JPH0583955B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a document recognition device that optically reads characters, and particularly to a document recognition device that reads handwritten documents written on paper without character frames.

(Prior art) Conventionally, as a feature extraction means for handwritten character recognition,
It has been widely reported that a method of creating a direction image in which each line element of an input image is classified according to direction, and performing matching with a character dictionary on this direction image is very effective.

As an example of a character recognition method that performs matching with a character dictionary on a directional image, see the document Michio Yasuda et al., “Improvement of correlation method for character recognition,” Journal of the Institute of Electronics and Communication Engineers.
79/3 Vol.J62-DNo.3 pp217-224 (1979),
References Noboru Babaguchi et al. “A method for extracting directional segments of character patterns without thinning” Transactions of the Institute of Electronics and Communication Engineers '82/7 Vol.J65-D No.7 pp874-881
(1982) etc.

There are various methods for creating a direction image as a feature extraction means for this purpose, but all of them involve checking the connection state of each black pixel on predetermined two-dimensional coordinates and creating a direction image based on the connection direction. It is something to create. Examples of directional image creation methods include the above-mentioned documents.

In addition, as a method for reading a character string by directly inputting a character string image without separating each character from the character string, Japanese Patent Application No. 62-11630 filed by the same applicant as the present application
There is a ``character string recognition device''.

(Problems to be Solved by the Invention) Creation of directional images is an important means in recognizing handwritten characters.

All conventional direction code images are created based on absolute coordinates. Since conventional handwritten character recognition targets characters written within a character frame, fluctuations in the direction of line segments are relatively small, making it possible to create a stable directional image. However, when characters are written on paper without a character frame, the direction of the line segments varies greatly due to the scribe's habits. Especially in horizontal line segments, there are many characters that slope upward to the right, so the range of variation is about 50 to 60 degrees. As an example, FIG. 2 shows an example of two character strings. The characters that humans perceive as horizontal lines are written horizontally in Figure 2 (a), but in Figure 2 (b) they are written approximately 45 degrees upward to the right.

FIGS. 3 and 4 show direction images in which a direction code has been added in a conventional manner to the character "upper" in FIG. 2. 3 is a direction image created from the character "above" in FIG. 2a, and FIG. 4 is a direction image created from the character "above" in FIG. 2b. In this way, the line segment corresponding to the line segment extracted in the horizontal direction in the character of FIG. 2a is extracted as a line segment pointing upward to the right in the character of FIG. 2b.

In order to perform high-performance correspondence between an input image and a character dictionary on a directional image, it is necessary to stably extract corresponding line segments on the same directional image. However, the conventional method of correcting the direction code for each character is unstable because there are a variety of characters, and it is necessary to correct each character correctly before reading the character from a freely written string. Difficulty in cutting it out makes it even more unstable.

In addition, even if the problem of processing amount is not considered, it is difficult to perform direction code processing on the entire document image, even if the document contains tables, figures, etc., printed characters, etc. They may contain characters written by multiple scribes, making them unstable.

Therefore, in order to stably read characters freely written on paper without a character frame, it is necessary to extract characters stably before reading them, and to extract characters in units of character strings whose internal character properties are stable. It is considered most desirable to treat as much as possible.

(Means for Solving the Problems) In order to solve the above-mentioned problems, the character string recognition method provided by the present invention inputs a two-dimensional image of a character string, which is a sequence of characters, and a direction assigning means for measuring which line segment a pixel belongs to in a predetermined direction and assigning the measured direction to each black pixel; and a direction assigned to each black pixel. means for obtaining a separate distribution; means for reassigning the direction assigned by the direction assignment means based on the relative coordinates for each character string determined from the distribution for each direction; and character string recognition means for sequentially reading the obtained character string direction image based on a standard character pattern.

(Function) When determining the direction of a character line segment, humans do not use the absolute direction based on fixed coordinates, but rather the direction of the line segment based on relative coordinates that vary from character to character. It is considered that the

Furthermore, it is thought that the variation in the direction of character line segments is largely due to the habits of the scribe, and the direction of line segments of characters written by the same scribe at the same time is considered to be stable. For example, it is highly unlikely that a scribe will write a character that slopes upward to the right as the first character and a character that slopes down to the right as the second character.

Therefore, in the present invention, the direction of a character is first measured on predetermined coordinates from a given image, and the absolute direction of a line segment of the input image is determined. Next, the distribution of character lines in each direction is determined. At this time, if we calculate the directional distribution of a single character, the distribution will not be stable depending on the type of character, and the directional distribution of the entire document will not be stable due to influences other than those to be read. However, if we calculate the distribution by character string, , Japanese characters are mostly composed of conceptual horizontal and vertical lines, and as mentioned above, due to the stability of writing by the same person, the horizontal and vertical directions that serve as the standard for the character string. The direction on the image can be determined.

By determining relative coordinates for each character string based on this distribution and reassigning direction codes based on these coordinates, it is possible to obtain a stable direction image.

The relative coordinates referred to here refer to the correspondence between a conceptual line segment direction and a direction on a two-dimensional image. The relative coordinates determined from FIGS. 2a and 2b are shown in FIGS. 5a and 5b, respectively.

Then, rather than cutting out characters from this character string image before reading them, we use the character string image as an input image and read the characters sequentially from the beginning, thereby determining the shape and character of the character. This makes it possible to read character strings written on sheets of paper without character frames, which have large variations in their arrangement.

(Example) Next, the present invention will be explained in detail with reference to the drawings.

FIG. 1 is a flowchart for explaining one embodiment of the present invention, and FIG. 2 is a diagram showing a sequence of input character string images to be recognized by this embodiment. Figure 2 shows
The perspective view is shown as a character line part, and the other parts are shown as a background part. The input character string image is quantized at appropriate intervals.

In this embodiment, the direction of each line segment is measured in step 2 for the image input in step 1.
The method of measuring the direction of each line segment is based on a conventional technique, and detailed explanation will be omitted here. The directions to be measured are also quantized at appropriate intervals corresponding to the number of direction images ultimately required. Below, as an example, the final result is horizontal, upward to the right, vertical, upward to the left.
Assuming that two directional images are to be created, here it is assumed that measurements are taken in eight quantized directions shown in FIG.

Among the direction images created for each direction measured with respect to the character "above" in the image a in FIG. 2, the direction images corresponding to the direction codes 1, 3, 5, and 7 in FIG. FIG. 3 shows an upward direction to the right, a vertical direction, and an upward direction to the left. Similarly, b in Figure 2
Figure 4 shows the measurement results for the character "upper" in the image. FIGS. 3 and 4 are direction images that are images of black pixels in different directions.

Based on the measurement results in the line segment direction, in step 3, the distribution in the line segment direction is determined for each quantization direction.

Then, the most distributed direction in quantization directions 1, 2, 3, and 8 is taken as the horizontal direction, and the most distributed direction in horizontal directions 4, 5, and 6 is taken as the vertical direction, and the intermediate direction is defined as upward-rightward and upward-leftward directions. direction.

For the image in Figure 2a, for 1, 2, 3, and 8, the distribution in the 1 direction is the largest, and in 4, 5, and 6, the distribution in the 5 direction is the largest, so the image shown in Figure 5a is The relative coordinates are obtained, and for the image in Figure 2 b, the distribution in the direction of 3 is the largest for 1, 2, 3, and 8, and the distribution in the direction of 5 is the largest for 4, 5, and 6. Relative coordinates as shown in FIG. 5b are obtained.
The numbers in parentheses represent the quantization direction in FIG.

Direction codes are assigned in step 4 based on this coordinate axis. In this case, the intermediate direction is an ambiguous direction, and the two adjacent directions are
Can be distributed in both directions.

In the example in Figure 2a, direction codes 1, 2, and 8 are horizontal, 2, 3, and 4 are upward to the right, 4, 5, and 6 are vertical, and 6, 7, and 8 are upward to the left. An image is created. The results are shown in FIG.

In the example in Figure 2b, direction codes 2 and 3 are horizontal, 4 is upward to the right, 5 and 6 are vertical, 6, 7,
8 and 1 are upward left directions, and respective direction images are created. The results are shown in FIG.

In this way, a stable directional image can be created. Then, in step 6, this character string direction image is recognized. The method of sequentially reading this character string direction image based on the standard character pattern is described in the above-mentioned Japanese Patent Application No. 11630/1983 "Character string recognition device".
Use the same method as . The details are omitted here.

(Effects of the Invention) As described above, according to the present invention, the line segment direction of each character can be created based on the relative coordinates of each character string, not as the absolute value of the coordinates given in advance. This makes it possible to create a stable directional image that is close to that of a human. Since character string images are mainly handled until character recognition, even character strings written on paper without character frames can be recognized stably and accurately.

Furthermore, in the present invention, since the measurement in the direction that takes the longest processing time is performed only once, the processing time is not much different from the conventional method.

As explained above, according to the present invention, when reading a handwritten character string written on paper without a character frame, in addition to fluctuations in the arrangement of characters and fluctuations in the position and length of each character line, By stably extracting fluctuations in the direction of character lines, which are thought to be a human habit, for each character string, it is possible to process character strings written relatively freely on paper without a character frame at the same processing speed as before. Character strings can be recognized with higher accuracy than before.

[Brief explanation of drawings]

FIG. 1 is a flowchart showing the operating procedure of an embodiment of the present invention, FIGS. 2 a and b are diagrams of a character string image showing an example of an input image, and FIG. Fig. 4 is a diagram of a directional image obtained by measuring the character image in Fig. 2 b using the conventional method, and Fig. 5 is an illustration of the input image of Fig. 2. FIG. 6 is a diagram showing an example of the definition of quantization for direction measurement. FIG. 7 is a diagram showing the direction obtained for the character image in FIG. 2 a by applying the present invention. The image diagram, FIG. 8, is the same as that shown in FIG.
FIG. 3 is a diagram of a direction image obtained for a character string image.

Claims (1)

[Claims] 1. Take a two-dimensional image of a character string as an input, measure to which line segment each black pixel of the input image belongs to a pre-given direction,
A direction assigning means for assigning the measured direction to each black pixel, a means for determining a direction distribution of the directions assigned to each black pixel, and a relative coordinate for each character string determined from the direction distribution. and means for reallocating the direction given by the direction giving means;
A character string recognition method comprising a character string recognition means for sequentially reading character string direction images obtained by reallocation by the reallocation means based on a standard character pattern.