JP2805476B2 - Character recognition method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a method for recognizing low-quality characters such as engraved characters of a product by image processing.

BACKGROUND OF THE INVENTION In the manufacturing process of various products, a product number or the like is written on a part or the like by, for example, an engraving, and the product is managed by reading the engraved character. Therefore, it is necessary to read and recognize characters in the course of such product management.

2. Description of the Related Art Character reading in the field of image processing has been put to practical use, for example, for reading printed characters. In reading printed characters, since there is sufficient contrast between the characters and the paper surface, characters can be extracted by simple binarization, and relatively accurate reading can be performed by feature extraction.

However, in the process of recognizing the engraved characters in the manufacturing process, for example, the engraved characters of a cast product, differences in the illuminance are likely to appear on the engraved image, and oil stains are locally generated. Becomes difficult. Also, during the engraving process, the engraved character may be partially formed deeply due to the unevenness of the engraved surface, or the opposite part may be partially shallow, and the line element portion of the character will be discontinuous. Therefore, it is difficult to extract characters, and erroneous recognition is likely to occur at the final character recognition stage.

By the way, in a typical matching method, a character template is superimposed on an unknown character, and the character of the character template having the highest matching degree is determined to be the input character by comparing each pixel. If there is a deformation, accurate matching cannot be obtained, which causes rejection or misreading.

OBJECTS OF THE INVENTION It is therefore an object of the present invention to enable relatively low quality characters to be accurately recognized in the field of image processing.

SUMMARY OF THE INVENTION In accordance with the above object, the present invention repairs a character image to be recognized by means of half-binarization and direction-dependent line segment enhancement for background image stains and scratches and low-quality characters. While removing, and matching the direction valve OR pattern template for variations in the position of the engraved characters, etc.
It allows accurate positioning.

In addition, a character template decomposed for each direction of a line segment is superimposed at the same time with respect to a change in the line width of the character or a slight deformation or displacement, and a window is set in the direction of the line width. By performing matching in consideration of the character deformation, characters can be stably recognized.

First, FIG. 1 shows the configuration of a character recognition device 1. This character recognition device 1 recognizes a character 2 to be recognized by a program based on the character recognition method of the present invention.

The character 2 to be recognized is converted into an electric image signal by an imaging camera 3 such as a CCD image sensor.
The imaging camera 3 scans the character 2 and performs amplification, clamping operation, and the like, and then converts the character 2 into an image signal of an electric quantity proportional to the brightness of the optical image for each pixel. Then, this image signal is stored as a one-frame image signal by the next image memory 4 and taken into the image processing device 6 through the interface 5 at the time of character recognition. The image processing apparatus 6 is configured by a computer, and executes a program necessary for character line segment reinforcement (repair) and recognition by an image processing program based on the character recognition method of the present invention. The image processing apparatus 6 is connected to a keyboard 7 for input, a printer 8 for output, and a display 9 for monitor.

Next, FIG. 2 shows a series of sequential steps of the low-quality character recognition method of the present invention.

In the first half-binarization process, half-threshold processing is performed on each dot of the character image in order to replace a dot having a predetermined luminance level or less with a background level luminance for each dot of the stored one-frame character image. Going on. First,
As shown in FIG. 3 for one frame of image information, for example, in a frame of 255 × 255 dots, an appropriate window is set at a position corresponding to the character image for each of the frame memory and the corresponding work table, As shown in FIG. 4, a brightness histogram is created for each dot of the character image in this window, and the average brightness is calculated. Then, each dot in the window is half-binarized by the following equation.

Kn = Ka−Ko + ΔK Kn: half-binary data stored in the corresponding work table from the frame memory Ka: average luminance in the window of the frame memory Ko: raw data of the frame memory ΔK: correction value, but half-binary When the digitized data Kn is smaller than the “0 (white)” level, the half-binary data of the dot is “0”
Is set as

As described above, since the average luminance Ka in the window is involved as a different value for each imaging in the half binarization, the half binarized data Kn is used for each image data, such as illuminance difference and oil stain. Is binarized in a state where the noise is not included. Note that, in a general simple binarization process, the threshold value is set uniformly, so that the data after the binarization is subjected to the process of the binarization process due to the influence of oil stains and illuminance differences. Thus, the character information disappears, and dirt and the like are easily extracted as character information. In this way, the noise in the background portion is removed, and the luminance (density) is compressed.

