关键词: 风格迁移, 服饰纹样, 辅助设计, 生成对抗网络, 清代官补, 艺术评价

Abstract:

Objective An improved cycle generative adversarial network (Cycle GAN) style transfer method is proposed to address insufficient color preservation and loss of multi-layered pattern details when transferring Qing dynasty mandarin square badge styles, aiming to preserve these imperial rank insignia's cultural heritage while effectively reproducing their complex pattern systems through deep learning techniques.

Method The Qing dynasty mandarin square patterns were analyzed through pattern elements, color schemes, and compositional structure. Five style transfer models, i.e., self-discriminative cycle generative adversarial networks (SD GAN), no-independent-component-for-encoding GAN (Nice GAN), dual generative adversarial network (Dual GAN), generative adversarial networks that learns to discover relations between different domains (Disco GAN), and cycle generative adversarial network (Cycle GAN), were compared. Based on the recognition of Cycle GAN as the optimal foundation, three modifications were implemented, including adding feature fusion modules to the generator to enhance pattern detail capture, and simplifying the discriminator structure to preserve compositional features, and introducing perceptual loss functions to prevent pattern-background color merging.

Results The modified model demonstrated significant technical enhancements. The structural similarity index measure (SSIM) increased by 0.041, peak signal-to-noise ratio (PSNR) improved by 2.17 dB, mean squared error (MSE) decreased by 0.031, learned perceptual image patch similarity (LPIPS) reduced by 0.042, and the frÉchet inception distance (FID) index dropped by 4.157.

The modified model also generated clearer traditional motifs, particularly auspicious cloud patterns and sea-mountain designs. It effectively separated background colors from foreground patterns, enhancing layering effects and edge definition crucial to these historical insignia. The feature fusion mechanism in the generator captured multi-scale pattern features more effectively, preserving intricate details of emblematic elements while maintaining proportional relationships. This improvement appeared most notably in the preservation of fine line work and distinctive gradations in cloud and water motifs. The simplified discriminator better preserved pattern distribution logic and spatial relationships between decorative elements, maintaining the symbolic hierarchy inherent in these rank-signifying textiles. This improvement particularly affected the compositional integrity of central animal motifs (such as cranes or lions) that designated specific ranks. The perceptual loss functions addressed color bleeding between foreground and background elements, resulting in more precise color boundaries while maintaining harmonious transitions characteristic of traditional silk embroidery. The improved color fidelity preserved symbolic color associations while allowing adaptability to contemporary palettes.

Visual assessment confirmed the improved model reproduced the distinctive qualities of Qing dynasty mandarin squares more accurately, including balanced composition, symbolic coherence, and intricate detailing.

Conclusion This study systematically analyzed style transfer models for Qing dynasty mandarin square patterns, revealing their unique artistic characteristics through content form, color composition, and layout arrangement. Through comprehensive evaluation of five mainstream image generation style transfer models using artistic and technical metrics, Cycle GAN was identified as the optimal choice. Effective improvements were developed addressing insufficient layering, pattern overlapping, and excessive color blending in generated images. The improved model achieved breakthroughs in three key areas, namely (i) the feature fusion module enhanced multi-scale pattern feature capture, particularly improving preservation of cloud motifs and sea-mountain designs, (ii) the optimized discriminator with 1×1 convolution kernels and reduced network layers prevented information loss, enhanced understanding of unique compositional layouts, and resolved pattern continuity issues, and (iii) the perceptual loss functions optimized background-pattern separation, creating natural color transitions consistent with traditional principles. These improvements enabled successful transfer of artistic characteristics while preserving modern image content, integrating traditional elements with contemporary aesthetics. This research provides an innovative pathway for digital inheritance of mandarin square patterns, enabling automated creation of modern animal images in traditional styles, improving design efficiency and meeting personalization requirements. Future research will explore this method's applicability to other complex traditional patterns.

Key words: style transfer, clothing pattern, auxiliary design, generative adversarial network, Qing dynasty mandarin square, artistic evaluation