JPH0127474B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a high-speed and highly accurate character recognition device.

In order to absorb variations in characteristics of handwritten characters in conventional character recognition devices, such as the inclination of character lines due to differences in writers, and for printed characters, due to the inclination of character lines due to the inclination of printed characters, the method uses multiple dictionary masks. Variations in the characteristics mentioned above were absorbed. However, with this method, the time required to compare the extracted features with the dictionary during identification increases in proportion to the number of dictionary masks, resulting in a reduction in the processing speed of the device.

In order to eliminate such conventional drawbacks, the present invention extracts the slope of the character line in each direction for each character,
This system selects a dictionary suitable for the character and performs recognition, and its purpose is to provide a high-speed and stable character recognition device. An embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 1 shows a block diagram of an embodiment of the character recognition device of the present invention. In the figure, an optical signal of a character is converted into a binary quantized digital electrical signal from an optical signal input 1 by a photoelectric converter 2, and is stored in a pattern register 3. At the same time, the line width W of the input pattern is calculated in the line width calculating section 4. The sub-pattern extraction unit 5 performs a vertical scan over the entire surface of the pattern register 3, and extracts vertical sub-patterns based on the relationship between the continuous length of black dots (character line portions are black dots) and the line width calculated by the line width calculation unit 4. Extract the pattern (VSP). Similarly, horizontal scan creates horizontal subpattern (HSP),
Right diagonal subpattern (RSP) is extracted by right diagonal 45° scan, and left diagonal subpattern (LSP) is extracted by left diagonal 45° scan. Figure 2 shows an example of the original pattern and each sub-pattern, where a is the original pattern, b is the vertical sub-pattern (VSP), c is the horizontal sub-pattern (HSP), d is the right diagonal sub-pattern (RSP),
e is a left diagonal subpattern (LSP).

The stroke extraction unit 6 performs a horizontal or vertical scan over the entire surface of each sub-pattern register, detects points of change from a white point to a black point, and from a black point to a white point,
Strokes are extracted from the relationship between the number of change points and change point coordinates in the scan of the previous column (or row) and the number of change points and change point coordinates in the current column (or row), and the strokes in each extracted subpattern register are The coordinates of both end points in the two-dimensional coordinate system defined by the pattern register 3 (with the lower left of the pattern register as the origin) are sent to the average angle calculation unit 7. The average angle calculation section 7 refers to the coordinates of both end points of the stroke in each subpattern register extracted by the stroke extraction section 6, and calculates the average angle for each subpattern. In other words, the coordinates of both end points of the stroke extracted from the horizontal sub-pattern are (HXS _p , HYS _p ),
(HXE _p , HYE _p ), where p=1, . . . , P, P is the number of strokes, and the average angle θ _H is calculated using equation (1). (However, HXE _p HXS _p ) Similarly, θ _V , θ _R , and θ _L are calculated using equations (2) to (4).
However, VYE _q VYS _q , RXE _s RXS _s , LXE _t
Let LXS _t .

Note that in the above formula, Q, S, and T are the numbers of strokes extracted from the vertical subpattern, right diagonal subpattern, and left diagonal subpattern, respectively. Furthermore, when the number of strokes is 0, the average angle is also 0. The average angle calculation unit 7 sends the average angle of each sub-pattern calculated using the above equations (1) to (4) to the dictionary selection unit 8.
The character frame detection section 9 detects a character frame circumscribing the character pattern in the pattern register 3, and sends the result to the character frame division determination section 10.

The character frame division determining unit 10 divides the detected character frame into an M×N area (M and N are integers; in this embodiment, M=
Coordinates of dividing points on the X-axis and Y-axis for dividing into N=5) are determined. Here, the X axis indicates the horizontal direction of the character frame, and the Y axis indicates the vertical direction.

