運用空間和頻譜分析與遞迴切片逆迴歸進行紋理切割

Abstract

在影像分析中，影像切割是一個基礎而重要的步驟。對電腦而言，要 自動做紋理切割及分類是很困難的，而人類的視覺系統卻有辦法完成這些 工作。因此我們運用進階的統計工具，模擬人類的視覺模型，進行紋理影 像的特徵截取並進一步切割和分類。在特徵向量的截取方面，我們同時考 慮空間和頻譜的分析。如果有事前的分類資訊時，我們運用切片逆迴歸分 析作進一步分類，否則我們運用主成份分析和獨立成立分析先做初始的分 割，再遞迴地運用切片分析作修正，以完成紋理切割及分類。本論文針對 不同的紋理影像作實證研究，發現這種新的方法實際可行。 Texture segmentation is an important step in image analysis. While human perception has great capacity in texture segmentationand recognition, it is difficult to segment texture images automatically by computers.Hence, we are motivated to apply advanced statistical tools together withthe vision model to mimic human perception.It is aimed to extract the features of textures for further segmentationand recognition based on these techniques.Different transforms ranging from the space domain, the Fourier transform in the frequency domain, and the Gabor filter banks in the space-frequencyanalysis are considered to generate the feature vectors of textures.Dimension reduction techniques are used to find out the projecteddirections of feature vectors.Based on these projected feature vectors, classification rules are selected by the sliced inverse regression when the training set isavailable.For unsupervised segmentation, initial clustering is suggested.Then, the technique of sliced inverse regression is applied iteratively to recluster the texture image by adjusting the projected feature vectorsdynamically.The simulation and empirical studies demonstrated the feasibility ofthese new approaches.