would update the pattern shape in real time, meeting the needs of fine-tuning the pattern in actual production.

Conclusion The proposed automatic generation method for T-shirt patterns integrating YOLOv8 key-point recognition and parametric dimension constraints, effectively solves the problems of low efficiency in traditional pattern making and the inability of existing AIGC methods to output results directly applicable to production. The optimized detection model realizes the accurate positioning of pattern key-points with an accuracy of 96.8%, and the introduction of adjacency matrices to construct topological relationships enables the generation of parametric digital patterns with reasonable structure and smooth contour under size driving. Experiments show that while maintaining high structural similarity and connection compliance rate, the output results of the proposed method can be directly connected to subsequent production links, realizing the effective connection from design to production. The core value of this study lies in providing an intelligent front-end solution for the conventional CAD parametric design process, significantly reducing the professional threshold and manual dependence. Although the current method still relies on predefined templates when dealing with garment categories with very different structures (which ensures the standardization and production usability of the output results), future research will focus on expanding the coverage of the template library and exploring template adaptive generation technology based on few-shot learning, so as to promote the development of intelligent garment pattern making systems towards more universal and efficient directions.

Key words: clothing pattern making, dimensional constraint, keypoint detection, parametric modeling, Bezier curves