應用神經網路分類器與金字塔狀特徵抽取技巧於三維物體辨識

Abstract

我們提出一新的方法，能夠正確辨識一物體，而不會受到該物體在三 維空間中，位置的改變，大小的改變，以及方位的改變等變化之影響。首 先我們利用一個模糊估測(fuzzymeasure) 技巧去決定一張影像的最佳臨 界值。利用此臨界值將待辨識的物體由背景分離 出來。然後我們建立一 個影像金字塔狀資料結構 (image pyramid data structure)，藉 由環形 和扇形視窗 (annular and sector windows) 的分割技巧，去抽取具有不 變性的 特徵 (invariant feature)。獲得待辨識物體的特徵之後，我們 採用一種神經網路稱之 為指導式模糊適應性 Hamming網路(supervised fuzzy adaptive Hamming net)作為分類 器，劃分特徵空間以決定每一物 體的種類。經由模擬的結果，所提出的方法可獲得不錯效果，這顯示我們 所提出的方法是適合於三維空間上的物體辨識。 We propose an approach to 3-D object recognition irrespective of its position, size, and orientation. We use a fuzzy measure technique to find an optimal threshold value and obtain the shape of the object from the background in an input image. We then build an image pyramid data structure to extract the invariant features. This is supported by a segmentation technique using annular and sector windows. After obtaining the features of the object, we adopt a neural network model, the supervised fuzzy adaptive Hamming net, as a classifier whose purpose is to partition the feature space into decision regions corresponding to each object class. The simulationresults show that the proposed method can obtain a satisfactory performance. So, the proposed method provides a suitable approach to 3-D object recognition.