JPH02100783A - Character recognizing method - Google Patents

Abstract

PURPOSE:To improve the character recognizing process efficiency by giving the positional correction to the feature extracted from a single character in response to an area which is turned into a blank space with one character against a square and collating the feature undergone the positional correction with the feature read out of a dictionary for selection of candidate characters. CONSTITUTION:A processor 1 corrects the extracted feature with a position correcting arithmetic operation in response to an area which is turned into blank space with one character against a square. Then, the processor 1 performs the collation between the feature of a dictionary 8 and the feature of one character that undergone the positional correction to select the candidate characters. Thus it is not required to give the conventional horizontal, vertical, and oblique scans to the area corresponding to the blank area for extraction of the feature in case the feature is extracted with scan of one character. Therefore a scanning area is reduced and the feature extracting time is also shortened.

Description

[Detailed description of the invention] [Table of contents] Overview Industrial field of application Conventional technology Problems to be solved by the invention Means for solving the problems Effects of the invention [Summary] Extraction of features of cut out characters Regarding character recognition devices that reduce the processing time when selecting candidate characters with a small degree of difference by comparing them with a dictionary, we cut out a single character from an unknown character string and The area is scanned to extract features, and the characters are compared with the features read from a dictionary that stores the features of characters inscribed in a square of a predetermined size, and characters in the dictionary with a small degree of difference are selected as candidate characters. In the processing of the character recognition device to be sent, the vertical and horizontal sizes of the cut out character are compared with the size of the square, and if the vertical and/or horizontal size of the character is larger than the square, the length of the character is Find the reduction ratio when the larger direction or width is inscribed in the square, reduce one character by the reduction ratio, scan the area of the reduced character to extract features, and apply it to the square. Corresponding to the area that becomes blank due to the reduced character, position correction is performed on the features extracted from the reduced character, and if the vertical and horizontal dimensions of the extracted character are smaller than a square, the area of the single character is Scan the square to extract features, correct the position of the features extracted from each character corresponding to the area that becomes blank due to one character in the square, and compare the position-corrected features with the features read from the dictionary. The configuration is such that candidate characters are selected.

[Industrial application field]

The present invention extracts the characteristics of cut out characters and compares them with a dictionary,
The present invention relates to a character recognition device that selects candidate characters with a small degree of dissimilarity, and particularly to a character recognition device that makes it possible to shorten processing time when extracting features.

With the development of an information-oriented society, it has become popular to read documents and have computer systems process them.

For this purpose, a character recognition device is used, and this character recognition device cuts out characters one by one from the read document and enlarges/reduces the cut out characters so that they are inscribed within a frame of a predetermined size. The normalized characters are scanned to extract features, compared with a dictionary, and characters with a small degree of difference are sent to the computer system as recognition results.

Scanning to extract features from normalized characters is performed many times from the top, bottom, left, right, and diagonal directions of the characters, so the larger the scanning area, the more time is required. There are different types of characters, and it is desirable to shorten the feature extraction processing time for small characters, vertically long characters, or horizontally long characters.

[Conventional technology]

FIG. 7 is a block diagram explaining the conventional technique, and FIG. 8 is a diagram explaining the operation of FIG. 7.

The processor 1 operates by reading the program memory 2 or the Q program, for example, controls the scanner 10 to store the character string read from the document in the image memory 5, and controls the character cutting section 3 to store the character string in the image memory 5. One character is cut out from the character string and stored in the image memory 5 as a character image.

Under the control of the processor 1, the normalization unit 4 stores the image memory 5.
A normalization process is performed to enlarge/reduce the character image stored in the character image so that it is inscribed within a frame of a predetermined size, and the character image is stored in the image memory 5.

That is, the normalization unit 4 normalizes the characters as shown in FIG.
The characters are inscribed in a square frame (■) with a size of 48 x 48 dots.

Under the control of the processor 1, the feature extracting unit 6 extracts the characters stored in the image memory 5 in the diagonal lines shown in FIG.
As shown by the arrow ■■, the image is scanned from a diagonal direction, and 432-dimensional features, for example, are extracted using a known method and stored in the image memory 5.
Store in.

Under the control of the processor 1, the collation unit 7 stores the image memory 5.
The characteristics of the characters stored in the dictionary 8 are compared with the characteristics of the characters read from the dictionary 8, and the character read from the dictionary 8 and having the characteristics with the smallest degree of difference is selected as a candidate character and sent to, for example, a host device via the interface section 9. Send. Generally, a plurality of candidate characters are selected and sent in descending order of dissimilarity, starting with those with the greatest dissimilarity.

[Problem to be solved by the invention]

As mentioned above, conventionally, when characters are normalized by the normalization unit 4,
As shown in Figure 8 (bl), characters larger than the frame ■ are reduced, and as shown in Figure 8 (C), characters smaller than the frame ■ are enlarged so that they are inscribed in the frame ■.

