JPH041870A - Method and device for recognizing pattern - Google Patents

Abstract

PURPOSE:To improve reliability for the result of recognition by comparing assurance with a threshold value to define an output result as the result of recognition when the assurance is larger than the threshold value and to reject the result of recognition vice versa. CONSTITUTION:An output signal corresponding to each category is obtained from a pattern to be inputted or the pattern feature value with use of a multi- layer type neural circuit network. Then the maximum category is defined as the result of recognition. Furthermore the assurance R is calculated from an entire vector of the output value. When the assurance R is larger than the threshold value, the category having the maximum output value is selected as the result of recognition. Meanwhile a rejection signal is outputted when the R is smaller than the threshold value. Thus the uncertain result of recognition can be rejected and the reliability of the result of recognition is improved.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention provides a pattern recognition device using a multi-layered neural network model, in which not only the recognition result but also the recognition result is calculated according to the certainty factor calculated from the output signal of the neural network. The present invention relates to a pattern recognition method and device that outputs a reject signal.

[Conventional technology]

Conventionally, in pattern recognition using multilayer neural networks, patterns and feature quantities are input, and the literature (“A Neu-ral
Network Digit Recognition
” by D, J, Burr: JEERIntern
ational Conference on Sys.
tem, Manand Cybernetics:
1986. The output signal is obtained as a vector quantity in which each element expresses the degree of similarity to a category, as shown in (pp, pp. 1621-1625). Often, the elements of the output signal are expressed as numbers between 0 and 1 or between -1 and 1. The recognition result is the category represented by the element that takes the maximum value among the elements of the vector quantity.

In the pattern vlWA device, a pattern that is not a recognition target may be input, or a pattern that cannot be recognized as one of two categories may be input. In such a case, it is better to reject the recognition result as unrecognizable than to forcefully obtain the recognition result, thereby increasing the credibility of the recognition result.

Therefore, in conventional pattern recognition devices, there is a method in which the degree of matching of patterns is evaluated by the degree of similarity between an input pattern and a standard pattern for each category, and the category with the highest degree of similarity is determined as a recognition result. In such a method, a candidate is rejected if the similarity of the first candidate does not reach a certain value or more, or it is rejected if the difference in similarity between the first candidate and the second candidate does not exceed a certain value, etc. A method is being taken.

[Problem to be solved by the invention]

In pattern recognition methods using neural networks, the output signal is a vector value whose elements are values with upper and lower limits, but this value is different from the similarity seen in conventional pattern recognition methods; When using the above rejection method, it is difficult to set a threshold to reduce both recognition errors and rejections, and many patterns that are originally correctly recognized end up being rejected.

In the rejection method in the conventional pattern recognition method as described above, only the value of the first candidate or the values up to the second candidate are used.

This results in many rejections when classifying multiple very similar patterns. Further, with such a method, it is not possible to uniformly evaluate the values such as similarity of two or more candidates and reject them.

An object of the present invention is to obtain the confidence level of a recognition result from a vector value that is an output signal of a neural network, compare that value with a threshold value, and output a recognition result or output a reject signal. The object of the present invention is to provide a pattern recognition method and device that can perform the following steps.

[Means to solve the problem]

The first invention is a pattern recognition method using a multilayer neural network, in which an input pattern or a feature extracted from the input pattern is input and converted into an output vector representing a category by the multilayer neural network. And from that output vector, 1
Calculate the confidence such that the confidence is maximum when only one element is near the upper bound of the range and the other elements are near the lower bound of the range, and compare that confidence with a threshold The second invention is characterized in that the output result is recognized as a recognition result when the confidence level is larger than the threshold value, and is rejected when the confidence level is smaller than the threshold value. In a pattern recognition device using a network, means for realizing a neural network that converts input data into an output vector signal representing a category, a means for detecting a category having a maximum value among the output vector signals, and an output vector. means for calculating the confidence such that the confidence is maximized when only one element of the signal is near the upper limit of the range and the other elements are near the lower limit of the range; It is characterized by comprising means for determining that the output of the maximum value detection unit is the recognition result when it is larger than the threshold value, and as a rejection when the confidence level is less than the threshold value, and outputting the recognition result or a rejection signal. shall be.

