JPH028348B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention provides a method for processing the input pattern including the recognition target portions of the input pattern that could not be inspected and the recognition target units that could not be recognized among the inspected recognition target units. The present invention relates to a highly practical pattern processing device that can effectively recognize patterns.

[Technical background of the invention and its problems]

2. Description of the Related Art In recent years, as a simple means of inputting information into computers and the like, various technologies have been researched, such as character recognition for printed and handwritten characters, speech recognition for input speech, etc., and some of them have been put into practical use. This type of pattern recognition processing is basically performed by cutting an input pattern into individual recognition target units, such as characters and phonemes, and comparing them with a recognition dictionary. However, when recognizing handwritten characters written on a blank sheet of paper, for example, if multiple characters forming the character string are connected or overlap each other, it becomes difficult to inspect each character, and the recognition dictionary ends up being difficult. There was a problem that the recognition could not be performed before the verification was performed. Conventionally, therefore, a frame is provided in advance on the paper to restrict the position where characters are written in order to prevent the above-mentioned problems in the inspection process. For this reason, the scope of use and application of pattern recognition technology is restricted, and furthermore, restrictions are placed on the position of writing characters on paper, which poses a problem of placing a considerable burden on users. Furthermore, in the case of handwritten characters, individual characters are often deformed so much that they cannot be recognized. For this reason, it is still insufficient as a means for inputting information to the computers and the like.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to provide a pattern processing device that can reduce the constraints on recognition objects such as character frames and improve the practicality of pattern recognition. It's about doing.

[Summary of the invention]

The present invention attaches a first pointer to a recognition target portion of an input pattern that could not be inspected, and also attaches a first pointer to a recognition target portion of an input pattern for which a recognition result has not been obtained from the input pattern. 2 pointers are attached, and for the parts with these pointers, the information held by the recognition target part is processed by using the recognition results obtained for other recognition target units and knowledge information about the input pattern as post-processing. It is designed to estimate. for example,
Utilize the meaning of the input pattern as knowledge from part of the recognition information about the input pattern,
The entire content of the input pattern is recognized by estimating the content of the portions of the input pattern that cannot be inspected or recognized. In particular, for recognition target parts for which no recognition target units have been extracted from the input pattern, they are subjected to estimation processing by the post-processing unit without being subjected to dictionary matching processing in the recognition unit, and there is a possibility that recognition target units may be extracted from there. The present invention is characterized in that the entire content of the input pattern is recognized by estimating the recognition target units and the recognition results for these recognition target units.

〔Effect of the invention〕

Thus, according to the present invention, since the input pattern to be recognized is relatively freely expressed,
It becomes possible to effectively recognize input patterns that could not be recognized by conventional recognition processing devices. In other words, even if there are parts of the input pattern that cannot be inspected or recognition target units that are unrecognizable, these parts can be processed using the recognition results of other parts and knowledge information about the input pattern as described above. Since the input pattern is estimated based on the input pattern, it is possible to effectively and efficiently create the input pattern in a short time. Therefore, the constraints on the recognition target (input pattern), such as a character entry frame, can be significantly relaxed, and the field of application of the pattern recognition device can be expanded, resulting in great practical effects.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings. Here, recognition of an address consisting of a handwritten character string will be explained as an example of pattern recognition.

