Attention-enhanced CycleGAN with multi-scale VGG feature fusion for ceramic decorative pattern style transfer

Ceramic decorative patterns (CDP) are essential for enhancing ceramic artistic value, but traditional handmade design faces low efficiency and difficulties in maintaining structural integrity, color reproduction, and texture consistency during style transfer. To address these challenges, this study proposes an improved CycleGAN framework for high-quality CDP style transfer. The original convolution module is replaced with VGG19 to capture multi-scale semantic and fine style features. A self-attention mechanism (SAM) is integrated into the generator to model global pixel dependencies and improve pattern continuity and style consistency. In addition, an adaptive sharpening preprocessing mechanism is designed to enhance image clarity and reduce noise interference. Experimental results show that the proposed model achieves a PSNR of 25.10 dB, an SSIM of 0.87, and an FID of 12.89, outperforming the baseline model. The experiments are conducted on the Chinese Ceramic Pattern Dataset (CCPD) containing 12,000 high-quality ceramic pattern images. Practical evaluations demonstrate improved visual satisfaction, pattern integrity, and process feature restoration, while significantly reducing design time for both small and large batches. The proposed method effectively improves CDP design efficiency and artistic expression, providing technical support for ceramic art innovation.