Pattern recognition with a warping function decided for each reference pattern by the use of feature vector components of a few channels

• US 4,282,403 A
• Filed: 08/08/1979
• Issued: 08/04/1981
• Est. Priority Date: 08/10/1978
• Status: Expired due to Term

• Current Assignee: Nippon Electric Company Limited (NEC Corporation)
• Original Assignee: Nippon Electric Company Limited (NEC Corporation)
• Inventors: Sakoe, Hiroaki
• Primary Examiner(s): Nusbaum, Mark E.
• Assistant Examiner(s): KEMENY, EMANUEL
• Application Number:
  US06/064,965
• Time in Patent Office: 727 Days
• Field of Search: 179/1 SM, 179/1 SB, 364/728, 340/146.3 H, 340/146.3 AQ
• US Class Current: 704/234
• CPC Class Codes: G06V 10/754 involving a deformation of the sample pattern or of the reference pattern; Elastic matching
• CPC Class Codes: G10L 15/00 Speech recognition G10L17/00 takes precedence
• CPC Class Codes: G10L 15/12 using dynamic programming techniques, e.g. dynamic time warping [DTW]

Abstract

In a pattern recognition device according to pattern matching, one or more specific dimensions of vector components are memorized for each reference pattern feature vector sequence in a reference pattern memory for the reference pattern feature vector sequences. A warping function for time-normalizing input pattern feature vectors of a sequence and the vectors of each reference pattern feature vector sequence is determined so as to minimize the difference between a pattern represented by the specific vector components of the specific dimension or dimensions and another pattern represented by the vector components corresponding in the input pattern feature vector sequence to the specific reference pattern feature vector components as regards the dimensions of a space in which each input or reference pattern feature vector is defined. The input pattern feature vector sequence and each reference pattern feature vector sequence are subjected to nonlinear pattern matching with reference to the warping function. The pattern matching may be between the vector components of all dimensions or those of several dimensions including the specific dimension or dimensions. Preferably, one or more dimensions are specified as the specific one or ones by selecting each dimension for which a variation with time of a pattern represented by the reference pattern feature vector components is a maximum of similar variations of patterns represented by the vector components of other dimensions.

Claims

1. A pattern recognition device for recognizing an input pattern with reference to a plurality of reference patterns, said device comprising:

input pattern buffer means for memorizing a time sequence of input pattern feature vectors representative of said input pattern and defined by vector components in a space of a prescribed number of dimensions;

reference pattern memory means for memorizing time sequences of reference pattern feature vectors representative of the respective reference patterns and given by vector components in said space, the vector components of any one of said reference pattern feature vector sequences and said input pattern feature vector components of an associated one of said input pattern feature vectors thereby corresponding to one another, said reference pattern memory means being for further memorizing dimension specifying signals in one-to-one correspondence with said reference pattern feature vector sequence, each dimension specifying signal specifying specific vector components within a specific dimension of the reference pattern feature vector sequence;

time-normalizing means connected to said input pattern buffer means and said reference pattern memory means for carrying out time normalization between said input pattern feature vector sequence and each reference pattern feature vector sequence by determining a warping function for said each reference pattern feature vector sequence which time correlates said input pattern feature vectors and the feature vectors of said each reference pattern feature vector sequence to one another so as to minimize a difference between a pattern represented by the specific vector components specified by the dimension specifying signal corresponding to said each reference pattern feature vector sequence and another pattern represented by the input pattern feature vector components corresponding to said specific vector components; and

comparing means connected to said input pattern buffer means, said reference pattern memory means, and said time-normalizing means for comparing said input pattern feature vector sequence and said reference pattern feature vector sequences with one another with reference to the warping functions determined for the respective reference pattern feature vector sequences to decide which of said reference patterns is most similar to said input pattern.

2. A pattern recognition device as claimed in claim 1, wherein said comparing means comprises:

reading means connected to said input pattern buffer means, said reference pattern memory means, and said time-normalizing means for reading time correlated input pattern feature vectors and the feature vectors of said each reference pattern feature vector sequence;

nonlinear matching means connected to said reading means for carrying out nonlinear pattern matching between a pattern represented by the read-out input pattern feature vectors and another pattern represented by the read-out reference pattern feature vectors to derive for said each reference pattern feature vector sequence a similarity measure between said input pattern and the reference pattern represented by said each reference pattern feature vector sequence; and

decision means connected to said nonlinear matching means for deciding to which of said reference patterns, said input pattern corresponds in response to the maximum similarity measure.

3. A pattern recognition device as claimed in claim 1, wherein said comparing means comprises:

reading means connected to said input pattern buffer means, said reference pattern memory means, and said time-normalizing means for reading particular vector components of said input pattern feature vectors and particular vector components of the feature vectors of said each reference pattern feature vector sequence which are time correlated to one another by the warping function determined for said each reference pattern feature vector sequence, said particular reference pattern feature vector components being the components of a plurality of particular dimensions predetermined for each of said reference pattern feature vector sequences, said particular dimensions including the at least one specific dimension specified by said dimension specifying signal, the dimensions of said particular input pattern feature vector components corresponding to the dimensions of said particular reference pattern feature vector components;

nonlinear matching means connected to said reading means for carrying out nonlinear pattern matching between a pattern represented by the read-out particular input pattern feature vector components and another pattern represented by the read-out particular reference pattern feature vector components to derive for said each reference pattern feature vector sequence a similarity measure between said patterns; and

decision means connected to said nonlinear matching means for deciding which of said reference patterns said input pattern corresponds in response to the maximum similarity measure.

4. A pattern recognition device as claimed in claims 1, 2, or 3, further comprising:

input means connected to said reference pattern memory means for successively supplying said reference pattern feature vector sequences to said memory means to be memorized therein;

dimension specifying means connected to said input means for successively specifying the specific dimensions for the respective reference pattern feature vector sequences, said dimension specifying signals produced by comparing variations in time of patterns represented by the vector components of the respective dimensions for each of said reference pattern feature vector sequences with one another and by selecting the dimension corresponding to the reference pattern feature vector sequence for which the variation is a maximum; and

means between said dimension specifying means and said reference pattern memory means for successively supplying the produced dimension specifying signals to said reference pattern memory means to allow said reference pattern memory means to memorize the produced dimension specifying signals.