影像分割爲進行影像分析的重要前處理步驟,此步驟的成效對於後續影像分析的結果通常有顯著之影響。貝氏分類法與類神經網路則爲監督式學習自動影像分割方法中較爲被廣泛應用者。本研究之主要目的爲以類神經網路與貝氏分類法建立適用於不同背景與照明條件下的影像分割方法,同時進行兩種方法的比較分析及闡明其適用性。我們首先以物件導向程式建立了應用貝氏分類法與類神經網路的影像分割軟體元件,再以二維資料進行分析,利用空間座標的表示法分別比較兩種方法在可線性分離、非凸集、以及相互包含等三種資料類型的分類效果。在彩色影像的影像分割應用方面,則以蔬菜種苗影像爲對象,依據影像元素色彩資訊的空間分佈情形,探討兩種方法的分類效果。實驗結果顯示,對於照明良好與背景單純的彩色影像,貝氏分類法與類神經網路的影像分割效果接近,平均分類誤差均小於1%。但對於複雜背景之影像則以類神經網路的影像分割效果較佳,同時類神經網路對於學習過程中取樣均勻程度的要求較具有強健性。
Image segmentation is an important preprocesing procedure for image analysis. The result of image segmentation significantly affects the accuracy of subsequent image analysis. The Bayesian and neural network classification methods are widely used for automatic and supervised image segmentation. The objectives of this research were to establish image segmentation methods for various background and lighting conditions based on Bayesian and neural network methods, and to compare the efficacy and feasibility of applying these two methods under different conditions. Software components employing object oriented programming language for color image segmentation was initially built and tested with 2-dimensional data sets. These data sets were represented with 2-dimensional plots corresponding to linearly separable, non-convex, and mutually included classification patterns. The data sets were classified with both the Bayesian and neural network methods, and the results were compared and discussed. Finally, color images of selected vegetable seedling sunder different background and lighting conditions were tested. The results was compared along with the spatial distribution of pixels in RGB color coordinates. The experimental results indicated that under good lighting and uncomplicated background conditions, the use of Bayesian and neural network methods were not significantly different. The average errors of image segmentation were all below 1%.However, for images with complex background, image segmentation using neural network showed better result than the Bayesian method. In terms of the sampling requirement during the learning phase, the neural network method showed better robustness.