KR100361176B1 - Method for recognizing written multi-character - Google Patents

Abstract

PURPOSE: A method for recognizing a written multi-character is provided to recognize a character of at least one language by studying a morpheme, constructing a database based on the morpheme, searching a corresponding morpheme from the database when a copyist writes down a character, judging a related position between each morpheme, and combining the morpheme. CONSTITUTION: Duplicated dots and dots existing in a predetermined distance are removed in an input dot of a cell which is a group of inputted strokes comprising at least one stroke. An area is separated in accordance with a constitution of a stroke and an accumulated angle formed by a movement of each dot. A characteristic value capable of displaying a characteristic of each area is extracted. A dissimilarity is calculated by comparing the extracted characteristic value with a characteristic value of a reference morpheme in a database. A reference morpheme having the lowest dissimilarity is set as a recognized morpheme. Candidate characters are formed using all recognized morphemes as strokes. A character made by a morpheme having the lowest total dissimilarity is set as a recognition character.

Description

Cursive Multi-Character Recognition Method

The present invention relates to a method for recognizing handwritten characters, and in particular, by recognizing a dependent part and a non-dependent part of a handwritten character separately, one or more characters, that is, a Korean character, an English character, a Chinese character, a Chinese character, a number, a symbol, and the like can be simultaneously recognized. The present invention relates to a handwritten multi-character recognition method.

As a general field of character recognition, there is an online character recognition method, which recognizes a character by extracting the stroke character when the writer writes the character.

However, in this case, since only the characters of a single language are recognized, there is a problem in that a satisfactory recognition result cannot be obtained when the characters of one or more languages are mixed and used because of the unique characteristics of each language.

Therefore, in view of this problem, the present invention learns a semantic unit (minimum unit having meaning), that is, a language-dependent part, and forms a database, and then finds a matching semantic unit from the database when the writer writes. By determining whether the positional relationship between each semantic unit is correct and combining them, the present invention has the purpose of being able to recognize letters for one or more languages.

The handwritten multilingual recognition method of the present invention includes a first method of filtering distances by removing overlapping points and points existing within a predetermined distance from an input point of a cell, which is a set of input strokes of one or more strokes as shown in FIG. And a second process of separating regions according to the composition of the stroke and the cumulative angles created by moving the points, and then extracting feature values to make use of the characteristics of each region, and the feature values extracted by the second process. A third process of calculating a specific value and dissimilarity of a reference semantic unit previously stored in a database, and using the semantic unit recognized by the third process as a reference semantic unit having a low dissimilarity; Attempts to create a candidate character by trying all possible cases where the character may be composed, and then recognizes the character with the lowest semantic unit of dissimilarity. It takes place in a fourth step of characters.

The present invention thus constructed will be described in detail with reference to FIGS. 1 and 2.

A cell, which is a set of input strokes consisting of one or more strokes, is a unit that is compared with the minimum minimum unit of the learning database. The feature extraction and recognition process is based on this cell.

Therefore, first, the distance filtering process removes the overlapping points of the cell and the points existing within a certain distance, and the distance value, which is the reference of the distance filtering, may vary according to the resolution and the stroke size.

In addition, this distance value generally has a high resolution and uses a large value when the stroke size is large.

This distance filtering provides the effect of sampling the cell's input points at regular time intervals, regardless of the writing speed of the writer, thereby absorbing the change in the writing speed. The number can be reduced so that the amount of computation can be reduced.

Subsequently, the data having been subjected to the distance filtering process are separated and its feature is extracted, and feature extraction is performed. This feature extraction has the greatest influence on the recognition rate, and extracts only the optimal feature that can represent the character. It is important.

However, if we consider only the direction of the short stroke in this feature extraction process, there is a disadvantage that it cannot absorb the change of length in the stroke. On the other hand, if the feature is extracted with the absolute distance, the change in size or rotation of the baseline of the stroke occurs. There is a disadvantage of not absorbing the changes that occur.

Therefore, since feature extraction is generally used to compensate for these disadvantages, the present invention uses both the feature of the direction of stroke and the feature of absolute distance.

First of all, the stroke is separated by the stroke composition and the cumulative angle created by each point. In general, the letter is composed of vertical segments, horizontal segments, and circular segments. If you can distinguish it, you can recognize any handwritten character.

