TY - JOUR
T1 - Unsupervised Spectral Regression Learning for Pyramid HOG
AU - Qiang Li, Zhongli Peng & Xiaomei Lin
JO - Journal of Fiber Bioengineering and Informatics
VL - 1
SP - 117
EP - 124
PY - 2015
DA - 2015/08
SN - 8
DO - http://doi.org/10.3993/jfbi03201511
UR - https://global-sci.org/intro/article_detail/jfbi/4691.html
KW - Dimensionality Reduction
KW - Eigen-decomposition of Dense Matrix
KW - Three-tier Pyramid HOG
AB - Applying the original raw data to machine learning will bring in a poor performance, because so many
features are not necessary and redundant. Extracting a small number of good features will be an
important issue, and it can be solved by using dimensionality reduction techniques. However, the popular
dimensionality reduction method will suffer from the eigen-decomposition of dense matrix problem which
is expensive in memory and time. We adopt unsupervised (unlabeled) spectral regression method for
dimensionality reduction, which well avoids the problem of dense matrix eigen-decomposition problem
and can be applied on large scale data sets. Histograms of Oriented Gradients (HOG) are robust features
which not only well characterize the local shape and appearance but also show a certain degree of local
optical and geometry invariance. In order to characterize the local shape and appearance better, we
extract a three-tier pyramid HOG descriptor vector for one sample. Then we adopt the unsupervised
spectral regression method for dimensionality reduction on these descriptor vectors. Our algorithm
can be applied in the library entrance guard system of university and other research fields. Several
experiments on well-known face databases have shown good performance and good invariance against
illumination, occlusion and local deformation, etc.