Optimal dimensionality reduction of a single matrix is given by the truncated singular value decomposition, and optimal compression of a set of vector data is given by the principal component analysis. We present, in this paper, a dyadic singular value decomposition (DSVD) which gives a near-optimal dimensionality reduction of a set of matrix data and apply it to image compression and face recognition. The DSVD algorithm is derived from the higher-order singular value decomposition (HOSVD) of a third-order tensor and gives an analytical solution of a low-rank approximation problem for data matrices. The DSVD outperforms the other dimensionality reduction methods in the computational speed and accuracy in image compression. Its face recognition rate is higher than the eigenface method.
|Number of pages||4|
|Journal||Proceedings - IEEE International Symposium on Circuits and Systems|
|Publication status||Published - Dec 1 2005|
|Event||IEEE International Symposium on Circuits and Systems 2005, ISCAS 2005 - Kobe, Japan|
Duration: May 23 2005 → May 26 2005
All Science Journal Classification (ASJC) codes
- Electrical and Electronic Engineering