【作者】 邓卫燕；

【导师】 陆国栋；

【作者基本信息】 浙江大学 ， 机械设计及理论， 2008， 博士

【摘要】 服装个性化定制成为服装设计和制造的重要研究方向之一。三维人体建模是三维服装CAD系统的基础,决定了服装设计的质量和用户的满意度。三维人体建模得到的人体模型能否真正反映实际人体尺寸和外形取决于建模原理和实现方法。因此,对个性化人体建模方法进行分析和研究对推动服装个性化定制的发展有重要理论和应用意义。本文在分析国内外三维人体建模方法现状的基础上,提出面向服装设计的基于用户照片和神经网络的三维个性化人体建模方法。通过基于图像的人体特征区域和特征参数提取得到人体尺寸信息(尺寸、轮廓点和轮廓线);通过基于神经网络的三维人体特征曲线生成得到人体三维截面信息;以三维人体库中搜索的相似三维人体为体形信息载体,通过特征尺寸、曲线驱动相似三维人体变形,融合分治所得信息,快速生成三维个性人体。提出基于图像的二维人体特征区域和特征参数提取方法。通过图像预处理,利用差分法得到用户照片的二值化图像;以从用户照片图像识别出的特征点为参考,对用户照片图像的进行分割,实现人体特征区域的精确定位,为三维相似人体搜索提供依据;提出参照模型法特征尺寸计算方法,实现从图像像素中较高精度的特征参数提取。利用实际采集的图像得到人体特征参数,验证了所提方法快速可靠,为人体特征曲线的生成和三维个性化人体模型的建立提供二维尺寸信息。提出基于神经网络的三维人体特征曲线生成方法。阐述了训练样本数据的获取和神经网络三维人体曲线建模原理,给出了具体的神经网络结构模型。利用三维扫描人体数据得到胸部、颈部、腰部和臀部等特征截面的训练样本,通过训练确立人体特征部位的神经网络权值参数和特征曲线模型,弥补了从用户照片不能直接获取人体三维截面信息的不足。输入用户截面的宽度、厚度和围长等二维尺寸信息,可直接得到用户的三维截面的特征曲线,克服了以往通过曲线拟合方法需要大量三维测点信息的缺陷。提出基于扫描人体库的多层信息匹配相似人体搜索和曲面重建方法。以用户照片中识别的特征点、特征尺寸、轮廓线为信息输入,将三维相似人体搜索分为尺寸层、特征点层和轮廓线层三重筛选匹配过程,进而在三维扫描人体库中搜索到相似人体;利用曲面重建原理,对相似人体运用切片和分片技术进行表面重建,并进行光滑处理,得到具有光滑表面的三维标准人体。提出基于特征尺寸和特征曲线的混合驱动人体变形方法。利用从用户照片提取的特征尺寸和神经网络特征曲线生成模型获取的三维特征截面曲线信息,融合三维标准人体,混合驱动人体变形得到个性化三维人体模型。与前文基于图像的人体特征参数获取、神经网络特征曲线生成、三维相似人体的搜索和重建一起构成了个性化人体生成的完整体系。最后以本文的研究成果为技术基础,开发出个性化人体模型生成模块,并应用于三维服装设计系统LooksTailorX。本文给出了从照片中提取用户人体轮廓和特征参数、用户人体特征曲线生成、三维扫描人体模型重建、混合驱动人体变形和个性化人体生成等的实例。更多还原

【Abstract】 With the rapid popularization of the computer and network technology, MTM (Made to Measure) has become an important research direction for garment design. 3D (three-dimensional) human body modeling has an important role in MTM, and is the basis for 3D garment CAD system. If a 3D mannequin can truly reflect actual human size and shape depends on the modeling methods. Therefore, to stady and find new methods of modeling has important theoretical and applied significance.Based on current research status of human body modeling, this paper presents a method of 3D human personalized body modeling based on photo and neural network. 2D (two-dimensional) size information of human body is got by extraction of feature region and parameters based on user image. 3D cross-section information is got by generation of feature curve of human body based on neural network. 3D standard mannequins are generated by search for similar body from scanned human body database and by human surface reconstruction. Personalized mannequins are got through human deformation driven by feature size and curve of human body.A 2D human body feature region and feature parameters extraction method based on user image is proposed. On the basis of binary images of the user photo got through image pre-processing, a method of how to distinguish each feature region of the human body is given. This provides a basis for contour matching user body and similar body. According to the modeling method and the reference method, the way to recognize feature points and calculate feature sizes of human body deducted from image pixel is introduced. This method provides 2D size information for the generation of feature curve and 3D human body modeling.A general method for human body feature curves automatically generating based on neural network is proposed. The technique of training sample data acquisition and the model of neural network are given. By making use of real human body data, the training samples of the neck curve, bust curve, waist curve and hip curve are obtained. The weight and the feature curve of the neck, bust, waist and hip can be got respectively after training. It makes up the defect that 3D cross-section information can not be got from user photo. The error analysis is done and the results show that the method can approach user human body feature curve. Only with the thickness, size of girth, and other 2D information, cross-section information of the curve can be obtained.A method for searching for similar body from scanned human body database and human surface reconstruction is proposed. Firstly, scanned human body database is established, then making use of size, feature point and contour triple match, similar body best matched user body is searched for from scanned human body database, finally, according to the human surface reconstruction technique, a 3D standard mannequin is generated.A human body deformation method driven by feature size and curve is also proposed. According to size information obtained from user photo and cross-section information got from human body feature curve, 3D standard mannequin is deformed to personalized mannequin driven by feature size and curve. The above research constitutes a complete system for personalized human body modeling.A 3D personalized human body modeling module has been developed by exploiting the algorithms presented in this thesis and has been integrated into a garment design system named LookStailorX(LSX). Using the LSX, many examples including the technique for feature parameters extraction, human body feature curve generation, scanned body reconstruction, body deformation, combination drive, and personalized human body modeling are given.更多还原