JP2001147989A - Word recognizing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a word recognizing device, with which recognition is improved even in the case of a different character size such as the character width of a word, concerning the word recognizing device for recognizing a handwriting character string. SOLUTION: This word recognizing device is provided with an input word feature amount extracting mans for extracting a feature amount by segmenting and normalizing image data from the prescribed word area of an input image, a word dictionary for holding words, a character dictionary for holding the average value and variance value of characteristic values of respective characters and character images, a dictionary word feature amount extracting means for calculating plural characteristic value ratios among characters from the characteristic values of respective characters of a word, compositing plural word images in a prescribed size from the character images of respective characters of the word corresponding to the characteristic value ratios and extracting plural word feature amounts for each word from the plural word images, and a word collating and output means for collating the input word feature amount with the dictionary word feature amount, selecting and outputting the word of the closest feature amount.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a word recognition device for recognizing a handwritten character string.

[0002]

2. Description of the Related Art Among character recognition devices, besides a character recognition device for recognizing a character for each individual character, there is a word recognition device which focuses on a character string in order to improve a recognition rate.

FIG. 8 shows a configuration diagram of a conventional word recognition apparatus 1.

The word recognition device 1 comprises an image input unit 2, a normalization unit 3, a feature extraction unit 4, a feature comparison unit 5, a feature synthesis unit 8, a word dictionary 9, and a character dictionary 7.

FIG. 9 is a diagram illustrating the operation of a conventional word recognition apparatus.

FIG. 9A shows an example of an address.

Frames for handwritten entry of addresses, prefectures, municipalities, towns and villages, and addresses are printed in print. Also,
Characters that serve as references for determining the position of a prefecture are printed at prescribed coordinates. It is assumed that "Tokyo", "Tachikawa", and "Shibasaki" are written by hand in the address column.

FIG. 9B shows an operation example.

First, the image input unit 2 detects the position where the characters of the prefecture serving as the position reference are printed in the input image, and cuts out the character string "Tokyo" preceding it. Then, the extracted character string is normalized by the normalization unit 3.
The feature amount of the normalized word is extracted by the feature extracting unit 4,
The information is passed to the feature matching unit 5. The same applies to municipalities and towns and villages.

The feature synthesizing section 8 selects words in the word dictionary 9 such as prefectures one by one.

All words such as "Hokukai", "Tokyo" and "Okinawa" are registered in the word dictionary. In this example, assuming that the word "North Sea" is selected, the feature amounts of "North" and "Sea" are extracted from the character dictionary 7 in which the feature amounts of all characters are registered. Are synthesized. The result of the synthesis is compared with the feature amount of the input image by the feature matching unit 5, and a difference in feature amount between the two is obtained. Then, the same processing is performed for all the words in the word dictionary.
When the matching processing is completed, the word having the smallest difference in the feature amount among the words is output from the recognition result output unit 6 as a recognition result. The word recognition of the address example is completed by performing the same for the city group and the town.

[0012]

In a system in which words are collectively recognized without being separated into individual characters, a single word feature is calculated from a single feature amount of each character constituting the word for one word. The amounts are combined and matched against the input image. Therefore, the word recognition rate is low with respect to variations in character size such as the width of handwritten characters.

An object of the present invention is to increase the recognition rate of words having different character sizes such as the width of a character in a word recognition device.

[0014]

A word recognition apparatus includes an input word feature extracting means for extracting image data from a predetermined word area of an input image and normalizing and extracting feature data; a word dictionary for storing words; , A character dictionary that holds the average value, variance value, and character image of each character, and a word obtained from the word dictionary, and the average value of the characteristic value of each character of the obtained word,
Characteristic value ratio calculating means for calculating a plurality of characteristic value ratios between characters from the variance value, and based on the plurality of characteristic value ratios between characters,
Word image synthesizing means for synthesizing a plurality of word images of a predetermined size from a character image of each character of a word, and dictionary word feature amount extracting means for extracting a plurality of word feature amounts for one word from the synthesized word images And word matching output means for comparing the input word feature amount with the dictionary word feature amount, and selecting and outputting the word having the closest feature amount.

[0015] With this configuration, a plurality of characteristic value ratios can be calculated based on the average value and the variance value (square of the standard deviation) of the characteristic values (for example, line density) of each character constituting the dictionary word (word obtained from the dictionary). By generating multiple word images from a character image based on the generated characteristic value ratio, it is possible to extract multiple feature amounts for one word. Can be read.