A third invention is a pattern recognition device using a multilayered neural network, which includes means for realizing a neural network that converts input data into an output vector signal representing a category, and a means for sorting the output signals in descending order. and means for storing the category of the maximum value of the sorted results; and among the fixed number of high-order elements of the sorted output vector signal, only the first-ranked element is near the upper limit of the range; Other elements include means for calculating confidence so that the confidence is maximized when it is near the lower limit of the value range, and recognition based on the contents of the maximum value category storage unit when the confidence is greater than a threshold value. The method is characterized in that it includes a means for determining that the recognition result is rejected when the confidence level is less than a threshold value, and outputs a recognition result or a rejection signal.

[Effect]

The pattern recognition method, which is the first invention, uses a multilayer neural network to obtain an output signal corresponding to each category from the input pattern or feature amount of the pattern, and selects the category with the maximum value as the recognition result. Established as Furthermore, the confidence level is calculated from the entire vector of output values using the following formula. Here, the number of categories to be recognized is N,
The output vector from the neural network is Pi (1≦i≦N
, 0≦Pi≦1), α, and β are coefficients, the confidence level R is obtained as follows. This value takes the maximum value when only one value is close to the upper limit l and all other values are close to the lower limit 0, and if the element Pi with the maximum value is away from 1 or R becomes smaller when the element approaches its maximum value. If the value R is larger than the threshold value, the category with the maximum output value is selected as the recognition result, and if R is smaller than the threshold value, a reject signal is output, so that uncertain recognition can be avoided. Results can be rejected.

The pattern recognition device, which is the second invention, includes means for performing an operation according to formula (1) from all element values of the output vector of the neural network, and comparing the obtained confidence value with a threshold value. and means for determining whether to output the recognition result or reject the recognition result, thereby rejecting an uncertain recognition result.

The pattern recognition device, which is the third invention, includes means for arranging the output vector values of the neural network in descending order and calculating the confidence level from a certain number of higher-order values according to equation (1), and By providing a means for determining whether to output or reject the recognition result by comparing the value with a threshold value, uncertainty is calculated from the higher category values of the output value that may affect the recognition result. Reject recognition results.

〔Example〕

DESCRIPTION OF THE PREFERRED EMBODIMENTS A character recognition method and apparatus according to an embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows a flowchart illustrating the pattern recognition method of the present invention. A pattern to be recognized is input (step S101), and then features are extracted from the pattern (step S102) and converted into vector values. The extracted features may be of any type, and in the case of character recognition, they may include shading features using a Gaussian filter, features extracted from directional curvature, and the like.

Furthermore, when feature extraction is not performed, it is also possible to treat the input pattern itself as a feature vector value. - As an example, a method of character pattern feature extraction using a Gaussian filter will be explained with reference to FIG.

In the filtering unit 203, a Gaussian filter 202 given in a 5×5 matrix is convolved with the character pattern 201 input as a black and white binary image. The input value image is converted into a binary image B (i, j) of 32 pixels x 42 pixels.
, 5×5 filter as F(k,l), and the resulting multivalued image as G
(i, j), calculation is performed according to the formula G(i, j) (2). The resulting G(i, j
) is thinned out to 99 pieces of data (9×11) by the resampling unit 204, and becomes a grayscale feature. Fig. 2 shading @2
06 is a display of this data, and the value of each element is indicated by the size of the circle.

Next, the feature vector value is converted into an output value using a multilayer neural network (step S103). The calculation method will be explained using FIG. In FIG. 3, a calculation method for a neural network with a three-layer structure from the 0th layer to the 2nd layer will be explained.
It can be easily extended to more than just layers.

In FIG. 3, the feature vector value extracted in step 3102 of FIG. 1 is input to the multilayer neural network and is set as the vector value of the 0th layer (step 5301). Next, the first layer matrix product is calculated from the zero layer vector values according to the following equation (3) (step S302). The vector value of the 0th layer is o=' (1≦1≦N o ), and the first layer weight matrix value is W==' (1≦i≦N0.1≦j≦
N, ), the vector value of the calculation result is o='(1≦j≦N
, ), then 0.1 = Σ (o4'Xw point Jl) 10 01 rin (3). However, here, θ, 1 is a bias value. Next, use the following equation (4) for each of o and '(1≦j≦N,)
The S function is calculated according to (step 5303).

o'J' = 1/2 (1+tanh (oj')
) (4) Next, the second layer weight matrix is w,
," (1≦j≦N,, t≦≦NZ), and the bias value is θ-, okz=Σ (Oj,' X Wjk") + θ-(5)
A calculation is performed (step S304). Furthermore, o
, 2 (1≦and≦N z ), the S function is calculated according to each equation (6) (step S305).

o'-= 1/2 (1+tanh (oil")
(6) Finally, the second layer data is output as an output signal (step S306).

The weight matrices of the first and second layers used here are described in the literature (“An Introduction to
Cospu-ting with Neural Ne
ts″by R,P,Lippmann: IEEE.

ASSP: April 1987. pp, 4-22
) is determined in advance through learning.

Next, from the output data o'h'' (1≦to≦N,),
Find the maximum value. The number is set as the maximum value category (step S104).

Furthermore, the certainty factor R is calculated from the output data according to equation (1) (step S105). Here, the output data is an N-dimensional vector, and the vector element Pi (1≦i≦N
) take values from O to 1, respectively. When using this pattern recognition method for number recognition,
Since 10 characters are to be recognized, the value of N is set to 10, and when alphanumeric characters are to be recognized, 26 uppercase letters and 26 lowercase letters.

In order to recognize 10 numeric characters, the value of N is set to 62. Further, the coefficients α and β in equation (1) were each set to l.

This is not an essential problem, and other values are also possible. Also, in equation (1), each element of the output data is O
Although we assumed a value from -1 to 1, it can be easily extended to the case where each element of the output data takes a value from a to b in -boat fishing, as shown in equation (7) below. It can be expressed as

R=-a Σ (P 1-a) (b-P i)-β
(Σ ((P 1-a)/ (ba-a) )・・
・(7) Next, compare the calculated confidence R and the threshold T and calculate (
Step S106): If R is greater than or equal to the threshold, proceed to step 5107, and output the maximum category number obtained in step 5104; if R is less than the threshold, proceed to step 5108, which is a reject signal. Outputs 0. Although the value of the threshold T is assumed to be 0.1 for convenience, this is not essential and may be any value. The above steps complete the recognition process for one pattern.

Next, an embodiment of the pattern recognition device of the second invention will be described with reference to FIGS. 4, 5, and 6.

FIG. 4 shows an overall block diagram of the character recognition device. This character recognition device includes a feature extraction section 402 and a recognition section 403.
and a code generation unit 404.

A character pattern to be recognized is input from a terminal 401 and converted into a feature vector by a feature extraction unit 402. This feature extraction unit is not an essential problem in the present invention, and the direction feature, contour curvature direction feature, etc. used in conventional character recognition devices may be used. The feature data extracted by the feature extraction unit 402 and converted into a vector is sent to the recognition unit 403, where it undergoes recognition processing and is converted into an output signal. This output signal is the category number of the recognition result from 1 to N, or the number O representing reject. N is the number of categories to be recognized. This output signal is sent to a code generation section 404, and the recognition result is converted into a character code such as a JIS code and outputted from a terminal 405.

An example of the recognition unit 403 will be explained using FIG. 5.

This recognition unit includes a multilayer neural network calculation unit 501, an output data storage unit 502, a maximum category detection unit 503, a certainty calculation unit 504, and a recognition result/rejection determination unit 50.
It consists of 5.

The feature data obtained by the feature extraction unit 402 is transferred to the multilayer neural network calculation unit 501, where it is subjected to recognition processing and converted into an output vector. Each element of the output vector here represents the score of the category to be recognized, and the number of dimensions N of the output vector is equal to the number of categories to be recognized.

An example of the multilayer neural network calculation unit 501 will be explained using FIG. 6. In this example, a three-layer model is used, but
This is not an essential problem, and it can be easily extended even when using a two-layer model, a model with four or more layers, or a neural circuit model with other combination types.

This multilayer neural network calculation unit includes a matrix product calculation unit 60
1, the S function section 602, the intermediate layer storage section 603, the matrix product operation section 604, the S-function section 605, and the first
It is composed of a layer weight matrix storage section 606 and a second layer weight matrix storage section 607.

The vector value input from the feature extraction unit 402 is expressed as oi'(
1≦i≦N,), and the first layer weight matrix storage unit 60
Wij(1≦i≦N6,
1≦j≦N1), and the vector value held in the intermediate layer storage unit 603 is '=' (1≦j≦N,), then the matrix product calculation unit 601 performs calculation according to equation (3).

However, here, θJ1 is a bias value and is stored in the first layer weight matrix storage unit 606. Result oJ+
(1≦j≦N,) are each expressed by the formula (
4), and store it in the intermediate storage unit 603.

If the weight matrix stored in the second layer weight matrix storage section 607 is wJm'' (1≦j≦N, 1≦≦N2), and the bias value stored in this storage section 607 is θ-, matrix product operation is performed. The unit 604 performs calculations to realize equation (5).
In the function section 605, o' is calculated according to equation (6).
1 is calculated, outputted as an output signal, and stored in the output data storage section 502.

It is assumed that the weight matrix used here has been determined in advance through learning. Also, as the S function, equation (4) 1
When formula (6) is used, the output value of the function is a decimal value between 0 and 1.

The vector pattern read from the output data storage unit 502 is sent to the maximum category detection unit 503 to find the element with the maximum value, and the category number represented by that element is sent to the recognition result/rejection determination unit 505.

The vector data read from the output data storage unit 502 is also sent to the certainty calculation unit 504.

The reliability calculation unit 504 calculates the reliability R by calculation according to equation (1).

The value R obtained by the certainty calculation unit 504 is sent to the recognition result rejection determination unit 505 and compared with a threshold T. If it is larger than the threshold T, the category number of the recognition result is changed to the code generation unit 404. and the confidence R is the threshold value T
If it is smaller than , a reject signal is transferred to code generation section 404 .

Here, the category number is a numerical value from 1 to N (N is the number of categories), and the reject signal is a numerical value 0.

Next, a character recognition device which is an embodiment of the third invention will be described.

The overall structure of the character recognition device is according to FIG. 3, and the structure of the recognition section is different from that of the second invention. Recognition unit 700 in this embodiment
is shown in Figure 7. The recognition unit 700 includes a multilayer neural network calculation unit 701, an output data storage unit 702, a data sorting unit 703, a maximum category detection unit 704, a certainty calculation unit 705, and a recognition result/rejection determination unit 706. It consists of

The input pattern is converted into a feature vector by the feature extraction unit 402 and input to the multilayer neural network calculation unit 701 for recognition. The configuration of the multilayer neural network calculation section 701 is the same as the configuration of the multilayer neural network calculation section 501 shown in FIG. 6 described above. The N-dimensional output data therefrom is stored in the output data storage unit 702. Here, N represents the number of categories to be recognized.

The data read from the output data storage section 702 is rearranged in the data sorting section 703 in descending order of value. The category number with the maximum value is stored in maximum category register 704.

The M values output from the data sorting unit 703 are input to the certainty calculation unit 705 in descending order of value, and the calculation according to (8) is executed. Here, M is a smaller number than N; in the embodiment, M is set to 5 for handwritten numeric recognition, and M is set to 10 when alphanumeric characters are to be recognized.

The value of M is not essential and may be any other value. Pi is the output value of the data sorting section 703, and the smaller i is, the larger Pi takes. α and β were each set to 1.

The confidence value calculated by the confidence calculation unit 705 is sent to the recognition result/rejection determination unit 706, and if it is larger than a preset threshold T, the confidence value calculated in the maximum category register is determined as the recognition result. The category number or, if not, the input signal is transferred to the code generation section 404, where the character code of the recognition result is generated and output. Here, the threshold value T is set to 0.1, but other values can also be used.

〔Effect of the invention〕

As is clear from the above description, in the pattern recognition method and apparatus of the present invention, the score of each category output from the recognition unit in the neural network has a range of possible values (
a, b), if only one element takes the value a and all other elements take the value b, then the category that clearly takes the value a is output as the correct recognition result. be done. When the maximum value among all categories is far from the upper limit a of the range, or when there are multiple categories whose values are close to the category that takes the maximum value, the confidence level decreases and false recognition can be avoided. This recognition result is rejected as a recognition result with a high probability of occurrence.

If the recognition result is rejected, the reliability of the recognition result can be increased by starting another recognition method or manually correcting the recognition result.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of the pattern recognition method of the first invention of the present application, FIG. 2 is a diagram showing an example when a Gaussian filter is used as the feature extraction section, and FIG. 3 is a diagram showing a three-layer neural structure. 4 is a block diagram of a character recognition device; FIG. 5 is a block diagram showing an example of a recognition unit; and FIG. 6 is an example of a multilayer neural network calculation unit. Block Diagram FIG. 7 is a block diagram showing another example of the recognition unit. 402 ... Feature extraction section 403 ... Recognition section 404 ... Code generation section 501, 701 ... Multilayer neural network calculation section 502, 7
02...Output data storage unit 503...Maximum category detection unit 504, 705...Confidence calculation unit 505, 706...Recognition result/rejection determination unit 70
3 ... Data sorting section 704 ... Maximum category register agent Patent attorney Yoshi Iwasa Kozen 2 Figures Horse 1 Figure Thread 3 Figure vi) Figure 6 Figure 5 Thread 7

Claims (3)

[Claims] (1) In a pattern recognition method using a multilayered neural network, input patterns or features extracted from the input patterns are input, are converted into output vectors representing categories by the multilayered neural network, and are converted into output vectors representing categories. From the output vector, 1
Calculate the confidence such that the confidence is maximum when only one element is near the upper bound of the range and the other elements are near the lower bound of the range, and compare that confidence with a threshold A pattern recognition method characterized in that the output result is recognized as a recognition result when the confidence level is larger than a threshold value, and is rejected when the confidence level is smaller than the threshold value. (2) In a pattern recognition device using a multilayered neural network, means for realizing a neural network that converts input data into an output vector signal representing a category, and a method for determining a category having the maximum value among the output vector signals. means for detecting the output vector signal; and a means for determining that the output of the maximum value detection unit is the recognition result when the confidence level is greater than the threshold value, and as a rejection when the confidence level is less than the threshold value, and the recognition result or A pattern recognition device characterized by outputting a reject signal. (3) In a pattern recognition device using a multilayered neural network, means for realizing a neural network that converts input data into an output vector signal representing a category, a means for sorting the output signals in descending order, and sorting. means for storing the category of the maximum value of the sorted result, and among a fixed number of high-order elements of the sorted output vector signal, only the first-ranked element is near the upper limit of the range, and the other elements means for calculating the confidence so that the confidence is maximized when the confidence is near the lower limit of the value range, and when the confidence is greater than the threshold value, the content of the maximum value category storage unit is used as the recognition result; What is claimed is: 1. A pattern recognition device comprising means for determining a pattern as a reject when the degree of certainty is less than a threshold, and outputting a recognition result or a rejection signal.