Patent US5835633A successfully refreshed from Google Patents.
US5835633A
Expired / Lapsed
US
⏱️ 0.0 years remaining
Concurrent two-stage multi-network optical character recognition system
Assignee: International Business Machines Corp
•
Filing: 1995-11-20
•
Granted: 1998-11-10
•
Expires: 2015-11-20
Abstract
Abstract A multi-stage multi-network character recognition system decomposes the estimation of a posteriori probabilities into coarse-to-fine stages. Classification is then based on the estimated a posteriori probabilities. This classification process is especially suitable for the tasks that involve a large number of categories. The multi-network system is implemented in two stages: a soft pre-classifier and a bank of multiple specialized networks. The pre-classifier performs coarse evaluation of the input character, developing different probabilities that the input character falls into different predefined character groups. The bank of specialized networks, each corresponding to a single group of characters, performs fine evaluation of the input character, where each develops different probabilities that the input character represents each character in that specialized network's respective predefined character group. A network selector is employed to increase the system's efficiency by selectively invoking certain specialized networks selected, using a combination of prior external information and outputs of the pre-classifier. Relative to known single network or one-stage multiple network recognition systems, the invention provides improved recognition, accuracy, confidence measure, speed, and flexibility.
Inventors
- 👤 Tetsunosuke Fujisaki
- 👤 Jianchang Mao
- 👤 Kottappuram Mohamedali Mohiuddin
Bibliographic Details
Application Number
—
Jurisdiction
US
Data Source
google_patents
Last Synchronized
2026-11-09 10:53
CPC Classifications
G; G06; G06V; G06V30/00; G06V30/10; G06V30/24; G06V30/242; G06F; G06F18/00; G06F18/20; G06F18/25; G06F18/254; G06F18/256; G06V30/248; G06V30/2504
Claims Analysis
Total Claims: 45
Claim 1
Independent Claim
183 words
A character recognition system, comprising: a soft pre-classifier responsive to a representation of an input character to provide a first output signal comprising a coarse vector estimating probabilities of the input character belonging to each one of multiple predefined groups of recognized target characters; for each particular group of target characters, a specialized neural network to receive the input character and provide a second output signal comprising a fine vector estimating probabilities of the input character representing each one of multiple target characters in the particular group of target characters; a multiplier, coupled to the soft pre-classifier and each of the specialized neural networks, to weight each specialized neural network's second output signal in proportion to the first output signal to provide multiplier output lists, each multiplier list including a list of target characters each associated with a confidence value indicating a probability that the associated target character correctly represents the input character; and a candidate selector, coupled to the multiplier, to compile and rank contents of the multiplier output lists and provide an aggregate list of target characters and associated confidence values.
Claim 2
Dependent Claim
183 words
A character recognition system, comprising: a soft pre-classifier responsive to a representation of an input character to provide a first output signal comprising a coarse vector estimating probabilities of the input character belonging to each one of multiple predefined groups of recognized target characters; for each particular group of target characters, a specialized neural network to receive the input character and provide a second output signal comprising a fine vector estimating probabilities of the input character representing each one of multiple target characters in the particular group of target characters; a multiplier, coupled to the soft pre-classifier and each of the specialized neural networks, to weight each specialized neural network's second output signal in proportion to the first output signal to provide multiplier output lists, each multiplier list including a list of target characters each associated with a confidence value indicating a probability that the associated target character correctly represents the input character; and a candidate selector, coupled to the multiplier, to compile and rank contents of the multiplier output lists and provide an aggregate list of target characters and associated confidence values.
Claim 3
Dependent Claim
20 words
The character recognition system of claim 1, the aggregate list of target characters being ordered according to their confidence values.
Claim 4
Dependent Claim
23 words
The character recognition system of claim 1, the aggregate list being limited to target characters having a number of the highest confidence values.
Claim 5
Dependent Claim
22 words
The character recognition system of claim 1, the aggregate list being limited to target characters having confidence values exceeding a predetermined threshold.
Claim 6
Dependent Claim
34 words
The character recognition system of claim 1, the soft pre-classifier comprising a feedforward neural network to perform probability calculations to identify one group of target characters that the input character most likely belongs to.
Claim 7
Dependent Claim
38 words
The character recognition system of claim 1, each specialized neural network comprising a feedforward neural network to perform probability calculations to identify a target character in a particular group of characters that the input character most likely represents.
Claim 8
Dependent Claim
21 words
The character recognition system of claim 1, each of the multiple groups of target characters including non-overlapping sets of target characters.
Claim 9
Dependent Claim
19 words
The character recognition system of claim 1, the multiple groups of target characters including partially overlapping sets of characters.
Claim 10
Dependent Claim
34 words
The character recognition system of claim 1, the multiple groups of target characters including: a group of lower-case characters, a group of upper-case characters, a group of digits, and a group of special symbols.
Claim 11
Dependent Claim
35 words
The character recognition system of claim 1, further including a transducer coupled to the soft pre-classifier to receive the input character and provide the representation of the input character, said representation comprising a bit map.
