现代纺织技术 ›› 2023, Vol. 31 ›› Issue (4): 250-258.

• • 上一篇    下一篇

基于CycleGAN的服装图像混搭风格迁移

  

  1. 大连工业大学,a.服装学院; b.服装人因与智能设计研究中心,辽宁大连116034
  • 收稿日期:2022-11-11 出版日期:2023-07-10 网络出版日期:2023-09-12
  • 作者简介:张功(1995—),男,山东滕州人,硕士研究生,主要从事服装图像处理方面的研究
  • 基金资助:
    教育部社科规划基金项目(21YJAZH088);辽宁省教育厅高校基本科研重点攻关项目(LJKZZ20220069);教育部产学协同育人项目(220404211305120); 辽宁省教育厅项目(1010152); 中国纺织工业联合会项目(2021BKJGLX321)

Mix and match style transfer for the images of clothes with CycleGAN#br#

  1. a. School of Fashion; b. Clothing Human Factors and Intelligent Design Research Center, Dalian Polytechnic University, Dalian 116034, China
  • Received:2022-11-11 Published:2023-07-10 Online:2023-09-12

摘要: 为解决复杂背景下服装图像的风格迁移形式单一和局部细节失真问题,提出一种基于CycleGAN的服装图像混搭风格迁移的方法,用于实现服装款式和图案的多风格迁移。通过加入分割掩码,一方面,对特定区域的风格化形成空间约束,在判别器中加入谱归一化和引入背景优化损失保留了局部细节的真实度,实现服装风格款式的风格迁移;另一方面,提出图像融合的方式,将图案融入判别器输出的服装图像中,实现多风格迁移。最后,通过与CycleGAN和InstaGAN比较,依据生成图像的效果进行主观分析,使用图像质量评估指标IS和SSIM进行客观评估以验证该方法的有效性。

关键词: 混搭风格迁移, 生成对抗网络, 服装款式, 服装图案, 智能设计, 服装设计

Abstract: With the continuous integration of artificial intelligence (AI) technology and the fashion field, the use of style transfer technology to generate new images has become one of the research hotspots of aided intelligent clothing design. However, the use of current style transfer technology in the process of aided intelligent design still has great limitations. Only completing the transfer of a single style limits the diversity of generated clothing images, and the detail distortion of the clothing image with the character background reduces the authenticity of the generated clothing images.
Aiming at solving the problems of the undiversified transfer form and local detail distortion of clothing image styles in complicated conditions, a method of clothing image mixing and matching style transfer was proposed to realize the multi-style transfer of clothing styles and patterns. During the experiment, we took CycleGAN as the baseline model, with the advantage of improving the effect of style transfer without requiring pairwise training and cyclic consistency loss, used Resnet generator and PatchGAN discriminator for training, and introduced the segmentation mask. On the one hand, spatial constraints were formed for the stylization of specific areas, instance normalization was added to the discriminator to maintain the independence of image instances, spectral normalization was added to the first and last layers of the convolution layer to enhance the classification ability of the network, and background optimization loss was added to optimize the local details of the generated images, especially the boundary artifacts, which jointly promoted the generation effect and realized the style transfer of clothing styles. On the other hand, the method of image fusion was proposed. According to the pattern fusion mapping relationship, the pattern was integrated into the clothing image output by the discriminator to realize the multi-style transfer of clothing styles and patterns. In order to verify the effectiveness of the above method in the multi-style transfer of clothing images, the clothing image design sketches generated in the experiment were compared with the design sketches generated by CycleGAN and InstaGAN models. By subjectively analyzing the style diversity and detail differences of the design sketches, the IS and SSIM were used for quantitative analysis, and the subjective visual effect and objective numerical comparison both showed the advantages of this experimental method in the diversity and the authenticity of image details.
The multi-style transfer of clothing styles and patterns can provide designers with creative inspiration and shorten the time period required for effect presentation, making AI more suitable for assisting clothing design behavior. In subsequent experiments, we will explore the transfer of more types of clothing styles, so as to achieve a diversified and controllable style transfer method. In addition, emotional elements needed for perceptual design should be added in the process of image style transfer, so as to promote the integration of computational thinking and design thinking of future-oriented design paradigm.

Key words: mix and match style transfer, generative adversarial network, clothing style, clothing pattern, intelligent design, clothing design

中图分类号: