Recognition and judgement apparatus having various learning functions

• US 5,966,701 A
• Filed: 06/19/1996
• Issued: 10/12/1999
• Est. Priority Date: 06/21/1995
• Status: Expired due to Fees

• Current Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Original Assignee: Matsushita Electric Industrial Company Limited (Panasonic Holdings Corporation)
• Inventors: Kohda, Toshiyuki, Miyazaki, Keiichi
• Primary Examiner(s): Hafiz, Tariq R.
• Assistant Examiner(s): STARKS, WILBERT L
• Application Number:
  US08/668,588
• Time in Patent Office: 1,210 Days
• Field of Search: 706/16, 706/23, 706/244
• US Class Current: 706/20
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06N 3/043 based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS]
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06V 30/242 Division of the character sequences into groups prior to recognition; Selection of dictionaries

Abstract

A fuzzy rough classification section 7 classifies an input pattern into a plurality of category groups and calculates their attribution factors. A plurality of fine classification sections 8 calculate similarities in respective category groups. In each fine classification section 8, a weight change amount control section 82 controls the weight change amount in an associated similarity calculating section 81 in accordance with the group attribution factor selected by a second group selecting section (10), thereby enabling a cooperative learning operation among plural fine classification sections 8. A force learning control section 17 performs a supplemental learning operation, if the input pattern is rejected regardless of its correctness, until its reject judge value is reduced below a predetermined reject threshold.

Claims

1. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input pattern outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section,

wherein;

said similarity calculating section is constituted be a multi-layer structure comprising a most-lowest layer consisting of a plurality of third unit recognizers, a previous or preceding layer of said most-lowest layer consisting of a plurality of second unit recognizers, and another layer consisting of a plurality of first unit recognizers;

said first unit recognizer comprises;

a signal input section receiving an input signal;

a quantizer converting the input signal into a quantized level;

a neighboring section selector selecting a neighboring quantized section of a concerned quantized section corresponding to the input signal;

a weight table memorizing weight values corresponding to said quantized section of said input signal and its neighboring quantized section;

a route input section having at least one route input terminal;

a route output section having at least one route output terminal; and

a route weighting section varying a linkage intensity between said route input terminal of said route input section and said route output terminal of said route output section by changing a setting position of each weight to be applied to the route in accordance with a quantized result;

said second unit recognizer comprises;

a signal input section receiving an input signal;

a quantizer converting the input signal into a quantized level;

a route input section having at least one route input terminal;

a route output section having at least one route output terminal;

a transfer amount change section shifting a transmission value be several bits when said transmission value is entered from said route input terminal of said route input section, and a learning unit renewing linkage intensity between said route input terminal and a route output terminal of said route output terminal indicated by an output of said quantizer in accordance with an output of said transfer amount change section,

said third unit recognizer comprises;

an adder adding input signals entered through its plural route input terminals, and a threshold processing unit processing an output of said adder by a given threshold, and

said similarity calculating section firer comprises a teacher signal converting section which converts an output of said teacher signal generating section into a signal identifying a specific third unit recognizer which should generate a maximum output in said fine classification section, and an output of said teacher signal converting section is entered into the signal input section of said second unit recognizers.

2. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input patterm outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section,

wherein;

said similarity calculating section is constituted by a multi-layer structure comprising a most-lowest layer consisting of a plurality of third unit recognizers, a previous or preceding layer of said most-lowest layer consisting of a plurality of second unit recognizers, and another layer consisting of a plurality of first unit recognizers;

said first unit recognizer comprises;

a signal input section receiving an input signal;

a quantizer converting the input signal into a quantized level;

a neighboring section selector selecting a neighboring quantized section of a concerned quantized section corresponding to the input signal;

a weight table memorizing weight values corresponding to said quantized section of said input signal and its neighboring quantized section;

a route input section having at least one route input terminal;

a route output section having at least one route output terminal; and

a route weighting section varying a linkage intensity between said route input terminal of said route input section and said route output terminal of said route output section by changing a setting position of each weight to be applied to the route in accordance with a quantized result;

said second unit recognizer comprises;

a signal input section receiving an input signal;

a quantizer converting the input signal into a quantized level;

a route input section having at least one route input terminal;

a route output section having at least one route output terminal;

a transfer amount change section shifting a transmission value by several bits when said transmission value is entered from said route input terminal of said route input section, and a learning unit renewing linkage intensity between said route input terminal and a route output terminal of said route output terminal indicated by an output of said quantizer in accordance with an output of said transfer amount change section,

said third unit recognizer comprises;

an adder adding input signals entered through its plural route input terminals, and a threshold processing unit processing an output of said adder by a given threshold, and

said weight change amount control section controls a bit shift amount in said transfer amount change section of a corresponding similarity calculating section in response to the output of said second group selecting section.

3. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimiation signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger soup attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input pattern outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section,

wherein;

said similarity calculating section is constituted by a multi-layer structure comprising a most-lowest layer consisting of a plurality of third unit recognizers, a previous or preceding layer of said most-lowest layer consisting of a plurality of second unit recognizers, and another layer consisting of a plurality of first unit recognizers; and

further comprises a teacher signal converting section which converts an output of said teacher signal generating section into a signal identifying a specific third unit recognizer which should generate a maximum output in said fine classification section, and an output of said teacher signal converting section is entered into the signal input section of said second unit recognizers.

4. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input pattern outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section,

wherein said weight change amount control section controls a bit shift amount to change a transfer amount in a corresponding similarity calculating section in response to the output of said second group selecting section.

5. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input pattern outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section,

wherein said recognition result calculating section selects a maximum category similarity and a second-largest category similarity to obtain a reject judge value equivalent to a ratio of said selected two category similarities, and generates a reject signal when said reject judge value is larger than a predetermined reject threshold.

6. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input pattern outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section, further comprising;

a group number selecting section which selects a group number having a maximum group attribution factor among fine classification sections involving the category of an input data based on said category information memorized in said category information memory section;

an additional learning information memory section which memorizes an erroneously recognized input data, its corresponding category number and said selected group number;

an error recognition data counting section which counts the number of erroneous recognition data in connection with each group number and each category based on the information memorized in said additional learning information memory section; and

a first additional learning candidate selecting section which selects an input data and a category number corresponding to the group number and the category number which are large in the number of the erroneous recognition data counted by said error recognition data counting section.

7. The learning type recognition &

judgement apparatus in accordance with claim 6, further comprising;

a reference data memory section which memorizes reference data and a teacher signal used for evaluating a learning result relating to the data selected by said first additional learning candidate selecting section;

an additional learning result acceptability judging section which judges the acceptability of an additional learning operation by evaluating the recognition result relating to said reference data, and

a second additional learning candidate selecting section which selects only additional learning candidate data judged acceptable in said additional learning result acceptability judging section.

8. The learning type recognition &

judgement apparatus in accordance with claim 7, wherein said additional learning result acceptability judging section judges the acceptability of the additional learning operation based on a change of a sum of the number of rejected data and a multiple (integer not less than

     1) of the number of erroneous recognition data before and after the additional learning operation.

9. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input pattern outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section,

wherein said supplementary learning control section controls the learning operation of said fine classification section until a reject judge value is reduced below a predetermined threshold.

10. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input pattern outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section, further comprising;

a first group number selecting section which selects a group number having the maximum group attribution factor among fine classification sections involving the category of an input data based on said category information memorized in said category information memory section;

a force learning data memory section which memorizes a rejected correct-answer input data, its corresponding category number (teacher signal) and said selected group number;

a rejected correct-answer data number counting section which counts the number of rejected correct-answer data in connection with each of group number and category based on the information memorized in said force learning data memory section; and

a first force learning candidate data selecting section which selects input data and category number corresponding to the group number and the category number which are large in the count number of the rejected correct-answer data counted by said rejected correct-answer data number counting section.

11. The learning type recognition &

judgement apparatus in accordance with claim 10, further comprising;

a reference data memory section which memorizes reference data and a teacher signal used for evaluating a force learning result relating to the data selected by the first force learning candidate data selecting section;

a force learning result acceptability judging section which judges the acceptability of a force learning operation by evaluating a recognition result relating to said reference data; and

a second force learning candidate data selecting section which selects only force learning candidate data judged as acceptable in the force learning result acceptability judging section.

