%0 Journal Article
%T 无声语音接口中超声图像的混合特征提取<br>Hybrid feature extraction from ultrasound images for a silent speech interface
%A 路文焕
%A 曲悦欣
%A 杨亚龙
%A 王建荣
%A 党建武
%J 清华大学学报（自然科学版）
%D 2017
%R 10.16511/j.cnki.qhdxxb.2017.26.060
%X 在基于超声的无声语音接口实现中，通常使用主成分分析或离散余弦变换提取舌部超声图像的特征。为了保留图像的关键信息，该文提出3种混合特征提取方法：使用主成分分析从小波系数中提取特征（Wavelet PCA）、分块离散余弦变换主成分分析（block DCT-PCA）和分块Walsh Hadamard变换主成分分析（block WHT-PCA）。根据能量选取适量的离散余弦变换或WHT变换系数，使用主成分分析提取选定系数的特征。实验结果表明：该文提出的混合特征提取方法优于主成分分析或离散余弦变换，其中block DCT-PCA方法最优。<br>Abstract：Principal component analysis (PCA) and discrete cosine transform (DCT) are used to extract features from ultrasound images to build an ultrasound based silent speech interface. The critical information in the image is presented by using three hybrid feature extraction methods. The first method uses PCA to extract discrete wavelet transform coefficient features. The second and third methods truncate the DCT or Walsh-Hadamard transform coefficients to the appropriate number according to the energy with the truncated coefficients then used by PCA to extract the features. Tests show that this hybrid feature extraction method outperforms standalone PCA or DCT analyses. The block DCT-PCA method gives the best result among all the methods.
%K 无声语音接口
%K 超声
%K 舌部
%K 主成分分析
%K 离散余弦变换
%K Walsh-Hadamard变换
%K &lt
%K br&gt
%K silent speech interface
%K ultrasound
%K tongue
%K principal component analysis
%K discrete cosine transform
%K Walsh-Hadamard transform
%U http://jst.tsinghuajournals.com/CN/Y2017/V57/I11/1159