臺灣大學: 資訊網路與多媒體研究所李明穗楊逸民Yang, Ying-MingYing-MingYang2013-03-222018-07-052013-03-222018-07-052012http://ntur.lib.ntu.edu.tw//handle/246246/251115影像分割（Image Segmentation）是電腦視覺領域中的重要議題。此技術常用於前景物件選取（Foreground Extraction）以供後續上的處理，例如物件追蹤、分析辨識或是影像編輯。如何自動取得完美分割是相當困難的議題。因此，「互動式分割」的概念讓使用者提供先驗知識，並藉以輔助運算出更佳的結果。 本篇論文提出一自動化前景選取技術，架構在「蛇」，亦稱作「動態輪廓線」及「GrabCut」方法上來提供更健全的選取結果，同時去除所有使用者介入。針對輸入影像先產生其高斯金字塔，於各個層級上以動態輪廓線進行前、背景分類作為GrabCut的初始分割，在其迭代式能量最小化方法中加入兩項新機制讓過程更穩定且具備彈性。經由模擬「潛藏狄利克里分配」模型以及名為「中國餐廳程序」的非監督式機率隨機過程來優化「高斯混合模型」，並能動態決定所使用的高斯模型數目。 實驗結果顯示本論文提出的方法在主體明顯的影像中，可以有效地分割出前景物件而不需要任何使用者介入。Image segmentation is an essential topic in computer vision. It is generally applied to extract foreground objects out of an image for further usages, such as tracking, recognition or image editing. Retrieving perfect segmentation results automatically is a difficult issue. Therefore, the concept of “interactive segmentation” allows users to provide prior knowledge in order to obtain better outcome. This thesis presents an automatic extraction technique based on Snake (Active Contour) and GrabCut method. It aims to generate more robust result without user interaction. A Gaussian pyramid of the input image is created firstly, Snake algorithm then derives rough foreground and background separation on each level as the initial for GrabCut. Two novel schemes are exploited to provide more robustness and flexibility during GrabCut’s iterative energy minimization process. These mechanisms simulate the Latent Dirichlet Allocation (LDA) model and unsupervised probabilistic stochastic process called Chinese Restaurant Process. In other words, they assist to refine Gaussian Mixture Models (GMM) and adaptively determine the number of Gaussian components in used. The experimental result shows that proposed method can automatically generate satisfactory extraction output from images which contain obvious foreground objects.3994542 bytesapplication/pdfen-US影像分割前景選取動態輪廓線GrabCut高斯混合模型潛藏狄利克里分配Image SegmentationForeground ExtractionActive ContourGaussian Mixture ModelLatent Dirichlet Allocationthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/251115/1/ntu-101-R99944035-1.pdf