### Citations

6607 |
Neural Networks for Pattern Recognition
- Bishop
- 1996
(Show Context)
Citation Context ...er, MLPs are efficient tools for learning large databases [21]. Therefore, all classifiers presented in this work are MLPs trained with the gradient descent applied to a sum-of-squares error function =-=[2]-=-. The transfer function employed is the familiar sigmoid function. In order to monitor the generalization performance during learning and terminate the algorithm when there is no longer an improvement... |

316 |
Neural network classifiers estimate Bayesian a posteriori probabilities,
- Richard, Lippmann
- 1991
(Show Context)
Citation Context ... remains between zero and one because of the sigmoidal functions used, but the criterion used for training did not require the outputs to sum to one. Nevertheless, as shown by Richard and Lippmann in =-=[29]-=- the summed outputs of the MLP network are always close to one. As such, normalization techniques proposed to ensure that the outputs of an MLP network are true probabilities such as softmax [4] may b... |

196 |
The Viterbi algorithm.
- Jr
- 1973
(Show Context)
Citation Context ...ypothesis, which is the correct one, got a higher final probability. 4.4 Hypothesis Generation The generation of k best hypotheses of an amount is carried out by means of a Modified Viterbi Algorithm =-=[10]-=- which ensures the calculation of the k best paths of the segmentation-recognition graph. Fig. 12 shows the three best hypotheses of segmentation-recognition. Basically, the algorithm computes the ði ... |

122 |
Training stochastic model recognition algorithms as networks can lead to maximum mutual information estimation of parameters.
- Bridle
- 1989
(Show Context)
Citation Context ...n in [29] the summed outputs of the MLP network are always close to one. As such, normalization techniques proposed to ensure that the outputs of an MLP network are true probabilities such as softmax =-=[4]-=- may be unnecessary. This is further supported by results of experiments performed by Bourlard and Morgan [3], which demonstrated that the sum of the outputs of MLP networks is near one for large phon... |

102 |
Document image decoding using Markov source models,”
- Kopec, Chou
- 1994
(Show Context)
Citation Context ... NUMERICAL STRINGS: A RECOGNITION AND VERIFICATION STRATEGY 1439 result produced by the recognizer. The integration of all modules is done through a probabilistic model inspired by information theory =-=[20]-=-. The focus of this work is to show how the verification modules can improve the recognition rate and reliability of the system. We introduce a new scheme of verification where two verifiers are consi... |

87 |
Pattern classification using neural networks.
- Lippmann
- 1989
(Show Context)
Citation Context ...ers in the variant considered, and P ðM;VÞ is the joint probability of amounts and their variant. 3.2 Neural Classifiers Although many types of neural networks can be used for classification purposes =-=[25]-=-, we opted for a Multilayer Perceptron (MLP) which is the most widely studied and used neural network classifier. Moreover, MLPs are efficient tools for learning large databases [21]. Therefore, all c... |

55 | Reject Option with Multiple Thresholds.
- Fumera, Roli, et al.
- 2000
(Show Context)
Citation Context ...stic model used, a 10-digit string usually supplies a global probability (Pglobal) smaller than a 2-digit string. Among the different strategies that we have tested, the one proposed by Fumera et al. =-=[11]-=- provided the better error-reject tradeoff for our system. Basically, this technique suggests the use of multiple reject thresholds for the different data classes (T0; ...;Tn) to obtain the optimal de... |

52 |
Off-Line, Handwritten Numeral Recognition by Perturbation Method
- Ha, Bunke
- 1997
(Show Context)
Citation Context ...ognition approach where an explicit segmentation algorithm determines the cut regions and provides a multiple spatial representation. Contrary to the systems that process just isolated numerals [17], =-=[13]-=-, our system has to solve a crucial problem: distinguishing, at the recognition stage, a sequence corresponding to an intercharacter segmentation from another relative to an intracharacter segmentatio... |

40 |
Recognition of isolated and simply connected handwritten numerals”,
- SHRIDHAR, BADRELDIN
- 1986
(Show Context)
Citation Context ...o: tpami@computer.org, and reference IEEECS Log Number 114038. commas, periods, strokes, currency names, etc. Strategies for digit string recognition can be divided into segmentation-then-recognition =-=[30]-=- and segmentation-based recognition [26]. In the first approach, the segmentation module provides a single sequence hypothesis where each subsequence should contain an isolated character which is subm... |

39 |
Integration of Structural and Statistical Information for Unconstrained Handwritten Numeral Recognition”.
- Cai, Liu
- 1998
(Show Context)
Citation Context ...ed recognition approach where an explicit segmentation algorithm determines the cut regions and provides a multiple spatial representation. Contrary to the systems that process just isolated numerals =-=[17]-=-, [13], our system has to solve a crucial problem: distinguishing, at the recognition stage, a sequence corresponding to an intercharacter segmentation from another relative to an intracharacter segme... |

35 |
Automatic Bankcheck Processing: a New Engineered System. In: Impedovo,
- Dimauro, Impedovo, et al.
- 1997
(Show Context)
Citation Context ...us works that use some kind of syntactic analysis. Knerr et al. in [19] consider the segments below the baseline in order to obtain the final interpretation of the numerical amount. Dimauro et al. in =-=[9]-=- use nonnumerical symbols such as “#” in order to identify the beginning and the end of the numerical amount. Heutte et al. in [16] present a postprocessing module for French bank checks which takes i... |

34 |
An automatic reading system for handwritten numeral amounts on french checks. In:
- Lethelier, Leroux, et al.
- 1995
(Show Context)
Citation Context ...es used to infer a plausible interpretation. The probabilistic model that we are using has been applied to speech recognition [24], handwritten word recognition [6], and handwritten digit recognition =-=[23]-=-. Such a model estimates the most probable interpretation of the written amount M (noted bM). Its input corresponds to an image I after preprocessing. In a probabilistic framework, bM is given by the ... |

31 | Off-line Handwritten Numeral String Recognition By Combining Segmentation-based and Segmentation-free Methods,”
- Ha, Zimmermann, et al.
- 1998
(Show Context)
Citation Context ...roup a character composed of several components by detecting potential parts and grouping each of them to its nearest neighbor. We have adopted a strategy similar to the one presented by Ha et al. in =-=[14]-=-. 4.2 Segmentation The segmentation module that we have used in this system is based on the relationship of two complementary sets of structural features, namely, contour/profile and skeletal points. ... |

23 |
Merging Multilayer Perceptrons and Hidden Markov Models: Some Erperiments in Continuous Speech Recognition,
- Bourlard, Morgan
- 1989
(Show Context)
Citation Context ...oposed to ensure that the outputs of an MLP network are true probabilities such as softmax [4] may be unnecessary. This is further supported by results of experiments performed by Bourlard and Morgan =-=[3]-=-, which demonstrated that the sum of the outputs of MLP networks is near one for large phoneme-classification speech-recognition problems. Let S be a pattern space which consists of A mutually exclusi... |

22 |
Segmentation of Single- or Multiple-Touching Handwritten Numeral String Using Background and Foreground Analysis”,
- Chen, Wang
- 2000
(Show Context)
Citation Context ...nition module has to discriminate various configurations such as fragments, isolated characters, and connected characters. In this strategy, segmentation can be explicit when based on cut rules [27], =-=[7]-=- or implicit when each pixel column is a potential cut location [5], [22]. In this paper, we present a modular recognition system for handwritten numerical strings. This system takes a segmentation-ba... |

17 |
Integrated segmentation and recognition of handwritten numerals with cascade neural network,”
- Lee, Kim
- 1999
(Show Context)
Citation Context ...ts, isolated characters, and connected characters. In this strategy, segmentation can be explicit when based on cut rules [27], [7] or implicit when each pixel column is a potential cut location [5], =-=[22]-=-. In this paper, we present a modular recognition system for handwritten numerical strings. This system takes a segmentation-based recognition approach where an explicit segmentation algorithm determi... |

14 |
H.: Automated reading of cheque amounts
- Kaufmann, Bunke
- 2000
(Show Context)
Citation Context ...een two digits, (c) natural fragmentation, (d) fragmentation caused by noise, and (e) isolated digits classified as undersegmentation by vu. of 10,000 images of French bank checks. Kaufmann and Bunke =-=[18]-=- propose a system for Swiss postal checks that neither considers nonnumerical classes nor the cents portion. The result achieved by this system is 79.3 percent (zerorejection level) and 58.1 percent w... |

14 |
Speech recognition using stochastic explicit-segment moduling
- Leung
- 1991
(Show Context)
Citation Context ...ystem modules in order to allow a sound integration of all knowledge sources used to infer a plausible interpretation. The probabilistic model that we are using has been applied to speech recognition =-=[24]-=-, handwritten word recognition [6], and handwritten digit recognition [23]. Such a model estimates the most probable interpretation of the written amount M (noted bM). Its input corresponds to an imag... |

13 |
Recognizing overlapping hand-printed characters by centered-objects integrated segmentation and recognition
- Matan, Burges
- 1991
(Show Context)
Citation Context ...ECS Log Number 114038. commas, periods, strokes, currency names, etc. Strategies for digit string recognition can be divided into segmentation-then-recognition [30] and segmentation-based recognition =-=[26]-=-. In the first approach, the segmentation module provides a single sequence hypothesis where each subsequence should contain an isolated character which is submitted to the recognizer. This technique ... |

13 | A Modular System to Recognize Numerical Amounts on Brazilian Bank Cheques
- Oliveira, Sabourin, et al.
- 2001
(Show Context)
Citation Context ...low-level verifiers which is more robust and reliable. Initially, we developed a strategy based on an isolated verifier which was responsible for detecting both oversegmentation and undersegmentation =-=[28]-=-. After some experiments, we realized that a more specialized verifier could produce better results than a generic one. In the following sections, we present both verifiers, their respective feature s... |

10 | A two-stage HMM based system for recognizing handwritten numeral strings
- Jr, Sabourin, et al.
- 2001
(Show Context)
Citation Context ...agments, isolated characters, and connected characters. In this strategy, segmentation can be explicit when based on cut rules [27], [7] or implicit when each pixel column is a potential cut location =-=[5]-=-, [22]. In this paper, we present a modular recognition system for handwritten numerical strings. This system takes a segmentation-based recognition approach where an explicit segmentation algorithm d... |

10 | Verification of graphemes using neural networks in an HMM-based on-line Korean handwritting recognition system
- Cho, Kim, et al.
- 2000
(Show Context)
Citation Context ...acters and parts of them. The purpose of a low-level verifier is not to recognize a character, but rather to determine whether a hypothesis generated by the general-purpose recognizer is valid or not =-=[8]-=-. In this work, we propose two low-level verifiers to cope with the oversegmentation and undersegmentation problems. The objective of these verifiers is to validate the general-purpose recognizer hypo... |

9 |
A new segmentation approach for handwritten digits
- Oliveira, Lethelier, et al.
- 2000
(Show Context)
Citation Context ... recognition module has to discriminate various configurations such as fragments, isolated characters, and connected characters. In this strategy, segmentation can be explicit when based on cut rules =-=[27]-=-, [7] or implicit when each pixel column is a potential cut location [5], [22]. In this paper, we present a modular recognition system for handwritten numerical strings. This system takes a segmentati... |

8 |
Backpropagation Applied to Handwritten Zip
- LeCun, Boser, et al.
- 1989
(Show Context)
Citation Context ...fication purposes [25], we opted for a Multilayer Perceptron (MLP) which is the most widely studied and used neural network classifier. Moreover, MLPs are efficient tools for learning large databases =-=[21]-=-. Therefore, all classifiers presented in this work are MLPs trained with the gradient descent applied to a sum-of-squares error function [2]. The transfer function employed is the familiar sigmoid fu... |

7 |
Recogniton enhancement by linear tournament verification
- Takahashi, Griffin
- 1993
(Show Context)
Citation Context ...er is applied after a generalpurpose recognizer and it is designed to “plug and play,” i.e., it is used without knowing the implementation details of the recognition modules. Takahashi and Griffin in =-=[31]-=- define three kinds of verification: absolute verification for each class (Is it a “0”?), one-to-one verification between two categories (Is it a “4” or a “9”?), and verification in clustered, visuall... |

5 | Recognition and verification of touching handwritten numerals
- Zhou, Gan, et al.
- 2000
(Show Context)
Citation Context ...classes considered by the general-purpose recognizer. The goal of the verifiers at this level is to confirm or deny the hypotheses produced by the general-purpose recognizer by recognizing them [31], =-=[32]-=-. We define as low-level verifiers those that deal with metaclasses of the system such as characters and parts of them. The purpose of a low-level verifier is not to recognize a character, but rather ... |

4 |
A system for the off-line recognition of handwritten text
- Bruel
- 1994
(Show Context)
Citation Context ...ound integration of all knowledge sources used to infer a plausible interpretation. The probabilistic model that we are using has been applied to speech recognition [24], handwritten word recognition =-=[6]-=-, and handwritten digit recognition [23]. Such a model estimates the most probable interpretation of the written amount M (noted bM). Its input corresponds to an image I after preprocessing. In a prob... |

1 |
do Brasil, Manual de Normas e Instruções—Circular no 001825
- Central
- 1990
(Show Context)
Citation Context ... hypotheses of recognition, and (c) three best paths. numerical amount. In addition to these rules, the Central Bank of Brazil decided that the cents portion should be present in the numerical amount =-=[1]-=-. In order to deal with such rules we have developed a deterministic automaton which is associated with the term P ðM;VÞ of (10). This automaton is depicted in Fig. 13. Once such an automaton is forme... |