基于深度学习的三维人体估计方法研究

Method for 3D Human Body Estimation Based on Deep Learning

为提升虚拟试衣的视觉效果,针对三维试衣中服装下人体模型获取不易且通常精准度不高的问题,论文提出了一种基于深度卷积神经网络的人体体型估计方法。该方法以单张RGB图像的掩膜图作为输入,将整体图像剪裁为以关键点为中心的局部区域,通过依次提取每个局部区域的人体特征,逐步推断出整个人体模型。根据图像特性对传统VGG-16网络模型进行层间的增删,调整结构提高图像信息的学习效率和局部特征的检测速度。实验表明:论文方法可以有效提高对图像特征的检测,精确服装下的人体建模重建,使之在多种体型下都展现了较为理想的效果。

In order to improve the visual verisimilitude of virtual fitting,this paper presents a method to estimate human body based on deep convolutional neural networks for the problem that it is difficult to obtain the naked human body mesh in three-dimensional fitting and the accuracy of the model is not ideal usually.The method takes the mask map corresponding to a single RGB image and crops it into several local images centred on key points.Then it estimates the whole body model gradually by body features which are extracted in each local region.The structure of the VGG-16 network model is adjusted by augmented or deleted between layers according to the partial mask image characteristics,which improves efficiency in learning image information and detecting local features.Experiments demonstrate that the method can effectively improve the detection of image features and the accurate reconstruction of the human body under clothing,so that it shows more satisfactory results by a variety of body types.