The feature matrix extraction unit 11 extracts VSP, VSP, and
Divide the character frame area on each sub-pattern register of HSP, RSP, and LSP into M×N areas, count the number of black dots B _no in each area, and use the line width W calculated by the line width calculation unit 4. A feature indicating the character line length is calculated using equation (5), and an M×N×4-dimensional feature matrix is created.

L _no = B _no /W (5) After that, the VSP feature matrix is the length of the character frame in the Y-axis direction ΔY, the HSP feature matrix is the length in the X-axis direction ΔX, and the RSP and LSP feature matrices are (ΔX + ΔY). /2 and finally create an M×N×4-dimensional feature matrix _fi .

The dictionary selection unit 8 refers to the average angles θ _H , θ _V , θ _R , and θ _L output from the average angle calculation unit 7 and sends a selection signal to the dictionary memory unit 13 for selecting a dictionary suitable for the input character pattern. Send to. Dictionary memory section 13
Several types of dictionary masks are prepared for each feature, which are classified and created in advance for each feature. In this embodiment, the dictionary memory section includes a VSP feature matrix, an HSP feature matrix, an RSP feature matrix,
We prepared dictionary masks corresponding to three types of average angles for each LSP feature matrix. The average angle of each feature matrix is θ _H −0.25, −0.25<θ _H <0.25, 0.25θ _H …(7) θ _V −0.25, −0.25<θ _V <0.25, 0.25θ _V …(8) Three types of θ _R 0.7, 0.7<θ _R <1.4, 1.4θ _R …(9) Three types of θ _L −1.4, −1.4<θ _L <−0.7, −0.7θ _L
…(10) is supported. The dictionary selection unit 8 outputs a signal for selecting a dictionary for each feature matrix corresponding to the average angles θ _H , θ _V , θ _R , and θ _L obtained from the average angle calculation unit. The dictionary memory unit 13 causes the identification unit 12 to refer to the dictionary for each feature specified by the selection signal. Below, the dictionary selection section 8 and the dictionary memory section 1
The operation of step 3 will be explained in detail. First, dictionary selection section 8
are the average angles θ _H , θ _V , θ _R , θ _L of each sub-pattern obtained from the average angle calculating section 7, and
Apply (9) and (10) to determine the corresponding classification number.
For example, if the average angle is θ _H =0.3, θ _V =0.0, θ _R =1.0,
When θ _L = −1.0, the HSP feature matrix has classification number 3, the VSP feature matrix has classification number 2, and RSP
The feature matrix is determined to have classification number 2, and the LSP feature matrix is determined to be classification number 2. The determined classification number is sent to the dictionary memory section 13 as a selection signal 217. The operation of the dictionary selection section 8 will be explained in more detail with reference to the flowchart shown in FIG.

The dictionary selection unit 8 receives the average angles θ _H , θ _V , θ _R , θ _L from the average angle calculation unit 7 , first judges based on the magnitude of θ _H (step 301), and determines the classification number S of the HSP feature matrix. Determine _H (steps 302, 303,
304). Next, judge from the magnitude of θ _V (step
305) Determine the classification number S _V of the VSP feature matrix. Similarly, the classification number S _R of the RSP feature matrix and the classification number S _L of the LSP feature matrix are determined based on the magnitudes of θ _R and θ _L (steps 309 to 3).
316). The classification number S _H determined in this way,
S _V , S _R , and S _L are sent to the dictionary memory section 13 .

FIG. 4 shows the configuration of the dictionary memory section 13 in this embodiment. HSP feature matrix dictionary 201,
202 and 203 correspond to the classification numbers S _H 1, 2, and 3, respectively. Similarly, the VSP feature matrix dictionaries 204, 205, and 206 are 1, 2, and 3 of classification number S _V , and RSP feature matrix dictionary 2.
07, 208, 209 are classification numbers S _R 1, 2,
3. LSP feature matrix dictionary 210, 211, 2
12 corresponds to 1, 2, and 3 of the classification number S _L , respectively.

