聯合局部保留字典學習法研究;A Study of Locality Preserved Joint Dictionary Learning

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/81050

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/81050

题名:	聯合局部保留字典學習法研究;A Study of Locality Preserved Joint Dictionary Learning
作者:	沈正勝;Mathulaprangsan, Seksan
贡献者:	資訊工程學系
关键词:	字典學習;聯合詞典學習;局部特徵保留;非負矩陣分解;dictionary learning;joint dictionary learning;locality preserving;nonnegative matrix factorization
日期:	2019-05-01
上传时间:	2019-09-03 15:31:02 (UTC+8)
出版者:	國立中央大學
摘要:	本論文結合局部特徵保留(Locality preserving)技術與字典學習方法，並且藉此提升其應用在語音情緒辨識以及物件辨識上之效果。首先，針對影像物件辨識應用，我們提出二個新穎之局部保留字典學習方法，其一為具鑑別性(Discriminative)之局部保留KSVD(LP-KSVD)，將標籤資訊引入局部保留項。其二為標籤一致性(Label-consistent)之LP-KSVD(LCLP-KSVD)，利用標籤一致做為限制項來進一步強化不同類別間之鑑別性。接著，本論文針對語音情緒辨識應用，提出具局部保留之聯合非負矩陣分解(Joint nonnegative matrix factorization)方法(LP-JNMF)，透過同時重建語音特徵與訓練一簡單線性分類器，來學習具備高度鑑別力之共通特徵。此外，我們也引入局部保留限制項來使得學習出的特徵保留高維度特徵之流型(Manifold)。實驗結果顯示，所提出的方法在物件辨識與語音情緒辨識的應用上，優於多項先進字典學習方法。 ;This study focuses on using the locality preserving technique, which uses geometric information of data, to boost up the performance of dictionary learning approaches to a number of pattern recognition tasks including speech emotion recognition and object recognition. Firstly, to exploit fully the potential of the locality-preserving technique for the object recognition task, two novel dual-layer locality-preserving methods were developed. The former is the discriminative LP-KSVD (DLP-KSVD), which incorporates the label information into locality-preserving term. The latter is the label-consistent LP-KSVD (LCLP-KSVD), which applied the label-consistent constraint to the original LP-KSVD model to penalize the sparse codes from different classes to improve the discriminative power. Secondly, a novel approach for speech emotion recognition, named locality preserved joint NMF (LP-JNMF), is introduced. This study achieves two goals jointly; the first is to learn a dictionary for the reconstruction of input acoustic features and the second is to learn a simple linear classifier for annotation. Since the learned representations are shared between the learned dictionaries and annotation matrix, the discriminative power is promoted. Moreover, to preserve the manifold of input acoustic features, a locality penalty term is incorporated into the objective function of joint dictionary learning. Thus, the discriminability of the learned dictionary is further improved. Experimental results prove that the proposed methods outperform the baseline algorithms, which are state-of-the-art dictionary learning algorithms for object recognition and speech emotion recognition problems.
显示于类别:	[資訊工程研究所] 博碩士論文

文件中的档案:

档案	描述	大小	格式	浏览次数
index.html		0Kb	HTML	123	检视/开启

在NCUIR中所有的数据项都受到原著作权保护.

社群 sharing

数据加载中.....