FIG. 1 is a schematic configuration diagram of an embodiment apparatus, in which 1 is a character image input section, 2 is a character recognition section, 3 is a post-processing section, and 4 is a recognition dictionary. For example, the input section 1 is
A CCD scanner is used to observe an image of a letter or post with a handwritten address, and the input image is subjected to inspection processing to identify individual characters (character images), which are the units to be recognized.
This is to detect. The above-mentioned inspection process analyzes the distribution of black dots representing written characters 6 in the input image 5 as shown in FIG. Then, each character of the character string is examined as a unit to be recognized from the vertical projection pattern 6b of the black dots of the character string. At this time, for example, if it is determined that the character image contains multiple characters based on the length of the area cut out as one character image, a cutout is added to that area (recognition target part). A disable symbol (*) is attached as the first pointer.
The character image to which this first pointer is attached is not subjected to the next recognition process in the recognition unit 2, as will be described later. In the recognition unit 2, the recognition target unit (character image) obtained from the input unit 1 to which the first pointer is not attached is subjected to normalization processing and blurring operation to remove noise components. Then, a comparison process is performed with a standard pattern of the recognition unit registered in advance in the recognition dictionary 4, and the recognition result is determined. At this time, character images that are determined to be unrecognizable are marked with an unrecognizable symbol (?)
is attached as the second pointer. The information of the character image to which the second pointer (?) is attached and the information of the character image to which the first pointer is attached are post-processed together with the recognition result of the character image obtained by the recognition unit 2. Sent to Department 3. In this post-processing unit 3, the recognition result of each character image obtained by the recognition unit 2 and the relative position information of that character image in the input character string are treated as one piece of knowledge information, and the content of the recognition target (in this case, an address) is treated as another knowledge information, and the content (information) of the character image to which the first or second pointer is attached is estimated using this knowledge information. The recognition results for the entire input image (input pattern) are obtained by obtaining the estimation results for the inspection and unrecognizable portions of the input pattern by the post-processing unit 3.

Incidentally, the estimation process using the knowledge information regarding the inspection/unrecognizable portion of the input pattern in the post-processing section 3 is realized by using an automaton, for example, as shown in FIG. In other words, the input character string that has been subjected to verification and recognition processing changes its state one after another from the initial state, and when it reaches the final state, the indefinite symbols (* and ?) are added from the transition path leading to that state. This is to determine (estimate) the part of . As an example, a case where the input character string is given as "article?ku*en" will be explained with reference to FIG. First, when the input character "product" is input while in the initial state 11, the state changes to state 15.
Next, when "?" is input, it transitions to state 22 because it is an indefinite symbol, and when the next input "ku" is entered, state 17 is entered.
Transition to. At this point, based on the premise that the input character string indicates an address and the knowledge that the character string information up to that point indicates the ward, it is assumed that the "?" part is "river". Can be estimated (determined). Next, when "*" is input, the automaton moves from state 17 to state 24. Then, the end state 19 is reached with the next input "continue". in this case,
The above input "*" indicates a portion that cannot be inspected, and therefore, there is a possibility that this portion that cannot be inspected consists of a plurality of characters. Therefore, it is not possible to specify this uncuttable part to a single character, and based on the knowledge information mentioned above, candidates for the above "*" include, for example, "chu", "higashichu",
Three are estimated to be "Nishinaka". In this way, the input character string "Shinagawa-ku*Nobu" is converted (recognized; estimated) into three candidates: "Shinagawa-ku Nakanobu", "Shinagawa-ku Higashi-Nakanobu", and "Shinagawa-ku Nishinobu". For these three candidates, display this on the display,
The operator may use a keyboard or a light pen to instruct the selection. In addition, if there is an error in the characters written on the input form or in the recognition results, and characters that are not used as words to express an address are input to the post-processing unit 3, for example, in the initial state 11, " If a character other than ``品'' or ``東'' is input, it is processed as an indistinguishable state 20. In other words, it is rejected from the recognition process.

Note that this post-processing section 3 can also be realized by software as shown in FIG. 4, for example, if a knowledge dictionary in which addresses are registered is created in advance. That is, the process of determining whether the input character string matches the content of the knowledge dictionary may be performed while taking into account the role of the indefinite symbols (* and ?).

That is, first, in the initial setting, the knowledge dictionary and input character string pointers are positioned at the beginning, the correction storage flag is turned off, and the correction storage stack is cleared. Thereafter, when the first character in the dictionary is "$", the input character string pointer is advanced until the next character in the dictionary matches the input character. This "$" symbol represents any character string having 0 or more characters. Therefore, for example, by creating an entry in the dictionary as ``$Shinagawa-ku Nakanobu'', you can add ``Tokyo'' or ``Tokyo'' to higher-level address categories before ``Shinagawa-ku'' in the input character string. Even if the characters indicating the above are written, or even if the above-mentioned upper division notation is omitted and "Shinagawa-ku" is written, it is possible to obtain a match.

By the way, the process of matching the input character string and the characters indicated by the dictionary pointer is realized by the subroutine process shown in FIG. That is, if the input character is "?" or "*", the character indicated by the dictionary and input character string pointer is stored as a correction part in the correction storage stack, and the correction storage flag is turned on. In the case of other characters, the input character string and the character indicated by the dictionary pointer may be compared to extract a matching dictionary item.

