Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/31949
Title: | 利用整體和局部特徵辨識物件 Object Recognition Using Global and Local Features |
Authors: | Hsiang-Ting Lan 籃湘婷 |
Advisor: | 李瑞庭(Anthony J.T. Lee) |
Keyword: | 物件辨識,特徵樣式,小波轉換,角度直方圖, object recognition,feature pattern,wavelet transform,angle histogram, |
Publication Year : | 2006 |
Degree: | 碩士 |
Abstract: | 物件辨識的目的是識別圖片中的物件並且將物件歸類到具有相同特性的類別中。在本篇論文中,我們利用資料探勘方法找出特徵樣式以便辨識物件,我們提出的方法主要包括三個步驟。在預先處理階段,先正規化圖片以處理仿射轉換(affine transformation)後所帶來的影響。針對取出的輪廓點座標做小波轉換(wavelet transform)取得32個係數來當作整體特徵,接著將每一個輪廓點當作一個視窗的中心,計算角度直方圖(angle histogram)以取得一些向量當作局部特徵。在訓練階段,我們利用這些整體和局部特徵找出各種類別物件的特徵樣式(feature pattern)。在測試階段,利用一些圖片來檢驗各類別特徵樣式的有效性。針對每一張測試圖片,計算符合每一類整體和局部特徵樣式的比例,再利用權重將這兩個比例整合。最後,這張圖片會被歸類到擁有最高比例的類別。實驗結果顯示出,我們的方法在樹葉資料庫可達到98.15%的分類準確率,在ETH資料庫可達到97.62%的分類準確率,均勝過由Sun和Super所提出的方法。 The goal of object recognition is to identify the object in an image. In this thesis, we proposed a data mining approach to realize object recognition. Our proposed method consists of three phases. In the preprocessing phase, we normalize the image to make our method invariant to translation, scale and rotation. We do Haar discrete wavelet transform on the coordinates of the extracted contour and get 32 coefficients to be the global feature. Then we use each contour point to be the centroid of a window, and calculate the angle histogram to get some vectors as local features. In the training phase, we use these global and local features to find the patterns of each class. In the testing phase, some images are used to test the effectiveness of the representative patterns. For each test image, we calculate the ratios of the global and local patterns of the test image conformed to each class, and use a weight to combine both ratios. Finally, the image is classified into the class with the highest ratio. The experimental results show that the classification accuracy rate of our method achieves 98.15% in the leaves database and 97.62% in the ETH object database and outperforms the method proposed by Sun and Super. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/31949 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 資訊管理學系 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-95-1.pdf Restricted Access | 631.7 kB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.