A novel DAGAN for synthesizing garment images based on design attribute disentangled representation. (April 2023)

• Record Type: Journal Article
• Authors: Fang, Naiyu; Qiu, Lemiao; Zhang, Shuyou; Wang, Zili; Hu, Kerui; Wang, Kang
• Is Part Of: Pattern recognition. Volume 136(2023)
• Publication Date: 2023-04
• Subjects: DAGAN -- Garment design attributes -- Disentangled representation -- Garment image synthesis -- Online costume design; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.109248
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: The various clothing styles are disentangled into general four design attributes. DAGAN is proposed to synthesize garment images by entangling design attributes. DA-dataset about garment design attributes has been released. Abstract: In online costume design, it is vital to preview the design effect rapidly by entangling design attributes from reference images. This paper proposes a novel method, named design attributes generative adversarial network (DAGAN) for synthesizing garment images based on design attribute disentangled representation. The garment style is disentangled into the shape, texture, shadow, and decoration design attributes. The shape mask, repeating texture region, Laplace image gradient, and local logo are leveraged as visual representations for clothing design attributes from reference images. Following the design sequence from global to local, GDA-Net and LDA-Net in DAGAN entangle global and local design attributes, respectively, in the latent space. Then, the desired garment image is synthesized to represent design intentions explicitly. The DA-dataset for clothing design attributes is released. Extensive experiments demonstrate that the DAGAN is robust to various instances of design attributes on Design Attributes dataset (DA-dataset), and that is superior to the cross-domain transfer models in entangling design attributes from reference images.