基于感知一致性的过程纹理生成研究

发布时间：2018-04-10 01:22

本文选题：过程纹理　切入点：纹理生成　出处：《中国海洋大学》2015年博士论文

【摘要】：基于过程纹理模型产生和表示纹理是一个重要的研究课题,已经广泛的应用于计算机游戏,动画,艺术设计等领域。通过变化数学模型的输入参数可以有效的产生不同外观的纹理表面,然而,如何选择合适的过程纹理模型及其参数产生符合感知一致性的纹理一直是一个具有挑战性的问题。本文将视觉感知与机器学习相结合,从纹理感知研究入手,提出符合感知一致性的过程纹理生成框架,根据输入纹理,自动确定过程纹理模型及对应参数,以产生与输入纹理外观相似但是具有不同感知量的新纹理。论文主要的工作包括：一、构建了过程纹理数据库,数据库中包含23种过程纹理模型产生的450幅纹理高度图,以及在相同光照条件和视角下进行渲染得到的具有真实感的三维纹理图像。通过心理物理学实验获取了库中样本的感知数据,包括人对纹理样本的相似性判断及12种视觉感知特征的量化值。二、确定了过程纹理生成模型的感知特征及过程纹理的主要感知维度。通过构建纹理感知空间,确定了过程纹理的三个主要维度,每个维度是不同的特征组合。实验表明对于不同的感知维度,组合特征的权重根据纹理上下文环境发生变化,然而,感知维度却是不变的。另外,实验验证了构建的纹理感知空间的感知一致性,在纹理感知空间中的相似性度量是与人的感知相符的。三、提出了基于主成分分析(PCA)的深层卷积网络结构(PCN)的特征提取方法。PCN中包括级联的特征提取阶段和非线性输出阶段。卷积层中通过PCA学习滤波器组,低层卷积层输出的特征图根据不同的规则进行组合,下一层的滤波器基于组合的特征图学习得到。同时,利用间隔的图像块采样和特征池化单元加快了特征学习的速度。实验证明,该算法可以有效的提取出图像鲁棒特征表示,在不同的图像分类任务中都能达到与现有水平相当或者更优的性能。使用PCN算法提取纹理特征,可以准确地估计出纹理的感知特征量化值及对应的产生模型。四、提出了符合感知一致性的过程纹理生成算法框架,实现了基于纹理样例感知特征量化值的过程纹理生成。算法通过提取输入纹理的计算特征,估计其感知特征量化值和能够产生类似纹理的过程纹理模型。当用户在估计出的感知特征量化值基础上做相应变化时,通过在纹理感知空间中的感知相似性度量,确定参数最终产生符合用户提出的感知特征的新纹理。实验验证了算法可以适用于多种类型的纹理,能够有效的根据感知特征找到模型及对应的参数产生纹理。
[Abstract]:The process of texture model generation and texture is an important research topic based on, has been widely used in computer games, animation, art design field. The texture surface, through the input parameters of mathematical model can effectively produce different appearance however, such as how to choose the appropriate parameters of the process model and texture which conform with the perceptual consistency the texture is always a challenging problem. The visual perception combined with machine learning, starting from the study of texture perception, perception of the consistency of the process of texture generation frame, according to the input texture, texture model and the corresponding automatic process parameters, a new texture similar to produce with the input texture appearance but with different amount of perception. The main work includes: firstly, construct the process texture database contains 23 texture database process model The 450 texture height map, and realistic 3D texture image rendering in the same light conditions and perspective. Through psychophysical experiments to get the sensor data base samples, including quantification of the similarity of texture samples and 12 kinds of visual perception feature value. Two, the the main dimensions of texture perception characteristics and the process of texture model. By constructing texture perception space, identified three main dimensions of texture, each dimension is different combinations of features. The experimental results show that the different dimensions of the perceived weight, the combination of characteristics of the texture according to the context changes, however, the perception dimension is unchanged. In addition, experiments verify the consistency of the perception of texture perception space, similarity in texture perception in space is metric and the perception of the people are provided. Three. The principal component analysis (PCA) based on the deep convolutional network structure (PCN) feature extraction method in.PCN includes cascaded nonlinear feature extraction stage and output stage. Through the study of the PCA layer convolution filter, low layer convolution layer output feature map are combined according to different rules, the next layer of filter characteristics map based on the combination of learning. At the same time, the use of image block interval sampling and the characteristic of unit pool speed of feature learning speed. The experimental results show that this algorithm can effectively extract the robust image feature representation, in different image classification tasks to achieve the equivalent current level or better performance. The extraction the texture features using PCN algorithm can accurately estimate the quantitative texture perception feature value and corresponding model. Four, put forward with the process of texture generation algorithm of perceptual consistency of frame, solid The sample texture perception features quantized texture is generated based on the process by calculating the input texture feature extraction, texture feature estimation process model of the perception and quantitative values can produce similar texture. When the perceived characteristics of users in the quantitative estimate the value of the basis of the corresponding change, the texture perception in space the perception of similarity measure, determine the parameters eventually produce new texture with perceptual features presented by the user. Experiments show that the algorithm can be applied to many types of texture, can effectively according to the perceptual feature model and find the corresponding generated texture parameters.

【学位授予单位】：中国海洋大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】