Research on textile pattern structure reconstruction and innovative design method based on deep learning

Abstract

Textile patterns are important visual carriers in garment design, home textiles, digital printing, and cultural creative industries. However, traditional textile pattern design relies heavily on designers’ experience and often suffers from low efficiency in structural extraction, difficulty in reconstructing repeated motifs, and limited controllability in innovative generation. More importantly, most existing textile-pattern generation models mainly optimize visual plausibility at the image level, while the editable structural logic of textile design, including motif decomposition, repeated-unit organization, symmetry, and spatial rhythm, is rarely modeled explicitly. To address this research gap, this paper proposes a deep learning-based framework for textile pattern structure reconstruction and innovative design. The proposed method establishes a structure-first reconstruction paradigm that integrates dataset construction, multi-level structural representation, structural feature extraction, motif-level graph reconstruction, and structure-constrained generative design. Specifically, textile patterns are first annotated and represented by motif regions, contour boundaries, color areas, repeated units, and spatial relationships. Then, an improved structural feature extraction network with multi-scale fusion and attention mechanism is designed to enhance the recognition of complex motif boundaries and repeated structures. Motif units are further modeled as graph nodes, while spatial relationships such as adjacency, symmetry, translation, rotation, and repetition are modeled as graph edges to support controllable structure reconstruction through relation-aware attention. Finally, the reconstructed structure is used as an explicit condition for generative design, enabling innovative textile pattern generation while preserving structural regularity. Experimental results show that the proposed method achieves 84.96 % mIoU and 90.32 % Dice in structural recognition, 0.887 SSIM, 29.64 dB PSNR, 0.142 LPIPS, and 88.35 % structural preservation rate in reconstruction, as well as 21.56 FID, 3.72 IS, 0.867 diversity index, and 6.12 expert score in innovative generation. These results demonstrate that the proposed framework advances intelligent textile design from appearance-oriented image synthesis toward controllable, interpretable, and design-applicable structural innovation.