Apparatus and method of pattern recognition

A pattern recognition apparatus comprising:

• an input unit configured to input at least one input pattern;
• an input vector generating unit configured to generate an input vector from said input pattern;
• a reference subspace calculating unit configured to calculate a reference subspace, respectively, by analyzing a main component for a reference pattern relevant to a plurality of comparison targets to compare them with said recognition target;
• a constraint subspace storing unit configured to obtain a subspace for learning obtained by a main component analysis from a pattern including a variation resulting in declining pattern recognition performance and storing constraint subspaces obtained from said subspace for learning;
• a projecting unit configured to project said input vector and said plurality of reference subspaces onto each one of said plurality of constraint subspaces;
• a similarity calculating unit configured to calculate similarities between the input vector in said projected constraint subspaces and all of said plurality of reference subspaces, individually;
• a similarity combining unit configured to combine the similarities of said plurality of constraint spaces for each reference subspace and calculating a combined similarity for each reference subspace;
• and
• a determining unit configured to determine a comparison target of the reference subspace corresponding to a combined similarity having a high value among said combined similarities as said recognition target.

A pattern recognition method comprising:

• an input step of inputting at least one input pattern;
• an input subspace generating step for generating an input subspace by generating an input vector from said input pattern and analyzing a main component for said input vector;
• a reference subspace calculating step of calculating a reference subspace, respectively, by analyzing a main component for a reference pattern relevant to a plurality of comparison targets to compare them with said recognition target;
• a constraint subspace storing step of obtaining a subspace for learning obtained by a main component analysis from a pattern including a variation resulting in declining pattern recognition performance and storing a plurality of constraint subspaces obtained from said subspace for learning;
• a projecting step of projecting said input subspace and said plurality of reference subspaces onto each one of said plurality of constraint subspaces;
• a similarity calculating step of calculating similarities between said input subspace in the projected constraint subspaces and all of said plurality of reference subspaces, individually;
• a similarity combining step of combining the similarities of said plurality of constraint spaces for each reference subspace and calculating a combined similarity for each reference subspace;
• and
• a determining step of determining a comparison target of the reference subspace corresponding to a combined similarity having a high value among said combined similarities as said recognition target.

A pattern recognition method comprising:

• an input step of inputting at least one input pattern;
• an input vector generating step of generating an input vector from said input pattern;
• a reference subspace calculating step of calculating a reference subspace, respectively, by analyzing a main component for a reference pattern relevant to a plurality of comparison targets to compare them with said recognition target;
• a constraint subspace storing step of obtaining a subspace for learning obtained by a main component analysis from a pattern including a variation resulting in declining pattern recognition performance and storing constraint subspaces obtained from said subspace for learning;
• a projecting step of projecting said input vector and said plurality of reference subspaces onto each one of said plurality of constraint subspaces;
• a similarity calculating step of calculating similarities between the input vector in said projected constraint subspaces and all of said plurality of reference subspaces, individually;
• a similarity combining step of combining the similarities of said plurality of constraint spaces for each reference subspace and calculating a combined similarity for each reference subspace;
• and
• a determining step of determining a comparison target of the reference subspace corresponding to a combined similarity having a high value among said combined similarities as said recognition target.

A computer readable recording medium for storing a program for a pattern recognition method for causing a computer to implement:

• an input function of inputting at least one input pattern;
• an input subspace generating function of generating an input subspace by generating an input vector from said input pattern and analyzing a main component for said input vector;
• a reference subspace calculating step of calculating a reference subspace, respectively, by analyzing a main component for a reference pattern relevant to a plurality of comparison targets to compare them with said recognition target;
• a constraint subspace storing function of obtaining a subspace for learning obtained by a main component analysis from a pattern including a variation resulting in declining pattern recognition performance and storing a plurality of constraint subspaces obtained from said subspace for learning;
• a projecting function of projecting said input subspace and said plurality of reference subspaces onto each one of said plurality of constraint subspaces;
• a similarity calculating function of calculating similarities between said input subspace in the projected constraint subspaces and all of said plurality of reference subspaces, individually;
• a similarity combining function of combining the similarities of said plurality of constraint spaces for each reference subspace and calculating a combined similarity for each reference subspace;
• and
• a determining function of determining that a comparison target of the reference subspace corresponding to a combined similarity having a high value among said combined similarities as said recognition target.

A computer readable recording medium for storing a program for a pattern recognition method for causing a computer to implement:

• an input function of inputting at least one input pattern;
• an input vector generating function of generating an input vector from said input pattern;
• a reference subspace calculating function of calculating a reference subspace, respectively, by analyzing a main component for a reference pattern relevant to a plurality of comparison targets to compare them with said recognition target;
• a constraint subspace storing function of obtaining a subspace for learning obtained by a main component analysis from a pattern including a variation resulting in declining pattern recognition performance and storing a plurality of constraint subspaces obtained from said subspace for learning;
• a projecting function of projecting said input vector and said plurality of reference subspaces onto each one of said plurality of constraint subspaces;
• a similarity calculating function of calculating similarities between the input vector in said projected constraint subspaces and all of said plurality of reference subspaces, individually;
• a similarity combining function of combining the similarities of said plurality of constraint spaces for each reference subspace and calculating a combined similarity for each reference subspace;
• and
• a determining function of determining that a comparison target of the reference subspace corresponding to a combined similarity having a high value among said combined similarities as said recognition target.