| As the carrier of Chinese civilization,blue and white porcelain carries the mission of inheriting and promoting traditional culture.Ceramic craftsmen paint natural scenery,animals and plants on porcelain vases,plates,etc.to form blue and white color paper patterns,which are fired at high temperature to form a style of ink painting in shades of light and shade.In the field of deep learning,the development and application of style transfer technology is gradually becoming mature.Computers process abstract image styles based on pixel matrix,and cannot accurately distinguish images of different styles like the human eye.Therefore,extracting style features to solve the above problems has become a key research direction.The new generation of artificial intelligence technology is applied to the traditional ceramic field.Through deep learning,the characteristics of the style image of blue and white flower paper are extracted,and the content of the landscape image is expressed in the style of blue and white flower paper,which not only satisfies the diversified pursuit of contemporary people for art,and promote the development of ceramic art and culture.This paper studies and compares the three models of cycle GAN,GANILLA and improved SE-GANILLA for the style transfer of blue and white paper,because CycleGAN cannot distinguish the content and style of the generated blue and white color paper images.Therefore,this paper first uses GANILLA network for experiment.The network uses jump connection and up sampling to fuse low-level features and high-level features during down sampling.The purpose is to balance the style and content of the transferred image,and then use the improved SE-GANILLA network conducts comparative experiments.On the original GANILLA’s residual blocks,according to the channel attention mechanism theory of computer vision,SE block is added to improve the extraction of style and achieve the extraction effect of abstract style and tonal style.Finally,Ghost module is added to the generated network to improve the overall performance.With the SE-GANILLA model for blue-and-white paper style transfer,the experiment proves that at 5-10 epochs,when the pooling method Max-pool and the target optimization algorithm use Adam,the PSNR value of the test set generated image is19.5967,and the SSIM value is 0.7838.At 10-15 epochs,the PSNR value of the image generated by the test set is 20.9816,and the SSIM value is 0.8148.At this time,the "tone" effect of the style transfer of the blue and white flower paper is the best;The original GANILLA model has a PSNR value of 20.1192 and a SSIM value of 0.8161 at5-10 epochs,but its abstract subjective artistic effect is distorted.At 10-15 epochs,the PSNR value is 20.5047,the SSIM value is 0.8169,and the color tone In effect,SE-GANILLA is better for restoring the details of the original image.In this paper,we choose to artificially collect blue and white paper images to train style characteristics.The experiment found that the CycleGAN model takes about3600s-3900 s for each training,and the results are not ideal;the GANILLA model takes about 1800s-2700 s for each training time,which achieves the style transfer of blue and white paper.The basic requirements of the SE-GANILLA model,while the average training time per epoch is about 1700s-2500 s,and on the basis of the original blue and white flower paper style transfer,the generated patterns can be divided into partial"abstract" style and partial "pattern" according to the degree of training.Tone" style,the generated pattern is more diverse and practical.Based on the above research and analysis,to generate ceramic decorative products,users can transfer the style of blue and white flower paper according to their own landscape patterns to obtain personalized ceramic products with decorative patterns. |
