KR100204618B1 - Method and system for recognition of character or graphic - Google Patents

Abstract

The present invention provides a method of character recognition using information of information other than the writing order of characters input online.

A method of recognizing a character or figure input on a writing area online, the method comprising: sampling a character or figure input online and extracting sampling information; and a plurality of localities from the input character or figure based on the sampling information. Determining a region, obtaining a feature vector for each local region, obtaining a vector sequence based on the arrangement on the writing region of each feature vector, and recognizing input characters or figures based on the vector sequence How to have a step.

Description

Recognition method and system of character or figure

The present invention relates to a method and system for recognizing a character or a figure, and more particularly, to a method and system for recognizing a handwritten Chinese character inputted online.

As an automatic system for recognizing handwritten characters or figures input online, many techniques have been proposed in the past. For example, Japanese Patent Application No. 4-220410 (U.S. Patent No. 5,343,537) discloses the contribution of one or several feature vector spaces, Gaussian modeling in each space, and all associated prototypes in all spaces. Disclosed is a method and apparatus for automatically recognizing a written text on the basis of a proper display of a written text in the mixed composite considered.

Specifically, in response to the writing of the writer by the stylus on the electronic tablet, a known character including a handwriting input from the writer is sampled. A handwriting prototype is provided in response to the provision of the parameter-vector display in the handwriting space of the sampled known characters, which performs the parameter-vector display in the handwriting space of the sampled known characters. In response to the writing by the stylus on the tablet, unknown characters including the handwriting input recognized by the stylist are sampled to perform parameter vector display in the sample space of the sampled unknown characters. Writing a handwriting prototype based on a ballistic comparison between a handwriting prototype and a parameter-vector representation in the handwriting space of the sampled unknown character, including making a list of candidate characters. Similarity is evaluated in which the drawing is a candidate character recognized as an unknown character. Then, the ballistic analysis of the candidate character list is performed to recognize sampled unknown characters including the recognized text input.

In addition, Japanese Patent Application Laid-Open No. 4-328128 discloses a method and system for performing handwriting recognition using a hidden Markov model.

In general, online handwritten character recognition can recognize characters with a high probability when it can be assumed that the characters are written in the correct handwriting order. The above-described prior art also performs character recognition on the basis of such handwritten order, that is, temporal information called a stroke order. However, in characters with many strokes such as Chinese characters, the writing order varies depending on the person writing the letter, so that the same letter and the temporal information are different. Therefore, the same letter may be recognized as wrong in the writing order, or the character recognition may be performed by comparing the models extracted and written from the input letters with a plurality of models stored in the text recognition system in advance. Is achieved. Therefore, if the models of several patterns expected for any one character are stored in advance in consideration of the difference in the order of the above-described strokes, the dictionary contents inevitably expand. This is not preferable because it is related not only to the increase of the preliminary capacity but also to the decrease of the recognition speed.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and an object thereof is to propose a recognition method that is not limited to the input handwriting order and the number of conversation. This method is effective for the recognition of many Japanese, especially Chinese characters, compared to English, such as the number of characters.

Another object of the present invention is to perform character recognition using information other than the writing order of characters, that is, information other than temporal information.

1 is a block diagram of the handwriting-recognition system in this embodiment.

2 is a flowchart of a method of recognizing a handwriting in the present embodiment.

3 is a flowchart in which points of vector columns of feature vectors are obtained.

4A and 4B show a set of points and a local area when "to" is input to a tablet writing area.

FIG. 5 is a table showing points constituting in each local area. FIG.

6 is a table depicting the y coordinate of each local area.

* Explanation of symbols for main parts of the drawings

102: Recognition Program 104: Front End

108: Modeling Elements 110: Computer Platforms

112: operating system 114: micro instruction code

116 hardware element 118 random access memory (RAM)

120: central processing unit (CPU) 122: input and output interface

124: terminal 126: electronic input tablet

128: data storage device 130: printer

According to an aspect of the present invention, there is provided a method of recognizing a character or a figure input on a writing area, the method comprising: extracting sampling information by sampling the input character or figure; Determining a plurality of local areas from the input characters or figures based on the sampling information; Obtaining a feature vector for each local region; Obtaining a vector column based on the arrangement on the writing area of each feature vector; And recognizing an input character or figure based on the vector column. Here, the writing area refers to an area in which a character or figure to be recognized exists. Then, the vector sequence of the feature vectors can be obtained by arranging the feature vectors in order based on the position of the local region on the region. This writing area is generally a frame area or a underline area where characters are input for online character recognition.