In other words, it touches the four sides of the square frame (■).

Therefore, the feature extraction unit 6 must scan the entire area inside the frame (3) and extract 432-dimensional features.

FIG. 9 is a diagram illustrating the time required by the normalization section and the feature extraction section.

If we take time on the vertical axis and scanning area on the horizontal axis, the time required for normalization (not shown in If the size is inscribed, it is only necessary to move it within the 50 frame ■ of the image memory, and as shown by the movement in the figure, the time is minimal, but if it is larger than the frame ■, it will take time to scan and reduce this character. The larger the character, the more scanning time is required, and the time increases in proportion to the size of the character, as shown by the diagonal lines indicated by the reduction in the figure.

Also, if it is smaller than the frame ■, the time it takes to scan and enlarge this character is because the smaller the character is, the less time it takes to scan it.
The number decreases in proportion to the size of the character, as shown by the diagonal line shown in the enlarged figure.

The time required to extract the feature shown in (2) is a constant time, as it is necessary to scan the entire area inside the frame (2) as described above, and is independent of the size of the character cut out by the character cutout section 3.

Further, after the feature extraction is performed, the feature correction shown in (■) is performed so that the same characters have the same features, but this time is also constant.

Therefore, as shown by the dotted line ■, the total time increases in proportion to the size of the character cut out by the cutout section 3.

FIG. 10 is a diagram illustrating the distribution of character sizes.

If we take character size on the horizontal axis and frequency on the vertical axis, the frequency of character sizes handled by the character recognition device is 48 x 48 doso] - This shows that smaller characters are more common than larger characters. There is.

In other words, characters that are smaller than the frame■ are often handled, and the frame■
Larger characters also include characters that are larger in the horizontal direction and characters that are larger in the vertical direction, but conventionally, all of them are enlarged/reduced so that they are inscribed in the four sides of the frame ■, so the feature extraction time is constant. , there is a problem that the normalization time for enlarged characters increases in proportion to the size of the characters.

In view of this problem, the present invention moves characters smaller than the frame ■ so that they touch two sides of the frame ■ without enlarging them, thereby avoiding scanning of blank areas within the frame ■. This eliminates the scanning time required for enlargement during normalization, and also reduces the scanning time required for feature extraction. In addition, for characters larger than the frame ■, the larger one in the horizontal or vertical direction touches the two horizontal or vertical sides of the frame ■ψ, eliminating the need to scan blank areas and shortening the feature extraction time. The purpose is to increase recognition processing efficiency.

[Means to solve the problem]

FIG. 1 is a process flowchart showing the configuration of the present invention.

FIG. 1 shows the flow of processing by a processor that controls a character recognition device, and 1 to 19 are processing steps.

[Effect]

Cut out one character from an unknown character string, scan the area of the cut out character to extract features, and compare them with features read from a dictionary that stores features obtained from characters inscribed in a square of a predetermined size, The processor of the character recognition device that selects a character in the dictionary with a small degree of difference as a candidate character and sends out the selected candidate character as a recognition result calculates the vertical and horizontal size of the character cut out in processing step 11. Check whether the size is larger than the square.

The vertical and/or horizontal size of one character is
If it is larger than the square, in step 12, a reduction rate is calculated to reduce this one character so that the larger horizontal or vertical side is inscribed in the two sides of the square.

Then, using the reduction ratio calculated in step 13, one character is reduced and normalized, and in step 14, the reduced character area is scanned to extract features.

FIG. 2 is a diagram illustrating an example of feature extraction.

For example, after reducing and normalizing the characters as shown in Figure 8 (bl or (C1), for example, the width is It is an inscribed character.

Brosset 4no. controls the feature extraction unit to scan the characters indicated by diagonal lines in Figure 2 ta+ in the same manner as above, but at this time, the characters are moved within the frame ■, and for example, the upper part of the character is moved into the frame ■.
Create a blank area [phase] by inscribing this blank area [
Phase] scanning is not performed. Then, features are extracted using a known method.

In step 15, the processor performs an operation to correct the position of the extracted feature corresponding to the area that becomes blank due to the reduced character with respect to the square, that is, the blank area [phase].

In other words, when normalizing a single character, the features extracted by the feature extraction unit are not reduced so that they are inscribed in the four sides of the frame ■, so it is assumed that the part corresponding to the blank area [phase] of the frame ■ has been reduced. Although it is a feature based on similar character images, the dictionary has a frame
Since the features extracted from the character images inscribed on the four sides of are stored, the positions are corrected so that the features correspond.

Next, in step 18, the processor compares the features of the dictionary with the features of the position-corrected character, and in step 19, selects candidate characters.

Further, in step 11, if the length and width of one character are smaller than a square, the processor executes step 16.
The feature extraction unit is controlled to scan an area of one character and extract features.