However, unfortunately, letters near sperm can be recognized to a satisfactory level. On the other hand, the more severe the shedding, the lower the recognition rate will be.

Therefore, the present invention utilizes the characteristics of the characters to the extent that the domains can be separated, and solves the semantic unit recognition process for the characters from which the original characteristics can be extracted due to heavy shedding.

Next, referring to the feature extraction process, if the character recognition is performed only by the data according to the region separation process, it is impossible to detect a small change in each region, and thus it is impossible to expect an improvement in the recognition rate. Several characteristic values can be determined to save the characteristic. The characteristic value representing a region uses the following defined values.

First, the process of extracting information on the movement of a stroke in a region is characterized by the process of coding the shape of the stroke in the region. Therefore, each region is divided into eight equal parts regardless of the length of each point. The difference between the angle consisting of the beginning and the end of the area is used as the feature value.

Thus, it is possible to distinguish the approximate shape of the letter by area separation and to detect the minute shape of the letter in eight equal parts of the area.

Next, the process of extracting information on the size of the region is characterized by using the size and position of the separated region in the cell as a feature value, which is equal to the x-y value of the top-left of the region. It is the x, y value of the bottom-right.

Finally, the process of extracting information on the start and end points of a stroke in a region is characterized by the direction of the stroke's movement in the region, which indicates the start and end points of the stroke in the region. .

If the cell consists of several strokes, each stroke consists of several regions, and each feature value is designated according to the input stroke order.

That is, if the first stroke is divided into three regions and each feature value is composed of F11, F12, and F13, and the second stroke is divided into two regions, and the feature value is composed of F21 and F22, the entire stroke is composed of two strokes. The feature values of the region may be expressed by five feature values as shown in Equation (1) below.

F = {F11, F12, F13, F21, F22} ------------ Formula (1)

The data extracted from the feature extraction process is subjected to a normalization process that can reproduce the shape of the handwritten characters while minimizing the memory capacity. In addition, the information about the size of the region can be completely changed depending on the writing position, and the feature value also requires too much memory and capacity for the cursor operation.

Accordingly, the diagonal length of the cell is set to 15 so that the handwritten character form can be reproduced as it is while minimizing the memory capacity, and each feature value is set to be proportional to the size of the cell.

The reason is that only 4 bits of memory are required to display values from 0 to 5, and through this normalization process, the negligible small changes in the feature can be absorbed.

Next, the process of recognizing a semantic unit that calculates the feature value and dissimilarity of the reference semantic unit previously stored in the database as the normalized feature value as the recognized semantic unit will be described. Same as

The feature values of the semantic units that are the reference are stored in a number of different database formats according to the number of regions.

Recognition of semantic units is a process of comparing the feature values of each semantic unit in the database with the feature values of the extracted cells. First, the recognition of semantic units is performed with the number of cell regions extracted from multiple databases separated by the number of feature regions. Choose a database whose difference is between -2 and +2.

This process is often called a major classification process in letter recognition.

In this database selection process, we first look for a database with the same number of regions, then a database with one less region, then two smaller databases, one large database, and two larger databases. The reason for this is that if the handwriting is correct and separation of regions is desired, the probability of having the same number of regions is the same as that of the reference database, and then hooking: If there is one input device at the beginning or the end of the input device and the noise inputted by the writer's mistake, the probability is the same as that of the reference database with the smallest number of areas, and the probability that the next two smaller reference databases are the same. Becomes high.

However, the probability that the number of regions is the same as that of one large database and two large databases is searched later because the probability is low because the handwriting with hooking is learned in the reference database and there is data about it.

When a database is selected by this process, the semantic unit is recognized by comparing all semantic units and feature values of the cell. The algorithm is described as follows.

First, we define some variables used in the expression.

n; Input pattern is number of areas

m: number of areas of comparison pattern

: i-th area of input pattern with n number of areas

j j area of the comparison pattern with m number of areas

: I-th area of the input pattern and j-th area of the comparison pattern

: Dissimilarity between i i + 1 th region relation and comparison pattern j and j + 1 th region relation of input pattern

Looking at the case where the number of regions partitioned by the defined variable and the number of regions of the selected database are the same (n = m), the total dissimilarity penalty can be expressed by the following equation (2). .

Here, n calculations are not always performed according to the above formula (2). If the values of Dist1 and Dist2 are equal to or greater than a certain value, that is, the area is determined to be different from each other, the next item is not calculated.

Next, look at the case where the area number of the cell partitioned and the number of area of the selected database are different (n = m). If the number of areas appears different even though they are the same characters, the area is divided by hooking. In this case, there are two cases in which the region is divided by the error of judging vertical and horizontal components.

Therefore, if it is determined that the number of areas is different by hooking, then the area must be removed after the hooking is removed. On the other hand, if the error of the vertical and horizontal components is judged, the two divided areas are put together again and then the number of areas is compared. shall.

First, when the number of areas is judged to be different by hooking, the difference between the size of the two areas to be compared and the direction of the two areas is more than a certain level, and at the same time, the difference between the sizes of the two areas to be compared and If the difference in direction is not constant, the area currently compared is considered as hooking, and the similarity by hooking is added to the total dissimilarity.

In addition, in the case where it is determined that the area is separated by the error of judging the vertical and horizontal components, the area that should be one originally is divided into two areas because the angle change between two points is more than a certain level. If the direction angle of the area compared to the direction angle between the starting point and the end point of the two areas is a certain difference or less, it is determined that the areas are separated, the two areas are combined, and then the dissimilarity between the areas is calculated. To be recognized.

The calculation of dissimilarity thereof is based on the following equation (3).

In this case, p means the number of regions after the removal of the hooking components from the n input patterns and the m comparison patterns, and after combining the divided main regions with each other, in the two patterns, that is, the number of modified regions. P <min (n, m) is satisfied.

The following is a process of combining semantic units by making a candidate character by trying all possible cases where the character may consist of the strokes of the last recognized semantic unit, and then using the character consisting of the semantic unit with the lowest dissimilarity as the recognition character. Looking at it as follows.

All possible cases in which characters may be composed of strokes recognized by the semantic unit recognition process are attempted to create a tree, and as a result, a recognition character and a candidate character are generated.

In other words, as shown in FIG. 2, the process of recognizing the input character 'ga' consisting of three strokes is performed by first making one stroke into a cell, recognizing the first tree, and then the first stroke and two strokes. The process of creating a second tree by making strokes into cells is performed sequentially.

If a node recognizes a plurality of nodes, a new tree is created.

If the semantic unit recognition of the nodes of each tree fails, or the positional relationship between the current node and the previous node is not correct, no further node expansion and no identifier are made.

In this way a tree is created for all cases, and candidate characters are recognized in the tree reaching the last stroke.

As a result, when several candidate strings are formed by the tree, the character that is composed of semantic units having the lowest dissimilarity among the recognized results is recognized as the character of recognition.

As described above, in the present invention, since the characters are recognized in the smallest unit in meaning, characters in multiple languages can be recognized, and the recognition result and the candidate result are taken instead of only one character recognition result. As a result, it has a function to correct misperceptions.

1 is a diagram showing the process of the present invention handwritten multi-character recognition method.

2 is a diagram showing a combination process of FIG.

Claims (7)

In the first process of distance filtering, the overlapping points and the points existing within a certain distance are removed from the input point of the cell, which is a set of input strokes consisting of one or more strokes. Therefore, a second process of extracting feature values that can make use of the characteristics of each region after separating the regions, and feature values and dissimilarity diagrams of reference meaning units that have previously stored the feature values extracted by the second process in a database The candidate process by attempting the third process of using the meaning of the reference unit having the lowest dissimilarity as the recognized semantic unit, and all possible cases in which the characters may be composed of the strokes of the semantic unit recognized by the third process. Handwriting, characterized in that the fourth step of making the recognition character of the letter consisting of the lowest semantic unit after the sum of dissimilarity Multiple character recognition method. The method of claim 1, further comprising a normalization process for reproducing a handwritten character form while minimizing a memory capacity after the second process. The method of claim 1, wherein the feature value extracted in the second process is information about a stroke movement in the region. The method of claim 1, wherein the feature value extracted in the second process is information about a location of a size of an area. The method of claim 1, wherein the feature value extracted in the second process is information about a start point and an end point of a stroke in a region. The method of claim 1, wherein the database is a plurality of databases having a number of areas corresponding to the number of areas divided by the second process. 4. The handwritten multi-character recognition of claim 3, wherein the information on the movement of the stroke is a difference value between an angle of each point and an angle formed by the beginning and the end of the area by dividing an area into predetermined portions regardless of length. Way.