Traditional methods face challenges such as weak feature extraction, content distortion, and unstable training in the style transfer of Liao, Jin, and Yuan ceramic patterns.This study proposes a style transfer model based on an improved VGG16 and a cyclic consistency adversarial network, integrating multi-channel input, spatial attention, and a multi-scale feature pyramid, with a three-scale discriminator and structural similarity constraints.Experiments show that in Liao-to-Jin transfer, style accuracy reached 89.7%, content retention 91.2%, SSIM 0.801, and PSNR 27.9 dB.In Jin-to-Yuan and Yuan-to-Liao transfers, average accuracy and retention were 88.4% and 90.5%, with SSIM 0.792 and PSNR 27.5 dB.The three-scale discriminator improved style fidelity by 6.1%, and expert ratings averaged 8.7-9.1.While the method preserves semantic content and enhances stability for flat painted patterns, performance decreases for engraved designs with strong 3D features (e.g., Yaozhou kiln), with a success rate of 78.3%.
Zhao et al. (Thu,) studied this question.