Home > Publication > International Conference Papers  
Title The effect of mean value extraction on HRTF analysis
Author Daehyuk SON, Youngjin PARK
Conference MoViC2018
Year of Pub. 2018
File MoViC2018_DSon_Full_paper.pdf
Many researchers applied the Principal Component Analysis (PCA) to analyze HRTFs because this method can observe possibly correlated variables by orthogonal transformation procedure. In case of PCA on HRTF, it is commonly accepted to extract empirical mean vector of a dataset as a pre-processing to eliminate the first basis function which represents the mean value. On the other hand, in case of multi-dimensional SVD of HRTF, if any tensor has an order larger than 2, the empirical mean can be defined in various ways, that is; a first-order, second-order, or larger-order. Therefore, it is necessary to carefully define the mean for multi-dimensional analysis. In this work, the HRTF tensor is set as the log-magnitude of HRTF for frequency, azimuth, elevation, and subject dimensions. The Frobenius norm and spectral distortion are analyzed to compare energy reduction and basis functions. Extraction of the mean value from the azimuth dimension leads to the largest energy reduction in the view of the Frobenius norm. Furthermore, extraction of the mean value from the subject axis is the most efficient way in terms of the spectral distortion score of the reconstructed HRTF.