摘要
Abstract
Blue calico is a traditional craft printing and dyeing product in China with a long history.It is a simple and primitive blue and white fabric that reflects people's preferences and embodies their longing for a happy life and simple aesthetic taste.It is famous for its unique pattern design style and broken lines.However,the lack of an algorithm for the automatic generation of blue calico patterns has hindered the innovative research of patterns.For this reason,an end-to-end automatic generation method of the blue calico pattern was proposed to realize the automatic conversion from a simple stroke image to a single blue calico pattern.
Our solution belongs to CycleGAN-based methods,which are.popular to stylize images in artistic forms such as painting.CycleGAN is a generative adversarial network based on dual learning,overcoming the the limitations of requiring corresponding datasets of other methods,but it is slightly insufficient for geometric shape changes.Given the abstract style and small dataset of blue calico,the coder and decoder in the CycleGAN generation network are reconstructed,and the SE(squeeze and exception)attention module and residual module are connected in series with the original convolution module to improve the ability of feature extraction and network learning.SE attention mechanism gives different weights to different positions of images from the perspective of the channel domain,with the network focusing on key information.At the same time,reducing the number of residual block layers of the converter in the generative network to reduce over-fitting is implemented.Besides,to carry out this experiment,we have made the dataset of a single pattern of blue calico and the dataset of simple strokes.
Due to the lack of previous research on this issue and the lack of appropriate measurement indicators for such experiments,the study used the generated image for visual comparison.In the experiment,to prove the superiority of the proposed algorithm,we compared our algorithm with the original CycleGAN and other algorithms based on the dual idea,like DuaIGAN and DiscoGAN.The experimental results show that our proposed algorithm can effectively combine the content of simple strokes with the style of blue calico,while the original CycleGAN algorithm has insufficient geometric deformation ability,and the other two algorithms save too much original information,which does not conform to the concise characteristics of blue calico.At the same time,the study also compared the effect of using different attention mechanisms,and compared the ECA(efficient channel attention)mechanism with CBAM(convolutional block attention module)mechanism.Among various attention mechanisms,the SE attention mechanism still has the best effect.The image generated by using ECA attention lacks the content of the original image;the image style conversion generated by using ECA attention is not enough.Besides,compared with the other two methods,the image generated by using the SE attention mechanism has smoother color block edges and less noise.The reason for the gap caused by using different attention mechanisms may be that different attention mechanisms increase the learning ability of generators to different degrees,and some make it difficult for generators and discriminators to form effective competition.Through the above experiments,the blue calico automatically generated by the method based on the SE attention CycleGAN network is closer to the original style.Through this algorithm,the design process of the blue calico can be simplified,which is beneficial to the digital inheritance and innovation of the blue calico.
As a national intangible cultural heritage,the blue calico is of important value and significance.In this paper,an end-to-end network SE-CycleGAN based on the SE attention mechanism is proposed.Compared with other networks,the content of the image is better integrated with the style of the blue calico.The proposed network does not need matching datasets,so it can be well migrated to the generation of other similar patterns.Nevertheless,due to the limitation of the dataset and resolution,the patterns generated at present are still relatively simple.The next research work will focus on studying the method of generating multi-class and complex blue calico.关键词
蓝印花布/SE注意力/风格迁移/CycleGAN/单纹样/半监督学习/图像生成Key words
blue calico/SE attention/style transfer/CycleGAN/single pattern/semi-supervised learning/image generation分类
轻工纺织