Further, the word recognition device cuts out image data from a predetermined word region of the input image, normalizes the image data, and extracts a feature value, a word dictionary for holding words, a word dictionary for storing the words, The image size is divided into n (integer) regions in the horizontal or vertical direction, and each divided region is used as a start position to set the word image size to m / n (m <n, n, m).
Words are acquired from a word dictionary, which holds a plurality of m / n feature amounts extracted by normalizing a character image to a size and the average value and variance value of the characteristic value of each character, and a word dictionary. A characteristic value ratio calculating means for calculating a plurality of characteristic value ratios between characters from an average value and a variance value of characteristic values of each character of the word; and a character dictionary for each word based on the plurality of characteristic value ratios between characters. A dictionary word feature synthesizing means for acquiring a plurality of m / n features of a character, and synthesizing a plurality of word features for one word from the acquired m / n features of each character; And a word collation output unit for collating the quantity with the dictionary word feature quantity and selecting and outputting a word having the closest feature quantity.

In this configuration, a plurality of m / n feature amounts of characters are mounted in advance in the character dictionary for each region position obtained by dividing the word image size into n in the horizontal or vertical direction. And
A plurality of word characteristic value ratios are determined based on an average value and a variance value (standard deviation squared) of characteristic values (for example, line density) of characters constituting a dictionary word (a word obtained from the dictionary), and the word characteristic value is determined. By synthesizing a plurality of feature amounts for one word according to the ratio, high-precision reading can be performed with respect to fluctuations such as the width of a character.

The dictionary word feature amount synthesizing means moves the m / n feature amount of each character obtained from the character dictionary by a predetermined amount in a predetermined direction based on a plurality of characteristic value ratios between characters constituting the word. And a feature amount transforming unit that regenerates the feature amount of the word by using the m / n feature amount of each character regenerated by the feature amount transforming unit. With this configuration, by moving the feature amount by a predetermined amount in a predetermined direction of the character (for example, in the case of a horizontally written document, upward or downward of the character), the characteristic amount of a word having a slope (a downward-sloping word, for example) can be obtained. Since it can be generated, high-precision reading is possible even for inclined words and the like.

Further, the word recognition device divides a predetermined word image size into n (integer) regions in the horizontal or vertical direction, and sets a specific region in the divided regions as a start position to m of the word image size. / N (m <n, where n and m are integers) a plurality of m / n extracted by normalizing the character image to the size
A character dictionary that holds a feature amount, an average value and a variance value of a characteristic value of each character, and a plurality of m / n feature amounts of each character of the word based on a plurality of characteristic value ratios between characters constituting the word. Is obtained from the character dictionary, the plurality of m / n features of the obtained characters are converted into the features of the area on the word image based on the characteristic value ratio, and the m / n features of the converted characters are used as the basis. And a dictionary word feature amount synthesizing means for synthesizing a feature amount of a word.

According to this configuration, each m / n feature amount of a character image constituting a word is provided at a specific region position divided into n on the word image, so that each m / n feature amount is provided for each region position.
Since there is no need to have a feature amount, the dictionary capacity of characters can be reduced.

[0021]

FIG. 1 is a block diagram of a word recognition apparatus 1 according to a first embodiment.

The word recognition device 1 comprises an image input unit 2, a normalization unit 3, a feature extraction unit 4, and a recognition processing unit 11. The recognition processing unit 11 includes a feature matching unit 5, a feature synthesizing unit 8, a recognition result output unit 6, a word dictionary 9, and a character dictionary 7.

The feature synthesizing unit 8 includes a word feature amount extracting unit 81 or a word feature amount synthesizing unit 82.

FIG. 2 is an explanatory diagram of the operation of the word recognition device.

FIG. 2A shows an example of an address.

The address, name, prefecture, city and ward group, town and village, and the frame for handwritten entry of the address are printed in print.
Characters that serve as references for determining the position of a prefecture are printed at prescribed coordinates. It is assumed that "Tokyo", "Tachikawa", and "Shibasaki" are written by hand in the address column.

FIG. 2B shows an operation example.

First, the image input unit 2 detects the position where the character of the prefecture serving as the position reference is printed in the input image, and cuts out the character string "Tokyo" preceding it. Then, the extracted character string is normalized by the normalization unit 3. The feature amount of the normalized word is extracted by the feature extracting unit 4 and passed to the feature matching unit 5.

The same processing is performed for the municipalities and towns and villages.

