Please use this identifier to cite or link to this item:
標題: 基於影像的細胞計量與基於特徵的影像切割
Image-Based Cell Counting and Feature-Based Image Segmentation
作者: 徐安眉
Hsu, An-Mei
關鍵字: Image segmentation
cell counting
fluorescence images
phase contrast images
texture segmentation
textural image
feature extraction
出版社: 資訊管理學系所
摘要: 在目前的影像切割技術應用中,可以分為兩類,一種是應用到平滑物件的影像;另一種則是屬於紋理切割的技術,辨識具有相同紋理的區域。基於上述第一類的應用,在本論文中,我們首先提出了針對顯微鏡影像為基礎的細胞計量。針對細胞計量,我們應用到兩種不同類型的影像:螢光顯微影像(fluorescence image)與相位差影像(phase contrast image),分別提出演算法(FIS與PCIS)。主要採用bi-group技術加強細胞邊緣,GDW強化邊緣並抑制雜訊,利用適應性門檻值方法(Adaptive threshold method)針對每個像素分別取適合的門檻值,並利用基因演算法決定實驗中最適合的參數。 由於上述方法基於物件是平滑的假設,並不適用於紋理影像的切割。因此在實際應用上限制其發展空間,所以本篇論文接著提出一套是針對紋路影像做基於特徵的影像切割技術(FBS)。我們擷取GLCM、Tamura與梯度特徵,透過加強物件邊緣與二值化等切割處理,再經由合併階段,將影像中紋理特徵相近卻過度分割的區塊,透過判斷的機制,達到合併的結果,以達到切割的結果。 總結我們所提出的主要兩種演算法,主要是利用針對具有不同特性的影像,根據其不同特性,分別利用不同特徵作為影像切割的依據。實驗結果證明,我們提出的方法(FIS、PCIS與FBS)在準確度方面優於其他方法。
Today's image segmentation techniques can be divided into two types. The first type is applied to the images which have uniformity of intensities in local regions. The second type is about texture segmentation for segment the objects out from an image so that each of the objects has similar texture patterns. For the application of first type, we proposed an unsupervised scheme (FIS and PCIS) of image-based cell counting for two types of images: fluorescence images and phase contrast images. In cell counter, a bi-group filter is proposed to enhance the boundaries of cells, a gradient-direction-weight (GDW) enhancer to suppress the boundaries of noises while strengthening the boundaries of cells, and an adaptive threshold method to give the proper threshold for each pixel. The thesis also adopts a generic algorithm to determine the most suitable parameters in the experiment. Based on the assumption that objects have uniformity of intensities, the techniques for cell counter are not suitable for texture segmentation. Therefore, we later proposed an algorithm of feature-based image segmentation (FBS). First, we extract three features: GLCM feature, Tamura feature, and gradient feature). Second, segment the regions according to the feature we extracted. Finally, merge the over-splitting regions with similar texture. To summarize our method, we utilize different features to segment objects in an image according to its properties. The obtained result is better than other methods in accuracy.
Appears in Collections:資訊管理學系



Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.