Claim 12
Dependent Claim
14 words
The character recognition system of claim 10, the transducer comprising a digital optical scanner.
Claim 13
Dependent Claim
17 words
The character recognition system of claim 10, the transducer comprising a pen and an associated digitizing pad.
Claim 14
Dependent Claim
20 words
The character recognition system of claim 10, further including a feature extractor interposed between the transducer and the soft pre-classifier.
Claim 15
Dependent Claim
60 words
The character recognition system of claim 1, further comprising: a transducer to receive a tangible medium bearing the input character and create a machine-readable matrix representative of the input character; and a feature extractor, coupled between the transducer and the soft pre-classifier, responsive to the matrix to generate a feature vector containing representations of selected characteristics of the input character.
Claim 16
Dependent Claim
14 words
The character recognition system of claim 14, the transducer comprising a digitizing optical scanner.
Claim 17
Dependent Claim
17 words
The character recognition system of claim 14, the transducer comprising a pen and an associated digitizing pad.
Claim 18
Dependent Claim
34 words
The character recognition system of claim 14, the feature extractor comprising a mechanism responsive to the matrix to generate a feature vector containing representations of characteristics including contour direction features of the input character.
Claim 19
Dependent Claim
34 words
The character recognition system of claim 14, the feature extractor comprising a mechanism responsive to the matrix to generate a feature vector containing representations of characteristics including bending point features of the input character.
Claim 20
Dependent Claim
85 words
The character recognition system of claim 1, the soft pre-classifier including a first input to receive an external information signal associating the input character with one or more specialized neural networks, a second input to receive the representation of the input character, and a network selector responsive to the first output signal and the external information signal to select all specialized neural networks identified by the external information signal and all specialized neural networks whose component of the first output signal exceeds a predetermined threshold.
Claim 21
Dependent Claim
21 words
The character recognition system of claim 19, the network selector further including circuitry to deactivate all specialized neural networks not selected.
Claim 22
Dependent Claim
87 words
The character recognition system of claim 1, further including a network selector including a third input to receive the first output signal from the soft pre-classifier, a fourth input to receive an external information signal associating the input character with one or more specialized neural networks, and circuitry responsive to the first output signal and the external information signal to select all specialized neural networks identified by the external information signal and all specialized neural networks whose component of the first output signal exceeds a predetermined threshold.
Claim 23
Dependent Claim
18 words
The character recognition system of claim 21, the network selector inhibiting outputs of the selected specialized neural networks.
Claim 24
Dependent Claim
18 words
The character recognition system of claim 21, the network selector disabling outputs of the selected specialized neural networks.
Claim 25
Dependent Claim
119 words
The character recognition system of claim 1, further including a network selector including a third input to receive the coarse vector from the soft pre-classifier, the coarse vector including multiple components each estimating a probability that the input character belongs to a different one of the predefined groups of recognized target characters, a fourth input to receive external information associating the input character with one or more specialized neural networks, and circuitry programmed to select one or more specialized neural networks by performing steps comprising: selecting all specialized neural networks identified by the external information; normalizing the components of the coarse vector; and selecting all specialized neural networks whose normalized component of the coarse vector exceeds a predetermined threshold.
Claim 26
Dependent Claim
22 words
The character recognition system of claim 24, the network selector further being programmed to inhibit output of the selected specialized neural networks.
Claim 27
Dependent Claim
29 words
The character recognition system of claim 1, including a programmed digital data processing device to implement at least one of the pre-classifier, specialized neural networks, multiplier, and candidate selector.
Claim 28
Dependent Claim
119 words
The character recognition system of claim 1, further including a network selector including a third input to receive the coarse vector from the soft pre-classifier, the coarse vector including multiple components each estimating a probability that the input character belongs to a different one of the predefined groups of recognized target characters, a fourth input to receive external information associating the input character with one or more specialized neural networks, and circuitry programmed to select one or more specialized neural networks by performing steps comprising: selecting all specialized neural networks identified by the external information; normalizing the components of the coarse vector; and selecting all specialized neural networks whose normalized component of the coarse vector exceeds a predetermined threshold.
Claim 29
Independent Claim
142 words
A method for character recognition, comprising the steps of: responsive to a representation of an input character, providing a first output signal comprising a coarse vector estimating probabilities of the input character belonging to each one of multiple groups of characters; responsive to the input character, providing multiple second output signals each corresponding to a particular group of characters and comprising a fine vector estimating probabilities of the input character representing each one of multiple characters in the particular group of characters; weighting each second output signal in proportion to the first output signal to provide multiplier output lists each comprising an estimate of individual probabilities that the input character corresponds to each of the characters of a different specialized neural network; and assembling and ranking contents of the multiplier output lists to provide individual ranked estimates according to a selected criteria.
