A pattern recognition apparatus and method for doing the same

Claim 2

A pattern recognition apparatus for recognizing an input pattern by extracting feature quantities from the input pattern and matching the input pattern with representative feature quantities each obtained by classifying a plurality of learning patterns, comprising: feature quantity extracting means for extracting feature quantities from each of a plurality of regions of said input pattern; classifying means for classifying said plurality of learning patterns of said regions based on the feature quantities of said regions extracted from said feature quantity extracting means for each of said plurality of learning patterns, representative feature quantity operating means for arithmetically operating representative feature quantities of respective classes, a dictionary storing means for storing said representative feature quantities of said respective classes and the kind of learning patterns in said respective classes, similarity operating means for arithmetically operating the similarity between the feature quantities of respective regions extracted from the feature quantity extracting means for the input patterns which is unknown, and the representative feature quantity of the corresponding respective classes in said dictionary storing means; point accumulating means for accumulating points corresponding to the order of the class of respective learning patterns included in the classes of predetermined upper orders of similarity with regard to the respective regions, after searching said dictionary storing means; and recognition result candidate output means for outputting the recognition result candidates from the highest to a lower number of accumulated points of the learning pattern after an accumulation operation is completed by said point accumulating means for respective regions.

Claim 3

The pattern recognition apparatus according to claim 1, wherein, said input pattern is of two dimensions, and said feature quantity extracting means comprises: direction component extracting means for extracting the direction component of the dot on which the pattern exists and is selected from the adjacent dot, for each dot of the input pattern, region dividing means for dividing said input pattern into a plurality of regions which overlap each other, and partial feature vector operating means for totalling the direction components of respective dots within the respective regions divided by said region dividing means, said direction component being obtained from said direction component extracting means by applying a weight to said direction component such that the weight decreases from the central portion of the region to the peripheral portion of the region, the totalling being conducting for respective directions, and arithmetically operating the partial feature vector having the totalling result for respective direction result as an element of the vector as the feature quantity of the corresponding regions of the input pattern.

Claim 4

The pattern recognition apparatus according to claim 2, wherein, said input pattern of two dimensions comprises a rectangular type pattern, and said region dividing means divides the input pattern into a plurality of rectangular regions which overlap each other.

Claim 5

The pattern recognition apparatus according to claim 2, wherein, said direction component extracting means has windows in the 45-degree, 90-degree, and 135-degree directions, scans respective windows for respective dots of the input pattern and extracts the direction component as the direction corresponding to the window when the pattern exists in the respective windows.

Claim 6

The pattern recognition apparatus according to claim 2, wherein, said similarity operation means arithmetically operates a similarity for determining a small value obtained as a sum of the second powers of the euclidean distance between respective elements of the partial feature vector of respective regions extracted from said partial feature vector arithmetic operation means from said unknown input pattern and the respective element of the representative partial feature vectors of the representative feature quantity of the corresponding respective classes in said dictionary storing means, as a large similarity.

Claim 7

The pattern recognition apparatus according to claim 2 further comprising: similar-pattern dictionary storing means for, after extracting the partial feature vector corresponding to at least one predetermined region from said partial feature vector operating means for respective similar learning patterns within a plurality of similar learning patterns, storing the kinds of said respective similar learning patterns, the discrimination data for said predetermined region, and the partial feature vector of said similar learning patterns of said predetermined region, predetermined-region similarity operating means for where the recognition result candidate outputted from said recognition result candidate output means corresponds to any one of the group of said similarity learning patterns stored in said similarity pattern dictionary storing means, arithmetically operating the similarity between a partial feature vector which is for the unknown input pattern corresponding to said recognition result candidate output, is obtained by said partial feature vector operating means and corresponds to the predetermined region obtained from said discriminating data stored in said similarity pattern dictionary storing means, and a partial feature vector of respective similar learning patterns of the group corresponding to said recognition result candidate output stored in said similar pattern dictionary storing means and a fine classification result candidate output means for outputting the similar learning pattern in the order of the highest to a lower similarity as the fine classification result candidate corresponding to said recognition result candidate output with a rank.

Claim 8

The pattern recognition appartus according to claim 1, wherein, said clustering means carries out a clustering of the feature quantity of respective regions extracted from the feature quantity extracting means for a plurality of respective learning patterns and classifies a plurality of learning patterns for respective regions, and said representative feature quantity arithmetical operating means for arithmetically operating the representative feature quantity of respective classes by arithmetically operating the average of respective feature quantities corresponding to said respective learning pattern in respective classes.

Claim 9

The pattern recognition apparatus according to claim 1, wherein, said point accumulaing means searches said dictionary storing means and accumulates points obtained by adding "1" " to the value obtained by subtracting the order of the class from the predetermined order, for respective learning patterns included in classes up to the predetermined order from the highest to a lower similarity for respective regions, the point corresponding to the class.

Claim 10

A pattern recognition apparatus for recognizing an input word pattern by extracting a feature quantity from the input word pattern and matching the input word pattern with representative feature quantities each obtained by classifying a plurality of learning patterns, comprising feature quantity extracting means for extracting a feature quantity from each of a plurality of regions of the input word pattern, a classifying means for classifying said plurality of learning word patterns in respective parts, based on a feature quantity of each part extracted from the feature quantity extracting means for said plurality of respective learning word patterns, a representative feature quantity operating means for arithmetically operating the representative feature quantities of respective classes, a dictionary storing means for storing the representative feature quantities of respective classes and the kinds of learning word patterns included in respective classes for respective parts as a dictionary, similarity operating means for arithmetically operating a similarity between a feature quantity of respective parts extracted from the feature quantity extracting means for the input word pattern which is unknown and representative feature quantities of corresponding classes in said dictionary storing means, a rough classification means for arithmetically operating a similarity between said respective learning patterns stored in the dictionary storing means and said unknown input pattern based on the similarity obtained by the similarity operating means and outputting a rough classification word candidate in the order of the highest to a lower similarity of the learning word pattern, and a vowel sequence recognition means for recognizing a vowel sequence of said unknown input word pattern, a vowel sequence dictionary means for storing the vowel sequence of respective learning patterns, and a fine classification means for perfoming DP matching of a vowel sequence of respective learning word pattern corresponding to respective rough classification word candidates extracted from said vowel sequence dictionary storing means and the vowel sequence of said unknown input word pattern outputted from said vowel sequence recognition means and outputting fine classification candidates in the order of the highest to a lower matching degree of the rough classification word candidate.

Claim 11

A pattern recognition apparatus for recognizing an input word pattern by extracting a feature quantity from the input word pattern and matching the input word pattern with representative feature quantities each obtained by classifying a plurality of learning patterns, comprising feature quantity extracting means for extracting a feature quantity from each of a plurality of regions of the input word pattern, a classifying means for classifying said plurality of learning word patterns in respective parts, based on a feature quantity of each part extracted from the feature quantity extracting means for said plurality of respective learning word patterns, a representative feature quantity operating means for arithmetically operating the representative feature quantities of respective classes, a dictionary storing means for storing the representative feature quantities of respective classes and the kinds of learning word patterns included in respective classes for respective parts as a dictionary, similarity operating means for arithmetically operating a similarity between a feature quantity of respective parts extracted from the feature quantity extracting means for the input word pattern which is unknown and representative feature quantities of corresponding classes in said dictionary storing means, a rough classification means for arithmetically operating a similarity between said respective learning patterns stored in the dictionary storing means and said unknown input pattern based on the similarity obtained by the similarity operating means and outputting a rough classification word candidate in the order of the highest to a lower similarity of the learning word pattern, and a vowel sequence recognition means for recognizing a vowel sequence of said unknown input word pattern, a vowel sequence dictionary means for storing the vowel sequence of respective learning patterns, and a fine classification means for perfoming DP matching of a vowel sequence of respective learning word pattern corresponding to respective rough classification word candidates extracted from said vowel sequence dictionary storing means and the vowel sequence of said unknown input word pattern outputted from said vowel sequence recognition means and outputting fine classification candidates in the order of the highest to a lower matching degree of the rough classification word candidate.

Claim 12

The pattern recognition apparatus according to claim 9, wherein said feature quantity extracting means comprises normalizing means for normalizing said input word pattern to have a predetermined time period, cepstrum operating means for dividing said normalized input word pattern into a plurality of frames with the same time period and arithmetically operating the cepstrum coefficient of the predetermined dimensions through a cepstrum analysis, partial feature quantity extracting means for collecting every plurality of frames having a continuous cepstrum coefficient and extracting the feature quantity corresponding to each one of the plurality of words.

Claim 13

The pattern recognition apparatus according to claim 10, wherein the similarity is operated by determining a low value obtained as the second power of the sum of the euclidean distances between cepstrum coefficients of respective dimensions of respective parts extracted from said partial feature quantity extracting means from unknown input word patterns and the representative cepstrum coefficient of respective dimensions, which are representative feature quantities of the corresponding respective classes in the dictionary storing means.

