虚拟试衣技术探索性评价研究

【作者】 张莉；

【导师】 马大力； 许君； 章莹；

【作者基本信息】 天津工业大学 ， 材料与化工， 2023， 硕士

【摘要】 三维虚拟服装是当前虚拟技术研究中的一个热门课题,由“元宇宙”、“虚拟世界”、“数字资产”等词感知到虚拟服装对虚拟技术的强烈影响。随着虚拟技术的发展,人们对于虚拟服装造型真实程度要求越来越高。所以本文为建立虚拟服装和真实服装造型相似评估模型,提高对虚拟环境中的服装的造型相似性评估效率,降低人工审核的时间和经费成本。主要做了如下研究:首先,通过文献研究法系统分析了影响虚拟服装造型的重要面料性能参数,选择KES织物风格仪、主成分分析法和BP神经网络评估模型作为探究虚拟服装和真实服装造型相似性的重要科研手段。而后,运用熵权-TOPSIS法和专家评价法,综合主客观结果抉择了制作虚拟服装的载体Vidya。基于虚拟载体的确定,选取36款的面料进行悬垂形态观察,主观判断面料俯视静态悬垂效果差异明显的面料,得到共24款面料并对其进行拉伸、弯曲和压缩等物理机械性能的KES测试,通过对比KES与Vidya的面料相关性能测试原理,建立基于KES测试数据的Vidya虚拟面料,作为虚拟服装的储备面料库。其次,由选中的面料制成真实T恤和对应的虚拟T恤。通过交互改变Vidya的两个面料参数,即弯曲刚度参数、折叠体积和形状参数,增加测试样品数,获得虚拟面料96种,由此制得96件虚拟T恤样品。使用专家问卷调查法确立了真实T恤和虚拟T恤造型相似性的评估指标,分别针对T恤的领部、衣身轮廓和褶皱的差异显著性,对真实T恤和虚拟T恤在正面、侧面和背面进行感官评估,并进行综合打分。基于轮廓和褶皱差异共设计两个模块指标。其中,轮廓差异性指标是T恤在领部、衣身的宽度、长度、腰高比和腋下夹角和肩部夹角,褶皱差异性指标是褶皱数量、褶皱宽度和褶皱深度,共得到21项指标。接下来是基于计算机图像处理技术,建立图像的二维直角坐标系,通过截取图像指定部位的算法提取和量化轮廓指标,并使用灰度直方图均衡化,图像分割技术等提取和量化褶皱评估指标。最后,对21项量化完成的评估指标进行主成分分析(PCA),提取出具有代表性12项指标,分别为肩领夹角α、宽度W2、夹角β1、夹角γ、前腰高比B1、褶皱数量N1、褶皱宽度K1、夹角β2、褶皱宽度K2、后腰高比B2、褶皱数量N3、褶皱宽度K3。运用Pearson相关性分析量化指标与感官评估指标的相关程度。依据PCA提取出的量化指标作为真实T恤和虚拟T造型相似性评估模型(BP模型)的输入端,感官评估结果作为评估模型的输出端,再分别建立训练集、测试集和验证集,使用MATLAB2021a编程实现了真实T恤和虚拟T造型相似性评估模型。本文训练的模型在以隐藏层神经元个数为4时效果最好,模型整体相关系数R达到0.95962。上述研究成果都通过严谨的实验和论证。表明使用计算机图像处理技术、感官评价感知技术、BP神经网络模型和其他处理方法,实现了虚拟服装和虚拟服装造型相似性评价模型的建立,能够进一步优化虚拟试衣系统效果,并推动虚拟服装造型相似性智能评价程序或者软件的开发设想。更多还原

【Abstract】 Three-dimensional virtual clothing is a hot topic in the current research on virtual technology,as perceived by words such as "metaverse","virtual world" and "digital assets".The strong influence of virtual clothing on virtual technology is evident from terms such as "metaverse","virtual world" and "digital assets".With the development of virtual technology,people are demanding more and more realism in the modeling of virtual clothing.Therefore,this paper,to establish a model for assessing the similarity between virtual and real garment styling,to improve the efficiency of assessing the styling similarity of garments in virtual environments,and to reduce the time and cost of manual auditing.The following research is mainly done.Firstly,the important fabric performance parameters affecting the virtual garment shape were systematically analyzed through literature research,and the KES fabric style meter,principal component analysis,and BP neural network evaluation model were selected as important research tools to investigate the similarity between the virtual and real garment shapes.Based on the determination of the virtual carrier,36 types of fabrics were selected for draping observation and subjective judgment was made on fabrics with significant differences in draping effect in top view.24 types of fabrics were obtained and tested for physical and mechanical properties such as tensile,bending and compression.By comparing KES with Vidya’s fabric-related performance testing principles,a Vidya virtual fabric based on KES test data was established as a reserve fabric library for the virtual garment.Secondly,real T-shirts and corresponding virtual T-shirts were made from selected knitted fabrics.By interactively changing two fabric parameters of Vidya,i.e.bending stiffness parameter,folding volume,and shape parameter,96 virtual fabrics with adjusted parameters were obtained to increase the number of test samples,thus 96 virtual T-shirt samples were produced.An expert questionnaire was used to assess the similarity between the real T-shirt and the virtual T-shirt.The T-shirt was evaluated on the front,side,and back of the T-shirt and scored for the significance of the differences between the collar,the silhouette,and the pleats.Two module indicators are designed based on silhouette and crease differences.The indicators of silhouette variability are the width,length,waist-height ratio,underarm angle,and shoulder angle of the T-shirt at the collar and body,and the indicators of crease variability are the number of creases,crease width,and crease depth,resulting in a total of 21 indicators.The next step is to establish a two-dimensional rectangular coordinate system of the image based on computer image processing techniques,extract and quantify the contour indicators by an algorithm that intercepts the specified parts of the image,and extract and quantify the pleat assessment indicators using grey-scale histogram equalization,image segmentation techniques and so on.Finally,a principal component analysis(PCA)was performed on the 21 quantitatively completed assessment indicators to extract 12 representative indicators,namely shoulder collar angle α,width W2,angle β1,angle γ,front waist height ratio B1,number of pleats N1,pleat width K1,angle β2,pleat width K2,back waist height ratio B2,number of pleats N3 and pleat width K3.Pearson The correlation between the quantitative indicators and the sensory evaluation indicators was analyzed using Pearson correlation.The quantitative indicators extracted based on PCA were used as the input of the real T-shirt and virtual T-styling similarity assessment model(BP model),and the sensory assessment results were used as the output of the assessment model,then a training set,a test set and a validation set were created respectively,and the real T-shirt and virtual T-styling similarity assessment model was implemented using MATLAB2021 a programming.The model trained in this paper works best when the number of neurons in the hidden layer is 4,and the overall correlation coefficient R of the model reaches 0.95962.The above research results have been rigorously experimented with and validated.It is shown that the use of computer image processing technology,sensory evaluation perception technology,BP neural network model,and other processing methods to achieve the establishment of virtual clothing and virtual clothing shape similarity evaluation model can further optimize the effect of the virtual fitting system and promote the development of virtual clothing shape similarity intelligent evaluation program or software idea.更多还原