請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/31120
標題: | 基於外觀的高效率物件辨識方法 An Efficient Method for Appearance-based Object Recognition |
作者: | Wei-Hao Chang 張緯浩 |
指導教授: | 洪一平(Yi-Ping Hung) |
關鍵字: | 外觀,物件辨識,模式剃除,特徵抽取,模式匹配, appearance,object recognition,pattern rejection,feature extraction,pattern matching,Walsh-Hadamard, |
出版年 : | 2007 |
學位: | 碩士 |
摘要: | 物件辨識一直是過去電腦視覺領域想要解決的重要目標。為了讓物件辨識技術 可以更快速地應用在我們的生活當中,我們發展出了一套高效率的物件辨識方法 可以讓電腦系統經由快速的訓練過程,使電腦系統可以快速地進行物件辨識工作 並達到很好的辨識率。
我們利用 Image Retrieval 的概念來設計所提出的物件辨識技術,將欲辨識的 物件影像轉換為特徵向量,進而在資料庫中找尋與其最為相似的特徵向量集合, 統計出現次數最多的物件類別即為辨識結果。 為了達成我們提出的辨識方法,我們提出了以下幾種新的技術:一個階層化剔 除樹 (Hierarchical Rejection Tree),這個技術可以快速地尋找與被辨識物件 影像最為相似的特徵向量,在這個技術當中,我們改良Hel-Or 2005 年9 月在PAMI 發佈的研究成果(Projection Scheme) [3]。我們所提出的Hierarchical Rejection Tree 提供了高效率的搜尋機制,除了用在本論文的應用之外,也可以 應用在需要高維度特徵空間的搜尋應用加速上。此外 我們也發展了兩個有效率的 特徵萃取方法(feature extraction method):Hue mean/Intensity DCT-WH 影像 特徵及Hue mean/Hue histogram-WH 影像特徵,利用影像的頻譜特性及色彩分布 當作影像特徵,將其投影至Walsh-Hadamard basis vectors 並加入色彩資訊(色 調均值)以增進影像特徵的代表性。 結合上述的新技術,我們發展出一套兼具速度與準確度的物件辨識方法 並經由 實驗數據証明結合這些技術所發展出來的物件辨識系統具有快速訓練,快速辨識 及高辨識率等特色。 In computer vision, the ways to make computers being capable of seeing and understanding the world have been intensely studied more than three decades. In this thesis, an efficient method for appearance-based object recognition is proposed. The proposed method including both the training phase and recognition phase is very efficient. The concept of image retrieval is applied in our method. It tries to find the most similar feature vector set in image databases and classify the testing image by using the most frequent class label in the feature vector set. For accomplishing the proposed object recognition method, we developed several techniques: an efficient pattern rejection scheme - Hierarchical Rejection Tree is first proposed here. Hierarchical rejection tree can find the most similar feature vector in database efficiently. Based on Hel-Or’s Projection Scheme [3], we enhanced its performance by iterative indexing tree structure. The experimental result shows the performance of Hierarchical Rejection Tree is faster 2.6 times than projection scheme. In this thesis, two new effective feature extraction methods are also be introduced. For an image, the hue mean/intensity DCT-WH image feature captures color information and DCT spectrum as its characteristics. For an image, the other feature extraction, the hue mean/hue histogram-WH image feature captures color information and color distribution as its characteristics. We combine the Hierarchical Rejection Tree technique with feature extraction methods to develop our appearance-based object recognition system. Three image databases which contain large number of object images are used as testing databases. The experimental results show the proposed method has the following features: fast training speed, fast recognition speed and high recognition rate. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/31120 |
全文授權: | 有償授權 |
顯示於系所單位: | 資訊工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-96-1.pdf 目前未授權公開取用 | 1.61 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。