The feature extraction unit scans the characters indicated by diagonal lines in Fig. 2 (bl) in the same manner as described above, but at this time, the characters are moved within the frame ■, for example, the upper and left side of the characters are inscribed in the frame ■, A blank area (■) is created, this blank area (■) is not scanned, and features are extracted using a known method.

Here, in step 17, the processor selects an area that becomes blank due to one character for the square, that is, a blank area 0.
Corresponding to this, a calculation is performed to correct the position of the extracted feature.

In other words, the features extracted by the feature extraction unit are based on character images that are similar to those in which one character is not expanded to be inscribed in the four sides of the frame ■, but the part corresponding to the blank area ■ of the frame ■ is reduced. , Since the dictionary stores the features extracted from the character images inscribed on the four sides of the frame ■, the positions are corrected so that the features correspond.

Next, in step 18, the processor compares the features of the dictionary with the features of the position-corrected character, and in step 19, selects candidate characters.

As described above, when scanning one character and extracting features, the blank area [phase] or the area corresponding to Since there is no need to extract 432-dimensional features and the scanning area is reduced, the feature extraction processing time can be shortened.

Then, the processor corrects the position of the feature extracted by the feature extractor based on the ratio between the size of the cut out character and the size of the frame ■, but since the calculation time is short, the feature extraction processing time of the feature extractor is The shortening effect is significant.

Further, during normalization, if the cut out character is smaller than the box (■), the enlargement process is not performed, which makes it possible to shorten the normalization time.

Therefore, the processing efficiency of the character recognition device can be improved.

〔Example〕

FIG. 3 is a block diagram of a circuit showing one embodiment of the present invention.
FIG. 4 is a flowchart 1- explaining the operation of FIG.
FIG. 5 is a diagram illustrating the operation of FIG. 3.

The same reference numerals as in FIG. 7 indicate the same functions. Processor '1 reads a program stored in program memory 12 and operates. That is, as shown in FIG. 4 Φφ,
When character recognition is instructed from the host device via the interface unit 9, the scanner 10 is controlled to read one page of character strings from the document, and store them in the image area of the image memory 5. Then, as shown in FIG. 5 (al), the character position on the image area of the image memory 5 is detected.

That is, the coordinate values x and y of tal in FIG. let

Then, the cut out characters are compared with 48 dots.

That is, as shown in FIG. 4, it is checked whether W>48 or not, and then it is checked whether h>48.

If W and 11 are smaller than 48 dots, the cut out character is moved to the normalized image area of the image memory 5, as shown in the lip in FIG. 4, but if W and/or h are larger than 48 dots, , to find the normalized reduction ratio, use Fig. 4 0
As shown in , check whether W≧h or not, and if W≧h, k
= w, and if W≧h, then k = h, p =
48/k to find the reduction rate p.

The processor 1 controls the normalization unit 4 to scan a single character area stored in the image memory 5 from side to side as shown in FIG. 5 tbl, and creates a table as shown in FIG. 5 (C1).

That is, for example, the positions a and b of the starting and ending points of a black pixel on the scanning line indicated by @ in FIG.

Determine r, g, and j and create a table as shown in Figure 5 (C1).

Next, the processor 1 creates a conversion table by multiplying each pixel position in the table shown in FIG. 5 (C1) by p as shown in FIG.
The following calculations are performed to create a conversion table as shown in FIG. 5 fdl.

Next, the processor 1 performs conversion processing on the same data, as shown in FIG. This means that when reducing a character with many line segments, adjacent line segments will overlap, but to avoid this overlap, either delete the overlapping line segments or
5 (dl conversion table).

Next, processor 1 uses this conversion table to create characters based on the conversion table. That is, FIG. 5(e)
As shown in FIG. 3, for example, if W≧h, a character with a width of 48 dots and a height of h is created and stored in the image memory 5.

Then, scan this character up and down to create a chiffle indicating the positions of the start and end points of black pixels, and multiply each pixel position in this table by p to create a square as shown in Figure 5 (fl).
A conversion table is created that creates characters with a width of 48 dots and a height of l-1=hXp.

Then, as in i7, the same data is converted, and the normalization unit 4 is controlled from this conversion table as shown in Fig. 5(f).
A normalized image as shown in FIG. 1 is created and stored in the normalized image area of the image memory 5.

Next, the processor 1 calculates the smaller height and width of the character by H=pXh or W=p×W as shown in FIG. 4 As shown in Figure 0, feature extraction is performed within the range of 48 x H or 48 x W. That is, for example, if the width is 48 dots and the height is H
Scan the character range and extract the features, or scan the character range of width W and height I that was moved to the normalized image area in Figure 4 0 and extract the features. Ru.

Therefore, in this case, as shown in FIG. 2 (al and (bl), the blank area [phase] or the area of ■ is not scanned.

The processor 1 stores the features extracted by the feature extractor 6 in the image memory 5, and extracts the features from the characters inscribed in the 48 x 48 square frame and compares them with the features in the dictionary 8 stored therein. , as shown in Figure 4 [Phase], 48/H
Alternatively, the positions are corrected by multiplying by 48/W, and the special scenery is also corrected so that the same characters have the same characteristics.

Then, the matching unit 7 is controlled to match the features of the dictionary 8,
Candidate characters are selected and stored in the image memory 5. Then, it is checked whether all the characters for one page have been selected, and if not, the routine returns to cutting out the next character, and if it is, the answer is sent to the host device via the interface section 9. Send.

Then, it is checked whether there is a next page, and if there is, the routine returns to the routine of controlling the scanner 10 to read one page's worth of character strings from the original image, and if there is no next page, the operation is ended.

FIG. 6 is a diagram explaining the invention in detail.

Compared to Fig. 9, the time required for normalization indicated by [phase] is
Only the shaded area corresponding to enlargement needs to be moved, which saves time. As for the time required for feature extraction indicated by ■, since the scanning range required for feature extraction becomes smaller, the smaller the character, the shorter the feature extraction time becomes.

In the feature correction and position correction shown in (2), the time required for position correction is longer than in FIG. 9. Therefore, as shown in [phase] (
, the range shown by diagonal lines is effective for the total time.

Note that when the cut out characters are reduced in size in Figure 6, scanning of the blank areas shown in Figure 2 is no longer necessary, so the time required for feature extraction is saved. Since it cannot be understood accurately, illustration is omitted.

〔Effect of the invention〕

As described above, since the present invention can shorten the feature extraction processing time, it is possible to improve the character recognition processing efficiency of the character recognition device.

[Brief explanation of the drawing]

FIG. 1 is a process flowchart showing the configuration of the present invention, FIG. 2 is a diagram explaining an example of feature extraction, FIG. 3 is a block diagram of a circuit showing an embodiment of the present invention, and FIG. Feature I-Character I, which explains the operation of the figure.
Figure 5 is a diagram explaining the operation of Figure 3, Figure 6 is a diagram explaining the invention in detail, Figure 7 is a professional diagram explaining the conventional technology, and Figure 8 is the operation of Figure 7. FIG. 9 is a diagram illustrating the time required by the normalization section and the feature extraction section. FIG. 10 is a diagram illustrating the distribution of character sizes. In the figure, ■ is a processor, 2 and 12 are program memories,
3 is a character extraction section, 4 is a normalization section, 5 is an image memory, 6 is a feature extraction section, 7 is a matching section,
Reference numeral 8 indicates a retractor, 9 indicates an interface unit, 10 indicates a scanner, and 11-19 indicates a processing step. Book a, @ (to Njo; 11 hari t・, Tare Shouge 1
2 White Paper on Snowbird Mosquitoes with Guko A> -1゛1 Σ3 Sunset Moon Maro Composition 2! 1' 箮 - September 1st Sarasatsu → h Ikutsu and i 卜 Ding? The front of the road should be 2 (3) and 3 (kiss 3)) 1; Self 2 Heather 5 (Part 1) (d) (, z) Brown-3 Gi Naka 11χ Maya Kinu 1 Self/Tagumi vine 5 Ship
ζ Yan's 2) Hands up the talented writer i' digit and the north bank 5 rod q 2! l Goro Kinto! ! 8 Pa sword moxibustion Στ beg B month 9 Ka Ru Figure (θ) (C)] 47 days 1 Hi υ lower and 1 yuan 9UT do Otsu GO ヰ 6
When you open your mouth, you can say that you have a father's field, power, and residence. Inquiry ward matters 10 ■

Claims (1)

[Claims] Cut out one character from an unknown character string, scan the area of the cut out character to extract features, and read out from a dictionary that stores features obtained from characters inscribed in a square of a predetermined size. In the process of a character recognition device that selects a character in the dictionary with a small degree of difference as a candidate character, and sends out the selected candidate character as a recognition result, Compare the size of the character with the size of the square (11), and if the vertical and/or horizontal size of the character is larger than the square, the larger vertical or horizontal size of the character is placed in the square. Find the reduction ratio when inscribed (12), reduce the character by the reduction ratio (13), scan the area of the reduced character to extract features (14), Corresponding to the area that becomes blank due to the reduced character, position correction is performed on the features extracted from the reduced character (15), and the vertical and horizontal sizes of the extracted character are determined to be the same. If it is smaller than a square, the area of the one character is scanned to extract the feature (16), and the position of the feature extracted from the one character is corrected in accordance with the area that becomes blank due to the one character with respect to the square. (
17) A character recognition method, characterized in that the position-corrected feature is compared with the feature read from the dictionary (18), and candidate characters are selected (19).