JPH03282985A - Noise processing system for hand-written numerical recognition - Google Patents

Abstract

PURPOSE:To surely recognize a numerical by removing a noise element only when the length of a part corresponding to the noise element is less than a prescribed value and comparing the numeral with a sorting code again. CONSTITUTION:A hand-written character is thinned like a straight line P1, a dot P2, a straight line P3, a loop P4, a dot P5, and a straight line P6. Since the dot P2 can be determined as a noise frequently appearing on a folded point or the like, the dot P2 is previously registered as a dot branch to be a noise. Since the dot P5 is not an end point branch, it is not applied to noise processing. Then, the dot branch number of a noise element is specified, whether the length of the corresponding branch is less than a prescribed value or not is decided, and when the length is less than the prescribed value, the end point dot branch less than the prescribed length is removed. Consequently, the generation of misrecognition due to noise can be prevented.

Description

DETAILED DESCRIPTION OF THE INVENTION Object of the Invention: (Industrial Application Field) The present invention relates to a noise processing method in number recognition for improving the recognition rate of handwritten numbers on checks and the like.

(Prior art) As a recognition method for recognizing handwritten numbers, etc.,
There are methods such as those shown in Japanese Patent Publication No. 116781 and Japanese Patent Application Laid-Open No. 1-116782. These methods involve pre-processing handwritten digits to thin them, extracting features using elements of loops, straight lines, and arcs, and comparing the extracted features with pre-registered classification codes to recognize the handwritten digits.

(Problem to be Solved by the Invention) In the conventional recognition method described above, short lines are also removed as noise through preprocessing. Therefore, for example, the first
In the handwritten character shown in Figure 2 (8), whisker part 1 is removed as noise in the preprocessing stage and processed as a number as shown in Figure 2 (B), so it cannot be recognized as "8". could not.

This invention was made due to the above-mentioned circumstances.
SUMMARY OF THE INVENTION An object of the present invention is to provide a noise processing method in a handwritten digit recognition method that enables reliable number recognition by performing noise processing only when certain conditions are met.

Structure of the Invention (Means for Solving the Problems) This invention preprocesses handwritten digits to thin them, and then extracts features using loop, straight line, and arc elements.
Regarding a noise processing method in handwritten digit recognition that is recognized by comparison with pre-registered classification codes, the above object of the present invention is to leave the noise without removing it in the pre-processing and to remove the noise from the pre-registered classification codes. A code containing noise is prepared as the classification code, and when the recognition result corresponds to the code containing noise, it is removed only when the length of the part corresponding to the noise element is less than a predetermined value, and the classification is performed again. This is achieved by comparing with the code.

(Operation) This invention reads handwritten digits on checks, etc., preprocesses them to make them thinner, extracts features based on loop, straight line, and arc elements, and compares the handwritten digits with pre-registered classification codes. This is a noise processing method when recognizing 1ffl. Then, preprocessing that unconditionally removes parts with a length within a fixed value as noise is applied in Japanese Patent Application Laid-Open No. 1-11678.
It is used in combination with dot processing as shown in Publication No. 2, and small loop processing, connecting processing,
This is a noise processing method for number recognition that introduces N-bar processing and further anti-blur processing using binary information.

(Example) This invention is based on the character recognition disclosed in Japanese Patent Application Laid-Open No. 1-116782 or 1-116781 as described above, and FIG. 1 shows the details of the processing operation of this invention. A preprocessed thinned character prototype is obtained (step 51), character candidates for major classifications of "0" to "9" are determined (step S2), and it is determined whether there are any candidates after that. (Step S3).

Then, if there are candidates, it is checked whether there is a dot branch that is a noise target among the candidates (step 5).
4). The dots themselves are appropriately defined when creating the reference code pattern. Here, dot branches are thin whiskers that occur at bending points that are unique to line thinning, or parts that are unnecessary as numbers. For example, in the handwritten character shown in Figure 3 (A), PI is a straight line and P2 is a dot. , P3 is a straight line,
P4 is a loop, P5 is a dot, P6 is a straight line, and the third
The lines are thinned as shown in Figure (B). And dot P
2 is known to be noise that often appears at bending points, etc., so this is registered in advance as a dot branch to be the noise target. Note that the dots indicate the type of element (
arcs, straight lines) are not relevant. Similarly, the type of dot P5 does not matter, but in this case, dot P5 is not an end point branch (branch with one end being an end point) and is therefore not subject to noise processing.

If the dot branch to be the noise target is among the candidates in step S4, the dot branch number to be the noise target is specified (step 55) (Fig. 3(B)).
In step P2), it is determined whether the length of the branch corresponding to the dot branch number is less than or equal to a predetermined length (step S6),
If the noise is below a predetermined value, noise processing 1 is performed (step 57).
). Noise processing 1 here refers to the removal of the end point dot branches mentioned in the specification of the dot branch numbers in step S5 above and branches having a fixed length or less.

