一個基於區域色彩分布的全自動快速切割演算法

Abstract

在本論文中，我們提出一快速自動切割物體的新方法。基於人類 的視覺系統，我們定義出切割中所需之特徵向量。首先我們先根據區 域色彩的分布情形將相似的色彩紋理樣式歸為區域。接著便分析各區 域的特性：色彩組成、大小、相對位置、及粗糙程度。最後，基於若 為相同物體的兩相鄰區域其有著相似特性的事實，我們將特性相似的 相鄰區域合併。大部分的紋理切割研究中透過小波轉換來取得紋理的 特徵來產生切割結果；然而，在自然影像中的不規則紋理通常導致其 有著過度切割的問題。我們所提出的特徵向量可較有效的表示自然影 像中的不規則色彩紋理，而其運算上亦較小波轉換為簡單。我們的方 法能在合理時間內得到符合人類視覺的可靠切割結果。本文所呈現的 實驗結果和現有的切割演算法之比較亦顯現出此法的有效性。

In this thesis, we proposed a new fast method in automatic object segmentation. Based on human visual system, we define some features needed in segmentation. According to local color distribution, similar color-texture patterns are grouped into a region first. Then we analysis properties of each region: color composition, size, relative position, and coarseness. Finally, similar neighboring regions based on the fact that two neighboring regions belonging to the same object will have similar properties are merged. Most texture segmentation algorithms via wavelet transform adopt textural features to achieve segmentation proposes. However, it often has over-segmentation problem for irregular texture in natural images. Our propose feature vector can represent irregular color-texture in natural scenes. It is also computationally more feasible than wavelet transform.aOur method can generate acceptable result based on human vision in a reasonable time. Some experiments and comparisons with some existing algorithms are also given to show the effectiveness of the proposed method.