Here, it is preferable that the step of obtaining the vector columns is determined based on the order in which the local regions in the opposite directions from one direction on the write region are arranged. More specifically, it is preferable to obtain the writing area represented by the x and y coordinates by referring to the local area corresponding to the vector column and arranging the y coordinates sequentially from the larger one.

In another aspect, the present invention provides a method for recognizing a character or a figure input on an online writing area, the method comprising: extracting sampling information by sampling a character or a figure input online; Determining a plurality of local areas from the input characters or figures based on the sampling information; Obtaining a feature vector for each local region; Obtaining a first vector sequence consisting of a plurality of feature vectors based on a sequence of strokes of characters or figures written online; Obtaining a second vector column consisting of a plurality of feature vectors based on the arrangement on the writing area of each feature vector; And recognizing an input character or figure based on the first vector column and the second vector column.

Here, it is preferable that the step of obtaining the second vector column is determined based on the order in which the local regions are arranged in opposite directions from one direction on the writing region.

The recognizing may include recognizing the input character or figure using a hidden Markov model method or a DP matching method.

In another aspect, the present invention provides a step of recognizing a character or a figure input on a writing area online, comprising: means for sampling the character or figure input online and extracting sampling information; Means for determining a plurality of local areas from input characters or figures based on sampling information and obtaining a feature vector for each local area; Means for obtaining a vector column based on the arrangement on the writing area of each feature vector; A system having means for recognizing input characters or figures based on vector strings is provided.

Here, it is preferable that the means for obtaining the vector column obtains the vector column based on the order in which the local regions are arranged in opposite directions from one direction on the write area.

Hereinafter, as an exemplary embodiment of the present invention, online character recognition will be described as an example. 1 is a block diagram of a handwriting recognition system according to the present invention. The system includes a computer platform 110. Computer platform 110 has hardware elements 116 consisting of random access memory (RAM) 118, central processing unit (CPU) 120, input / output interface 122, and the like. Computer platform 110 has operating system 112 and may have microinstruction code 114.

The computer platform 110 is connected to a means for extracting sampling information by sampling characters or graphics input on the writing area online, for example, an electrical writing tablet 126. This tablet 126 is for a user to write a desired character or figure in a writing area using an input pen. Many peripheral devices, such as the terminal 124, the data storage device 128, and the printer 130, are also connected.

In the platform device 110, the recipient recognition program 102 operates. The aquatic recognition program 102 operates the front edge 104, the relocation processing mechanism 106, and the modeling element 108. Here, the front edge 104 is used to determine a plurality of local areas from the input characters or figures based on sampling information, and to obtain a feature vector for each local area. The relocation processing mechanism 106 is for obtaining a vector column based on the arrangement on the writing area of each feature vector. This can be obtained, for example, on the basis of the order in which the local areas in the opposite directions from one direction on the writing area are arranged. In addition, the modeling element 108 is for recognizing a character or a figure input by a user based on the vector column.

2 is a flowchart of a method of recognizing a handwriting in the present embodiment. First, sampling information is extracted by sampling characters or graphics input online (step 201). This is because the character or figure written by the user on the writing area of the electronic input tablet 126 is subjected to sampling.

Based on this sampling information, a plurality of local areas are determined from the input characters or figures (step 202). That is, each sample point of the input extracted by step 201 is a point defined by the coordinates (X _n , Y _n ) in the writing area. The intervals between these points are usually displayed as a function of the speed written by the user because they are not necessarily constant. Here, each sample point is a point p (X _m, Y _m) gyeokin normalized deunggan. As such, each time-dependent sample point captured by tablet 126 is transformed into a time-dependent representation where all points are equidistant.

Next, the local region is determined based on the point p which is the normalized equal interval. Here, the local area means a characteristic part which is required when recognizing arbitrary characters and figures. In general, the local area is often an area including the maximum value and the minimum value of the start point, the end point, or the x and y coordinates of the stroke. The local area is determined to be composed of the same number of dots (for example, 2K + 1), and the character and figure have a plurality of local areas.

For each of the plurality of local regions obtained by step 202, a corresponding feature vector is obtained (step 203). This step is also performed with the step 202 on the front end 104 of the system shown in FIG.

Here, the feature vector refers to a vector having a parameter such as a characteristic of an arbitrary local region, for example, a relationship between points such as coordinates of each point in the region or a curve of a stroke. Parameters including how many points in the local area are displaced from other points in the same area are included.

In the case where the number of feature vectors in an arbitrary character is N, a vector sequence showing the order of the N feature vectors is also determined in this step. However, the order of this vector column is changed based on the arrangement on the writing area as described below. Therefore, since what is obtained in this step is the primary vector series, the order of the vectors may be in any order. The initial vector strings obtained here are generally arranged based on the stroke order, i.e., the temporal order, of characters or figures input online.

