JP3449392B2 - Discriminant function learning method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for learning and determining a discriminant function used for pattern recognition from a learning pattern. For details, refer to many characters such as printed kanji, handwritten kanji, alphanumeric characters, symbols, mathematical expressions, and figures, which are typical examples of patterns.
In the case of targeting a figure category, it can be applied to a method of learning a discriminant function for recognizing an image or the like in which noise such as blurring or crushing is added. Also, it can be applied to a method of recognizing a plurality of categories such as obstacles and illnesses that occur in the system based on the diagnostic results of many items used in disasters, medical care, and the like. It should be noted that the category is a unit to be recognized, and in character recognition, "character type" usually corresponds to this. For example, there are 10 categories from 0 to 9 for numbers, and the first level for Kanji. There are about 3300 categories for kanji.

[0002]

2. Description of the Related Art Traditionally, Kanji OCR (Optical Character)
A character recognition processing device such as r Reader) or a drawing recognition device extracts a feature for recognition in the form of a vector from a character or figure pattern expressed as a binary or multivalued pattern, and creates a standard dictionary created in advance. There is known a method in which a discriminant function such as a similarity or a dissimilarity is obtained from the standard pattern vector of each category in the above, and the most similar character or figure category is used as the recognition result.

[0003] In these methods, it is general to empirically design and design appropriate recognition features and discriminant functions, but the number of categories with the data input by the user such as OCR. Since the appearance frequency of each category and the degree of deformation occurring in each category do not necessarily match, there is a problem that the expected recognition performance cannot be obtained depending on the user.

Further, the feature amount reflecting the character line structure such as the direction of the character line, the connection function, and the position function is widely used as the element of the feature vector in the conventional character recognition, etc. It has been almost impossible to obtain sufficient recognition accuracy because these feature values greatly change in an image with severe background noise. Before the recognition process, it has been attempted to remove such noise by image processing, but the character line itself is considered to be part of the noise and removed, or part of the noise is the character line. As a result, a contaminated character pattern is output, and the feature amount also varies and cannot be recognized sufficiently.

[0005]

As described above, the conventional recognition techniques such as OCR have learned the discriminant function capable of correctly recognizing a character or the like which is erroneously recognized due to faintness, crushing, character background noise or the like. There was a problem that the method of asking was not fully established.

An object of the present invention is not only to solve the above problem of character recognition, but also to recursively learn a discrimination function that is robust against noise from learning data (learning patterns) for various pattern recognition. To provide.

[0007]

In order to achieve the above object, according to the present invention, there is provided a learning pattern based on information of a portion having a "figure" value or a portion having a "ground" value. A value obtained when matching with a reference pattern is used as a feature amount, and a function that determines a feature amount obtained from two patterns of different categories and a feature amount obtained from two patterns of different categories is used as a discriminant function. The main feature is that. Specifically, the "Figure" area, "Figure" area, and "Figure" area
Area and "ground" area, "ground" area and "figure" area, "ground" area
Match each pixel with each combination of
And the number of matched pixels is a, c, b, and e, respectively,
The first-order terms of a, c, b, e and the a, c, b, e of
By terms up to the Nth order (N ≧ 2) of all combinations of values
The features are constructed and obtained from the patterns of the same category.
Features obtained from patterns in different categories from the features
The function that discriminates is the discrimination function.

Further, the present invention is characterized in that a discriminant function obtained by learning by adding a pattern in which noise is artificially added to a learning pattern is used as a discriminant function to be obtained.

[0009]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, as one embodiment of the present invention,
The character pattern will be described with reference to the drawings. Note that other patterns can be similarly realized.

FIG. 1 is a block diagram showing an embodiment of an apparatus configuration to which the method of the present invention is applied. A pattern storage unit 1, a noise adding unit 2, a figure background pattern creating unit 3, a feature quantity counting unit 4, and , The discriminant function learning unit 5.

The pattern storage unit 1 stores image patterns (learning patterns) used for learning the discrimination function. The image pattern consists of n pixels. For example, in the case of a binary image pattern, each pixel is indicated by "0" or "1". Further, for example, in the case of a multi-valued pattern, each pixel has 0 to 1
It is shown by the real value between.

The noise adding section 2 artificially adds noise to the image pattern. For example, in the case of a binary image, the pixel value “0” is converted to “1” in the case of dirt noise, and the pixel value “1” is converted to “0” in the case of blur noise. Further, for example, in the case of multi-valued pattern, the pixel value is increased in the case of dirt noise, and the pixel value is decreased in the case of blur noise. The amount of noise added is indicated by the number of converted pixels. In the case of a multi-valued pattern, the amount of noise may be calculated using the amount of change in pixel value.

The figure background pattern creating section 3 divides each pixel of the image pattern into a "picture" area and a "ground" area. For example, in the case of a binary image pattern, a region having a pixel value of "1" is a "figure" region and a region having a pixel value of "0" is a "ground" region. Further, for example, in the case of a multi-valued image pattern, the “figure” region and the “ground” region are divided by threshold processing. Note that the figure / ground may be determined manually. A figure pattern and a ground pattern are created from the areas thus divided. This figure pattern and ground pattern are image patterns with noise added (learning patterns).
And each of the image patterns (reference patterns) to which noise is not added.

The feature quantity counting unit 4 outputs the figure pattern and the ground pattern of the image pattern (learning pattern) of the noise-added n pixel and the figure pattern and the ground pattern of the reference pattern from the figure background pattern creating unit 3. By inputting, collating these patterns and combining the values obtained at the time of collation,
Obtain a plurality of feature quantities.

The discriminant function learning unit 5 generates a discriminant function based on the feature amount input from the feature amount counting unit 4 by using a known discriminant analysis method or a rule finding learning method by a computer.

Next, the learning method of the discriminant function according to the present invention in the embodiment shown in FIG. 1 will be described. FIG. 2 is an example of the processing flowchart.

The pattern storage unit 1 stores image patterns of n pixels in each category. The pixel value of the image pattern is, for example, “0” or “1” in the case of a binary image pattern, and a real value between 0 and 1 in the case of a multivalued pattern. In the present embodiment, the image patterns stored in the pattern storage unit 1 are used as both learning patterns and reference patterns.

First, an image pattern (learning pattern) of each category is read from the pattern storage unit 1 and input to the noise adding unit 2 (step 201). The noise adding unit 2 adds the amount of noise designated by the operator to this image pattern (step 202). For example, in the case of a binary image pattern, "0" is converted into "1" in the case of dirt noise, and "1" is converted into "0" in the case of blur noise. The position of the pixel to be converted is set at random, for example. The amount of noise can be quantified, for example, by the number of changed pixels. In the case of multi-valued pattern, for example, the pixel value is increased in the case of dirt noise and the pixel value is decreased in the case of blur noise.
The image pattern (learning pattern) with this noise added is Xno.
Let ise = (x ₁ , x ₂ , ..., X _i , ..., X _n ).

Image pattern Xnoise added with the noise
= (X ₁ , x ₂ , ..., X _i , ..., X _n ) is input to the figure-ground pattern creating unit 3. In figure-ground pattern creation section 3, with respect to the image pattern Xnoise, it makes a determination "Figure", "land", FIG pattern _{Xnoise_fig = (xf 1, xf 2} , ..., xf i, ... x
f _n ), and the ground pattern Xnoise_back = (xb ₁ , x
b ₂ , ..., Xb _i , ... Xb _n ) are created (step 20)
3). For example, in the case of a binary image pattern, an area having a pixel value of "1" is determined as "figure", and an area having a pixel value of "0" is determined as "ground". Figure pattern Xnoise_fi
For g, the drawing area of the image pattern is set to "1" and the ground area is set to "0". The ground pattern Xnoise_back is created with the ground area of the image pattern set to “1” and the drawing area set to “0”. Further, in the case of a multi-valued pattern, the "figure" and "ground" are determined using a certain threshold value, and a figure pattern and a ground pattern are created. In that case, the multivalued level of the original multivalued pattern is used as it is for the drawing area of the drawing pattern, or the multivalued level is normalized, and the background area is set to, for example, "0". Also, for the ground area of the ground pattern, the multi-valued level of the original multi-valued pattern is used as it is, or the multi-valued level is normalized,
The drawing area is set to "0", for example. In addition, "figure""ground"
The determination of may be performed manually.

Next, the reference pattern of each category is read from the pattern storage unit 1 and input to the figure-ground pattern creating unit 3 (step 204). Here, the reference pattern is T =
(T ₁ , t ₂ , ..., T _i , ..., T _n ). The reference pattern T is an image pattern with zero noise amount. In figure-ground pattern creation section 3, with respect to the reference pattern T, it makes a determination "Figure", "land", FIG pattern _{T_fig = (tf 1, tf 2} , ..., tf i, ...
tf _n ) and the ground pattern T_back = (tb ₁ , tb ₂ , ...,
tb _i, ..., to create a tb _n) (step 205).

A noise-added image pattern (learning pattern) Xno of each category created by the figure background pattern creating section 3
ise figure pattern Xnoise_fig and ground pattern Xnoise_bac
The figure pattern T_fig and the ground pattern T_back of the reference pattern T with k and the noise amount of 0 are input to the feature amount counting unit 4. In the feature quantity counting unit 4, the image pattern X to which these noises are added is added.
noise pattern Xnoise_fig = (xf ₁ , xf ₂ , ..., x
f _i , ... Xf _n ) and the ground pattern Xnoise_back = (x
_{b 1, xb 2, ...,} xb i, ... xb n) and in FIG pattern T of the reference pattern _{T T_fig = (tf 1, tf} 2, ..., tf i, ... t
f _n ) and the ground pattern T_back = (tb ₁ , tb ₂ , ..., T)
b _i , ..., Tb _n ) are collated (step 206). Here, when i is a pixel position (i = 1, ..., N), in the case of a binary image pattern, the total number of pixels of xf _i = 1 and tf _i = 1 is a, xf _i = 1 and tb _i. = 1, the total number of pixels is c, x
Let b be the total number of pixels with b _i = 1 and tf _i = 1 and e be the total number of pixels with xb _i = 1 and tb _i = 1. That is, the total number (a) in which both Xnoise and T patterns are black points at each pixel position, the total number (b) in which Xnoise is a white point and T is a black point,
Four types of variables are obtained: a total number (c) in which Xnoise is a black point and T is a white point, and a total number (e) in which both Xnoise and t are white points. In the case of a multivalued pattern, for example, the inner product value of the figure pattern Xnoise_fig of the image pattern (learning pattern) Xnoise and the figure pattern T_fig of the reference pattern T is a, and the image pattern X is
The inner product value of the noise figure pattern Xnoise_fig and the ground pattern T1 of the reference pattern T is c, and the ground pattern Xno of the image pattern Xnoise.
These values are determined with the inner product value of ise_back and the figure pattern T_fig of the reference pattern T being b, and the inner product value of the ground pattern Xnoise_back of the image pattern Xnoise and the ground pattern T_fig of the reference pattern T being e.

Next, the characteristic amount calculation section 4 obtains a characteristic amount from the combination of the counted values of a, e, b, and c (step 207). Here, for example, the following 14 types, which are first-order and second-order terms, are characterized by feature amounts F = (f ₁ , f ₂ , ...,
f _j , ..., F ₁₄ ). F = (a, e, b, c, a * a, e * e, b * b, c *
c, a * e, a * b, a * c, e * b, e * c, b *
c) As the feature quantity, it is also possible to use a higher-order term of the third order or higher and a value obtained by nonlinearly converting them, for example, a logarithmic value.

When the image pattern (learning pattern) Xnoise and the reference pattern T are in the same category or different categories
Then, a sufficient number of feature quantities F are obtained for a plurality of pattern pairs (Xnoise, T) (step 208) and input to the discriminant function learning unit 5.

The discriminant function learning section 5 learns a discriminant function using the input feature amount F (step 209). The discriminant function is obtained by obtaining a function that discriminates between two classes, F when Xnoise and T are in the same category and F when Xnoise and T are in different categories.

Here, as an example, by Fisher's linear discriminant analysis (for example, it is described in detail in "Multivariate analysis method <revised edition>" by Okuno et al. (Nikka Giren Publishing Co., Ltd.), etc.) The formula for learning the discriminant function is shown below. Previous feature amount F = (f ₁ , f ₂ , ..., F _j , ..., F ₁₄ ) = (a, e, b, c, a * a, e * e, b * b, c * c, a * E, a * b, a * c, e * b, e * c, b * c)

[0026]

[Equation 1]

Define Here, the learning of the weighting factor w _i is performed by the noise-added image pattern (learning pattern) Xnois.
Regarding the reference pattern T of e and noise amount 0, when Xnoise and T are in the same category, they are determined by performing Fisher's discriminant analysis by classifying them into two classes in the case of different categories.

For learning the function for discriminating between the two classes, it is possible to use a method of learning for finding rules by a computer using a method by a neural network.

Next, referring to FIG. 3, the feature quantity counting unit 4 of FIG.
A specific operation example of is described. In FIG. 3, n = 16 (pixels)
As an example, a binary image pattern Xnoise to which noise is added and a reference pattern T, and their figure pattern and ground pattern are shown. In FIG. 3, 11 is an example of a binary image pattern Xnoise to which noise is added, 12 is an example of a binary reference pattern T, 13 is a diagram of a binary image pattern Xnoise 11, pattern Xnoise_fig, 14 is a diagram of a binary reference pattern T12. A pattern T_fig, 15 is a ground pattern Xnoise_back of the binary image pattern Xnoise11, and 16 is a ground pattern T_back of the binary reference pattern T12. Here, the pixel position i (i = 1, ... 1
In 6), the upper left corner is i = 1, and the order is from left to right and from top to bottom.

Xnoise_fig13, T_fig14, and Xn in FIG.
The patterns of oise_back15 and T_back16 are collated, and the values of a, e, b, and c are obtained from the respective pixel values. a is x
f _i = 1 and a number of pixels of tf _i = 1, patterns of FIG 1
In the examples of 3 and 14, i = 2, 3, 7, 11, 15, 16
6 pixels, and a = 6. e is xb _i = 1 and t
The number of pixels is b _i = 1 and, in the example of patterns 15 and 16 in FIG. 3, there are 7 pixels of i = 1, 4, 5, 6, 10, 12, 13 and e = 7. b is the number of pixels of xb _i = 1 and tf _i = 1. In the example of patterns 15 and 14 of FIG.
There are 14 pixels, and b = 1. c is the number of pixels xf _i = 1 and tb _i = 1, pattern 13 of FIG 3 and 16
In the example, there are two pixels of i = 8 and 9, and c = 2.

The feature amount F is obtained from the combination of the counted values of a, e, b, and c. Here, for example, assuming that the above-mentioned 14 types of first-order and second-order terms are the feature amount F, the following values are obtained as the feature amount.

F = (a, e, b, c, a * a, e * e, b * b, c * c, a * e, a * b, a * c, e * b, e * c, b * c) = (6,7,1,2,36,49,1,4,42,6,12,7,14,2) The embodiment of the present invention has been described above. In the description of the embodiment, the image pattern stored in the pattern storage unit 1 of FIG. 1 is used for both the learning pattern and the reference pattern, and the learning pattern is the image pattern X to which noise is added by the noise adding unit 2.
The reference pattern T is noise, and the reference pattern T is the image pattern before the noise of the Xnoise is added, but the reference pattern may be a pattern (standard pattern) input from another. In this case, for example, an image pattern (learning pattern) stored in the pattern storage unit 1
If noise is added in advance, the noise adding unit 2 can be omitted.

[0033]

According to the first aspect of the present invention, the discriminant function can be inductively learned from the learning pattern set, and an appropriate discriminant function can be obtained more easily than the conventional empirical method.

According to the invention of claim 2, by forcibly adding the noise-added pattern to the learning pattern set, a noise pattern in which superimposition such as fading or dirt, which has been difficult to recognize in the past, is also generated. A discriminant discriminant function can be easily obtained.

According to the invention of claim 3, claim 1 or 2
It is possible to obtain a more accurate discriminant function than the discriminant function obtained in.

[Brief description of drawings]

FIG. 1 is a block diagram showing an example of a device configuration to which a discriminant function learning method according to the present invention is applied.

FIG. 2 is an example of a flowchart of a discriminant function learning process according to the present invention.

FIG. 3 is a diagram showing a pattern example for explaining a specific operation example in the feature amount counting section of FIG. 1.

[Explanation of symbols]

1 pattern recording section 2 Noise addition section 3 Figure ground pattern creation department 4 Feature counter 5 Discriminant function learning section

─────────────────────────────────────────────────── --- Continuation of the front page (56) References JP-A-7-282199 (JP, A) JP-A-5-54195 (JP, A) JP-A-5-242252 (JP, A) (58) Field (Int.Cl. ⁷ , DB name) G06K 9/00-9/76

Claims (3)

(57) [Claims] 1. A discriminant function learning method for obtaining a discriminant function used for pattern recognition by learning from a learning pattern,
Each element of the learning pattern is a "Figure" area (recognition category area)
And "ground" area (area other than recognition category area),
Between the learning pattern and the reference pattern, the "Figure" area and the "Figure" area
Area, "figure" area and "ground" area, "ground" area and "figure" area
Area, in each combination of "ground" area and "ground" area
Matching is performed for each pixel, and the number of matching pixels is a,
c, b, e, and the first-order terms of the a, c, b, e and the
N-th order (N ≥ N of all combinations of values of a, c, b, and e)
A discriminant function learning method characterized in that the discriminant function is a function for discriminating a feature amount obtained from a pattern of the same category and a feature amount obtained from a pattern of a different category by constructing the feature amount from the terms up to 2). . 2. The discriminant function learning method according to claim 1, wherein a discriminant function obtained by performing learning by adding a pattern on which noise is superimposed to a learning pattern is a discriminant function to be obtained. Function learning method. 3. The discriminant function learning method according to claim 1, wherein a discriminant function is obtained for each category.