JPH0644409A - Character recognizing device - Google Patents

Abstract

PURPOSE:To reduce the device scale, to execute the recognition at a high speed, and improve the recognition rate by adopting a recurrent type neural network for the character recognizing device using a neural network. CONSTITUTION:With respect to the character recognizing device consisting of a character input means 101, a neural network 102, and a recognition result output means 103, a recurrent type neural network in which each node 104 for constituting a neural network 102 has a feedback in the inside is adopted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a character recognition device used for inputting characters and the like into a computer.

[0002]

2. Description of the Related Art Conventional character recognition devices are classified into two types: online handwritten character recognition devices and offline character recognition devices.

In online handwriting recognition,
In many cases, character recognition is performed by the basic stroke method as described in PRU82-37 (title "Online handwritten Chinese character / Hiragana recognition", hereinafter referred to as Conventional Example 1a).

In the basic stroke method, the strokes of characters input as time-series coordinate data are classified into predetermined basic strokes (hereinafter referred to as stroke recognition), and each character of the recognition candidate is the basic. This is a method in which the dictionary described by strokes is compared with the strokes after the classification, and the character having the highest similarity is used as the recognition result. In stroke recognition, the stroke from the time the pen descends on the input surface to the time it rises is treated as one stroke, and after performing line approximation and direction coding, classification is performed by methods such as cumulative length of different direction codes and DP matching. Was there.

Since this method has the features of a relatively small amount of calculation, a small dictionary capacity, and a high recognition rate, it is adopted by many of the online handwritten character recognition devices currently on the market.

[0006] Further, there is an electronic theory D-
111, No. 12, pp 1047-1056 (Title "Online handwritten Chinese character recognition by improved perceptron considering identification order", hereinafter referred to as Conventional Example 1b) has been reported. This solves the problem that it is difficult to learn a neural network by the three-layer structure and the error back-propagation learning method by a simple perceptron and a new learning method.

Next, various types of off-line character recognition devices have been proposed. Particularly, if a neural network is used, it has an advantage that one character can be cut out from a character string at the same time as character recognition. In such an earlier application, Japanese Patent Application Laid-Open No. 3-163684 (hereinafter referred to as Conventional Example 2a
Call. ) And the theory of theory Vol. J74-D-2 No. 1
1 pp. The ones described in 1556-1564 (hereinafter referred to as Conventional Example 2b) are typical.

The neural network used in Conventional Example 2a is of Hopfield type. The input field of view of the neural network is fixed, and the coupling weight is calculated and determined based on the energy function in advance.

The neural network used in the conventional example 2b is a feedforward type having a hierarchical structure. In order to handle time series data, a special node is provided in the input layer and feedback is performed from the output layer to the input layer.

[0010]

However, the conventional character recognition device has the following problems.

First, the problems in the online handwritten character recognition device are as follows.

With the recent spread of information processing by a computer, an online handwritten character recognition device is often used as the most familiar computer input means. However, the conventional online handwritten character recognition device has the limitation that it cannot be recognized unless the stroke number and stroke order are accurately written in a regular pattern, and in order to reduce the burden on humans, it is essential to be able to recognize a certain number of consecutive characters. It is said that. It is also essential in the English-speaking world to recognize continuous cursive English cursive character strings.

In the conventional example 1a, the basic stroke is
It is specified to correspond to one section of a regular print. Therefore, when a plurality of basic strokes, which should be written separately with the pen-up inserted in between, are written down in succession (hereinafter,
Call it a continuous stroke. ), Stroke misrecognition easily occurs. That is, in principle, the basic stroke method cannot recognize characters with continuous strokes (hereinafter, referred to as continuous characters).

Therefore, several methods are conceivable in order to overcome the above-mentioned drawbacks.

One is a method in which a continuous stroke is registered as a new basic stroke, and the basic stroke up to that point and the new basic stroke are described in parallel in the dictionary. However, there is a limit to the number of new basic strokes that can be prepared, and there is a trade-off between the allowance of consecutive characters and the dictionary description amount, and the allowance of consecutive characters and stroke recognition time.

The second is Japanese Patent Laid-Open No. 4-32431 (hereinafter,
This is called Conventional Example 1c. ), A pattern in which all strokes of one character are connected in the dictionary (hereinafter,
It is called a one-stroke pattern. ) Is registered, one writing pattern is generated from the input handwriting by complementing, and these are written as DP.
It is a method of comparing by matching.

However, with this method, it is necessary to prepare as many basic patterns for DP matching as one character in the expected stroke order, and the dictionary capacity and calculation amount tend to be enormous.

