Authors
Golubev A. S.
|
Degree
|
PhD, the senior lecturer of chair of Physics and the applied mathematics of VlGU |
|---|---|
|
E-mail
|
andrey.golubev@vlsu.ru |
|
Articles
|
Using laws maps and discreet cosine transform for face recognitionThis article contains a short survey for problems encountered in practical face recognition systems and foreign solutions for decision this problem. There are come to conclusions about the inapplicability their solutions for decision this problem with one sample per person. The recognition model is briefly describes, which have good results for face recognition for One Sample per Person problem. A mathematical model based on subspaces of the attributes is proposed to solve the stability problem of recognition coupled with various distortions of recognizable images that arise in real systems. The work of the proposed model is checked and recognition errors are analyzed. Read more...Recognition features using closest element method in estimates computing algorithmsA problem of threshold values selection for pattern recognition of compact sets with one sample per object using algorithms for calculating estimates is considered. We study an option to extend the set of samples by ad — hoc set of reference samples. Experimental results and sample set generation options are provided.
Read more...
Using prior training in pattern recognition system based on descriptorsIn the publication we discuss the enhancements of image recognition algorithms in two directions.
First, we define the binary descriptors of images as a second-lever features for image recognition
algorithm. They are constructed by a set of feature extraction functions together with an auxiliary set
of «quasi-etalon» image samples. Second, we use a novel prior-learning procedure named «Correction
by Noise Orthogonal Projection». By means of building special subspace of image features this
method leads to considerable decrease of intra-class distance while inter-class distance is practically
unchanged. Thus the probability of valid class partitioning is highly increased.
The subspace is built in two steps: 1) localize a feature space partition which contains the majority
of intra-class differences; 2) construct an orthogonal complement for the partition. This complement
is a target («noise») subspace. To enhance recognition, an input feature descriptor should be
projected to the subspace to reduce noise components.
The article provides the method’s mathematical formalization and experimental implementation.
The implementation is built upon a custom facial recognition system. We compare results of recognition
for several image sets, including well-known open face databases along with our own databases
captured from security video cameras located at places with high rate of people flow.
Read more...
|
