JPH03113583A - Character recognizing method - Google Patents

Abstract

PURPOSE:To reduce misrecognition rate by rejecting a first order candidate when the categories of the first order candidates obtained by matching plural feature vectors with each dictionary pattern do not coincide with each other. CONSTITUTION:In correlation stages 31, 32, the correlation of the feature vector generated from an inputted character picture with each dictionary pattern is acquired, and the order of similarity (coefficient of correlation) is determined, and the category of the high order of this similarity is outputted as the candidate. Then, in a reject decision stage 4, the candidate categories of the first order respectively among the categories of the high order outputted from the correlation stages 31, 32 are collated with each other. If two candidate categories do not coincide with each other, they are decided to be rejected, and heir codes are outputted as recognized results. Thus, the misrecognition rate can be reduced even under the same reject rate.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a character recognition method in an optical document capture device.

(Prior art and its problems) Conventionally, in character recognition methods for optical document reading devices, difficult-to-read character paste jets (failure judgments) due to scratches, scratches, etc. in character images are judged based on the degree of similarity with dictionary patterns. Ta. A typical method is to reject a candidate if the similarity of the first candidate is less than a reference value. 1 In this type of rejection determination based on similarity, it is possible to balance the rejection rate and false recognition rate (false recognition rate that is not rejected) to some extent by setting a reference value, but it is not possible to balance both the rejection rate and false recognition rate. It was difficult to lower it.

(Means for Solving the Problems) The present invention was made in view of the above problems, and is a character recognition method for classifying character images into known categories.
This method proposes a character recognition method that determines difficult-to-read characters instead of simply rejecting them if the categories of the first-ranked candidates, which are determined by matching multiple feature vectors with their respective dictionary patterns, do not match each other. It is.

(Function) In the present invention, by comparing character recognition results using multiple feature vectors, the false recognition rate can be significantly lowered than when only one feature vector is used, even for the same rejection rate. be.

(Embodiment) An embodiment of the present invention will be described with reference to the block diagram of FIG. The character recognition device 1 to which the present invention is applied includes a feature extraction step 21.22, a correlation step 31.32, and a rejection determination step 4.

Feature extraction steps 21 and 22 generate feature vectors from each character image 11 manually generated by a predetermined device (not shown) such as a scanning device. In this embodiment these two feature extraction steps 21 and 22 generate different feature vectors. That is, the feature extraction step 21 extracts the direction vector for the thinned character image, and the feature extraction step 2
2 is the direction vector for the character image subjected to contour extraction processing.

Each is generated as a feature vector.

Here, the direction vector is a distribution density matrix in the connection direction (4-connection or 8-connection) of black pixels normalized by the pixel size.

In the correlation steps 31 and 32, the feature vectors generated from the input character images are correlated with the respective dictionary buttons to determine the ranking of similarity (correlation coefficient), and the categories with the higher ranking of the similarity are determined. Output as a candidate.

Rejection determination step 4 is one of the higher-ranking candidate categories output from two correlation steps 31 and 32.

Compare the candidate categories that came in first place.

If the two candidate categories match, that candidate category is output as the recognition result. Furthermore, if the two candidate categories do not match, it is determined to be rejected, and its code is output as the recognition result.

In this embodiment, the feature extraction process and the correlation process are explained using two sets, but the present invention is not limited to this, and the number of sets may be further increased to improve the recognition rate. Of course.

(Effects of the Invention) As explained above, in the character recognition method of the present invention, by comparing the character recognition results using multiple feature vectors, it is possible to improve the accuracy by comparing the character recognition results using multiple feature vectors. This can significantly reduce the recognition rate.

[Brief explanation of drawings]

4- The drawings show one embodiment of the present invention, and FIG. 1 is a block diagram of the present invention. 1... Character recognition device, 21, 22...
・Feature extraction process, 31.32...Correlation process, 4
...Reject judgment step 11 ... Character image, 121, 122 ... Dictionary pattern, 13
・・・・・・Recognition result

Claims (1)

[Claims] A character recognition method for classifying character images into known categories, characterized in that if the categories of first-place candidates determined by matching a plurality of feature vectors with respective dictionary patterns do not match each other, the characters are rejected. Method.