Apparatus for generating a pattern recognition dictionary, a method thereof, a pattern recognition apparatus and a method thereof

A pattern recognition apparatus comprising:

• a constraint dictionary subspace memory section configured to store a constraint dictionary subspace which corresponds to a dictionary pattern;
• an input section configured to input a recognition-subject input pattern;
• a subspace generation section configured to generate an input subspace from the input pattern;
• a constraint subspace memory section configured to store beforehand a constraint subspace which is generated using a total sum of matrixes projected onto a dictionary subspace which corresponds to the dictionary pattern;
• a subspace projection section configured to project the input subspace to the constraint subspace to thereby obtain a constraint input subspace;
• and
• a discrimination section configured to obtain a canonical angle between the constraint input subspace and the constraint dictionary subspace and uses the canonical angle, to discriminate the recognition-subject input pattern.

A method of generating a pattern recognition dictionary, comprising:

• inputting a dictionary pattern;
• generating a dictionary subspace from the dictionary pattern;
• generating a constraint subspace from a sum matrix of matrixes projected onto the dictionary subspace;
• projecting the dictionary subspace to the constraint subspace to thereby generate a constraint dictionary subspace;
• and
• outputting the constraint subspace and the constraint dictionary subspace.

A method of recognizing a pattern, employing a constrained mutual subspace method by which an input subspace and a dictionary subspace in which respectively a recognition-subject input pattern and a dictionary pattern related to the recognition-subject input pattern are represented as a dictionary pattern are discriminated after being projected onto a constraint subspace which serves to suppress unnecessary pattern conditions, the method comprising:

• integrating, into one subspace, dictionary subspaces which are represented from dictionary patterns which belong to the same category and then obtaining a projection matrix;
• calculating a sum matrix for all or some of categories of the projection matrix;
• and
• using as the constraint subspace a subspace which is spanned by c number of eigenvectors, of eigenvectors of the sum matrix, which are selected starting from the eigenvectors having a smaller eigenvalue.