After such noise processing 1, or if NO is determined in steps S4 and S6 above, character recognition processing is performed (step S8), and it is determined whether the recognition result is OK or not (step S9), OK. In this case, the recognition result is output. Is the character recognition process described in the above-mentioned publication?
The process is shown in FIG. That is,
First, data structure and feature extraction are performed on the read data, and then the numbers "O" to "9" are recognized based on the feature extraction, and the recognized numbers are roughly classified and weight checked (as characters). For example, in the case of Figure 1O (8), check the ratio of each element shown in Figure 1 (B) to the whole.), check the distortion (straight line) ), position check (evaluate the position of each center point of the character. For example, in the case of Fig. 11 (A),
Check the vertical position of the intersection point 4 shown in FIG. 4(B). ) and detailed evaluation, then check for crushed characters and make a comprehensive judgment.

On the other hand, if there are no candidates in step S3, or if the recognition result is not OK in step S9, noise processing 2 is performed (step 510). This noise processing 2 means unconditionally removing endpoint branches of a certain length. After this noise processing 2, character recognition is performed again (step 5ll), and if the recognition result is OK, the result is output, and if the recognition result is not OK, a small loop process is performed (step 512).

The small loop processing (step 520) is to remove small loops whose loop length is less than a certain level from the small loop 2 shown in FIG. 4, and to connect the two points with one line. In addition, when the ink in the small loop 2 does not come out well with a hole vent, the inside often falls out and becomes a thin line, forming a loop. After this small loop processing, character recognition is performed again (
Step 521), determine whether the recognition result is OK or not.Step 522), if OK, output the result, and
If it is not K, a connection process is performed (step 523).
. This connecting process means connecting two points with a distance between end points of a certain length with a line. After this connecting process, character recognition is performed again (step 524), and it is determined whether the recognition result is OK or not. 525), the result is output if OK, and if it is not 0, the vertical bar is removed (step 526). Vertical bar removal is the process of removing vertical bar 3 in Figure 8 (B), which is a shared branch of two loops as shown in Figure 8 (A), and after this vertical bar removal, character recognition is performed again. Next, in step 527), it is determined whether the recognition result is 0 or not, and in either case, the result is output (step 528).

Next, countermeasures against erroneous recognition of blurred characters will be explained with reference to the flowchart shown in FIG. Typical examples of collapse include those shown in the binarized data of FIGS. 5(A) to 5(C). In other words, Figures 5(A) to (C) are 'O', '8°', and '0' or crushed, respectively, but when these are thinned, they become 1, '9', and '6, respectively.
Therefore, if you check whether it is the width of the original binarized data only when it is recognized as 1""9°°,°"6゛°, etc., you will avoid misidentification. For example, in the example shown in Fig. 5 (CB), when the character ``9°'' is recognized, first, the extraction of the average 3-point bulge of the upper loop is set to α, and the result is obtained as shown in Fig. 7 (step 5
30), and further extracts the line width near the collapse as β (step 531). Then, it is determined whether the neighboring line width β is larger than a·α using a as a coefficient (step 532).
, if it is larger, check out the binary level and invalidate the result recognized as "9" (step 533).
, if the neighboring line width β is smaller, the result is output. In other words, it may be recognized as "9".

According to this invention, for a handwritten numeral having whiskers 1 as shown in FIG. “I recognize it as 8゛.

However, in the conventional method, that is, in the process of removing whiskers whose length is less than a certain value, character recognition is performed after noise processing, so there is a possibility that "6'°" or "reject" will result.

Effects of the invention: As described above, according to the noise processing method of the present invention, if the recognition result corresponds to a code containing noise, it can be removed and classified only when the length of the part corresponding to the noise element is less than a predetermined value. This makes it possible to prevent erroneous recognition due to noise.

[Brief explanation of drawings]

FIG. 1 is a flowchart showing an embodiment of the present invention, FIG. 2 is a flowchart showing a general operation of character recognition,
Figures 3 (A) and (B) are diagrams for explaining character feature extraction, Figures 4 to 9 are diagrams for explaining this invention, and Figures 10 (A) and (B) are diagrams for explaining character feature extraction. Figures (A) and (B) are diagrams to explain the weight check; Figures 12 (8) and (B) illustrate noise processing by thinning handwritten characters. This is a diagram for 1... Beard part, 2... Small loop, 3... Vertical bar.

Claims (1)

[Claims] 1. After preprocessing handwritten numbers and thinning them, features are extracted using elements of loops, straight lines, and arcs, and the numbers are recognized by comparing them with pre-registered classification codes. is left without being removed, a code containing noise is prepared as the classification code registered in advance, and when the recognition result corresponds to the code containing noise, the part corresponding to the noise element is 1. A noise processing method for handwritten digit recognition, characterized in that the noise processing method is removed only when the length of the numerals is less than a predetermined value and compared with the classification code again.