應用神經網路分類器與多邊形近似技巧於靜態手勢辨識之研究

Abstract

我們在此論文中，提出一種新的架構和特徵抽取的方法，能夠正確而 有效地辨識靜態手勢。藉由我們所提的方法，即可有效地辨識我們所定義 的$32$張具有部份相似的靜態手勢，我們所使用的方法主要是利用拐角( corner)去抓取手勢的特徵，我們使用了二元化(binary)、邊界偵測(edge detection)、細線化(thinning)、與拐角偵測(cornertracking)等技巧去 獲得邊界外形，然後選擇最佳的12個拐角近似一個與原來外形相似的多邊 形。 在我們的系統裡，必須先找出影像是否屬於特殊手勢和以伸出手指數目來 作為特徵向量的第一個分量，若屬於特殊手勢則抽取原來$12$個最佳拐角 的長度與角度作為其特徵，否則，便抽取這些伸出手指拐角的長度與角度 作為特徵向量。如此，這些特徵即可不受位置、大小、與平面的旋轉而改 變。最後，我們選擇指導式乏析適應性Hamming網路(supervised fuzzy adaptive Hamming net)作為分類器，實驗的結果顯示我們的辨識率可達 到 $95\%$。 In this thesis, we propose a new, static-hand-gesture recognition system. We define the part similarities among 32 static hand gestures as standardgestures. Applying our method, we can effectively recognize those gestures.A method for recognizing gestures by extracting gesture boundary- shapefeatures using corners is also described. In order to obtain boundary shapes,our methoduses several algorithms that perform binary, edge detection,thinning, andcorner tracking. We then select a total of 12 optimal cornersthat represent apolygon approximating the target shape. Our system first determines the first component of our feature vector. Todo this, we perform finding the particular gesture and determining the numberof fingers extended. If it is a particular gesture, we then extract the angles and lengths of the original 12 optimal corners. Otherwise, the angles and lengths of the extended fingers are extracted as our features. The extracted features are invariant to translation, rotation and scaling. We choose supervised fuzzy-adaptive Hamming net as a gesture classifier. Experimentalresults show our recognition rate is $95\%$.