Abstract:

Objective This study aimed to address critical technical challenges in floral printed pattern style transfer, including pattern distortion, texture discontinuity, and checkerboard artifacts that commonly occur during the style transfer process. The research focused on developing an advanced style transfer model specifically designed for printed patterns to preserve structural integrity while seamlessly integrating target artistic styles, thereby enhancing creative expression capabilities and production efficiency in pattern design applications.

Method A multi-scale feature fusion neural style transfer model (MFFF-NST) based on decoupling architecture was proposed using an encoder-decoder framework, which introduced a Line Art Extraction Module(LAEM) in the content image preprocessing stage to resolve geometric deformation issues of pattern core structural features during the transfer process. The encoder incorporated three key modules, i.e., an adaptive feature extraction range adjustment module that dynamically adjusts receptive field sizes according to local pattern complexity, a shifted window attention mechanism for capturing multi-scale features, and an adaptive channel weighting module that intelligently emphasizes important feature channels while suppressing secondary ones. The decoder employed a style-aware decoder complemented by a VGG-19 convolutional network optimizer to eliminate checkerboard artifacts and enhance visual coherence in the generated images.

Results Comprehensive experiments were conducted on floral printed pattern datasets containing 5 100 images and WikiArt style datasets with 80 000 images. Comparative analysis against four representative baseline methods (STTR, WCT, S2WAT, and AdaAttN) demonstrated superior performance across all evaluation metrics. The proposed MFFF-NST achieved a structural similarity index (SSIM) of 0.758, representing an improvement of 0.117 over the best baseline method (STTR). Mean squared error (MSE) reached 0.032, reducing error by 0.016 compared to STTR, while learned perceptual image patch similarity (LPIPS) achieved 0.365, showing a reduction of 0.062. The model successfully preserved geometric boundaries and maintained symmetry properties of printed patterns while achieving uniform style distribution and eliminating checkerboard effects. Processing efficiency analysis revealed significant improvements, with training time reduced to 8 h compared to 11-14 h for baseline methods. Expert evaluation by art and design professionals confirmed superior performance in both content fidelity and style integration across different artistic styles, including pointillism portrait style, impressionist landscape style, and abstract modern art style. Ablation studies validated the effectiveness of key components, with the complete model outperforming simplified versions in all quantitative metrics.

Conclusion The proposed MFFF-NST model effectively resolves fundamental technical limitations in floral printed pattern style transfer through systematic architectural improvements. The integration of line art extraction preprocessing, adaptive feature extraction mechanisms, and style-aware decoding successfully eliminates semantic ambiguity and feature discontinuity while preserving essential structural characteristics of printed patterns. This research provides a robust technical foundation for intelligent floral printed pattern design, significantly enhancing creative expression capabilities and production efficiency in textile manufacturing. The findings contribute to the advancement of AI-driven pattern design and offer practical solutions for modern fashion industry requirements.

Key words: style transfer, checkerboard artifact, shifted window attention mechanism, feature fusion, pattern design, printed pattern