Claim 30
Independent Claim
142 words
A method for character recognition, comprising the steps of: responsive to a representation of an input character, providing a first output signal comprising a coarse vector estimating probabilities of the input character belonging to each one of multiple groups of characters; responsive to the input character, providing multiple second output signals each corresponding to a particular group of characters and comprising a fine vector estimating probabilities of the input character representing each one of multiple characters in the particular group of characters; weighting each second output signal in proportion to the first output signal to provide multiplier output lists each comprising an estimate of individual probabilities that the input character corresponds to each of the characters of a different specialized neural network; and assembling and ranking contents of the multiplier output lists to provide individual ranked estimates according to a selected criteria.
Claim 31
Dependent Claim
13 words
The method of claim 28, the selected criteria comprising all individual ranked estimates.
Claim 32
Dependent Claim
17 words
The method of claim 28, the selected criteria comprising all individual ranked estimates exceeding a predetermined threshold.
Claim 33
Dependent Claim
40 words
The method of claim 28, the step of providing the first output signal comprising the steps of employing a feedforward neural network to perform probability calculations to identify a group of characters to which the input character most likely belongs.
Claim 34
Dependent Claim
52 words
The method of claim 28, the step of providing multiple second output signals comprising the steps of employing multiple feedforward neural networks each corresponding to a different group of characters, to perform probability calculations to identify a character in the corresponding group of characters that the input character most likely corresponds to.
Claim 35
Dependent Claim
16 words
The method of claim 28, each of the multiple groups including a non-overlapping set of characters.
Claim 36
Dependent Claim
17 words
The method of claim 28, at least two of the multiple groups including characters of overlapping sets.
Claim 37
Dependent Claim
33 words
The method of claim 33, the multiple groups including a first group of upper-case characters, a second group of lower-case characters, a third group of digits, and a fourth group of special symbols.
Claim 38
Dependent Claim
76 words
The method of claim 28, the step of providing the first output signal comprising the steps of receiving external information associating the input character with one or more specialized neural networks, receiving the representation of the input character, and responsive to the first output signal and the external information selecting all specialized neural networks identified by the external information and all specialized neural networks whose component of the first output signal exceeds a first predetermined threshold.
Claim 39
Dependent Claim
17 words
The method of claim 36, further comprising a step of deactivating the specialized neural networks not selected.
Claim 40
Dependent Claim
70 words
The method of claim 28, the weighting step comprising the steps of: responsive to the first output signal, selecting second output signals having coarse estimates greater than a predetermined threshold; and weighting each selected second output signal in proportion to the first output signal to provide multiplier output lists each comprising an estimate of individual probabilities that the input character corresponds to characters of a different selected specialized neural network.
Claim 41
Dependent Claim
74 words
The method of claim 28, the step of providing multiple second output signals comprising the steps of, responsive to the first output signal and the input character, providing multiple second output signals each corresponding to a particular group of characters having coarse estimates greater than a predetermined threshold, each second output signal comprising a fine vector estimating probabilities of the input character representing each one of multiple characters in the particular group of characters.
Claim 42
Dependent Claim
90 words
The method of claim 28, the step of providing a first output signal comprising the steps of: employing a transducer to receive a tangible medium bearing the input character and create machine-readable matrix representative of the input character; utilizing a feature extractor responsive to the matrix to generate a feature vector containing representations of selected characteristics of the input character; and responsive to the feature vector, providing a first output signal comprising a coarse vector estimating probabilities of the input character belonging to each one of multiple groups of characters.
Claim 43
Dependent Claim
32 words
The method of claim 40, the employing step being performed by operating a digitizing optical scanner to optically scan the tangible medium and create a machine-readable matrix representative of the input character.
Claim 44
Dependent Claim
15 words
The method of claim 40, the selected characteristics including contour directions of the input character.
Claim 45
Dependent Claim
16 words
The method of claim 40, the selected characteristics including bending point features of the input character.
References Cited (Backward Citations)
US4180779A
Cited
US4958377A
Cited
US5020117A
Cited
US5195169A
Cited
US5237661A
Cited
US5150424A
Cited
US5212767A
Cited
US5327522A
Cited
US5214717A
Cited
US5303311A
Cited
US5257342A
Cited
US5239619A
Cited
US5052043A
Cited
US5166539A
Cited
US5239594A
Cited
US5248873A
Cited
US5361379A
Cited
US5596657A
Cited
US5276771A
Cited
EP0555023A2
Cited
US5361311A
Cited
US5544264A
Cited
USPTO Annuity & Maintenance Windows
| Annuity Stage | Window Opens | Payment Deadline | Surcharge Period Until | Estimated Fee | Status |
|---|---|---|---|---|---|
| 3.5 Years (Year 4) | 2001-11-09 | 2002-05-11 | 2002-11-10 | $2000.00 | Window Closed / Lapsed |
| 7.5 Years (Year 8) | 2005-11-09 | 2006-05-11 | 2006-11-10 | $3760.00 | Window Closed / Lapsed |
| 11.5 Years (Year 12) | 2009-11-09 | 2010-05-11 | 2010-11-10 | $7700.00 | Window Closed / Lapsed |