Traditional clothing pattern extraction considering attention mechanism and image data enhancement processing

Area of Science:

• Computer Vision
• Digital Heritage
• Pattern Recognition

Background:

• Traditional clothing pattern extraction faces challenges including poor image quality, complex textures, and limited labeled data.
• Existing methods struggle with precise segmentation and generalization across diverse cultural patterns.

Purpose of the Study:

• To propose a multi-scale data augmentation with polarized self-attention (MSDA-PSA) method for accurate traditional costume pattern extraction.
• To address image degradation, complex textures, and data scarcity in traditional clothing pattern segmentation.

Main Methods:

• A lightweight segmentation network fusing wavelet transform and generative adversarial networks for multi-scale enhancement.
• A dual attention mechanism using polarized self-attention (PSA) to focus on discriminative color channels and spatial details.
• Multi-scale data augmentation (MSDA) to combat image degradation and preserve texture details.

Main Results:

• Achieved 89.7% mean intersection over union (mIoU) and 0.87 Boundary F-score, outperforming DeepLabv3+ by over 10% mIoU.
• Demonstrated strong small-sample adaptability (72.4% IoU with 10% data) and cross-cultural generalization (83% accuracy).
• Real-time processing at 100FPS with low latency (35.4ms) and robust performance under noise (0.82 IoU).

Conclusions:

• The MSDA-PSA method offers a high-precision, interpretable framework for clothing pattern segmentation, balancing computational efficiency and cultural specificity.
• This approach provides a viable solution for digitizing cultural heritage, particularly in data-scarce and culturally diverse scenarios.
• The study establishes a robust method for enhancing feature consistency and critical region segmentation in complex pattern extraction tasks.