【作者】 张静；

【导师】 姜延；

【作者基本信息】 北京服装学院 ， 服装设计与工程， 2023， 硕士

【摘要】 苗族蜡染图案是苗族文化的一种重要载体,具有丰富多样的形态以及独特的美学特征。由于蜡染工艺本身的复杂特性,以及市场需求的变动,使其在传承与创新中面临了许多挑战。而随着深度学习的快速发展,计算机可以对各类图像进行学习,并进行图像生成。因此,将深度学习与苗族蜡染纹样相结合,生成不同风格的新颖图片,不仅能够辅助设计师进行设计工作,提高创作效率,也能够帮助苗族服饰文化得到更好的传播。首先,本文通过收集、整理并分析苗族蜡染图案的分类、寓意、造型特点与构成规则,构建了带有文本标注的蜡染纹样数据集,为模型训练和纹样创新奠定了基础。其次,为了解决当前风格迁移算法在纹样风格迁移中存在的边缘模糊问题,本文提出了具有边缘增强的蜡染图案的局部风格迁移算法。损失函数由局部内容损失、局部风格损失和拉普拉斯损失组成,实现了蜡染纹样与风格图像的融合,生成图像在细节纹理和颜色空间上具有良好的性能。再次,本文训练得到了蜡染纹样LoRA模型,在基于扩散模型的图像生成中使用该模型能够实现输入文本生成对应的蜡染纹样的功能,生成的图像符合文本描述且有一定的创新性,生成纹样细节清晰,能够实现蜡染风格图像的创新生成。最后,构建了蜡染纹样创新设计系统,使用户能够自主上传图片进行风格迁移,也可以输入一段文本描述进行蜡染纹样的生成,另外对创新纹样在服装服饰中的应用进行了展示,能够为纹样的应用提供参考,并提高用户对蜡染纹样的兴趣。更多还原

【Abstract】 Miao batik patterns are an important carrier of Miao culture,with rich and diverse forms and unique aesthetic characteristics.Due to the complex characteristics of batik technology itself and changes in market demand,it faces many challenges in inheritance and innovation.With the rapid development of deep learning,computers can learn various types of images and generate them.Therefore,combining deep learning with Miao batik patterns to generate novel images of different styles can not only assist designers in design work,improve creative efficiency,but also help spread Miao clothing culture better.Firstly,this article collects,organizes,and analyzes the classification,symbolism,styling characteristics,and composition rules of Miao batik patterns,and constructs a text annotated batik pattern dataset,laying a foundation for model training and pattern innovation.Secondly,in order to solve the edge blurring problem of current style transfer algorithms in pattern style transfer,this paper proposes a local style transfer algorithm for batik patterns with edge enhancement.The loss function consists of local content loss,local style loss,and Laplacian loss,achieving the fusion of batik patterns and style images,and generating images with good performance in detail texture and color space.Once again,this article trained a batik pattern LoRA model,which can be used in image generation based on diffusion models to achieve the function of generating corresponding batik patterns from input text.The generated images conform to the text description and have a certain degree of innovation.The generated pattern details are clear,which can achieve innovative generation of batik pattern style images.Finally,an innovative design system for batik patterns was constructed,allowing users to independently upload images for style transfer,or input a text description to generate batik patterns.In addition,the application of innovative patterns in clothing and apparel was demonstrated,providing reference for the application of patterns and increasing user interest in batik patterns.更多还原