Abstract
在目前現存的一些影像處理技術模組，以及其應用範疇內，『影像邊緣的搜尋』佔有十分重要的角色，主要是因邊緣為構成影像物體主要架構的基本因子。就一些現有的邊緣偵測法中，有的十分簡單但精度不夠，且對雜訊處理的能力有限；而有些雖能有效地處理雜訊，但卻易使邊緣的真正位置受到影響。有鑑於此，在本論文中將提出一個新的邊緣偵測方法（CED，Clustering Edge Detection），以期能改善上述的問題。CED 主要是先藉著一個區域分割理念來對局部影像做分割，然後再將分割完畢的結果利用類神經網路來判定該局部影像的中心點是否有可能為一邊點。至於所引用的類神經網路模式為反向傳遞網路（ Backpropagation Network ）。為了証實CED 可以找出品質不錯的邊緣，我們將CED與一些梯度運算子（ Gradient Operator ，如 Sobel Operators、Prewitt Operators、Robinson Operators以及Kirsch Operators等）做比較。根據實驗的結果可明顯發現在雜訊干擾的情況下，CED 的執行績效要比梯度運算子要得佳，另外，在邊緣愈傾斜和邊緣是曲線型態之影像，CED 的表現較之於梯梯度運算子，也是會愈來愈好。CED 總處理時間雖會略長，但就整體而言，偵測的精密度得以提高，時間雖略長仍不失其實用性。第一章為緒論，第二章為文獻探討，第三章則是有關 CED 的構建以及類神經網路的應用，第四章主要內容為績效評估，第五章是為結論。It is widely accepted that one of the prior processingstages in imachine vision is to extract the primitiveconstructs of the image, because one important primitiveconstruct in an image is an edge. In this paper, A new edgedetection (Clustering Edge Detection, CED) is proposed, whichuse Otsu's threshoding method (a technique about the regionalanalysis or binary) to cluster small local image into twogroups and Backpropagation Network to classify whether theclustered local image is one of those predefined edgepatterns. Here Backpropagation Network is one of ArtificalNeural Network, and the training set for BackpropagationNetwork is composed of the all possible clustered pattern. Ifthe input pattern matches one of the predefined edge patterns,the corresponding pixel is detedted as an edge pixel.Experimental results are shown where the proposed edgedetector is compared with gradient edge detectors in thevertical image and the oriented image; moreover, the WhiteGaussian noise is added to image to evaluate theperformance of each detector. Chapter 1 is introduction. Thecontext of Chapter 2 is paper review. Chapter 3 is relatedto the construct of CED and the application ofBackpropation Network. And Chapter 4 is to do some experimentsto evaluate the performance of CED and gradient detectors.The last chapter is conclusion.