12. The learning type recognition &

judgement apparatus in accordance with claim 11, wherein said force learning result acceptability judging section judges the acceptability of the force learning operation based on a change of a sum of the number of rejected data and a multiple (integer not less than

     1) of the number of erroneous reading data before and after the force learning operation.

13. A learning type recognition &

judgement apparatus comprising;

a group dictionary section for memorizing a plurality of group reference pattern signals representative of category groups consisting of a set of similar patterns;

a fuzzy rough classification section for calculating group attribution factors of each input pattern signal with reference to said group reference pattern signals, each of said group attribution factors representing an attribution degree of said input pattern signal to each category group;

a first group selecting section for selecting at least one of category groups based on said group attribution factors;

a plurality of fine classification sections, each including a similarity calculating section and a weight change amount control section, said similarity calculating section calculating an in-group similarity representing a similarity degree of the input pattern signal with respect to each category involved in the same category group, while said weight change amount control section controlling a weight change amount of an associated similarity calculating section in accordance with an output of a second group selecting section;

a discrimination signal weighting section which gives weights to said in-group similarities obtained by fine classification sections based on the group attribution factors of the category groups selected by first group selecting section;

a category similarity calculating section which obtains a category similarity representing a belonging degree to each category based on an output of said discrimination signal weighting section;

a recognition result calculating section which generates a recognition result or a reject signal based on the category similarity calculated by said category similarity calculating section;

a teacher signal generating section which generates a teacher signal required to learn said fine classification section;

a category information memory section which memorizes category information belonging to each fine classification section;

said second group selecting section selecting at least one larger group attribution factors in the fine classification sections comprising the category of the input pattern signal, based on said category information memorized in said category information memory section;

a learning control section which controls a learning operation of each fine classification section in accordance with a comparison between said recognition result of said recognition result calculating section and a category of input patterm outputted from said teacher signal generating section;

a recognition result display section which displays a judgement result of said recognition result calculating section;

a correct-answer input section which inputs a correct recognition result when said recognition result is erroneous or rejected regardless of its correctness; and

a supplementary learning control section which controls the learning operation of specific fine classification sections relating to erroneously recognized data or rejected correct-answer data with reference to said correct recognition result entered from said correct-answer input section, further comprising;

a first reject threshold control section which generates two kinds of reject thresholds by adding and subtracting a constant value to and from a preset reject threshold;

a second group number selecting section which selects a group number having the maximum group attribution factor with respect to evaluation data;

a recognition result accumulating section which accumulates the number of erroneous reading and rejected data in connection with each of said selected group number and category based on a recognition result of said evaluation data; and

a first reject threshold determining section which determines a reject threshold in connection with group number or category based on a comparison in each of the group number or the category between an accumulation result with respect to a predetermined reject threshold and an accumulation result with respect to a reject threshold obtained after the above-described change.

14. The learning type recognition &

judgement apparatus in accordance with claim 13, further comprising;

a threshold change candidate selecting section which selects at least one of group numbers or category numbers having a large difference between the number of erroneous reading data and the number of rejected data based on an accumulation result in said recognition result accumulating section, and

a second reject threshold control section which is substituted for said first reject threshold control section to add or subtract a constant value to or from the reject threshold of said group or category based on the largeness relationship between the number of erroneous reading data and the number of rejected data in the selected group or category.

15. The learning type recognition &

judgement apparatus in accordance with claim 14, wherein said reject threshold determining section adds a constant value to the reject threshold of said group or category when a sum of a multiple of an increased number of erroneous reading data and a reduced number of rejected data is less than a predetermined value, and subtracts a constant value from said rejected threshold of said group or category when a sum of a multiple of a reduced number of said erroneous reading data and an increased number of said rejected data is less than a predetermined value.

16. The learning type recognition &

judgement apparatus in accordance with claim 13, wherein said reject threshold determining section adds a constant value to the reject threshold of said group or category when a sum of a multiple of an increased number of erroneous reading data and a reduced number of rejected data is less than a predetermined value, and subtracts a constant value from said rejected threshold of said group or category when a sum of a multiple of a reduced number of said erroneous reading data and an increased number of said rejected data is less than a predetermined value.