In the word dictionary 9, all words of the prefectures such as Hokkai, Tokyo and Okinawa are registered.
In, the feature amounts of all characters are registered.

Next, FIG. 3 shows the flow of the processing of the recognition processing section 11.

First, the feature synthesizing unit 8 operates according to the word dictionary.
First, search for the word "North Sea". (S11 step in Fig. 3)
When the word "North Sea" is found, feature synthesis processing is performed on "North" and "Sea". (S12 step in FIG. 3) Next, two methods of feature synthesis which are features of the present invention will be described. (1) Word Feature Extraction Processing (Step S121 in FIG. 3) The processing of the word feature extraction unit 81 will be described.

Since it has a character image, it is used for a high-speed OCR device having a large memory capacity.

When a word image is synthesized from a character image, a sample image of the character is stored in the character dictionary 7 in advance. Then, a word image is generated from each character image constituting the word, and a word feature amount is extracted. For this reason, the characteristic value ratio between characters is calculated using the variation parameter based on the characteristic value of each character constituting the word, and a word image normalized based on the characteristic value ratio is generated. For example, the average value and the variance (square of the standard deviation) of the sum of the line densities in the vertical direction of the character based on a large number of sample images are obtained as the characteristic values of the width of the character.

Further, as the characteristic value of the character height, an average value and a variance value (square of the standard deviation) of the height are obtained based on a large number of sample images. These characteristic values are all installed in the character dictionary 7 in advance.

FIG. 4 shows a flowchart of the word feature value extraction processing. FIG. 5 is an explanatory diagram of the word image synthesizing process.

First, the average value and the variance of the sum of the line densities in the vertical direction of the characters constituting the word are obtained from the character dictionary 7. (S2
1 Step) The line density sum of “north” of the North Sea in FIG. 5 is set to, for example, an average value 30 and a variance value 36 (standard deviation 6). The sum of the linear densities of the “sea” in the North Sea is, for example, an average value 50 and a variance value 16 (standard deviation 4).

Next, the characteristic value ratio of a plurality of widths of the sum of the line densities between the characters constituting the word is calculated. (Step S22) From the average value and the variance value, the line density sum of “north” is 24 to 3
6. The sum of the linear densities of "sea" is a value from 46 to 54.

Therefore, the characteristic value ratio is 36:46.
(= 1: 1.27) to 24:54 (= 1: 2.25). For example, 1: 1.5, 1: 2.0, etc.

Next, an average value and a variance value of the heights of the characters constituting the word are obtained from the character dictionary 7. (Step S23) Next, the height ratio between the characters is calculated from the average value and the variance of the heights of the characters constituting the word. (Step S24) For example, when the number of pixels is 72 × 72, the average value of the height of the character “north” is 40, the variance value is 16 (standard deviation 4), the average value of the height of the “sea” is 60, and the variance value is 60. (Standard deviation 5) is set to 25.

From the mean and variance values, the height “north” is
36 to 46, the height of the “sea” is a value from 55 to 65.

Therefore, the height ratio is 46:55 (= 1: 1.2
7) Select multiple options from the range up to 36:65 (= 1: 1.8). For example, 1: 1.5, 1: 1.8, and so on.

Next, by combining the variation in the width and the variation in the height, a word image of a plurality of patterns (four patterns in the above example) having a normalized size is synthesized. (Step S25) Then, a plurality of word feature values are extracted. (Step S26) (2) Word Feature Amount Synthesis Process (Step S122 in FIG. 3) The process of the word feature amount synthesis unit 82 will be described.

This method is a method of synthesizing a feature amount of a word from a feature of each character constituting the word. Since a character image is not required, it is used for an OCR device having a small memory capacity. The character feature at this time is, for example, a weighted direction exponential histogram feature (IEICE Vol.J70DNo.7, pp.1390-1397)
Is used.

For example, if the width of the normalized image of a word is 7
To divide.

Next, the character image is normalized so as to occupy the width of m / 7 (m <7, m is an integer) of the divided horizontal width, and the weighted direction index histogram feature is applied to the normalized image. Extract. The obtained features are referred to as m / 7
Called features.

For example, if a certain character has a 1/7 feature, there are seven 1/7 features depending on the position of each of the seven divided areas.

The character dictionary 7 stores m / 7 for all characters in advance.
Generate and register features. Then, the word features are synthesized such that the sum of the m / 7 fractions of the character features constituting the word becomes 1.

For example, in the case of a two-letter word (for example, Tokyo), depending on the characteristic values of the characters registered in the character dictionary 7, 3/7 feature (east) +4/7 feature (K) or 2/7 feature
Combine with feature (east) + 5/7 feature (kyo).

As the characteristic value, for example, a linear density sum is used.

The characteristic value ratio in the width direction of each character constituting a word is determined based on, for example, the ratio of the sum of line densities, and m / 7 features are extracted from the character dictionary 7 and synthesized.

FIG. 6 shows a flowchart of the word feature amount synthesizing process. FIG. 7 is an explanatory diagram of the word feature amount synthesizing process.

First, the average value and the variance value of the sum of the line densities in the width direction, for example, of the characters constituting the word are obtained from the character dictionary 7. (S31 step) FIG. 7A shows an example of the characteristic value ratio.

The sum of the line densities of “north” of the North Sea is calculated, for example, as an average value of
0 and variance value 36 (standard deviation 6). The sum of the linear densities of the “sea” in the North Sea is, for example, an average value 50 and a variance value 16 (standard deviation 4).

From the average value and the variance value, the sum of the line densities for "north" takes values from 24 to 36, and the sum of line densities for "sea" takes values from 46 to 54.

Next, the characteristic value ratio of the sum of the line densities between the characters constituting the word is obtained. (S32 step) Then, as the characteristic value ratio, 36:46 (= 3.07: 3.92 mesh number 7) to 24:54 (= 2.15: 4.85 mesh number 7)
Select multiple integer values from the range up to.

For example, 3: 4, 2: 5. (S33 step) Next, it is determined whether the mode is the small-capacity dictionary use mode. (S34
Step ゜) When the small-capacity character dictionary is used, the characteristic amount of the character corresponding to the calculated characteristic value ratio is first obtained from the character dictionary 7. (S3
5) Next, the acquired feature amount is shifted to a predetermined area position on the word, and is calculated as a feature amount from the predetermined position. The method when using the small-capacity character dictionary is as follows.
Instead of having the / 7 feature, the m / 7 feature at a certain position is shifted laterally to substitute the m / 7 feature at a different position.

FIG. 7B shows an example of the feature amount movement. It is assumed that the character dictionary 7 has features from the 1/7 position.

In the case of the characteristic value ratio of “North Sea” 3: 4,
"North" is 3 mesh width from 1/7 position, "Sea" is 4
/ 4 position from the / 7 position.

Therefore, for “north”, the 3/7 feature from the 1/7 position is obtained from the character dictionary 7. For the “sea”, the 4/7 feature from the 1/7 position is stored in the character dictionary 7
Get more. Then, a feature amount having a width of 4 meshes from the 1/7 position of the “sea” is converted into a feature amount having a 4 mesh width from the 4/7 position by calculation. The other characteristic value ratios of 2: 5 are obtained in the same manner. (S36 step) Then, based on the acquired feature amounts, the feature amounts as words are synthesized for each characteristic value ratio. (S39 step) When the mode is not the character small-capacity dictionary use mode, the characteristic of the character corresponding to the calculated ratio is determined based on the region position.
Get more. (Step S37) In the case of the characteristic value ratio of the North Sea of 3: 4, for the “north”, the 3/7 feature from the 1/7 position is acquired from the character dictionary 7. For the “sea”, the 4/7 feature from the 4/7 position is obtained from the character dictionary 7. The other characteristic value ratios of 2: 5 are obtained in the same manner.

Based on the acquired feature amounts, feature amounts as words are synthesized for each characteristic value ratio. (S39 step) FIG. 7C shows a modification of the feature amount.

When adding the m / 7 characteristics of each character,
By changing each feature amount upward or downward and adding them together, it is possible to cope with fluctuations of characters constituting various words.

The following feature values are transformed with respect to the feature values of the respective characters obtained in steps S36 and S37.

For example, in consideration of the case where a two-letter word is written to the lower right, the feature of the first character is moved upward by one mesh, and the feature amount of the second character is left as it is. To deal with. At this time, the seventh mesh counted from the bottom of the first character is discarded by moving one mesh. Also, considering the case where the two-letter word is written to the upper right, the features of the first character are left as they are and the features of the second character are moved upward by one mesh to deal with the upward-sloping word. . (S38 step) The recalculated feature amount of the first character and the feature amount of the second character are combined for each characteristic value ratio. (S39 step) When the feature synthesis is completed, the feature is compared with the feature amount of the input image. (Step S13 in FIG. 3) This word feature is compared with the feature of the input image and the difference is stored.

These processes are performed on all the words, and finally, a feature having a minimum difference between the feature of the input image and the feature from the feature synthesizing section 8 is output as a recognition result. (S14 step in FIG. 3) Word recognition of the address example is completed by performing the same processing for the municipalities and towns and villages.

[0066]

According to the present invention, since a plurality of word feature amounts are synthesized for one word, recognition can be performed with high accuracy even if the height, width, inclination, and the like of the word fluctuate. In addition, it is possible to reduce the dictionary capacity by having one m / n feature amount of each character and calculating other region positions by calculation.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a word recognition device according to an embodiment.

FIG. 2 is an explanatory diagram of an operation of the word recognition device of the embodiment.

FIG. 3 is an explanatory diagram of a recognition processing unit.

FIG. 4 is a flowchart of a word feature amount extraction process.

FIG. 5 is an explanatory diagram of a word image combining process.

FIG. 6 is a flowchart of a word feature amount synthesizing process;

FIG. 7 is an explanatory diagram of a word feature amount synthesizing process;

FIG. 8 is a configuration diagram of a conventional word recognition device.

FIG. 9 is an explanatory diagram of an operation of a conventional word recognition device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Word recognition apparatus 2 Image input part 3 Normalization part 4 Feature extraction part 5 Feature collation part 6 Recognition result output part 7 Character dictionary 8 Feature synthesis part 9 Word dictionary 11 Recognition processing part 81 Word feature quantity extraction part 82 Word feature quantity synthesis Department

Claims (4)

[Claims] 1. An input word feature extracting means for extracting image data from a predetermined word area of an input image and normalizing to extract a feature, a word dictionary storing words, and an average value of characteristic values of each character. , A character dictionary that holds variance values and character images, and a characteristic value that acquires words from the word dictionary and calculates multiple characteristic value ratios between characters from the average value of the characteristic values of each character of the acquired words and the variance value Ratio calculating means, word image synthesizing means for synthesizing a plurality of word images of a predetermined size from a character image of each character of a word based on a plurality of characteristic value ratios between characters, and one word from the synthesized plurality of word images Dictionary word feature amount extraction means for extracting a plurality of word feature amounts, and word matching output means for comparing the input word feature amount with the dictionary word feature amount and selecting and outputting a word having the closest feature amount. Equipped A word recognition device characterized by the following. 2. An input word feature extracting means for extracting image data from a predetermined word area of an input image and normalizing and extracting a feature, a word dictionary for storing words, a horizontal or a predetermined word image size. N (integer) vertically
And a plurality of m / n extracted by normalizing the character image to m / n (m <n, n and m are integers) of a word image size with each of the divided regions as a start position.
A character dictionary that holds the feature values and the average and variance values of the characteristic values of each character; a word is obtained from the word dictionary; And a plurality of m / n feature amounts of each character of the word from the character dictionary based on the plurality of characteristic value ratios between the characters. A dictionary word feature amount synthesizing means for synthesizing a plurality of word feature amounts for one word from a plurality of m / n feature amounts, collating the input word feature amount with the dictionary word feature amount, and determining a word having the closest feature amount; A word recognizing device comprising: a word collating and outputting means for selecting and outputting. 3. The dictionary word feature amount synthesizing unit moves a m / n feature amount of each character obtained from a character dictionary by a predetermined amount in a predetermined direction based on a plurality of characteristic value ratios between characters constituting a word. 9. A method according to claim 8, further comprising: a feature amount transforming means for regenerating the feature amount of the word, wherein the feature amount of the word is synthesized based on the m / n feature amount of each character regenerated by the feature amount transforming means. 2. The word recognition device according to item 2. 4. A predetermined word image size is divided into n (integer) areas in the horizontal or vertical direction, and a specific area in the divided areas is used as a start position to obtain a word image size m / m.
a character dictionary holding a plurality of m / n feature amounts extracted by normalizing a character image to n (m <n, n and m are integers) size, and an average value and a variance value of a characteristic value of each character; Acquiring a plurality of m / n feature amounts of each character of the word from the character dictionary based on a plurality of characteristic value ratios between characters constituting the word;
A plurality of m / n feature values of each of the acquired characters are converted into feature values of a region on the word image based on the characteristic value ratio, and the feature values of the words are synthesized based on the converted m / n feature values of each character. 3. The word recognition device according to claim 2, further comprising: a dictionary word feature amount synthesizing unit.