請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/28520
標題: | 在個人相簿中以非臉部資訊輔助之人臉辨識 Context-Aided Face Identification in Personal Photo Albums |
作者: | Liang-Yu Chou 周亮瑜 |
指導教授: | 莊永裕(Yung-Yu Chuang) |
關鍵字: | 人臉分類,人臉辨識,人臉標註,以前後文輔助的,單一訓練樣本的, face classification,face identification,face annotation,context-aided,one training sample, |
出版年 : | 2007 |
學位: | 碩士 |
摘要: | 蒐集訓練樣本的繁複性,是目前人臉辨識技術所面對的最大困難之一。為了能涵蓋每張臉在不同光源、姿勢以及臉部表情下所經歷的變異量,我們需要有大量的訓練樣本,來涵蓋這些變異量。在本文中,我們提出了一個人臉分類架構,企圖減輕使用者花在蒐集訓練樣本上的煩瑣時間(亦即,加速註記人臉的過程)。使用者將被要求先註記少量被偵測到的人臉,使其能代表每一個將被用來分類的目標類別。接著基於一種所謂通用學習的技術,我們去計算所有偵測出的人臉之間,兩兩的相似度分數。在本文中,我們也提出了一個新穎的自動對齊人臉的演算法,用來進一步的減少使用者花費在標記每張臉的眼睛位置上的時間。為了在這單一訓練樣本的情況下提高人臉分類的識別性能,我們亦由輸入的照片集中,萃取出前後文訊息,來做為分類判斷的另一種線索。另外,我們也設計了一種新的使用者的反饋介面,能夠以一種新穎卻不失直觀的方式,讓使用者加入新的反饋訊息,並依此反覆的改善分類結果。在文章的後半段,多項針對消費者相冊的實驗結果,亦說明了我們所提出的,是一種有效、且有效率的人臉分類架構。 One of the main challenges faced by the current face recognition techniques lies in the difficulties of collecting samples. A large number of representative training samples are required to capture the considerable amounts of variations one face may undergo due to illumination, pose, expression, etc. In this paper, a classification framework is introduced as an attempt to alleviate the tedious effort spent on collecting training samples (i.e. speed up the annotation process). The user is required to annotate a small number of faces detected from the input photo set as different subjects (i.e. different clusters). Then, based on the technique of so-called generic learning, we classify the remaining faces into one of these clusters by comparing their pair-wise face similarities. A novel automatic face alignment algorithm is also devised for saving the user efforts on labeling eye coordinates. To boost the recognition performance in this one training sample application scenario, we extract context information as another cue for recognizing people. In addition, relevance feedback can also be added in a novel and intuitive way to iteratively refine the classifacation result. At the end of this article, experiments on real consumer photo albums are performed to demonstrate the effectiveness of the proposed framework. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/28520 |
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 | 1.56 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。