Claim 14

The pattern recognition apparatus according to claim 9, wherein said classifying means carries out a clustering of feature quantities of the respective parts extracted from the feature quantity extracting means from said plurality of learning word patterns and classifies said plurality of learning patterns for respective parts, said representative feature quantity arithmetic means arithmetically operates the representative feature quantities of respective classes by arithmetically operating the average of the respective feature quantities corresponding to the respective learning word patterns within respective classes.

Claim 15

The pattern recognition method according to claim 9, wherein said vowel sequence recognition means outputs a plurality of vowel sequence candidates as the vowel sequence of the unknown input word pattern, and said fine classification recognition means performs DP matching between said plurality of vowel sequence candidates and the vowel sequence of the rough classification word candidates.

Claim 16

The pattern recognition apparatus according to claim 9, wherein said vowel sequence recognition means outputs a plurality of vowel sequence candidates with ranks as the vowel sequence of said input word pattern, and said fine classification means performs a DP matching between a plurality of vowel sequence candidates and the vowel sequence of respective rough classification word candidates to which a weight is applied in accodance with the order.

Claim 17

A pattern recognition apparatus for recognizing the input word pattern by extracting a feature quantity from the input word pattern and matching the input word pattern with representative feature quantities each obtained by classifying a plurality of learning patterns, comprising: feature quantity extracting means for extracting feature quantities from respective parts of a plurality of parts of an input word pattern, classifying means for classifying said plurality of learning word patterns for respective parts, based on the feature quantity of respective parts extracted from said feature quantity extracting means for said plurality of respective learning word patterns, a representative feature quantity arithmetically operating means for arithmetically operating the representative feature quantity of respective classes, a dictionary storing means for storing the representative feature quantities of respective classes and the kinds of the learning word patterns included in respective classes, as a dictionary for respective parts, similarity operating means for arithmetically operating the similarity between the feature quantities of respective parts extracted from the feature quantity extracting means for the input word pattern which is unknown and the representative feature quantity of corresponding respective classes within the dictionary storing means, point accumulating means for accumulating points corresponding to the order of classes for respective learning word patterns included in the class of the predetermined upper order from the highest to a lower similarity for respective regions, rough classification means for outputting rough classification word candidates with a rank in accordance with the order from the highest to a lower number of accumulated points of the learning pattern after said point accumulating means completes the accumulating operation for respective regions, vowel recognition means for recognizing the vowel sequence of unknown input word patterns, vowel sequence dictionary storing means for storing a vowel sequence of respective learning word patterns as a dictionary and a fine classification means for performing a DP matching of the vowel sequence of respective learning word patterns corresponding to said rough classification word candidates extracted from said vowel sequence dictionary storing means, and a vowel sequence of said unknown input word pattern outputted from said vowel sequence recognition means which outputs fine classification word candidates with a rank in accordance with the order of from the highest to a lower matching degree of the rough classification word candidate.

Claim 18

A pattern recognition apparatus for recognizing the input word pattern by extracting a feature quantity from the input word pattern and matching the input word pattern with representative feature quantities each obtained by classifying a plurality of learning patterns, comprising: feature quantity extracting means for extracting feature quantities from respective parts of a plurality of parts of an input word pattern, classifying means for classifying said plurality of learning word patterns for respective parts, based on the feature quantity of respective parts extracted from said feature quantity extracting means for said plurality of respective learning word patterns, a representative feature quantity arithmetically operating means for arithmetically operating the representative feature quantity of respective classes, a dictionary storing means for storing the representative feature quantities of respective classes and the kinds of the learning word patterns included in respective classes, as a dictionary for respective parts, similarity operating means for arithmetically operating the similarity between the feature quantities of respective parts extracted from the feature quantity extracting means for the input word pattern which is unknown and the representative feature quantity of corresponding respective classes within the dictionary storing means, point accumulating means for accumulating points corresponding to the order of classes for respective learning word patterns included in the class of the predetermined upper order from the highest to a lower similarity for respective regions, rough classification means for outputting rough classification word candidates with a rank in accordance with the order from the highest to a lower number of accumulated points of the learning pattern after said point accumulating means completes the accumulating operation for respective regions, vowel recognition means for recognizing the vowel sequence of unknown input word patterns, vowel sequence dictionary storing means for storing a vowel sequence of respective learning word patterns as a dictionary and a fine classification means for performing a DP matching of the vowel sequence of respective learning word patterns corresponding to said rough classification word candidates extracted from said vowel sequence dictionary storing means, and a vowel sequence of said unknown input word pattern outputted from said vowel sequence recognition means which outputs fine classification word candidates with a rank in accordance with the order of from the highest to a lower matching degree of the rough classification word candidate.