JPH03262087A - Method for detecting position of character string or character line and character recognizing device - Google Patents

Abstract

PURPOSE:To make it possible to accurately segment a character string or a character line even when a noise exists or contrast is low by remarking the shape of the charac ter string or character line appearing on a projected waveform and detecting a segmenting position. CONSTITUTION:A waveform to be projected in a direction vertical to a character string or character line is found out and reference waveforms each of which has a pedestal part with a prescribed length set up on both the sides of two pulses corre sponding to the width of the character string or character line. The average value of the projected waveform part corresponding to the pedestal part of the reference waveform and the average value of the projected waveform part corresponding to the pulse width part of the reference waveform are computed to find out the pedestal value and peak value of the reference waveform and the integrated value of the square of the error between the computed reference waveform and the projected waveform is found within a range in which a difference from the reference waveform is larger than a previously determined threshold. The position minimizing the integrated value is set up as the position of the character string or character line. Even when a noise exists or the contrast between the characters and their background is low, the charac ter string or character line can be accurately segmented.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to a method for cutting out character strings or character lines in a character reading device. It relates to technology for accurately detecting the position of character strings or character lines. The present invention also relates to a character recognition device that extracts recognition target characters from an image containing the recognition target characters and performs recognition.

[Conventional technology] OCR (0pical character
Character recognition technology has been widely put into practical use in industry, as exemplified by the spread of optical character reading devices (optical character reading devices), and it has greatly contributed to improving office efficiency. After optically reading the target document to obtain an image, detecting the position of character strings or character lines in the image, and cutting out each character, various character recognition methods are applied. In general, single-character recognition methods (which are based on the premise of sequentially processing extracted characters one by one) make it difficult to detect the position of character strings or character lines even in the practical application of character reading devices. This is an extremely important technology.

The conventional method for detecting character strings or character lines is to add the density of each point on the image parallel to the character string or character line.
A method of detecting it from its distribution state is being used. Character strings and character lines differ only in the way the characters are arranged, but are essentially the same in terms of extraction, so the extraction of character lines will be explained below.

Figure 9 shows a method for detecting the position of character line 1 from a document image. First, the image density (pixel value) is added in the horizontal direction, and then projected on the vertical axis to form a projected waveform (projected value ) get 2. By comparing the projected waveform 2 with a predetermined threshold TH, the character line position can be detected.

However, this method has the disadvantage that it is easily affected by noise superimposed on the image.
If there is noise 3 as shown in the figure, if the threshold value is set as low as THI, the position of noise 3 will be mistaken as a character line position. If the threshold value is set as high as TH2, the influence of noise will be reduced (although it may result in part of one character line being missing or not being able to capture a character line).

On the other hand, in JP-A-63-129485,
A detection method for character segmentation is disclosed that prevents the influence of noise by setting the threshold value. This compares the horizontal or vertical projection value of the binary image with a threshold value to detect a character line or character, and determines the starting position of the character line or character in the character extraction process that extracts characters from the binary image. By setting the threshold for determining the end position of a character line or character lower than the threshold for determining the end position of a character line or character, noise is removed when determining the start position, and image loss is prevented when determining the end position. This is what I am trying to do.

[Problems to be Solved by the Invention] However, the above-described detection method for character segmentation can only reduce the effects of noise below the same level as the threshold value, and can only reduce the effects of noise at the same level or higher than the character line. If it is heavy, there is a problem in that it is not possible to distinguish between character lines and noise, and it is not possible to accurately cut out only character lines. In documents, examples of such noise include underlining, borders around text, and letterhead. In addition, when recognizing characters that are marked on products or intermediate products for management purposes rather than on printed matter, there are even more noises that impede extraction, such as background patterns, protruding paint on markings, and characters that are not recognized. (Also, since the difference in density (contrast) between the characters and the background is small, it is difficult to set an appropriate threshold value for the projection value, and there is a problem that good extraction cannot be performed.

In this way, with conventional technology, when there is noise of the same level or higher than the character string or character line to be extracted, or when the contrast between the characters and the background is low, the character string or character line cannot be cut out. There is no way to properly distinguish noise and perform accurate extraction, and as a result, noise can sometimes be mistaken for character strings or character strings can be overlooked, which has been a major obstacle in the practical application of character recognition.

The present invention has been made to solve the above-mentioned conventional problems, and when there is noise at a level equal to or higher than the character string or line to be extracted,
A method for detecting the position of character strings or character lines for appropriately distinguishing character strings or character lines from noise and performing accurate extraction even when the contrast between characters and the background is low, and a character recognition device suitably using this method. The challenge is to provide this information.

[Means for Solving the Problems] The present invention solves the above-mentioned problems. (1) A method invention includes determining the position of a character string or character line containing the recognition target characters from an image including the recognition target characters. The following method was applied to the detection method. In other words, ■ add the density of the image parallel to the character string or character line to obtain the projected waveform in the direction perpendicular to the character string or character line, and ■ determine the character string or character from the width and background of the character string or character line. Set a reference waveform with a pulse corresponding to the row width and a pedestal part of a predetermined length on both sides of the pulse. ■ Move the reference waveform along the projected waveform and compare it with the projected waveform at a predetermined period. The average value of the projected waveform portion corresponding to the pedestal portion of the waveform and the average value of the projected waveform portion corresponding to the pulse width portion of the reference waveform are calculated to obtain the pedestal value and peak value of the reference waveform, respectively, and the waveform of the reference waveform is calculated. In the range where the difference between the high value and the pedestal value is larger than a predetermined threshold, calculate the integral value of the square of the error between the calculated reference waveform and the projected waveform, and The position of the character line.

This is a method for detecting the position of a character string or character line.

The minimum value may be set to a value obtained by dividing the integral value of the square of the error by the difference between the peak value of the reference waveform and the pedestal value.

Further, the minimum value may be set to a value obtained by dividing the integral value of the square of the error by the square of the difference between the peak value of the reference waveform and the pedestal value.

Furthermore, the minimum value may be set to the integral value of the absolute value of the error.

In addition, the minimum value can be set to a value obtained by dividing the integral value of the absolute value of the error by the difference between the peak value of the reference waveform and the pedestal value.

(2) The device invention includes an image input section for inputting an image containing the recognition target character, and a character string or character line cutting section for cutting out a character string or character line containing the recognition target character from the image input by the image input section. a character extraction unit that extracts recognition target characters from an image of the character string or character line extracted by the character string or character line extraction unit; and a character recognition unit that recognizes the recognition target characters extracted by the character extraction unit. It was applied to a character recognition device, and the following technical measures were adopted.

Namely.

Character string or character line extraction part.

■A projected waveform forming unit that calculates a projected waveform in the direction perpendicular to the character string or character line obtained by adding the density of the image parallel to the character string or character line, and ■The width of the character string or character line. and a reference waveform setting unit that sets a reference waveform having a pulse corresponding to the width of a character string or character line from the background and a pedestal portion of a predetermined length on both sides of the pulse; The pedestal value and peak value of the reference waveform are calculated from the average value of the projected waveform part corresponding to the pedestal part and the average value of the projected waveform part corresponding to the pulse width part of the reference waveform, respectively. and (1) a comparison calculation unit that compares the calculated reference waveform and the projected waveform, performs a predetermined calculation on the error between the reference waveform and the projected waveform, and detects the minimum value. It is a character recognition device.

[Operation] In detecting the position of a character string or character line from a projected waveform, the present inventor has developed a method based on the information of only the density level, based on the premise that the density level of the character part is higher than the background part, as in the past. Instead of cutting out, the shape of the character string or character line that appears on the projected waveform is scanned, the projected waveform is scanned using a preset reference waveform, and the two are compared at a predetermined period to determine the reference waveform. Calculate the pedestal value and height to find the reference waveform at that point,
We have invented a new method and device for comparing the calculated projected waveform with its reference waveform and detecting the position most similar to the reference waveform in terms of shape as the cutting position of a character string or character line. .

A description will be given below regarding character lines, but the same applies to character strings.

Since the projected waveform is obtained by adding the density of the image parallel to the character line, a rectangular convex portion (pulse) corresponding to the width of the character line is generated on the projected waveform. Similar protrusions occur when there is noise, but in underlines, outer frames, etc., the width of the protrusions is different from the character line, and in cases of smudged printing, protruding paint, etc., the shape of the protrusions is different from that of the characters. Unlike lines, they have the characteristic that they are not rectangular. Therefore, a rectangular (pulse) reference waveform is set by predicting the character line waveform that appears in the projected waveform based on the width of the character, and the projected waveform and the reference waveform are compared to find the most similar waveform in terms of shape. If the position is the character line position, character lines and noise can be quickly distinguished.

In order to evaluate the similarity between the projected waveform and the reference waveform, it is conceivable to first use a cross-correlation function between the two. but,
The inventor's experiments have revealed that when the noise level is extremely high, the value of the noise portion becomes larger than that of the character line portion, resulting in erroneous detection.

Therefore, in the present invention, the square of the error between the projected waveform and the reference waveform or the integral of the absolute value of the error are used as the similarity evaluation value.

However, even if the shape of the character line waveform that appears in the projected waveform can be predicted in advance, the values of each part cannot be known in advance. ) and background part (pedestal) values cannot be set. Therefore, in the present invention, these values are calculated from the projected waveform itself at a certain point in time, and
This problem was solved by using the calculated reference waveform at that point in time and determining the character line position from the calculated reference waveform and the projected waveform.

The method of the present invention will be explained below with reference to FIG. Figure 1(a)
The projected waveform 2 is fixed, the reference waveform 4 is moved along the projected waveform 2, and both waveforms are compared when the rising position of the reference waveform 4 moves from ■ to ■. Figures 1(b) to 1(f) are projected waveforms 2 at each position.
This shows the relative positional relationship between the reference waveform 4 and the reference waveform 4. Note that the reference waveform 4 consists of a pulse portion corresponding to the width L2 of a character string or character line, and a pedestal portion L added to both ends of the pulse portion.
1 and L3.

As shown in FIG. 1(b) to FIG. 1(c), the projected waveform 2
and reference waveform 4 are superimposed.

The pedestal value a and the peak value of the reference waveform 4 are calculated and determined as follows, and are defined as the reference waveform 5.

Pedestal value a, pedestal portion L1 of reference waveform 4. Average value of the portion of projected waveform 2 that overlaps with L3.

Wave height value.

The average value of the portion of the projected waveform 2 that overlaps the pulse portion L2 of the reference waveform 4.

In FIG. 1(b), since the reference waveform 5 is not clear, it is shown inside the O mark at the top. In this way, the first
As shown in FIGS. 1B to 1F, the reference waveform 5 at that point in time is determined, the error between the projected waveform 2 and the reference waveform 5 is determined, and its square or absolute value is integrated.

As a result, as shown in FIG. 1(g), the character position can be detected using the minimum value of the calculation result 6 (when the rising position of the reference waveform 4 is ■ as shown in FIG. 1(e)). .

This operation is performed within the search area while moving the reference waveform 4 along the projected waveform 2, but the value a of the reference waveform 5
, b are submerged each time at a predetermined period and then the error is calculated, so this method focuses on the waveform shape rather than the level of the projected waveform 2. Note that the lengths of the pedestal portions L1 and L3 of the reference waveform 4 can be determined as appropriate, taking into consideration the spacing between character strings or character lines. That is, the background portion may be set so as not to overlap the adjacent character string or character line.

As shown in Fig. 1(b) to Fig. 1(d), in the part where the change in the projected waveform 2 is small, baa or b<a, so rb-a<threshold, and the part is outside the search area. . In such parts, the error becomes small in each part, and therefore the integral value also becomes small. In extreme cases,
If the projected waveform has a constant value, the values of the character string part and the pedestal part of the reference waveform 5 will be equal, and the projected waveform 2 and the reference waveform 5 will completely match, so the error will be zero.

In order to avoid determining this as a character line position, in the present invention, if the difference between the value of the character line part and the value of the pedestal part of reference waveform 5 is smaller than a predetermined threshold, the integral value is small. Even if the line is a character line, it is not recognized as a character line.

FIG. 2 is a flowchart showing the above-mentioned gist.

In order to detect the character position in this way, the present invention can be used in cases where there is noise at a level equal to or higher than the character string or character line to be extracted, or where the contrast between the characters and the background is low. However, it is possible to properly distinguish character strings or lines from noise and perform accurate extraction.

[Example] Hereinafter, as an example of the method invention, detection of the position of characters marked on the end face of a slab manufactured by continuous casting as shown in FIG. 3 will be described.

This character is actually marked with white paint on a black background, so the density level of the character is higher.
FIG. 3 shows the inverted and binarized version, where the characters are black and the background is white. In this case, printing is performed by punching out a character pattern from a long thin paper and applying it to the end face of the slab and spraying paint over it. Figure 3 shows the noise caused by the protruding paint above and below the character line l. 3 can be seen. When the images in Figure 3 are added in the horizontal direction, the result is as shown in Figure 4, and since the projected waveform 2 is larger in the protruding part of the paint than in the text line part, only the value of the projected waveform 2 in the conventional technology is With the method of processing character line 1, it is not possible to cut out character line 1.

Therefore, for the projected waveform 2 in FIG. 4, as shown in FIG. 5, a waveform expected from the width of the character line is set in advance as the reference waveform 4, and the projected waveform 2 and the reference waveform 4 are superimposed. However, by the method described above, the wave height value corresponding to the character line portion of the reference waveform 4 and the value a of the pedestal portion are calculated to obtain the digitized reference waveform 5, and the projected waveform 2 and the reference waveform 5 at that point are calculated. A predetermined calculation is performed on the error, and the point with the minimum value of the calculated value is set as the current character position.

The following values of Examples 1 to 5 can be used as the minimum calculated values.

In the first embodiment, the integral value is the square of the error between the reference waveform and the projected waveform.

In the second embodiment, the minimum value is the value obtained by dividing the value in the first embodiment by the difference between the peak value and the pedestal value of the reference waveform.

In the third embodiment, the minimum value is the value obtained by dividing the value in the first embodiment by the square of the difference between the peak value and the pedestal value of the reference waveform.

In the fourth embodiment, the minimum value is the integral value of the absolute value of the error in the first embodiment.

In the fifth embodiment, the minimum value is the value obtained by dividing the integral value of the absolute value of the error in the first embodiment by the difference between the peak value of the reference waveform and the pedestal value.

In this example, the calculation method presented in Example 2 is applied, and the value (calculation result 6) obtained by dividing the integral value of the square of the error by the difference (ba) between the peak value and pedestal value of reference waveform 5 is calculated. , reference waveform 5
Within a range where the difference (b-a) between the peak value and the pedestal value a is larger than a predetermined threshold value, the position where the minimum value is determined as the character line position was detected.

This is done by taking into account the peak value itself of the projected waveform, and using the known information that the text line part has a higher density level than the background part other than noise, and that the projected waveform protrudes into a rectangular shape. In such a part, the calculation result 6
is made smaller. This can further enhance the effect of preventing a character line position from being set in a portion where the calculation result 6 is small because the projected waveform is flat.

Calculation result 6 is shown in FIG. The part indicated by h is a part where the difference between the value a of the character line part of the reference waveform 5 and the value a of the pedestal part is smaller than a predetermined threshold, and is excluded from the character line candidate area. It is a part.

The position where the calculation result 6 in Fig. 6 is the minimum is the starting position of the character line detected by the present invention, which shows that the character line has been accurately cut out from the image with a very high noise level. has been done.

Example 1 does not perform standardization as in Example 2, and instead uses the square integral value of the error as the evaluation value, which corresponds to evaluating the projected waveform only based on its shape and not considering the value itself at all. It is something to do.

In the third embodiment, more emphasis is placed on the value of the projected waveform than in the second embodiment, and the evaluation value becomes smaller as the difference between the value of the character line portion and the value of the pedestal portion of the reference waveform becomes smaller. The first embodiment can be used when the noise level is very high, and the third embodiment can be used when the noise level is not so high.

In Example 4, the integral value of the absolute value of the error is used instead of the square integral of the error, and in the fifth embodiment, the integral value of the absolute value of the error is used as the value of the character line part and the value of the pedestal part of the reference waveform. It is normalized by dividing by the difference in , and can be used in almost the same way as in Example 1.3.

FIG. 7 shows a system diagram of an embodiment of the device of the present invention. This is an improved version of the character line cutting section 9 of the conventional device shown in FIG. 8, and as shown in 9a of FIG. Each of them is formed by a combination of ICs.

Various known image input devices such as a CCD camera and a scanner can be used as the image input section 7, and the input image is stored in the image memory 8. The character cutting part lO is from the cut out character line.

This is the part where characters are cut out one by one, and noise is completely removed by detecting the position of the character line according to the present invention. The 0 character recognition unit 11, which can perform character extraction by comparing a projection value obtained by adding in the direction and projecting it onto an axis horizontal to the character line, with a threshold value, includes various known recognition means, for example, It is possible to select from among pattern matching, neural network, etc., taking into consideration the font of the characters to be recognized. The result display and output unit 12 displays and records recognition results, generates an alarm when marking is not performed normally, and outputs results to the host computer.
This is the part that performs processing such as data transmission.

[Effects of the Invention] The present invention has the advantage that when there is noise at a level equal to or higher than the character string or character line to be extracted,
Furthermore, even when the contrast between characters and the background is low, it is possible to appropriately distinguish character strings or lines from noise and perform accurate extraction, resulting in extremely excellent effects on character recognition technology.

[Brief explanation of drawings]

Fig. 1 is an explanatory diagram of the method of the present invention, Fig. 2 is a flowchart showing the gist of Fig. 1, and Figs. Figure 3 shows the characters and background to be recognized, Figure 4 shows the projected waveform of Figure 3, Figure 5 shows the reference waveform of Figure 3, Figure 6 shows how character positions are determined from predetermined calculations, and Figure 7 shows the projected waveform of Figure 3. FIG. 8 is a system diagram of an embodiment of the character recognition device of the present invention, FIG. 8 is a system diagram of a conventional character recognition device, and FIGS. 9 to 1O are explanatory diagrams of a conventional character position detection method. 1-m-Character line 2...-Projected waveform (projection value) 3-Noise 4--Reference waveform 5...-Reference waveform (digitized) 6--Computation result 7--Image input section 8 --- Image memory 9.9a --- Character line cutting section 10 --- Character cutting section 11 --- Character recognition section 12 --- Result display and output section Ll, L3 --- Pedestal portion L2 --- Pulse portion No. Figure 3 Figure 4

Claims (1)

[Claims] 1. In detecting the position of a character string or character line containing the recognition target character from an image containing the recognition target character, the density of the image is added in parallel to the character string or character line to detect the character string or character line. A projected waveform in the direction perpendicular to the character line is obtained, and a pulse corresponding to the width of the character string or character line and a pedestal portion of a predetermined length on both sides of the pulse are obtained from the width of the character string or character line and the background part. Set a reference waveform with The average value of the projected waveform portion corresponding to the pulse width portion of the reference waveform is calculated to obtain the pedestal value and the peak value of the reference waveform, respectively, and the difference between the peak value and the pedestal value of the reference waveform is determined in advance. Find the integral value of the square of the error between the calculated reference waveform and the projected waveform in a range larger than the calculated threshold, and set the position where the integral value is the minimum as the position of the character string or character line. A method for detecting the position of a character string or line of characters. 2. The method for detecting the position of a character string or character line according to claim 1, wherein the minimum value is a value obtained by dividing the integral value of the square of the error by the difference between the peak value and the pedestal value of the reference waveform. 3. The position detection of a character string or character line according to claim 1, wherein the minimum value is a value obtained by dividing the integral value of the square of the error by the square of the difference between the peak value of the reference waveform and the pedestal value. Method. 4. The method for detecting the position of a character string or character line according to claim 1, wherein the minimum value is an integral value of the absolute value of the error. 5. The method for detecting the position of a character string or character line according to claim 1, wherein the minimum value is a value obtained by dividing the integral value of the absolute value of the error by the difference between the peak value and the pedestal value of the reference waveform. 6 an image input unit for inputting an image including characters to be recognized;
a character string or character line cutting section that cuts out a character string or character line containing a character to be recognized from the image input by the image input section; and an image of the character string or character line cut out by the character string or character line cutting section. In a character recognition device comprising a character cutting section that cuts out a character to be recognized, and a character recognition section that recognizes the character to be recognized cut out by the character cutting section, the character string or character line cutting section cuts out the character string or character. a projected waveform forming unit that calculates a projected waveform in a direction perpendicular to the character string or character line obtained by adding the density of the image parallel to the line; Alternatively, a reference waveform setting section that sets a reference waveform having a pulse corresponding to the width of a character line and a pedestal portion of a predetermined length on both sides of the pulse, and a reference waveform setting section that moves the reference waveform along the projected waveform to obtain a predetermined value. The pedestal value and waveform of the reference waveform are determined from the average value of the projected waveform portion corresponding to the pedestal portion and the average value of the projected waveform portion corresponding to the pulse width portion of the reference waveform by comparing the projected waveform with the period. a reference waveform calculation unit that calculates a high value; and a comparison calculation unit that compares the calculated reference waveform and the projected waveform, performs a predetermined calculation on the error between the reference waveform and the projected waveform, and detects the minimum value. A character recognition device characterized by: