基于单幅图像的三维发型建模技术及其应用

发布时间：2018-01-17 01:20

本文关键词：基于单幅图像的三维发型建模技术及其应用　出处：《浙江大学》2017年博士论文　论文类型：学位论文

【摘要】：发型是人物形象的重要组成部分,不同的发型颜色与风格可以反映出人的不同年龄、性格乃至职业。由于其复杂多变的几何结构,三维发型建模是图形学研究中极具挑战性的任务。现有三维发型建模方法存在着较大的局限性:一些依赖于手工交互的方法需要专业的建模技能以及繁杂的交互操作;另一些基于图像的发型重建方法只能处理受控环境下使用专业设备拍摄的大量图像。这些局限性阻碍了三维发型建模技术在实际应用中的普及。本文研究了基于单幅图像的三维发型建模这一难题。我们结合极少量用户辅助、机器学习方法以及先验几何模型,系统性地提出了一系列基于单幅肖像的高质量三维发型建模方法,并全面展示了这些方法在智能肖像操控中的应用。同时,针对应用中发型运动仿真速度慢的问题,我们设计了基于蒙皮的发型运动简化模型,用以在实时速度下完成高质量发型运动仿真及与环境物体交互的模拟。本文的具体研究内容包括如下五个部分:1.提出了一个基于单幅图像的交互式三维发型建模方法。该方法可以在少量用户辅助下生成高分辨的三维发型模型,这些发型模型在一定观察角度范围内都可以得到与原图相符的渲染效果,从而足以支持如肖像立体化、发型迁移以及发型几何编辑等肖像操控应用。此外,还对本方法进行了扩展以支持满足物理约束的动态发型建模及基于视频的发型建模。2.在交互式建模方法的基础上,我们提出了一个数据驱动的全自动单幅图像三维发型建模方法。其在保证结果质量的同时能完全去除繁琐的用户辅助步骤,从而将单幅肖像发型建模过程完全自动化,以便于更好地面向普通用户的实际应用以及大规模发型数据的生成。3.着眼于模型的几何质量,我们进一步提出了一个用于生成高质量模型的单幅图像肖像建模方法。该方法能够生成包含精细几何细节的完整肖像模型,用于支持一系列对模型质量有较高要求的肖像操控应用,包括基于图像的高质量重新照明、肖像浮雕三维打印等。4.针对实时发型运动仿真,我们提出了一个数据驱动的高效发型运动仿真方法。算法基于全新的发型蒙皮简化模型从训练数据中习得发型的运动特征以驱动高效的实时运动仿真,并通过在线碰撞修正的方式恢复因简化模型插值损失的运动细节。5.在发型蒙皮模型的基础上,我们进一步提出了自适应蒙皮简化模型,用以更好地解决发型与精细环境物体交互仿真过程中对蒙皮简化模型假设的冲突,从而在不影响效率的前提下实现高质量的发型物体仿真。
[Abstract]:Hair style is an important part of the character image, different hair color and style can reflect people's different age, character and even occupation, because of its complex and changeable geometry structure. 3D hairstyle modeling is a challenging task in graphic research. Some methods that rely on manual interaction require professional modeling skills and complex interactions; Other image-based hairstyle reconstruction methods can only deal with a large number of images captured by professional equipment in controlled environment. These limitations hinder the popularization of 3D hairstyle modeling technology in practical application. 3D hairstyle modeling of a single image is a difficult problem. We combine a very small number of user aids. A series of high quality 3D hairstyle modeling methods based on single portrait are proposed systematically by machine learning and a prior geometry model, and the applications of these methods in intelligent portrait manipulation are demonstrated comprehensively. At the same time. In order to solve the problem of slow simulation speed of hairstyle movement in application, we design a simplified model of hairstyle movement based on skin. It is used to simulate the high quality haircut motion and the interaction with the environment object at real time speed. The specific research contents of this paper include the following five parts:. 1. An interactive 3D hairstyle modeling method based on a single image is proposed, which can generate a high-resolution 3D hairstyle model with a small number of users. These hairstyles can be rendered in accordance with the original image within a certain range of viewing angles, which is sufficient to support portrait manipulation applications such as portrait three-dimensional, hairstyle migration and hairstyle geometric editing. The method is extended to support dynamic hairstyle modeling with physical constraints and video-based hairstyle modeling. 2. Based on the interactive modeling method. We propose a data-driven full-automatic 3D hairstyle modeling method for single image, which can completely eliminate the tedious user-aided steps while ensuring the quality of the results. Thus, the modeling process of single portrait hairstyle is completely automated, in order to better face the practical application of ordinary users and the generation of large-scale hairstyle data. 3. Focus on the geometric quality of the model. We further propose a single image portrait modeling method for generating high quality models, which can generate a complete portrait model with fine geometric details. It is used to support a series of portrait manipulation applications with high requirements for model quality, including high quality relighting based on images, 3D printing of portrait relief and so on. We propose a data-driven and efficient hairstyle motion simulation method, which is based on the new haircut skin simplification model to acquire the hairstyle movement characteristics from the training data to drive the efficient real-time motion simulation. And by online collision correction to restore the loss of the simplified model interpolation motion details. 5. On the basis of the haircut skin model, we further proposed an adaptive skin simplified model. In order to solve the conflict of the assumption of skin simplification model in the process of interaction simulation between hairstyle and fine environment objects, the high quality haircut object simulation can be realized without affecting the efficiency.
【学位授予单位】：浙江大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】