摘要
针对当前图像风格迁移算法在云锦线稿纹样合成中出现的轮廓模糊、色彩混杂的问题,本文提出了一种增加色彩损失计算的云锦局部风格迁移模型,即定义色彩像素在R、G、B通道的方差和为色彩损失,采用Adam函数优化色彩损失,以减小合成纹样的色彩差异,利用掩码图获得局部迁移轮廓,生成高识别度的云锦风格化效果图。实验结果表明:利用色彩优化的局部迁移算法获得的云锦效果图综合相似度(内容与风格)均值为0.38,相比传统的局部迁移算法提高了9.14%。优化模型在实现局部风格迁移的基础上,通过减小色彩差异,提高了合成纹样的语义辨识度,为云锦图案的创新设计提供了方法。
Nanjing brocade is a traditional handmade silk jacquard fabric made by the Dahualou loom with silk,peacock feather,gold and silver threads and other raw materials.As one of the three famous brocades,Nanjing brocade has exquisite patterns and rich colors.According to the survey,the number of brocade colors can reach dozens at most,so the patterns in the same latitude can show different colors.However,the traditional method of Nanjing brocade pattern design is time-consuming,inefficient,and it is easily affected by the skill level of inheritors,which greatly limits the innovation and inheritance of Nanjing brocade patterns.With the development of image processing and artificial intelligence technology,image style transfer algorithms provide new methods for pattern design.The algorithms utilize convolution neural networks(CNNs) to extract the contents and styles of images,and reorganize them to generate new art images.However,the present style transfer algorithms always lead to blur edges and mixed color confusion on transferred images when applying on Nanjing brocade.In order to inherit and innovate the pattern design of Nanjing brocade,we designed a local transfer model with color optimization function in this paper.Through many experiments,the transfer parameters were adjusted,and a better effect of generated image was obtained.Furthermore,in order to verify the effectiveness of the model,we compared the results between the proposed model and other state-of-the-art models.The testing results show that our method achieves ideal results of structure similarity and subjective evaluation.The contributions of our approach are three-folds:(i) utilizing the masks of the content image to isolate patterns from the background,in which the masks can avoid the blur of outlines.(ii) The sum of variance of pixels in the generated image in the three color channels is defined as color loss.Combining color loss with content loss and style loss,the new loss function obviously improves the semantic clarity of the pattern.(iii) Combining color loss and mask,we finally get a highly recognizable brocade stylized effect image with clear outline.As for the quality evaluation of effect images,we take advantage of subjective evaluation and objective evaluation separately.The subjective evaluation of the experimental results shows that the average scores of the original transfer,local transfer and color optimized local transfer models are 2.18,3.63 and 3.88,respectively.The data shows that the color optimized local transfer model not only realizes the local transfer but also improves the clarity of the semantic content of the pattern through color optimization.In the meanwhile,the structural similarity values of the effect images are calculated for objective evaluation.The grand mean of the comprehensive similarity of the effect images after the optimization is 0.382,which is 9.14% higher than that of local transfer algorithm.We find that the optimized transfer algorithm can obtain stylized images which are more similar to the structure of the content image,and it can be utilized to tackle the problem of style transfer on Nanjing brocade images.The local transfer model based on color optimization proposed in this paper not only improves the definition of contour by using mask,but also reduces the color difference and improves the semantic recognition of pattern by optimizing color loss.The optimized model is more suitable for the transfer of Nanjing brocade,and also provides an idea of method for other silk products.
作者
邱雪琳
孙迎
刘正
侯珏
QIU Xuelin;SUN Ying;LIU Zheng;HOU Jue(School of Fashion Design&Engineering,Zhejiang Sci-Tech University,Hangzhou 310018,China;MOC Key Laboratory of Silk Culture Heritage and Product Design Digital,Hangzhou 310018,China Technology,Zhejiang Sci-Tech University;Zhejiang Provincial Research Center of Fashion Engineering Technology,Zhejiang Sci-Tech University,Hangzhou 310018,China)
出处
《丝绸》
CAS
CSCD
北大核心
2022年第12期64-70,共7页
Journal of Silk
基金
文化和旅游部重点实验室资助项目(19076223-B)。
关键词
风格迁移
色彩损失
掩码图
云锦
语义清晰度
迁移效果
style transfer
color loss
mask graph
Nanjing brocade
semantic clarity
transfer effect