In the next line segment extraction process, the filters in the H, V, R, and L direction-dependent line segments are applied to the dots in the window of the work table after the half-binarization processing, and the half-binarized data is subjected to the direction-specific filtering. This is the process of extracting line segments. That is, as shown in FIG. 5, the character image data of the work table is
By passing through the line segment filter for each direction, it is stored in the line segment table for each direction for image processing. As shown in FIG. 6, such a line segment filter for each direction is composed of, for example, templates T1, T2, T3, and T4 of a 7 × 7 matrix. It has a line segment in the direction (V), a line segment in the 45 ° downward (R) direction, and a line segment in the right downward (L) direction.
Each line segment has a weight of “4”, a portion adjacent to the weighted portion of “4” has a weight of “0”, and the other portions have a weight of “−1”. doing. Of course, the sum of the weights is set to 0 for one template as a whole. And
For example, as shown in FIG. 7, the center of the first template T1 is each pixel A11, A12... A1n, A2 of the character image.
1, A22... A2n, An ₁ , An ₂ ... Ann, etc. are sequentially superimposed, and the elements T11 and T of the template T1 corresponding to these are superimposed.
Between T12 and T77, the multiplication is sequentially performed as follows, and each pixel A11, A12 ... A1n, A21, A22 ... A2n, An
It is replaced as the value of 1, An2 ... Ann. The multiplication for each pixel like this is done for all templates T1,
After performing for T2, T3, and T4, the total sum is calculated for each of them.

As a result, only each line segment is emphasized, and even if a part of the character 2 to be recognized has a discontinuous defect, the part is repaired so as to be continuous by weighting “4”. For this reason, even if a part of the character 2 is engraved during the engraving process, or a part of the character element is discontinuous due to a dent on the surface of the engraved product,
By this repair, the character 2 to be recognized is correctly corrected. Even if there is noise in a direction other than the line segment in each direction, the noise is almost corrected to a background level, that is, a white level by weighting “0” or “−1”.
In this process, random noise is almost completely removed.

The subsequent coarse character segmentation process is a process of performing character segmentation in the X direction and the Y direction for the line segment of the character 2 in each direction to perform coarse positioning.

First, as shown in FIG. 8, in the Y-direction cutout, a Y-direction feature amount graph is created for all direction-specific line segment tables, and the number of one-dimensional windows shifted by the character arrangement pitch in the Y-direction is set to Y. , The maximum value of the integrated feature amount in the window is obtained, and this position is set as the cutout position in the Y direction. The size of the positioning window is set to approximately 2 to 5 dots + pitch accuracy in consideration of the accuracy of the pitch.

Similarly, in the X-direction extraction, as shown in FIG. 9, a feature amount graph in the X-direction is created for all the line segment tables for each direction, and a one-dimensional window corresponding to the number of digits shifted by the pitch in the Y-direction. Is moved to obtain the maximum value of the integrated feature amount in the window, and the position of the maximum value is set as the character position. Here, the size of the one-dimensional window is
Taking into account the precision of the pitch, it is set to about 2 to 5 dots + pitch precision. In this manner, coarse positioning is performed by cutting out the line segment table for each direction in the X-axis direction and the Y-axis direction in common. here,
The feature amount graph is a histogram obtained by integrating the filter results for each direction in the Y-axis direction and the X-axis direction.

The next character size normalization process is a process of adjusting the size of the cut-out character to a prescribed template as described above. As shown in FIG. 10, the following process is performed for each character data. Repeat. Such normalization is performed for an area slightly larger than a prescribed character size in each direction-specific line segment table. This slightly larger area is for considering variation in position.

Next, in the process of extracting fine characters, the OR pattern template for each direction is superimposed on the normalized area as shown in Fig. 11 for the normalized line segment table for each direction, and the position accuracy is calculated while calculating the matching degree. Move as much as you want. This direction-specific OR pattern template is created as a template for each of the H, V, R, and L line segments with respect to all the characters 2 to be recognized, which are superimposed. In this movement process, the character is cut out from the line segment table for each direction at the position where the matching degree is the maximum, and the position of the character is determined.

The last character recognition (line segment matching) process is a process of sequentially superimposing all the character templates on the point after the positioning is completed, calculating the matching degree, and recognizing the maximum value of the matching degree as the character type. It is. Of course, this master side character template
It is stored in advance by a teaching process according to the type of the character to be recognized. If the matching degree is lower than the specified matching value, it is determined that reading is impossible, and a display indicating that reading is impossible is displayed. However, when the matching degree is larger than the specified value, the character of the pattern template having the maximum matching degree is recognized as the character to be specified.

Then, the actual line segment matching process is performed as follows. For all the characters 2 to be recognized in advance, a plurality of partial templates obtained by decomposing the line segments constituting each character in each direction are stored in a memory or the like as a set of direction-specific templates. The set of templates for each direction is prepared for each character type as shown in FIG. Only one line segment element in a certain direction is registered on one direction-specific template, and the line width of the element is one pixel. A set of directional templates are simultaneously superimposed on unknown characters of low quality, the similarity is measured, and the character having the maximum similarity by repeating the same for the character type is set as the recognition character. Here, both the size of the unknown character and the size of the template for each direction are l pixels horizontally and m pixels vertically, the character type is n, and the character type is horizontal (H), vertical ( V), diagonally right 45
It is assumed that the line segment is composed of line segments having four directions of {(R) and 45 ° left diagonal} (L). Further, it is assumed that the unknown character is binarized and takes only one of “0” and “1” together with the direction-specific template. Here, “1” corresponds to the existence of a line segment element. When the direction-specific template is superimposed on the unknown character, the coordinates are set to (i, j), 1 ≦ i ≦ l, and 1 ≦ j ≦ m in pixel units with the upper left pixel as the origin.

