基于扩散模型的印花图案生成方法设计  

Design of Printed Pattern Generation Method Based on Diffusion Models

作　　者：张佳伟 李华军 王秀丽 朱威[1,2] ZHANG Jiawei;LI Huajun;WANG Xiuli;ZHU Wei(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023,China;Joint Key Laboratory of Embedded Systems of Zhejiang Province,Hangzhou 310023,China)

机构地区：[1]浙江工业大学信息工程学院,杭州310023 [2]浙江省嵌入式系统联合重点实验室,杭州310023

出　　处：《计算机测量与控制》2024年第10期243-249,共7页Computer Measurement &Control

基　　金：国家自然科学基金青年项目(62303414);浙江省自然科学基金探索青年项目(LQ23F030016)。

摘　　要：印花图案的设计是服装生产领域非常关键的一环,但目前人工设计的印花图案存在内容相似化、设计效率低下的问题,因此设计并实现了一种基于扩散模型的印花图案生成方法;采用深度学习技术提取、扩充现有印花图案数据集,并从颜色和类别维度生成印花图案的文本描述,完成印花图案数据集制作;使用已制作的数据集微调扩散模型,并对图像特征空间进行平铺处理,使得生成的印花图案在边界过渡处满足纺织行业四方连续的要求。对局部扩散特性进行了分析,结合文本引导,实现细节可变的图文生图效果;实验结果表明,所设计的印花图案生成方法具备生成高质量印花图案的能力,并且其特征空间平铺方法使得印花图案边界过渡处较为自然。It is a crucial aspect to design printed patterns in the field of clothing production.However,manually designed patterns often have the shortages of content similarity and low design efficiency.Therefore,a printed pattern generation method based on a diffusion model was designed and implemented.This method employs a deep learning technology to extract and expand existing printed pattern datasets,generate the textual descriptions of these patterns from the dimensions of color and category,and complete the production of the dataset for printed patterns.The pre-made dataset is used to make small adjustments to the diffusion model and flatten the processing of the image feature space,which makes the generation patterns at the boundary transition edges meet the requirements of the textile industry.This paper analyzes the characteristics of local diffusion,and achieves a image generation effect with variable detail from images and text.Experimental results demonstrate that the designed pattern generation method is capable of producing high-quality patterns,and its feature space tiling method makes the transition of the printed pattern boundary more smooth.

关 键 词：印花图案 深度学习 多模态融合 扩散模型 图像生成 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]