Deep Learning-based 3D Human Perception And Dressing Image Generation

With the development of e-commerce,the online trade volume of clothing is increasing year by year.When shopping online,users can only see partial pictures or model fitting pictures provided by merchants,whose effect is far from actual fitting.This results in a high return rate and thus,virtual fitting has a wide market demand.The goal of virtual fitting is to help improve users' perception of the human body and dressing effect.In this paper,the author employs deep learning-based methods to conduct the research from the following three aspects: the association data-set of the human body and dressing image;3d human body perception;dressing image generation.The association data-set of multi-body and dressing image is constructed.The measurement errors of existing deep learning-based 3d human body perception methods are large,in addition,there is limited work on body shape in the field of dressing perception.One of the reasons is the lack of association data-set between dressing image and size of multi-body.In this paper,the fitting robot is used as a data acquisition platform to construct an association data-set of multi-body and dressing image.The data-set includes the clothing images of 100 typical body shapes,3 sets of clothing,14 clothing sizes and the key dimensions corresponding to each body shape.At present,no public data-set of real clothes worn by multi-body human body has been found.A 3D human perception method based on Convolutional Neural Networks is proposed.3D human body model based on dress image is a better way to experience human body perception.This chapter consists of two steps: human body key size prediction and 3D human body reconstruction.First,based on VGGNet,the author constructs a model whose input is the front and side dressing images and the output is bust,waist and hip.This model achieves better prediction performance than current image prediction methods.Second,the author utilizes the parametric human body model,establishes mapping relationship between the key dimensions of the body and the shape space of the human body,which realizes the efficient reconstruction of the 3d human model.A method of clothing image generation based on the Generative Adversarial Networks is proposed.Currently,most of the research on clothing effect perception is based on the attitude change rather than on the body shape change.In this chapter,first according to the user's key size,the fitting robot data with the closest body shape is found in the data-set,and its clothing image is taken as the deformation source image.Second,the deformation module in the GAN model is used to get the clothing deformation caused by the body shape change.Finally,the effect of dressing of the target user is obtained.