When a set of directional line segment templates are simultaneously superimposed on an unknown character, the direction of the line width, that is, a perpendicular line, is formed for each element on the directional line segment template because the line direction is known. In the direction, a window having a size of about 3 pixels in the line width direction and about 1 pixel in the line direction in consideration of the character deformation centering on the element is set, and if there is even one character element in this window, , The point (i, j)
Suppose that it matched with. In other words, the three pixels in the line width direction have a margin for the character deformation as shown in FIG.

The matching expression G _H (i, j) of the horizontal (H) component at this point (i, j) is defined by the following logical expression.

G _H (i, j) = H (i, j) ∧ ｛G (i, j) ∨G
(I, j-1) {G (i, j + 1)} where ∧ is a symbol representing a logical product and ∨ is a symbol representing a logical sum, respectively, and H is a point (i, j) of the horizontal template.
Where G is the point (i, j) of the unknown character,
The values at (i, j-1) and (i, j + 1) are shown.

 Similarly, the following equations hold for other components.

G _V (i, j) = V (i, j) ∧ ｛G (i, j) ∨G
(I-1, j) ∨G (i + 1, j)｝ G _R (i, j) = R (i, j) ∧ ｛G (i, j) ∨G
(I−1, j−1) ∨G (i + 1, j + 1)｝ _GL (i, j) = L (i, j) ∧ ｛G (i, j) ∨G
(I + 1, j-1) {G (i-1, j + 1)} Each element is added by the following equation.

The number of elements on each line segment template for each direction is T _H ,
Assuming T _V , T _R , and T _L , the similarity P is obtained by the following equation.

Effects of the Invention According to the present invention, the following specific effects can be obtained.

In the half-binarization process, pixels having a predetermined luminance (voltage) level or lower are replaced with background-level luminance. Therefore, even if the contrast of the recognition target is poor and there is an illuminance difference on the surface, the noise is reduced. Completely removed. Also, since the weighting process for each direction is performed on the data for each direction after the half-binarization, even if there is a discontinuous defect in a part of the character to be recognized, the part is repaired. Malfunctions in the cutting process of the horn can be prevented. Further, even if the character is slightly deformed or the engraving position is shifted, the character is cut out based on the line segment for each direction, so that the positioning is performed accurately and reliably.

Furthermore, in the matching process, final recognition judgment is made by integrating matching with the line segment template for each direction. Therefore, compared to the matching method for each template, it is more resistant to character deformation or displacement. .

Because of these features, the recognition method of the present invention is most suitable as a method for recognizing low-quality characters such as engraved characters, and is useful for quality control of products by computer image processing.

[Brief description of the drawings]

FIG. 1 is a block diagram of a character recognition device, FIG. 2 is a flowchart of a character recognition method of the present invention, FIG. 3 is a conceptual explanatory diagram of a half-binarization process, and FIG. FIG. 5 is a conceptual illustration of a line segment extraction process for each direction, FIG. 6 is an illustration of a line segment filter for each direction, FIG. 7 is an illustration of a line segment extraction process, and FIG. 9 is an explanatory diagram of a characteristic amount graph (Y direction), FIG. 9 is an explanatory diagram of a characteristic amount graph (X direction) in a coarse positioning process, FIG. 10 is an explanatory diagram of normalization, and FIG. FIG. 12 is an explanatory diagram of a line segment template for each direction, and FIG. 13 is an explanatory diagram of a matching state. 1 ... character identification device, 2 ... characters, 3 ... imaging camera,
4 ... Image memory, 5 ... Interface, 6 ... Image processing device, 7 ... Keyboard, 8 ... Printer, 9 ...
…display.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G06K 9/62 G06K 9/20 G06K 9/38

Claims (1)

(57) [Claims] 1. A half-binarization process in which, for each dot of a one-frame character image being stored, a dot having a luminance level lower than a predetermined luminance level is replaced with a luminance of a background level to erase a background part and leave only an engraved part; A line segment extraction process of extracting the character image after the half-binarization as a line segment for each direction, a coarse character extraction process of cutting out the position of the character from the integrated feature amounts in the X and Y directions for the line segment for each direction, A normalization process of converting one character of each direction line segment into a predetermined size, a fine positioning process of extracting a character by matching the OR pattern template of each direction line with each one character,
A character recognition process in which a template for each direction, which has been previously decomposed for each line segment, is simultaneously superimposed on an unknown character, and a window is set in the direction of the line width in consideration of character deformation for each line segment on the template to perform matching. A method for recognizing low-quality characters, comprising: