JPS5856909B2 - character recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a character recognition device that optically recognizes characters.

An example of a conventional character recognition device is shown in FIG.

A character pattern written on a form (not shown) is photoelectrically converted in a photoelectric conversion circuit 1, and subjected to preprocessing such as noise removal in a preprocessing circuit 2.

The preprocessed character pattern is supplied to a normalization circuit 3 and a feature extraction circuit 4.

The normalization circuit 3 performs normalization such as information compression on the preprocessed character pattern, and supplies the normalized character pattern to the superposition circuit 5.

The superimposition circuit 5 collects data of actually written character patterns, and superimposes them on the standard character patterns in the automatically designed similarity dictionary 6 based on the simulation, thereby generating a plurality of candidate characters and their similarity values. Output.

These pieces of information are supplied to the editing circuit 7.

The editing circuit 7 examines the similarity value of the first candidate character and the similarity value of the second candidate character, and determines whether to select the candidate character with the maximum similarity value as an answer, or a structure analysis circuit. Decide whether to leave the decision to 8.

The structure analysis circuit 8 is supplied with feature extraction data from the feature extraction circuit 4 that extracts features of preprocessed character patterns.

When the structure analysis circuit 8 is entrusted with the judgment by the editing circuit 7, the structure analysis circuit 8 generates a feature dictionary 9 that stores feature data of character patterns.
Character patterns are recognized by referring to the contents of the text.

In such a conventional character recognition device, for example, FIG.
The character pattern as shown in B and FIG. 3 is 1 for the reading designation of single numbers and mixed reading of numbers and alphabets, and is ``NO'' for the reading designation of single reading of kana and mixed reading of alphabetic characters and kana.

On the other hand, in mixed reading of numbers and kana, the character pattern shown in FIG. 2A is recognized as 1, and the character pattern shown in FIG. 2B is recognized as ``.''.

By the way, in the case of a character pattern that is indistinguishable from either 1 or ``,'' as shown in FIG.・The ability to distinguish between 1 and ノ in kana mixed reading is not that great.

For this reason, the editing circuit 7 often entrusted the decision to the structure analysis circuit 8 (note that reading specifications such as reading numbers alone, reading letters alone, reading numbers and letters mixed, etc. are collectively referred to as subsets hereinafter.

). However, in character recognition by structural analysis in the structural analysis circuit 8, it is possible to judge normal character patterns, but only characters with sharp curves are subject to judgment, such as 1 and ``1'' or 7 and ``Wa''. It was not easy to make a judgment.

This invention was made in view of the above circumstances,
It is an object of the present invention to provide a character recognition device capable of efficiently and accurately recognizing character patterns that are very similar.

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 4 is a schematic diagram of a character recognition device according to an embodiment of the present invention.

In FIG. 4, reference numeral 11 indicates a photoelectric conversion circuit.

This photoelectric conversion circuit 11 has a function of photoelectrically converting a character pattern written on a form (not shown) into a binary video signal of black and white.

Reference numeral 12 indicates a preprocessing circuit.

This preprocessing circuit 12 has a function of performing preprocessing such as noise removal on the photoelectrically converted character pattern.

Reference numeral 13 indicates a normalization circuit.

This normalization circuit 13 has a function of making constant the height, width, inclination, etc. of the character pattern preprocessed by the preprocessing circuit 12.

Reference numeral 14 indicates a first superposition circuit.

This first superimposition circuit 14 refers to the subset information and the stored contents of a main similarity dictionary 15 to be described later, calculates the similarity of the normalized character patterns sent from the normalization circuit 13, and calculates the similarity. It has a function that selects several candidate characters from the one with the highest value.

The main similarity dictionary 15 stores standard character patterns that are automatically designed through simulation by collecting data from character patterns actually written based on standard fonts.

Reference numeral 16 indicates a determination circuit.

This determination circuit 16 determines whether the character It has a function of entrusting recognition to the second superposition circuit 18, and entrusting processing to the editing circuit 19 if not recognized.

The feature similarity dictionary 17 is a similarity dictionary for discriminating between characters that are very similar, and is a first similarity dictionary using the main similarity dictionary 15.
This dictionary is used by a structure analysis circuit 20 using a superimposition circuit 14 and a feature dictionary 22 (to be described later) to discriminate between characters that are so similar that they may be misread or overprinted.

The second superimposition circuit 18 uses the drive signal from the determination circuit 16 to refer to the stored contents of the feature similarity dictionary 17, and
It has a function of discriminating the characters of the normalized character pattern output from the normalization circuit 13 and outputting the discrimination results as answers.

The editing circuit 19 examines the magnitude of the similarity values themselves and the differences between the several candidate characters obtained by the first superposition circuit 14, and outputs the first candidate character with the maximum similarity value as an answer. It has a function to decide whether to leave the character recognition judgment to the structure analysis circuit 20.

Reference numeral 21 indicates a feature extraction circuit.

This feature extraction circuit 21 has a function of extracting the features of the character pattern preprocessed by the preprocessing circuit 12.

Reference numeral 22 indicates a feature dictionary. This feature dictionary 22 is a dictionary that stores the features of each character.

The structure analysis circuit 20 compares the memory contents of the feature dictionary 22 with respect to the candidate characters entrusted with discrimination by the editing circuit 19;
It has a function of performing matching with the features extracted by the feature extraction circuit 21, selecting one candidate character, and outputting it as an answer.

Note that when creating the feature similarity dictionary 17, the character pattern shown in FIG. 2A is stored as 1.

Further, a character pattern as shown in FIG. 2B is stored as ``.

Furthermore, ambiguous patterns such as those shown in FIG. 3 are not stored.

Next, the operation of the above embodiment will be explained.

A character pattern written on a form (not shown) is photoelectrically converted by a photoelectric conversion circuit 11, and preprocessed by a preprocessing circuit 12.

Further, the preprocessed character pattern is normalized in a normalization circuit 13, and the normalized character pattern is sent to a first superposition circuit 14.

In this first superimposition circuit 14, the normalized character pattern and the standard character pattern stored in the main similarity dictionary 15 are superimposed.

As an evaluation function for this superposition, there is a composite similarity calculation formula shown below.

φl-dictionary vector, f=input cover turn vector Composite similarity calculation using the above equation is performed on all characters included in the subset, and based on the results, a plurality of characters with larger similarity values are selected as candidate characters.

Then, the similarity value between this candidate character and each candidate character is supplied to the determination circuit 16.

The determination circuit 16 checks whether the character matches the combination of subsets and candidate characters prepared in the feature similarity dictionary 17.

If they match, control is transferred to the second superposition circuit 18.

If they do not match, the candidate character and the similarity value of each candidate character are sent to the editing circuit 19.

When control is transferred to the second superposition circuit 18, the similarity between the plurality of candidate character patterns selected by the first superposition circuit and the normalized character pattern stored in the feature similarity dictionary 17 is calculated. Calculate each.

Then, character patterns that are very similar are recognized based on the magnitude of the similarity value and the difference between them, and the recognition result is output as an answer.

On the other hand, when control is transferred to the editing circuit 19, the size of the similarity value and the difference between the candidate characters sent from the first superposition circuit 18 are checked, and the similarity value is the maximum. A decision is made as to whether to output the first candidate character as the answer or to leave the decision to the structure analysis circuit 20.

Here, when entrusting the determination to the structural analysis circuit 20,
The editing circuit 19 sends the candidate characters and the similarity value of each candidate character to the structure analysis circuit 20 .

The structure analysis circuit 20 matches the feature dictionary 22 that stores the character pattern features of each character with the preprocessed character pattern feature results obtained by the feature extraction circuit 21, and if they match, outputs that character as the answer.

Otherwise, it will be rejected (reading will be refused).

In this embodiment, two superimposed circuits are provided, but one
Two superimposed circuits may be used for the same purpose.

In such a character recognition device, it is possible to perform discrimination between characters in which a sharp curve becomes a criterion for character recognition by referring to the stored contents of the feature similarity dictionary 17.
Not only can the accuracy of character recognition be improved, but also character recognition can be performed efficiently.

As described above, according to the present invention, it is possible to provide a character recognition device that can efficiently and accurately recognize closely similar character patterns.

[Brief explanation of the drawing]

FIG. 1 is a schematic configuration diagram showing an example of a conventional character recognition device.
FIGS. 2A and 3B and 3 are diagrams showing examples of character pattern entry, respectively, and FIG. 4 is a schematic diagram of a character recognition device according to an embodiment of the present invention. 11...Photoelectric conversion circuit, 12...Preprocessing circuit, 13...Normalization circuit, 14...
・First superposition circuit, 15... Main similarity dictionary, 16... Judgment circuit, 17... Feature similarity dictionary, 18... Second Superposition circuit, 19
... Editing circuit, 20 ... Structure analysis circuit, 21 ... Feature extraction circuit, 22 ... Feature dictionary.

Claims (1)

[Claims] 1. A photoelectric conversion means that photoelectrically converts characters recorded on a form and outputs a character pattern, a main similarity dictionary created based on standard fonts, and a photoelectric conversion means that refers to the contents of this main similarity dictionary. a first superimposition means that calculates similarity values for the character patterns obtained by the method and outputs a plurality of candidate characters; and a feature similarity dictionary created to be able to discriminate between closely similar characters. a second superimposition means for discriminating the character pattern obtained by the photoelectric conversion means with reference to the contents of the feature similarity dictionary and outputting the discrimination result as an answer; and the first superposition means. Determine whether to output one candidate character as an answer from among the plurality of candidate characters output from the above, or leave the character discrimination to others, and output the answer from among the above candidate characters, or decide whether to leave the character discrimination to others. and a determining means that determines a plurality of candidate characters outputted from the first superposition means and entrusts the judgment to the editing means or the second superposition means, a feature extraction means for extracting features of the character pattern output from the photoelectric conversion means; a feature dictionary storing character features;
When the editing means determines that character discrimination is to be entrusted to someone else, a structure analysis means refers to the contents of the feature dictionary, performs matching with the features of the character pattern extracted by the feature extraction means, and outputs the result as an answer. A character recognition device characterized by comprising: