協同表示之聯合核化字典學習及其於聲音事件辨識;Joint Kernel Dictionary Learning via Collaborative Representation and Its Application to Sound Event Classification

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/65801

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/65801

Title:	協同表示之聯合核化字典學習及其於聲音事件辨識;Joint Kernel Dictionary Learning via Collaborative Representation and Its Application to Sound Event Classification
Authors:	廖唯鈞;Liao,Wei-Chung
Contributors:	資訊工程學系
Keywords:	字典學習;稀疏表示;dictionary learning;sparse representation
Date:	2014-08-26
Issue Date:	2014-10-15 17:10:39 (UTC+8)
Publisher:	國立中央大學
Abstract:	聲音事件辨識的應用在當今人類社會已逐漸成為一個重要的課題，舉凡安全監測系統、環境聲音辨識、家庭看護系統等等皆與人們的日常生活息息相關，為了達到精準辨識的目的，從傳統的辨識器如支持向量機(SVM)、高斯混合模型(GMM)、到近年來火紅的稀疏表示分類器(SRC)，都能獲得不錯的結果。基於稀疏表示分類器，本論文針對其所使用的稀疏字典提出改良的訓練方法。在訓練的目標函式中加入辨識誤差項並在稀疏限制項改採ℓ2-norm 而非ℓ1-norm，訓練字典的過程同時訓練一個簡單的線性分類器達到增強辨識能力且節省時間的效果。除此之外，我們使用核化方法將訓練資料投射至高維特徵空間以增強字典的辨識及重建能力及提升系統彈性。線上學習 (online learning)的應用讓演算法在訓練資料依時間而變動時能有較高的效率。基於一個17 類別的聲音資料庫，實驗結果上以80.56%的辨識率優於其他演算法，證明本論文所提出之改進方法確實對字典的辨識能力有所提升。另外在執行測試的時間上也遠比SRC 和CRC 來的有效率。;Environment sound classification is become more and more popular in humans daily life, such as security surveillance, environment detection, human health care. To accurately classify the sound from different event, we can use the traditional SVM, GMM and the popular SRC to obtain well classification result. In this paper, we present a joint kernel dictionary learning (JKDL) method base on sparse representation. Using ℓ2-norm instead of ℓ1-norm can reserve the performance but reduce the computation time massively. Adding the classification error term into the objective function to train a simple linear classifier enhanced the relationship between the classifier and dictionary. Kernel method plays an important role which efficiently strengthen the reconstructive and discriminative ability. The dictionary update step is iteratively performed by taking partial derivatives on objective function in feature space. Online dictionary learning approach handle dynamic training data more efficient than batch approach. Experiments on a 17 classes sound database indicates that the proposed method can achieve an high accuracy rate about 80.56%. Also, the average executing time of a testing data is notably faster than SRC and CRC.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	474	View/Open

社群 sharing

Loading...