【信息技术】【1998.06】变方向光源照射下纹理表面的分类研究

本文为英国赫瑞·瓦特大学（作者：G. McGunnigle）的博士论文，共268页。

本文在物理背景下进行纹理分析。所用模型从相关文献中获得，并集成到将粗糙表面链接到基于分类的特征集的过程描述中。第一类是粗糙表面，从摩擦学和散射学的角度选择表面形貌模型。考虑各种反射率模型，评价并讨论了相关文献的表面/图像关系的光谱模型。研究并描述了入射图像与捕获数据集之间的关系。该模型与特征测度的谱描述相结合，形成从表面到特征集的过渡模型。

从该模型中可以清楚地看出，照明方向可以影响从给定表面获得图像的方向性。光照方向的变化将导致特征输出的变化。如果光照方向在训练和分类之间发生改变，则分类规则可能不适当，从而导致分类效果较差。本文考虑了几种方案来解决这个问题。其中一种技术通过选择使用物理表面的表示作为生成适当训练数据的基础来完成进一步评估。利用测光技术估计训练表面的表面导数场。渲染算法使用这些估计来模拟训练表面从任意方向照射时的外观。结果表明，在照明方向变化的情况下，该系统能够显著地优于朴素分类器，并且在一些情况下接近于在执行分类条件下训练分类器所获得的精度水平。

纹理分析是机器视觉中的一个重要领域，这在很大程度上是由于纹理特征在早期视觉系统中的重要作用。由此可见，纹理对于在无约束环境中工作的一般视觉系统而言是至关重要的，因此，较少强调更受控的检验查视任务。在无约束系统中，采用建模方法是不切实际的，纹理分析中的大多数工作都是以图像为出发点。本文对粗糙纹理表面的检测进行了研究。通过明确分类发生的环境，我们能够采用系统的建模，并在产生纹理图像的物理系统环境中描述纹理特征分类。

This thesis sets texture analysis in aphysical context. Models of the system components are obtained from theliterature and integrated into a description of the process linking the roughsurface to the feature set on which classification is based. The firstcomponent is the rough surface, models of the surface topography are selectedfrom the fields of tribology and scattering. Various reflectance models areconsidered and a spectral model of the surface/image relationship from theliterature, is evaluated and discussed. The relationship between the incident imageand the captured data set is investigated and described. This model isintegrated with the spectral description of the feature measures to form amodel of the transition from surface to feature set. It is clear from thismodel that the direction of illumination can affect the directionality of animage obtained from a given surface. Changes in the illuminant direction willresult in changes in the feature outputs. If the illuminant direction isaltered between training and classification, the classification rule may beinappropriate and classification poor. Several schemes are considered to combatthis problem. A technique which uses a representation of the physical surfaceas the basis for the generation of appropriate training data is selected for furtherevaluation. The surface derivative fields of the training surface are estimatedusing photometric techniques. A rendering algorithm uses these estimates tosimulate the appearance of the training surface when it is illuminated from anarbitrary direction. It is shown that where illuminant direction is varied thissystem is able to perform significantly better than a naive classifier, and insome cases approaches the level of accuracy obtained from training theclassifier under the conditions at which classification is performed. Textureanalysis is a significant area in the field of machine vision, this is in largepart due to the important role of texture in the early visual system. Itfollows from this that texture has been seen as being critical to generalvisual systems working in unconstrained environments, consequently, lessemphasis has been placed on more controlled inspection tasks. In anunconstrained system it is impractical to adopt a modelling approach and mostwork in texture analysis takes the image as its starting point. This thesis isconcerned with the inspection of rough textured surfaces. By making explicitthe circumstances under which classification occurs we are able to employmodelling of the system and describe texture classification in the context ofthe physical system which gives rise to a textured image.

1 引言
2 粗糙表面建模
3 图像形成
4 成像过程
5 分类系统
6 分类器倾斜响应建模
7 问题的解决
8 解决倾斜效应的基于仿真的方法
9 总结与结论
附录A 符号术语表示
附录B 纹理特征测试

下载英文原文地址：

http://page5.dfpan.com/fs/el0c9ja27221f2c9163/

更多精彩文章请关注微信号：