在電腦視覺以及圖形識別等影像處理技術中，前處理技術扮演很重要的角色，其中包括脈衝雜訊濾除（impulse noise removal）、邊緣偵測（edge detection）以及影像分割（image segmentation）等技術。這些技術的運算結果對於後續影像處理程序的效能有非常大的影響。然而，影像資訊包含千百種不同的內容，想要設計一個理想的前處理技術來適用於所有的影像內容非常困難。值得一提的是，模糊邏輯（fuzzy logic）對於具不確定性的資訊具有優異的處理能力且廣泛地被應用於工程領域。因此，在本論文中，我們嘗試設計數個基於模糊邏輯之新影像前處理技術。針對脈衝雜訊，我們提出一個新的模糊濾波器，並應用粒子群最佳化（particle swarm optimization, PSO）來訓練調整濾波器參數，使其具有最佳的雜訊濾除效果。針對邊緣偵測，我們應用模糊邏輯設計一個最大化目標函數（maximizing objective function）計算每一個邊緣方向的邊緣強度，在影像中偵測出更正確且更細緻的邊緣資訊。最後，針對影像分割，我們簡化所提出的邊緣偵測法並延伸應用於改良種子區塊成長（seeded region growing, SRG）影像分割法，使得影像可以被更準確且更適宜地被分割。 Image pre-processing techniques play a key role in image processing applications such as computer vision and pattern recognition. The common image pre-processing techniques contain impulse noise removal, edge detection, image segmentation, and etc. Good performances of those techniques lead to the better results of the consequent image processing techniques. However, the variety of the image contents is too numerous to count. It is difficult to design an exact image pre-processing technique that can cover all possible image contents and can obtain good results in all images. Fuzzy logic is very popular in engineering applications with its capability to deal with the typical uncertainty that characterizes any physical system. Therefore, this dissertation tries to design some fuzzy logic based image pre-processing techniques such that the performances of the conventional pre-processing are improved. In impulse noise removal, the fuzzy logic based impulse noise filter is proposed. The particle swarm optimization learning process is used to optimize the parameters of the filter such that the best noise removal capability is obtained. In edge detection, the fuzzy logic is used to define a maximizing objective function, in which the edge intensities of each edge direction are determined. Hence, the more correct edge without the double edge and thick edge are detected. In image segmentation, fuzzy logic and the simplified proposed edge detection method are used to improve the conventional seeded region growing for image segmentation. Consequently, the more correct and the more appropriate segmentation results than other existing methods are obtained.