A third method is to recognize by extracting a basic stroke from continuous strokes by including a stroke extraction instruction in the dictionary in addition to the normal basic stroke description. Japanese Patent Application Laid-Open No. 2-56689 is a prior application having such contents. This method can solve the problem that the second method has a huge amount of calculation.

However, when DP matching is used as the stroke extraction method, there is a problem of where to set the matching start point and end point in the continuous stroke. If all the points in the continuous stroke are targeted for the matching start and end points, the processing time becomes extremely long.

When DP matching is not used,
It is difficult to extract strokes with curved parts while absorbing handwriting deformation. In the previous application, only linear strokes are extracted, but this is insufficient for recognition of characters other than Chinese characters.

In the prior art example 1b as well, it can be inferred that the stroke order is free, but fluctuations in the number of strokes and continuous writing are not possible due to the limitation of the characteristics used. By learning the one-stroke writing pattern as a basic pattern and converting the input into the one-stroke writing pattern, it may be possible to deal with continuous characters. However, the problem that there are as many continuous writing patterns as there are stroke orders to deal with This is the same as the second case described above.

Further, none of the above-mentioned methods is suitable for recognizing a continuous writing English cursive character string.

The conventional online handwritten character recognition device has the above-mentioned problems, and online handwriting that does not increase the calculation amount and the dictionary capacity and can recognize continuous cursive and continuous cursive English cursive character strings. There was a problem of proposing a character recognition device.

Next, the problems in the off-line character recognition device are as follows.

In the conventional example 2a, it is impossible to optimize the neural network and improve the recognition performance by learning. Also, the local minimum problem, which is generally pointed out for Hopfield neural networks, cannot be ignored.

The conventional example 2b has a problem that the convergence of learning is extremely difficult. This is a problem generally pointed out for a hierarchical feed-forward neural network having feedback from output to input. Further, since the recognition result is obtained by providing a plurality of nodes in the output layer and one of the nodes is excited, if a new recognition target character needs to be added to the neural network, It takes time because the entire network has to be relearned. Further, because of the hierarchical structure, the associative ability is small, and there is a problem that the recognition performance is significantly reduced when there is noise or omission in the input character. The conventional off-line character recognition device has the above-mentioned problems, and has a problem of proposing an off-line character recognition device that can easily generate a neural network and has noise resistance due to the associative ability.

[0027]

The present invention is characterized by the following points in order to solve the above-mentioned problems.

1) A character recognition device using a neural network: a) character input means for generating time-series data of a character to be recognized; b) the neural network, a storage means for storing internal state values, and a memory. The internal state value and the external input value are used to update the internal state value, and a neuron-like element having a generating means for generating the external output value from the internal state value. C) The time series data Use as input to the neural network. 2) In the character recognition device of 1), a) stroke extracting means for extracting one or more predetermined basic strokes from time series data, and b) each character of the recognition candidate is described by the basic strokes. A dictionary that includes: c) a comparison unit that outputs the recognition result by comparing the extracted stroke with the dictionary; d) the character input unit converts the handwriting of the recognition target character into time series data in the order of occurrence; The stroke extraction means includes a neural network, and f) the neural network is trained to detect one or more basic strokes from a plurality of continuous strokes.

3) In the character recognition device of 2), the character input means approximates the handwriting of the recognition target character with a polygonal line, and then quantizes the direction of each polygonal line with a predetermined division number to generate a series of directional code strings. 4) In the character recognition device of 2), the character input means should approximate the handwriting of the recognition target character with a polygonal line and then generate the nodal coordinates of the polygonal line.

5) In the character recognition device of 1), the character input means inputs a partial handwriting appearing in the window as time series data while scanning a window of a predetermined size on the handwriting of the character to be recognized. Thing.

[0031]

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

(Embodiment 1) First, FIG. 1 is a basic block diagram of a character recognition apparatus of the present invention.

The time-series character data input by the character input means 101 is input to the neural network 102, and the output thereof is analyzed by the recognition result output means 103 and becomes a recognition result character code. Reference numeral 104 is one of the nodes forming the neural network.

A tablet is generally used as the character input means, and from this, time-series XY coordinate data of the character handwriting is output. Generally, since it is too redundant as the recognition data as it is, the polygonal line approximation is performed. After that, the coordinates of the start point and the end point of each polygonal line may be sent to the stroke extraction means as new time series data, or the direction of each polygonal line may be quantized by a predetermined number of divisions and then converted into a series of direction code strings. You can send it. The direction coding enables recognition regardless of the position where the character is written.

An example of the configuration of the neural network is that the number of nodes in the input layer is equal to the number of dimensions of the input character data, the number of hidden layer nodes is 8 to 32, and the number of output layer nodes is 2 (positive output and negative output). Are listed in. One neural network corresponds to one recognition target character, and if the input character data matches the corresponding character of the neural network, the output layer node raises the positive output and lowers the negative output. , If they do not match, it is learned to lower the positive output and raise the negative output. Therefore, at the time of character recognition, the connection weight of the neural network is reset for each recognition candidate character, the input character data is input, and the output is repeated, and the character for which the positive output is raised and the negative output is dropped is set as the recognition result. If the device scale permits, a neural network can be provided for each recognition candidate character, and input character data can be input to all neural networks at the same time and processed in parallel.

FIG. 2 is an explanatory diagram of one node of the neural network of the present invention.

The node 114 comprises an internal state value updating means 111, an internal state value storing means 112, and an output value generating means 113, and has a feedback connection from the internal state value storing means 112 to the internal state value updating means 111. ing. With this configuration, the inputs from the other i nodes to the node are I (i), the bias input is D, the coupling weight for the input I (i) is W (i), and the current internal state value is Is X, a constant is k, an output value of this node is Y, an output function is f (x), and time is t, the operation of this node is described by Equations 1 and 2.

[0038]

[Equation 1]

[0039]

[Equation 2]

Further, as f (x), a sigmoid function as expressed by the equation 3 is suitable.

[0041]

[Equation 3]

The neural network according to the present invention has a feedback connection inside the node, which is a conventional example 1.
b, 2a, and 2b, which is a significant difference, and with such a configuration, a single node can have the ability to process time-series data, so that the processing capability for time-series data is higher than that of a conventional device. Of. (The node of the conventional example itself does not have the ability to process time-series data.) FIG. 3 shows the character recognition device of the present invention at the same time as cutting out a character "b" from a continuously written English cursive. It is a figure explaining how to give the learning data to a neural network, when applying it to the online handwritten character recognition which performs. The learning data is created so that a character that should output a positive output is represented by O and a character that should output a negative output by X, so that XX, XX, and XX cover all recognition target characters. In the present embodiment, the case where the circle is “b” 123 and the circle is “a” 124 is shown.

FIG. 4 shows the character recognition device of the present invention.
It is a figure which shows the output example at the time of giving the continuously written English cursive "cba" to a neural network. Since the positive output 132 rises and the negative output 131 falls almost at the center of the input "b" in the character string, it is detected that "b" exists at this position.

As described above, according to the present invention, it is possible to recognize a continuous writing English cursive character string, which was impossible in the conventional invention.

In order to recognize the continuous characters of the kanji, the conventional example 1c is used.
Similarly, it is necessary to use a one-stroke writing pattern for the character data at the time of learning and input the handwriting of the recognition target character in one-stroke writing pattern. In this case, it is necessary to prepare the neural networks in the expected stroke order.

(Embodiment 2) FIG. 5 is a block diagram showing a case in which the character recognition device of the present invention is applied to online handwritten character recognition using the basic stroke method.

Continuous characters are input by the character input means 201, and the data is sent to the stroke recognition means 202. The stroke recognizing means uses the neural network 204 to extract basic strokes from continuous strokes in consecutive characters, and sends the basic strokes to the comparing means 203. The comparing means compares the basic stroke description with the extracted basic stroke while scanning the dictionary 205, and outputs the best matching character as a recognition result.

Extraction of basic strokes from continuous strokes will be described in detail with reference to FIGS. 5 and 6.

The neural network is trained to extract the third stroke 211 of "ha". The learning method is the same as that in the first embodiment and FIG. When the character "212" is continuously input to such a neural network, the positive output 221 and the negative output 22 as shown in FIG.
2 comes out. The stroke recognizing means recognizes from the output that there is a third stroke of "ha", and at the same time, reversely estimates the position of the third stroke of "ha" in the input stroke from the position where the output is standing and compares. Sends to the means a flag indicating that there was a third stroke of "wa" and said position.

(Embodiment 3) FIG. 8 is a diagram showing a character input means and a neural network when the character recognition device of the present invention is applied to an off-line character recognition device. Also,
The overall configuration of the device is similar to that of FIG.

In FIG. 8, the window 301 provided in the character input means scans the character string 302 from the left side.
The window is further divided into subsections, each subsection converting the portion of the character visible underneath it into numerical data. For this conversion, for example, there is a method of counting the number of black pixels included in a small section, but various character features in a general character recognition device may be extracted. If an image scanner is used as the character input means, the windows and subsections are some solids of the image sensor.

