三維方形工件分類方法之研究

Abstract

本論文提出一種對三維方形工件作自動分類的方法，分類的主要據為工件 的整體外形資訊。在此分類方法中，首先將一個三維方形工件以其三個二 維投影視圖表示，接著將此二維視圖之整體外形輪廓以近似其外形之正交 多邊形代表，然後將此正交多邊形以一樹狀結構之簡化線骨架表示，再將 簡化線骨架轉換成以向量的方式表達，此向量為類神經網路的輸入向量， 最後再藉由此網路對簡化線骨架分類，並且根據上述分的類的結果，將工 件形成工件族。此工件分類方法依研究過程的演進，可以分成三個階段： 在第一階段中，將一樹狀結構之簡化線骨架以多組的向量表示，並且以逆 傳遞類神經網路為分類的工具；可是若以多組向量表達一簡化線骨架，會 以許多不必要的資訊重複表達它。因此在第二階段中，另外提出一新的骨 架表達法，將簡化線骨架薇以唯一的一組向量表示，再利用此向量進行分 類；但是在前兩個階段中所使用的逆傳遞類神經網路為一監督式的神經網 路，所以在使用之前，必須先經由人為選擇幾個工件做為訓練時的樣本， 因為此人為的選擇因素，可能會促使分類的結果不一致。所以在第三階段 中，再提出一結合逆傳遞類神經網路及適應共振理論神經網路之分類系統 架構，應用此分類系統架構，可輔助選取訓練逆傳遞類神經網路之樣本， 使人為的參與程度降低，而且分類的效果頗佳。 This research presents a neural network approach to the classification of 3D prismatic parts based on their global shape information. In this approach, a 3D prismatic part is modeled by the contours of its three rectilinear polygons. The global shape information of each polygon is modeled by its simplified skeleton, which origionally is of a tree structure and can be represented by vectors by a conversion method. The vectors are the input to a polygon classifier which is constructed on the basis of the neural network. The classification results of polygons can be used to group the 3D prismatic parts into families in a hierarchical manners, by setting different levels of similarity criteria. According to the research progress, the proposed classification method evolved in three stages. In the first stage, the simplified skeleton is moodeled by multiple vectors. And the polygon classifier is constructed on a back-propagation neural network model. Such a skeleton representation method is deficient in using too much memory. In the second stage, a single vector representation method for modeling the simplified skeleton is proposed, which shows satisficatory result in reducing memory requirement. In the third stage, the polygon classifier is enhanced by integrating the ART and BP neural network models in a cascaded architecture. Such an architecture is distinct in relieving the load of manually selecting templates for training a BP network. That is, the ART network would first perform a rough clustering procedure to faciliate the manual selection process.