When there is no "$" symbol, or after processing routine A for the "$" symbol is completed, the processing proceeds according to the character indicated by the input character string pointer. First, if the input character is "*",
The dictionary pointer is advanced until the next input character and the character indicated by the dictionary pointer match, and if they match, the process moves on to the next character (routine B). If the input character is "?", the above-mentioned storage process of the corrected part is performed and the process moves on to the next character (routine C). If the input character is any other character, the matching process is performed as is, and if there is a matching item, the correction process up to that point is considered correct and the correction save stack is cleared, the correction save flag is turned off, and the next Let's move on to processing the characters. (Routine D). If there is no matching item, the correction save flag is referred to, and if the flag is on, the correction save stack is used to carry out backward verification (routine E). If there is a matching item as a result of this backward verification, it is assumed that a match has been achieved up to the character currently indicated by the pointer, and the correction save flag and stack are cleared, respectively, and processing moves on to the next character. still,
If there are no matching dictionary entries during this series of processing, it is assumed that the "*" and "?" cannot be corrected, and elimination processing F is performed, and the processing then ends. Through this series of post-processing, it becomes possible to estimate the information on the entire character string written on the paper, that is, the address, from the information on some of the recognized characters.

As explained above, according to the present invention, even when characters in an input pattern are connected and cannot be inspected, or even when there are so-called cursive or broken characters that cannot be recognized, these can be easily recognized. It becomes possible to effectively recognize the entire input pattern including the input pattern. Therefore, considerable flexibility in the format of letters, mail, etc. can be allowed. In addition, by performing such pattern processing, pattern recognition technology can be effectively applied even in fields where pattern recognition technology could not be used in the past, and various recognition devices can be put into practical use. becomes possible. Therefore, its practical advantages are enormous.

Note that the present invention is not limited to the embodiments described above. For example, the post-processing unit 3 displays multiple characters that are determined to be similar based on the recognition results along with "?"
or the length of the connected part of the letters,
By feeding the post-processing unit 3 with the number of characters that can be estimated from the ratio of the lengths of the parts that could be cut out for each character, even if there are multiple recognition candidates, one It is also possible to specify an address or a word. Furthermore, although the processing for recognizing an address consisting of a handwritten character string has been described here, the present invention can also be applied to recognition of continuously uttered voices and pattern recognition processing of other categories. In short, the present invention can be implemented with various modifications without changing the gist thereof.

[Brief explanation of drawings]

The drawings show one embodiment of the present invention. Fig. 1 is a schematic configuration diagram of the embodiment device, Fig. 2 is a diagram for explaining the inspection process, and Fig. 3 is a post-processing section using an automaton. FIG. 4 is a diagram showing the processing routine of the post-processing section using the knowledge dictionary, and FIG. 5 is a diagram showing the collation processing routine. 1... Input section, 2... Recognition section, 3... Post-processing section, 4... Recognition dictionary.

Claims (1)

[Scope of Claims] 1. An input unit that processes an input pattern to extract a predetermined recognition target unit and attaches a first pointer to a recognition target portion for which the recognition target unit could not be verified; a recognition unit that compares each recognition target unit extracted by the input unit with a recognition dictionary to obtain a recognition result, and attaches a second pointer to a recognition target unit for which a recognition result is not determined;
Estimates the recognition result for the recognition target unit to which the second pointer is attached by using the recognition result for the recognition target unit obtained by the recognition unit and the knowledge information regarding the input pattern, and For the recognition target portion to which the first pointer is attached, the post-processing unit estimates a recognition target unit that may be extracted from the recognition target portion and a recognition result for the recognition target unit, respectively. . A pattern processing device, wherein the recognition target portion to which the first pointer is attached is directly subjected to estimation processing by the post-processing unit to obtain a recognition result without performing dictionary matching by the recognition unit. 2. The post-processing unit converts the information on the recognition results of the recognition target unit for which recognition results have already been obtained for the input pattern and the information regarding the relative position of the recognition target unit in the input pattern into knowledge information regarding the input pattern. The pattern processing device according to claim 1, which is used as a pattern processing device.