JPS6355105B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character recognition device having a standard pattern correction function that gradually corrects a standard pattern in a dictionary using sample characters.

Since handwritten characters are deformed in a variety of ways, it is difficult to estimate the deformations from a small number of samples, as is the case with printed characters. Therefore, in order to improve the reading performance of a handwritten character recognition device, it is necessary to use a large number of sample characters and create standard patterns for a dictionary to be used for identification from these characters. However, it is difficult to collect a large number of sample characters in advance. Therefore, it is reasonable to first create a dictionary pattern using a small number of sample characters, and then gradually increase the number of sample characters and modify the dictionary pattern using the sample characters each time. In this case, it is not true that increasing the number of samples will result in a better standard pattern; rather, a standard pattern that incorporates all kinds of modified characters and calculates them from their average value will become far removed from the target character. Put it away.

The present invention attempts to deal with such a problem, and is characterized by an average _{value μ} ^C _N , standard deviation σ〓 ^C _N , and the number of samples N are stored in the dictionary, and the standard deviation σ〓 ^C _{N+1 including the newly input sample pattern X〓 N+ 1} becomes the threshold d. On the other hand ^, when ‖σ〓 ^C _{N +1 ‖d ,} the sample pattern > d, the average value of the category C is μ〓 ^C _N+1 , and the sample pattern X〓 _N+1 is the subcategory.
The main purpose of the present invention is to have a dictionary correction circuit for generating C', which will be explained in detail below with reference to the drawings.

Let xj be the series of sample values, j=1, ..., N
If the average value of N samples is μ _N and the standard deviation is σ _N , then μ _N = 1/N _N 〓 ^j=1 xj Therefore, _N 〓 ^j=1 xj=Nμ _N ……(1) σ ² _N = 1/N _N 〓 ^j=1 (xj-μ _N ) ² Therefore, _N 〓 ^j=1 x ² _j = N (σ ² _N + μ ² _N ) ...(2). Assuming that μ _N , σ _N , and N are known and a new (N+1)th sample value x _N+1 is obtained, the mean value μ _N and variance σ ² _N+1 are μ _N+1 = 1/ N+1 _N 〓 ^j=1 xj=1/N+1 ( _N 〓 ^j=1 xj+x _N+1 )=1/N+1 (Nμ _N +x _N+1 ) ...(3) σ ² _N+1 = 1/N+1 _N 〓 ^j=1 (xj−μ _N+1 ) ² =1/N+1 _N 〓 ^j=1 x ² _j −μ ² _N+1 =1/N+1 ( _N 〓 ^j=1 x ² _j +x ² _N+1 )− μ ² _N+1 = 1/N+1 {N(σ ² _N + μ ² _N ) x ² _N+1 }−μ ² _N+1 …
...(4), and μ _N+1 and σ _N+1 can be found from μ _N , σ _N , N, and x _N+1 . That is, when μ _N and σ _N for N samples are known, μ _N+1 for (N+1) samples with one sample added,
σ _N+1 can be obtained from equations (3) and (4). The present invention utilizes this relationship to start recognition with a standard pattern made up of N samples, modify the standard pattern with samples provided for each recognition process, and gradually change it to a more appropriate standard pattern. This is what we are trying to do.

When the feature extracted from a character sample i belonging to a certain category C is X〓 ^C _i , the average value of X〓 ^C _i , μ〓
^C _N
is the standard pattern of category C. That is,
When N character samples belonging to category C are given, μ〓 ^C _N is μ〓 ^C _N =1/N _N 〓 ⁱ⁼¹ X〓 ^C _i ...(5). Here, category C indicates types of characters such as "kanji" and "characters," and sample characters are those handwritten in various ways, and the characteristics X〓 ^C _i
is the line density obtained by projecting a handwritten character pattern in the X or Y direction (for example, a group of numerical values in which the number of black dots included in each horizontal scanning line of the video signal of the character "Kan" is arranged in vertical scanning order) This is an n-dimensional vector). By the way, same category C
When μ〓 ^C _N is created from the characteristic patterns of all sample characters belonging to (hereinafter referred to as sample patterns), since the sample pattern includes all the transformations of handwritten characters, conversely it forms a standard pattern μ〓 ^C _N with high accuracy. Can not do it. Therefore, the present invention features
Using the standard deviation σ of X〓 ^C i〓 ^C _{N ,} limit the category C so that ‖σ〓 ^C _N ‖d ...(6), and those that do not fall within this range are separated into subcategories (that is, categories C into multiple subcategories),
Construct a highly accurate standard pattern. Note that in the above equation, ‖·‖ indicates conversion from a vector to a scalar, and d is a threshold value. Category C has already been created from N sample patterns X〓 ₁ , X〓 ₂ , ...X〓 _N
Assuming that there exists a standard pattern μ〓 ^C _N , and its standard deviation is σ〓 ^C _N , when a new sample pattern X〓 _N+1 is given, first from equations (3) and (4), μ〓 ^C _N+1 , σ〓 ^C _N
Find ₊₁ and determine whether ‖σ〓 ^C _N+1 ‖d ...(7). If the above formula is satisfied, X〓 _N+1 is included in category C and its standard pattern (average value) is corrected from μ〓 ^C _N to μ〓 ^C _N+1 . On the other hand, if ‖σ〓 ^C _N+1 ‖>d (8), then X〓 _N+1 is not included in category C and is newly registered as a standard pattern of subcategory C'.

Figure 1 conceptually explains this. When category C is expressed in multidimensional feature space A, sample pattern x _N+1 that satisfies equation (7) is included in category C, and its standard Modify the pattern μ〓 ^C _N to μ〓 ^C _N+1 . However, it is harmful to include other sample patterns x _N+2 , x _N+3 , etc. that have the relationship in equation (8) in category C and modify the standard pattern, so for example
X _N+3 is included in subcategory C′. Sample patterns that belong to the same category but have _different characteristics, such as sample pattern Since Hiko's standard pattern correction operation is performed every time a new sample pattern appears while the device is in operation, it starts with a few samples and gradually becomes a qualified one that collects the characteristics of many samples, enabling highly accurate character recognition. becomes. Moreover, since a single standard pattern is not created from all the collected features, recognition accuracy will not be reduced by a simple collection (average value) of miscellaneous features.

FIG. 2 is a block diagram showing one embodiment of the present invention. In the figure, reference numeral 1 denotes a control section, which controls the entire pattern recognition apparatus. The observation section 2 is
The input character M on the form is converted into a video signal V and sent to the feature extraction section 3. In the feature extraction unit 3, the observation unit 2
The feature pattern X is extracted from the video signal V obtained by The identification unit 4 calculates the distance between the feature ^pattern Dictionary creation section 6, dictionary registration section 7
operates during dictionary registration. The dictionary creation unit 6 calculates a dictionary (standard) pattern using equations (3) and (4) by referring to the input feature pattern X and the dictionary 5. The dictionary registration unit 7 stores the calculated dictionary pattern μ _N+1
By comparing the standard deviation σ〓 ^C _N+1 with the pre-given threshold d as shown in equations (7) and (8), the existing dictionary patterns can be corrected or new dictionary patterns can be added and registered. .

FIG. 3 shows an example of the structure of the dictionary 5. The dictionary for one subcategory consists of four items: mean value, standard deviation, number of samples, and category name. The number of samples
N ₁ , N ₂ is the number of samples used to create the dictionary for each subcategory C ₁ , C ₂ , ..., and the average value and standard deviation are the average value and standard deviation of the characteristic pattern of the N samples. It is a deviation. According to the explanation of this subcategory, C and C' in Figure 2 are, for example,
Corresponds to C ₁ and C ₂ , and the whole category C is the same
belongs to

As described above, according to the present invention, there is an advantage that handwritten characters that are variously transformed can be sequentially recognized with high precision as various sample patterns are learned.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram of a dictionary pattern according to the present invention,
FIG. 2 is a block diagram showing one embodiment of the present invention, and FIG. 3 is an explanatory diagram of a dictionary classified into subcategories. In the figure, 5 is a dictionary, 6 is a dictionary creation section, and 7 is a dictionary registration section.

Claims (1)

[Claims] 1 Sample pattern X belonging to category C ₁ ~
The mean value μ〓 ^C _N , which is derived from X〓 _N and becomes the standard pattern for pattern recognition, and the standard deviation σ〓 ^C _N
,
In addition, the standard deviation σ〓 ^C _{N+1 including the newly input sample pattern X〓 N+ 1} is ‖σ〓 ^C _N+1 ‖d with respect to the threshold d. _If so _, ^include _the sample pattern The value remains μ〓 ^C _N+1, and the subcategory is set to the sample pattern X〓 _N+1.
A pattern recognition device characterized by having a dictionary correction circuit that generates C′.