| The integration of artificial intelligence technology and the fashion industry has become a hot topic in recent years.It aims to apply artificial intelligence technology to help the traditional fashion industry and facilitate the digital transition of the fashion industry.Artificial intelligence technologies have been researched and deployed in different application scenarios in the fashion sector,with two popular directions being virtual try-on and intelligent fashion design.The thesis focuses on studying controllable person image synthesis algorithms based on deep generative models.Specifically,we propose two controllable person image synthesis algorithms for virtual try-on and intelligent fashion design,respectively.(1)Firstly,this thesis proposes a new pose-guided controllable person image synthesis algorithm for virtual try-on.Existing works on pose-guided virtual try-on usually use paired source-target images to supervise the training,which significantly increases the data preparation effort and limits the application of the models.To this end,this thesis proposes a novel self-supervised per-region feature normalization for pose-guided virtual try-on.It only uses the source person image as supervised information and allows the user to flexibly manipulate the pose and appearance attributes of the person by decoupling the pose and appearance for achieving pose transfer and controlling the appearance attributes of the person.In comparison to conventional supervised learning-based pose transfer methods and unsupervised learning-based methods,the experimental results demonstrate the effectiveness of the proposed method,which provides a new solution for virtual try-on tasks.(2)Secondly,this thesis proposes a multimodal condition-guided controllable person image synthesis algorithm for fashion intelligent design.Existing work on fashion intelligent design focuses on single-view generation of a single person picture and uses end-to-end convolutional neural networks to achieve one-to-one mapping,which has issues including single image generation effect and limited controllability.To this end,this thesis proposes a new approach for fashion intelligent design that can generate a variety of multi-view fashion designs conditioned on a human pose and texture examples of arbitrary sizes,which can replace the repetitive and low-level design work for fashion designers.Firstly,employ a layout generative network to transform an input human pose into a series of person semantic layouts.Secondly,propose a texture synthesis network to synthesize textures on all transformed semantic layouts.Finally,leverage an appearance flow network to generate the fashion design images of other viewpoints from a single-view observation by learning 2D multi-scale appearance flow fields.Experimental results demonstrate that the proposed method can effectively solve challenging multi-modal image translation problems,which provides a new solution for fashion intelligent design tasks.In conclusion,this thesis proposes two practical and new deep generative models for virtual try-on and fashion intelligent design,respectively.Extensive experimental results demonstrate the efficacy and practicality of the proposed method. |