Selectors 213, 214, 215, and 216 receive selection signal 217 and switch the feature matrix dictionary for each feature. Therefore, the identification unit 12 can refer to the feature matrix dictionaries corresponding to the classification numbers S _H , _SV , _SR , and _SL determined by the dictionary selection unit 8 .

For example, the classification number S _H of the HSP feature matrix is 3, the classification number S _V of the VSP feature matrix is 2,
If the classification number S _L of the LSP feature matrix is 2 and the classification number S _R of the RSP feature matrix is 2, the selectors 213 to 216 are the HSP feature matrix dictionary 203 with classification number 3, the VSP feature matrix dictionary 205 with classification number 2, and the classification Switching is performed so that the RSP feature matrix dictionary 208 with number 2 and the LSP feature matrix dictionary 211 with classification number 2 are output.

Note that the four types of feature matrix dictionaries store a plurality of dictionary masks for each type, each corresponding to the four types of feature matrices for each character to be identified. By switching with the selector, the identification section 12 can read out the dictionary mask of the classification number determined by the dictionary selection section 8 for each feature. Therefore, the identification unit 12 can refer to the dictionary mask f _j for each character to be identified. The identification unit 12 sequentially reads the dictionary mask for each character and performs the identification operation described below.
The identification unit 12 uses a dictionary mask specified for each feature.
The expression between f _j and the extracted feature matrix f _i is
The distance D defined in (6) is applied, and the category name of the dictionary mask that minimizes D is output to the character name output 14.

D=√( _i − _j ) ² (6) In this way, the slope of the character line in each direction of the input character is calculated, and a dictionary suitable for the character is selected from a group of dictionaries prepared in advance by classifying each character. Since a mask is selected for each feature and the dictionary mask is used for identification, it is possible to identify, for example, changes in features due to the slope of character lines, such as the tendency of upward-sloping or downward-sloping characters characteristic of handwritten characters. This can be absorbed without increasing the dictionary mask that is referenced at the time.

As explained in detail above, the present invention extracts the slope of the character line for each input character and selects the most suitable dictionary to recognize the character, thereby achieving highly accurate recognition without increasing the recognition time. It is possible,
This has the effect of providing a character recognition device with high speed and high recognition accuracy.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the character recognition device of the present invention, FIG. 2 is a diagram showing an example of an original pattern and each sub-pattern, and FIG. 3 is a flowchart showing the operation of the dictionary selection section 8. FIG. 4 is a configuration diagram of the dictionary memory section 13. DESCRIPTION OF SYMBOLS 1... Optical signal input, 2... Photoelectric conversion section, 3... Pattern register, 4... Line width calculation section, 5... Sub pattern extraction section, 6... Stroke extraction section, 7... Average angle calculation section, 8... Dictionary selection section, 9...Character frame detection unit, 10
...Character frame division determination unit, 11...Feature matrix extraction unit, 12...Identification unit, 13...Dictionary memory unit, 14...
Character name output.

Claims (1)

[Scope of Claims] 1. A means for extracting a plurality of features from a character image to be recognized; and a means for extracting a stroke component in at least one of the horizontal or vertical directions from the character image to extract the inclination of a character line in the direction. means for extracting the amount; means for determining classification for each feature of the character image based on the amount of inclination of the character line in the extracted direction;
a dictionary memory unit having a dictionary for each feature, which is composed of a dictionary mask for each classification, created by performing the character line slope amount extraction and the classification determination; and a dictionary for each feature of the character image; means for selecting a dictionary mask of a corresponding classification from the dictionary for each feature based on the dictionary; and means for recognizing a character by comparing the feature extracted from the character image with the selected dictionary mask. A character recognition device comprising: 2. The means for extracting the amount of inclination of the character line extracts stroke components in a plurality of directions including a horizontal direction or a vertical direction from the character image, and extracts the amount of inclination of the character line in each of the directions. A character recognition device according to claim 1, characterized in that: