以幾何特徵頻譜於紋理分析之研究

Abstract

對於許多不同形式的影像分析，紋理是一項重要的特性，紋理分析在檢視 物體的表面上扮演很重要的角色，而且已經被廣泛地應用在各種不同領域 的影像分析；包含醫學影像、遙測影像、顯微影像等。紋理分析的方法主 要可分為兩大類:結構方法及統計方法，結構方法著重紋理基本元素及取 代法則的製定；統計的方法主要的問題在於紋理特徵的抽取，大部分抽取 特徵的方法是從紋理影像中量測某種特徵的統計值(如能量的平均值標準 差等)或是直接以特徵的統計分佈來做紋理的分析及分類。本文主要是提 出新的紋理特徵配合特徵的統計分佈來作紋理分析及分類，內容可分三部 分，第一部分是Triangle Mesh Model的方法，我們將紋理影像視為三維 度的三角形所組成，然後利用這些三角形的法向量、面積的分佈來分析紋 理，三角形的法向量分佈可以用來分析紋理的方向；而三角形的面積分佈 可以用來分析紋理的顆粒大小。第二部分是以形態運算來定義紋理的區域 特徵，Morphological Gradient 和 Modified Hit-or-Miss運算被使用來 作紋理的區域特徵，紋理的Morphological Gradient分佈可以用來分析此 紋理的粗糙程度；而Modified Hit-or-Miss運算在作紋理分析的觀念是先 定義一組較小的特殊圖案(結構元素)，利用這些特殊圖案與待分類紋理之 間match的程度高低來作紋理分類。但是在作形態運算時，結構元素的形 狀及內容會影響到運算的結果，所以結構元素的形狀及內容的選定是一個 問題。在第三部分，我們利用遺傳演算法(Genetic Algorithms)來學習並 決定較佳的結構元素以提高紋理分類的辨識率。上述方法在作紋理分類時 皆直接以特徵的統計分佈來做紋理的分析及分類，並以特徵的統計分佈之 距離來量測紋理之間的差距。最後，我們嘗試作碳鋼顯微影像的分析。碳 鋼所建構的材料早以被人類所廣汎使用，從事碳鋼顯微影像之金相研究不 只對鋼材特性有所了解，亦有助於對其它金屬合金固態反應的認知，但目 前利用碳鋼顯微影像的分析來判斷其所含碳含量時，須利用標準顯微影像 圖鑑來比對，既不客觀又費時間。因此有其自動化的需求。我們試著利用 所提出的統計分佈來作不同含碳量碳鋼顯微影像的分類，實驗結果得到相 當高的辨識率，說明了此方法在建立碳鋼顯微影像自動化辨識系統的可行 性。 Texture is an important characteristic for the analysis of many different types of images. Texture analysis plays a critical role in inspecting surfaces and provides important techniques in a variety of applications ranging from medical imaging to remote sensing. Texture analysis can be divided into two components: structural and statistical approaches. Structural approaches try to describe a repetitive texture in terms of the primitive elements and placement rules that describe geometrical relationships between these elements. In statistical approach, the major problem is the extraction of texture features. Most of the approaches to texture analysis quantify the texture measures by single values (means, variances etc.) or directly use the feature distribution. The thesis proposes new texture features and uses the corresponding feature distribution for texture analysis and classification. Three parts are included in the thesis. The first one is the Triangle Mesh Model approach. The main concept of this approach is that texture images can be considered as a set of three dimensional triangles. The normals and area distribution of these triangles are used as texture features. In the second part, we use morphological operations to extract geometrical texture features. Morphological gradient and modified hit-or- miss operations are applied in our experiments. In order to demonstrate the discrimination performance of the proposed texture features, natural textures from Brodatz album are used for classification. Experimental results show that our methods have good performances in texture classification. Since the content of the structuring elements will effect the features computed by morphological operations, in part three, genetic algorithms are used to dynamically select near optimal structuring elements to achieve high classification rates. In addition, we also focus on the analysis for carbon steel micrographs. The current practice of carbon steel sample analysis with