JPH0365781A - Pattern normalizing system - Google Patents

Abstract

PURPOSE:To improve a recognition rate by defining either maximum height or width in a row or a column, where a character symbol is positioned, as a reference side, defining the other side as dot existing width in a rectangular area, segmenting attention character symbol pattern in this rectangular area and variably magnifying the height and width with the ratio of the reference side to a normalizing size. CONSTITUTION:The character row is detected by a character row detection part 3 and the maximum height of the character row is calculated by a maximum height detection part 5. A character rectangle is detected by a character rectangle detection part 6 and the width of the character is detected by a character width detection part 8. For the rectangular area, either the maximum height or width in the row or column, where the character symbol as the object of normalization is positioned, is defined as the reference side and the other side is defined as the dot existing width of the attention character symbol pattern. Then, when the attention character symbol pattern is segmented in such a rectangular area and normalized by a character normalization part 9, the size of the attention character symbol pattern in height and width direction is variably magnified by the ratio of the reference side to the normalizing size. Thus, since the character is normalized while keeping original feature quantity, the recognition rate of the character is improved.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a pattern normalization method, and more particularly to a method for normalizing a character pattern to be recognized, which is performed at a stage prior to character recognition.

[Prior art] Conventionally, when normalizing a character pattern to be recognized,
The white pixel portions above and below or on the left and right sides of the rectangular character pattern are removed, and normalized ratios are calculated separately for each length and width.

[Problem to be solved by the invention] However, when normalization is performed using the conventional processing described above,
That is, if white pixel portions are removed and magnification is changed vertically and horizontally, the information of characters, symbols, etc. that are offset vertically or horizontally will be lost, which poses a problem.

For example, overline "-", minus sign underline ", etc. all became "■". Also, the difference between an uppercase letter "tsu" and a lowercase letter is large in the recognition results after conventional normalization processing. Because the information was lost, it was not possible to distinguish and recognize them.

The present invention has been made in view of such problems, and it is an object of the present invention to provide a pattern normalization method that makes it possible to obtain good recognition results.

[Means for Solving the Problem] and [Operation] The pattern normalization method of the present invention that solves this problem has the configuration shown below. In other words, in the normalization method of extracted character patterns performed in character recognition preprocessing, one of the maximum height or width in the row or column where the character symbol to be normalized is located is used as the reference side, and the other side is used as the target character symbol. When cutting out the character symbol pattern of interest in a rectangular area set as the dot width of the pattern and normalizing this cut out character symbol pattern, the height of the character symbol pattern of interest is determined by the ratio of the reference side to the normalized size. Change the size in the width direction.

[Embodiments] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a schematic configuration diagram of an apparatus according to an embodiment, and the present embodiment targets character recognition of horizontally written documents.

First, a document image is read by an input unit 1 such as a scanner and stored in an image memory 2. From this image memory 2, a character line is detected by a character line detecting section 3 and stored in a character line buffer 4. Furthermore, the maximum height of the character line is determined from the detected character line by the maximum height detection unit 5. Further, a character rectangle is detected from the character line buffer by the character rectangle detection unit 6 and stored in the character rectangle buffer 7. When a character rectangle is detected, the character width detection unit 8 detects the width of the character,
On the other hand, the height of the detected character line is inputted to the character normalization unit 9, and character normalization processing is performed. Furthermore, the feature extraction unit 10 extracts parameters representing the characteristics of the normalized characters, and the recognition unit 11 recognizes the characters by comparing them with a dictionary 12 that stores the characteristics of each character that have already been obtained. Do this.

FIG. 2 is a flowchart showing the processing of the feature extraction section.

In step S1, the maximum height of the character line is h, and in step S
2, the rectangular width of the character is set as d, and from these h and d, in step S3, the one that takes the maximum value is set as the reference side W. Then, in step S4, the normalized ratio r
In step S5, image normalization processing is performed to obtain a normalized character rectangle by determining = fl / w and multiplying each of the vertical and horizontal directions of the character rectangle by this r.

Note that, in the above-described processing, a known method is used for detecting character string lines, and a detailed description thereof will be omitted.

FIG. 3 shows an example of the detection results of character rectangles. The height of each character rectangle was set to "h" (in the illustrated case, the height of the character r degrees J), which is the maximum height of the line.

That is, this maximum height h is used as the height of other characters. Note that the maximum height h can be determined by examining the distribution of dots forming the area in front of the object of interest.

Furthermore, the width of each character was made to match the horizontal range of the dots that make up each character. This can also be determined from the dot distribution.

Now, suppose that the rectangle of the period symbol is determined to have a height h and a width d, as shown in FIG. 4(a), through the process described above. Between h and d, h is larger. Therefore, in order to make this the normalized size β×β, the ratio n/h is used as the normalized ratio and the scaling process is performed.

That is, for the height, of course, the height h of the rectangle is multiplied by the normalized ratio fl/h, and the width d is also multiplied by β/
By multiplying by h, a pattern of normalized size as shown in FIG. 4(b) is obtained (normalization processing is shown). The pattern scaling process itself may be performed by interpolating between dots.

The width d' (=axg/h) after scaling is d'
<Since this is a problem, a blank area (consisting of white pixels) of (g-d')/2 was added to the left and right sides of it to obtain a normalized character rectangle size of ℃×4.

Incidentally, in the conventional normalization processing, the normalization processing is performed on the character rectangle excluding the blank portions on the top, bottom, left, and right of the pattern of interest. Therefore, the pattern after the normalization process becomes as shown in FIG. 4(C), and it is difficult to tell whether it is a period or not.

However, according to the normalization process of this embodiment, as shown in FIG.
), with a blank area at the top.

exists, so by extracting the feature amount from this image, it becomes possible to recognize it as a period without adding any other special information.

FIGS. 5A to 5C show the overline, minus sign, and underline patterns normalized to the normalized size, respectively. As shown in the figure, even if each pattern is normalized, each characteristic position (one of the characteristic amounts) within the normalized size is maintained. In other words, characters can be recognized by extracting feature amounts from this normalized size. still,
In the case of FIGS. 5(a) to 5(C) as well, the vertical and horizontal lengths are compared and normalized based on the ratio of the longer length to the normalized size, as described in FIG. 4 above.

However, for character shapes with significantly different lengths and widths, if normalization is performed using the magnification of each height and width, the original shape and position will be
They become distorted and cannot be distinguished from other letters and symbols. For example, if overlines, minus signs, underlines, etc. are normalized using this process, they all end up as shown in Figure 5(d), but according to the normalization process of this embodiment, such things do not occur. do not.

As described above, according to this embodiment, it is possible to maintain the original feature amount of the normalized pattern, and therefore it is possible to improve the recognition efficiency in character recognition processing.

In the embodiment, the document to be recognized is written horizontally, but if the document is written vertically, the maximum width may be detected first, and this maximum width may be cut out as the width of each character. .

[Effects of the Invention] As explained above, according to the present invention, since normalization is performed while maintaining the original feature amount, it is possible to improve the character recognition rate.

[Brief explanation of drawings]

Fig. 1 is a plot configuration diagram of the character recognition device in the embodiment, Fig. 2 is a flowchart showing the processing contents of the normalization section, Fig. 3 is a diagram showing the principle related to character rectangle detection, Fig. 4 (a
) is a diagram showing the results of the normalization process of the example, Figure 4(b) is a diagram showing the results of the conventional normalization process, and Figures 5(a) to (c) are the results of the normalization process of the example. Figure 5(d) shows the results of conventional normalization processing. In the figure, 1... document image, 2... character line detection processing, 3
. . . Character rectangle detection processing; 4 . . . Each character rectangle normalization processing. Figure 2 Figure 3 Figure 4 Figure 5

Claims (1)

[Claims] In a normalization method for extracted character patterns performed in preprocessing for character recognition, one of the maximum height or width in a row or column in which a character symbol to be normalized is located is used as a reference side, and the other side is set as a reference side. When cutting out the character symbol pattern of interest in a rectangular area with the dot existence width of the character symbol pattern of interest and normalizing this extracted character symbol pattern, the ratio of the reference side to the normalized size is A pattern normalization method characterized by scaling the size in the height and width directions.