| With the rapid development of artificial intelligence in the fashion industry,there is a growing interest in fashion design and generation technologies.In addition to the large demand for fashion items on a daily basis,people are also seeking for personalization,artistry,diversity,etc.of fashion items.However,for most people,the time,resources and expertise required to perform fashion design and generation are unaffordable.And with the continuous development of deep learning technologies,fashion generation algorithms have excelled in tasks such as personalized fashion design,art creation,and virtual fitting,which can not only facilitate the development of emerging fields such as fashion art design,but also drive the fashion industry toward a more innovative and sustainable direction.In recent years,neural networks have been widely used in work related to computer vision and fashion generation,and in particular,the proposal of generative adversarial networks has led to a series of image generation tasks related to them receiving a lot of attention from researchers.The content and style transfer of fashion items based on generative adversarial networks is a very interesting and meaningful task.In this thesis,we focus on the style transfer of fashion items through generative adversarial networks and the generation of fashion items with multi-domain feature fusion.The research content and innovation points of this thesis are as follows:(1)A fashion style transfer method based on generative adversarial networks is proposed.Traditional style transfer methods require separate training for each style image,and the static normalization methods used(such as normalization or instance normalization)cannot handle the dynamic changes between different styles.To address these problems,this thesis proposes a fashion style transfer method based on generative adversarial networks,which can achieve arbitrary style transfer for fashion items.In particular,this thesis designs a layer-consistent dynamic convolutional model based on generative adversarial networks,which mainly contains two parts: dynamic instance normalization and fashion style fusion.Among them,dynamic instance normalization can dynamically adjust the normalization parameters according to different style images to achieve smooth transition between different styles;fashion style fusion is to dynamically convolve style features with fashion items,thus improving the flexibility and adaptability of the model.The experimental results show that the method has significantly improved the generation quality of fashion items,the information retention of global structure and the style transfer effect.(2)A fashionable content and style transfer method based on generative adversarial networks is proposed.Currently,most image generation tasks are limited to conversion between only two domains(e.g.,horse to zebra,natural landscape to oil painting style,winter to summer),ignoring the effective feature fusion of multiple image domains.To address this problem,this thesis proposes a fashion content and style transfer method based on generative adversarial networks,i.e.,given a fashion item,a content image(e.g.,maple forest)and a style image(e.g.,Monet style),a fashion item fusing maple forest landscape pattern and Monet style can be generated,thus achieving feature fusion in multiple image domains.In this thesis,based on the style transfer of fashion items,we propose a fashion content generation method based on contrast learning,which maps fashion items and generated images to feature vector space after encoding,and learns feature representation by maximizing similarity and minimizing difference,so as to make the features of content image domain more obvious,and then realize the content transfer of fashion items.Further,we construct a two-branch discriminator that can more fully integrate content and style features.The experimental results show that the method can effectively achieve feature fusion of multiple image domains,enhance the diversity of fashion item generation,and improve the generation quality. |