The small section is a neural network 30.
There is a one-to-one correspondence with each node in the input layer of No. 3, and the data of the character string which is time-series by scanning the window is input to each node. The neural network is learned to detect one character of all recognition candidate characters in the same manner as in the first embodiment.

As an example, the neural network is "
C has been learned to detect "A"
The output of the neural network when the character string "D" is input is as shown in Fig. 9. When the window hits "C", the positive output 311 rises and the negative output 312 falls, so "C" is detected. A recognition result output means analyzes such an output and converts it into a recognition result character code and a recognition position.

If the neural network is fully connected in the above apparatus, the associative ability of the neural network greatly reduces the recognition rate of the conventional recognition apparatus such that the recognition rate is extremely low for the handwriting with noise or missing characters. Can be improved.

[0055]

The effects of the present invention are listed below.

Regarding the on-line handwritten character recognition device: 1) Character recognition or stroke extraction can be performed by giving an input character to the neural network only once, so the amount of calculation is smaller than that of DP matching.

2) The method of causing the neural network to extract strokes only needs to extend the conventional basic stroke method. The basic stroke method has the advantages of high speed and small dictionary capacity, and continuous character recognition is possible. to enable. In particular, the feature is that strokes with curved parts can be extracted at high speed.

Regarding the off-line character recognition device: 1) The neural network of the present invention converges learning more than a hierarchical feed-forward type net configured to handle time series data by feedback from the output layer to the input layer. Is fast and easy to converge.

2) If a different neural network is used for each character to be recognized, it is only necessary to add a new neural network even when the number of characters to be recognized increases, and it is not necessary to re-learn the entire learned neural network. Absent.

3) Since a fully-combined structure can be adopted, the associative ability is not inferior to that of the Hopfield type, and the recognition rate is less likely to drop even when the input character is lost or noise is superimposed.

[Brief description of drawings]

FIG. 1 is a basic configuration diagram of a character recognition device of the present invention.

FIG. 2 is an explanatory diagram of one node of the neural network of the present invention.

FIG. 3 is a diagram for explaining how to provide learning data to the neural network of the present invention.

FIG. 4 is a diagram showing an output example of the neural network of the present invention.

FIG. 5 is a configuration diagram when the character recognition device of the present invention is applied to online handwritten character recognition using the basic stroke method.

FIG. 6 is a diagram illustrating a basic stroke for learning and a continuous stroke for input.

FIG. 7 is a diagram illustrating a method of extracting a basic stroke from a continuous stroke according to the present invention.

FIG. 8 is a diagram showing a character input unit and a neural network when the character recognition device of the present invention is applied to an off-line character recognition device.

FIG. 9 is a diagram showing an output example of the neural network of the present invention.

[Explanation of symbols]

 101 character input means 102 neural network 103 recognition result output means 104 node 111 internal state value update means 112 internal state value storage means 113 output value generation means 114 node 121 positive output 122 negative output 123 characters "b" (cursive) 124 characters "A" (cursive) 131 Negative output 132 Positive output 201 Character input means 202 Stroke recognition means 203 Comparison means 204 Neural network 205 Dictionary 211 For learning "(regular) 212 For input" (continuation) 221 Positive Output 222 Negative output 301 Window 302 Input string 303 Neural network 311 Positive output 312 Negative output

Claims (5)

[Claims] 1. A character recognition device using a neural network, comprising: a) character input means for generating time series data of a character to be recognized, and b) the neural network, storage means for storing internal state values, and storage. The internal state value and the external input value are used to update the internal state value, and a neuron-like element having a generating means for generating the external output value from the internal state value. C) The time series data A character recognition device which is used as an input to the neural network. 2. A) stroke extracting means for extracting one or more predetermined basic strokes from time-series data, b) a dictionary in which each character of a recognition candidate is described by the basic strokes, and c. ) Comparing the stroke after the extraction with a dictionary and outputting a recognition result, d) The character inputting means converts the handwriting of the recognition target character into time series data in the order of occurrence, and e) The stroke extracting means The character recognition device according to claim 1, further comprising: a neural network; and f) the neural network is trained to detect one or a plurality of basic strokes from a plurality of continuous strokes. 3. The character input means, after approximating the handwriting of the recognition target character with a polygonal line, quantizes the direction of each polygonal line with a predetermined division number to generate a series of directional code strings. 1. The character recognition device according to 1. 4. The character recognition device according to claim 1, wherein the character input means approximates the handwriting of the recognition target character with a broken line and then generates the nodal coordinates of the broken line. 5. The character input means inputs a partial handwriting appearing in the window as time series data while scanning a window of a predetermined size on the handwriting of the recognition target character. Character recognition device described in.