紋理影像切割:利用一個新的演算架構整合多重紋理特徵的分析結果

Abstract

紋理分割在影像分析及電腦視覺上扮演著相當重要的角色。紋理分割的主要目的在於根據紋理特徵將影像分割成許多有意義的區域(在同一區域內的像點具有類似的紋理特徵)。在過去十年間，有很多不同的擷取紋理特徵的方法不斷的被提出。紋理分割的效能除了和擷取紋理特徵的方法有關外，也和處理分割的演算法有關。 本論文提出了一個三階段的演算架構來處理紋理分割，包含了一些新的想法陳述於下: 一、此架構適用於不同的紋理特徵擷取方法。 二、此架構以三階段的程序來處理像點的分類， (1)第一階段，在獨立單一的紋理特徵空間裡進行分群的工作。 (2)第二階段，利用第一階段所得到的分析結果，以一種合併的程序得到一個整體的分群結果，此結果遠優於單一紋理特徵的分群結果。 (3)最後的階段，根據第二階段的合併結果，我們以一種類似nearest-neighborhood 的方法在多維向量空間進行像點重新分類的工作，但此方法在計算時間上的代價遠低於傳統的方法。 此演算架構的優點在於解決了同時找尋群集的中心並進行分群的困難，這是傳統方法所無法達到的，此三階段的演算架構中，第二階段用來決定群集的中心，第三階段則利用此群集中心進行分群的工作。更重要的是，它是一個one-pass的分割程序，而非傳統演算法的recursive程序。 本論文提出的架構曾經由不同種類的紋理影像加以測試，包含了均勻的紋理樣本及一般的自然景物影像，最後的分割效果是令人滿意的。

Texture segmentation plays an important role in image analysis and computer vision. The major goal of texture segmentation is to partition an image into meaningful regions of homogeneous texture properties, i.e. pixels in the same region share similar texture attributes. Many different approaches have been proposed to extract texture features in decades .The performance of texture segmentation depends on not only texture feature extraction methods which discriminate texture variations but also segmentation algorithms which demarcate regions of homogeneous distribution of texture property. A three-phase algorithmic framework for texture segmentation is proposed in the thesis, with some innovation to be mentioned below. (1)It allows the embedding of texture extractors at users’ disposal. Any set of texture extractors demanded by concerns of various aspects can be fit into the first phase of the framework. (2)The classification is handled by a 3-staged process, (a) clustering in each individual feature space associated with each feature extractor in phase I of the framework, (b) followed by a merging process in phase II for exploiting any complementary behaviors exhibited among texture extractors in order to obtain an overall result of clustering much better than any of the individual clustering by single texture alone, and (c) in the final phase, the merged (and refined) clustering result from phase II is reclassified in a way similar to conventional nearest-neighborhood problem in multidimensional vector space but at much lower computational cost for the final refinement of the segmentation. The strength of the proposed framework for texture segmentation lies in that the difficulty of resolving both the determination cluster centroids (as representative for each class) and the clustering operation simultaneously, as encountered in most conventional clustering problem in multidimensional vector space, has been circumvented by the three-phased design in which potential cluster centroids are determined in phase two, followed by the final staged clustering in phase three. Further more, it is a straight forward one-pass segmentation process, as opposed to conventional ones which are mostly iterative in nature. The proposed framework has been tested against textured images of various kinds, with regular texture patterns and natural scenes, and demonstrates satisfactory performance.