The obtained vector sequence is obtained in order based on the obtained plurality of feature vectors (step 204). In the step 203, when a vector column has already been obtained according to the order of the strokes, the case of replacing the order based on the arrangement on the writing area is also included. This step will be described later. This step is also performed for the relocation processing mechanism 106.

Finally, based on the vector sequence, the input character or figure is recognized (step 205). In step 204, the system recognizes the input characters and figures by modeling a vector sequence based on the arrangement on the writing area with a hidden Markov model HMM). That is, recognition of characters and the like is achieved by sequentially comparing the vector sequence of feature vectors with a hidden Markov model of arbitrary characters stored in the system's dictionary, and specifying the best of the scores. In other words, it is also possible to recognize using the DP matching method. Thus, by applying the present invention to an HMM that performs character recognition based on the writing order, it is possible to recognize a character with a high probability even for characters not written in the correct writing order or written continuously.

3 is described in more detail with respect to step 204 above. First, the y-coordinate of one local area of any character / shape obtained in step 202 is obtained (step 301). However, since there are a plurality of points constituting an arbitrary local area, and the y-coordinate also has a predetermined range, several algorithms are required to specify the y-coordinate of the local area as one. Therefore, since the y coordinate y _n of the local region corresponding to the feature vector f _n is specified, when the local region is composed of 2K + 1 points, the K + 1st point is represented by the y coordinate of the local region. define. For example, when the local area consists of five points, the y coordinate of the third point p is the y coordinate of the local area. By defining in this way, the y-coordinate of an arbitrary local region can be easily specified.

The y-coordinate of the local region is defined in this way because it gives an evaluation criterion for replacing the vector column of the feature vector based on the arrangement of the writing region. Accordingly, the present invention may be considered in various ways in addition to this definition. For example, y coordinates of 2K + 1 points may be obtained, and the evaluation value may be defined as the y coordinate of the local area.

Incidentally, the replacement of the vector column is based on the y coordinate in this embodiment, but of course it may be based on the x coordinate. It is an important feature of the present invention to perform character recognition with a vector column based on the arrangement on the writing area, instead of a vector column of feature vectors based on the stroke order of characters written online. Therefore, it is not limited to this method as long as the magnitude relationship of the coordinates of a local area | region is grasped | ascertained from one direction to an opposite direction on a writing area | region.

It is determined whether the y coordinates of all local regions are to be obtained (step 302). In the case of "no", the process returns to step 301 and is repeated until the y coordinates of all local regions are determined. In the case where a character or figure is composed of N local areas by this series of steps, all y-coordinates of those local areas are determined.

If " Yes " in step 302, i.e., the y-coordinates of all local regions are obtained, array A is created (step 303). The array A has parameters of the feature vector and the set (f _n , y _n ) of the y coordinate of the local region corresponding thereto.

The array A is replaced in large order of the y coordinate of the local area (step 304). This sort can be done with existing sorting algorithms.

Determine the vector column (step 305). Extracting and enumerating the feature vectors from the sorted array A in that order is called a vector column.

The character can be recognized by comparing the vector sequence obtained by the above procedure (steps 301 to 305) with the pre-stored contents of the system of the step 205.

Next, specifically as an example of character recognition, as shown in FIG. 4, the case where "(to)" is input to the writing area of a tablet is demonstrated. As shown in Fig. 4A, the characters input to the writing area are normalized as a set of 34 points specified according to the x-coordinate in the transverse direction and the y-coordinate in the longitudinal direction, not shown. The system determines _six local regions (r ₁ to r ₆ ) as shown in FIG. 4b based on a set of 34 normalized points. The local area is configured to include five points, and is determined to include an area where the start point, the end point, the x-coordinate value, or the y-coordinate value of each stroke of the input character becomes maximum and minimum.

FIG. 5 shows the relationship between the six local regions (r ₁ to r ₆ ) specified in this way and the point p _n constituting the local region. Further, here, the _{r 1, r 2, r 3} , and r _6, includes a point p _o claim 4a are not shown in Fig. Such a processing point is set so that the start point and the end point of the stroke become the center of the local area, and the algorithm for specifying the local area including the start point and the end point and the other areas are not distinguished. This is because it is treated as. Therefore, as can be understood from FIG. 5, the y-coordinate of the local area is the n + 1st point, that is, the third point which is a true point because n = 2, so that the third point is specified The specific value of the point p ₀ does not matter.

The feature vector corresponding to the local regions r ₁ to r ₆ is obtained from f ₁ to f ₆ , and a vector sequence based on the sequence of input strokes is referred to as (f ₁ , f ₂ , f ₃ , f ₄ , f ₅ , f ₆ ). do. In this case, the y-coordinate of each local area is shown in FIG. Therefore, the arrangement A consisting of the feature vector and the y-coordinate element of the local region is as follows.

Array A: (f ₁ , 100), (f ₂ , 59), (f ₃ , 73), (f ₄ , 27), (f ₅ , 0), (f ₆ , 19)

This is sorted in the order of the largest y coordinate using the array algorithm. It becomes as follows.

Array A: (f ₁ , 100), (f ₃ , 73), (f ₂ , 59), (f ₄ , 27), (f ₆ , 19), (f ₅ , 0)

In this way, only the feature vectors are extracted to obtain vector columns f ₁ , f ₃ , f ₂ , f ₄ , f ₆ , f ₅ based on the magnitude of the y coordinate.

This obtained vector column is the same also in the case where the order of the "to" strokes of FIG. 4A is reversed. Conventionally, since a separate vector sequence is generated according to the order of stroke even for the same character, a plurality of models according to the order have to be prepared in advance even for one character. However, in this method, since a vector sequence of the same feature vector is obtained irrespective of the order of strokes, one model in advance may be used. Therefore, the storage capacity of the dictionary can be reduced, and the recognition time can be shortened.

Moreover, in actual character recognition, many kinds of letters and small kinds of writing order are mixed in many cases. In such a scene, the character recognition rate can be improved by taking into account both scores by using two methods, a recognition method of the present invention based on the arrangement of writing areas and a recognition method based on the order of strokes. Moreover, the method of making emphasis on the score obtained by the method of this invention with respect to the character which the writing order most likely differs with a person is important, and the method of making importance on the score based on stroke order for the character with low possibility. In this case, both the vector column in which the feature vectors are arranged based on the stroke order of the characters or figures written online and the vector column in which the feature vectors are arranged based on the arrangement of the local regions (corresponding to the feature vectors) on the writing area are obtained. For each, we need to add a score against the model.

In the present invention, since the same feature vector is obtained even if the user inputs characters in any writing order, one character model can be reduced. Therefore, the capacity of the dictionary in which the model is stored can be reduced, and character recognition can be performed at high speed.

Claims (11)

A method of recognizing a character or a figure, the method comprising: extracting sampling information by sampling a character or a figure present in an arbitrary area; Determining a plurality of local areas in the character or figure based on the sampling information; Obtaining a feature vector for each said local region; Obtaining a vector column of the feature vector by arranging each of the feature vectors in order based on the position of the local region on the region where the character or figure is present; And recognizing the character or the figure based on the vector sequence. The method of claim 1, wherein the obtaining of the vector sequence is based on a position order of each of the local areas from one direction on the plane area where the text exists to the opposite direction. Way. A method of recognizing a character or figure input online, the method comprising: extracting sampling information by sampling a character or figure input in a writing area; Determining a plurality of local areas from the input characters or figures based on the sampling information; Obtaining a feature vector for each said local region; Obtaining the vector column based on an arrangement on the writing area of each feature vector; And recognizing the character or the figure based on the vector sequence. The method of claim 3, wherein the step of obtaining the vector sequence is based on the order in which the local regions are arranged from one direction on the writing area to the opposite direction. The method of claim 1 or 2, wherein the character is a Chinese character. A method of recognizing a character or a figure input on a writing area online, the method comprising: sampling the input character or figure and extracting sampling information; Determining a plurality of local areas from the input characters or figures based on the sampling information; Obtaining a feature vector for each said local region; Obtaining a first vector sequence consisting of a plurality of said feature vectors based on stroke orders of characters or figures written online; Obtaining a second vector column consisting of a plurality of said feature vectors based on an arrangement on said writing area of each said feature vector; And recognizing the input character or figure based on the first vector column and the second vector column. 7. The method of claim 6, wherein the step of obtaining the second vector column is based on the order in which the local areas are arranged from one direction on the writing area to the opposite direction. The character or character of claim 1 or 2, wherein the recognizing comprises recognizing the input character or figure using a hidden Markov model method or a DP matching method. Recognition method of figure. CLAIMS 1. A system for recognizing text or graphics input on a writing area, comprising: means for sampling input text or graphics to extract sampling information; Means for determining a plurality of local regions from the input characters or figures based on the sampling information, and obtaining a feature vector for each of the local regions; Means for obtaining the vector column based on the placement of each feature vector on the write area; And means for recognizing the input character or figure based on the vector sequence. 10. The character or figure recognition according to claim 9, wherein the means for obtaining the vector sequence obtains the vector sequence based on the order in which the local regions are arranged from one direction on the writing area to the opposite direction. system. 10. The system of claim 9, wherein said means for recognition is means for recognition using hidden Markov models or DP matching.