JPH0565913B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to a pattern recognition device having a function of learning standard patterns according to individual characteristics.

[Background of the invention]

Conventionally, in the fields of speech recognition and character recognition, a method is known in which a recognition dictionary, that is, a set of standard patterns provided for each category, is learned and created for each user individually. When studying, (1)
A method of modifying a standard pattern based on a pattern for each individual (Japanese Patent Application Laid-open No. 103940/1983, etc.), (2)
A method of adding non-standard patterns according to the individual's habits (Japanese Patent Application Laid-Open No. 1987-87536, etc.); (3) A method of first registering each individual's standard pattern and using it from then on (Japanese Patent Application Laid-Open No. 1973-87536); 85048, etc.) are used. Among these conventional learning methods, methods (1) and (2) have the problem that information representing individual habits is buried in the average information of everyone, making it difficult to achieve highly accurate recognition. There is. On the other hand, method (3) may be able to achieve high recognition accuracy for a specific individual, but when the target category reaches 2000 to 3000 characters, such as in handwritten kanji recognition, each use It is extremely difficult to have people write sample patterns for all categories. Furthermore, in order to increase recognition reliability, there is a demand for increasing the number of times each standard pattern is learned, and if this is to be met, the above-mentioned difficulties will be doubled.

[Object of the Invention] The object of the present invention is to create a standard pattern in a pattern recognition device that can be used without a large amount of learning prior to use, and whose recognition accuracy improves as the learning continues. The purpose is to provide a method.

[Summary of the invention]

The present invention includes a first storage means for storing a plurality of standard patterns for each character category, a distance between an input character pattern and each of the plurality of standard patterns stored for each character category, and a distance between the input character pattern and each of the plurality of standard patterns stored for each character category. is the smallest, and if the selected distance is smaller than a predetermined value, the character category to which the standard pattern with the smallest distance from the input character pattern belongs is output as a recognition result, and the selected distance is greater than or equal to the predetermined value. In the pattern recognition device, the pattern recognition device has a recognition means for determining that the input character pattern is unrecognizable if the character pattern is unrecognizable, and a display means for displaying the input character pattern and the recognition result of the recognition means. If the recognition result displayed on the display means is an erroneous recognition,
Or, if the above recognition means determines that it is unrecognizable, input the correct character code for the above input character pattern, and select the above from among multiple standard patterns belonging to the character category corresponding to the input character code. A standard pattern having the smallest distance from the input character pattern is selected, and if the learning frequency of the selected standard pattern is a predetermined value, the selected standard pattern is deleted from the first storage means, and the input character pattern is deleted from the first storage means. Storing the pattern in the first storage means as a standard pattern of the character category, increasing the learning frequency of the stored standard pattern from a predetermined value, storing the increased value in the second storage means, and selecting. If the learning frequency of the selected standard pattern is greater than a predetermined value, the selected standard pattern is deleted from the first storage means, and the selected standard pattern and the input character pattern are divided into ratios according to the learning frequency. further increasing the learning frequency of the stored standard pattern and storing the increased value in the second storage means. Features.

In the present invention, in the initial stage of use of the pattern recognition device, each standard pattern stored in the first storage means, which is a recognition dictionary, is an average standard pattern suitable for everyone, but the second Since the learning frequency stored in the storage means is used to replace the standard pattern in the first storage means with a standard pattern suitable for the individual user, it is possible to improve recognition accuracy without increasing storage capacity. .

[Embodiments of the invention]

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

Characters written on the form 1 are converted into a digital pattern by the photoelectric conversion section 2. The photoelectric conversion unit 2 and various parts of the device to be described later are connected to a bus 3, and are connected to a control unit 4 composed of an ordinary microprocessor.
operates under the control of The operation flow of 4 is the second
The output of the photoelectric conversion section 2, which is shown in the figure and will be described in detail later, is cut out character by character by the cutout section 5 (step 29 in FIG. 2). The cut out pattern feature extractor 6 converts it into a set of character features (feature pattern) used in recognition, and stores it in the buffer memory 7 (step 30 in FIG. 2).
The above characteristic pattern is compared with the standard pattern stored in the memory 9 in the recognition section 8 and recognized (steps 31 to 35 in FIG. 2). Here, the memory 9 is a so-called recognition dictionary, which stores feature patterns determined in advance for each category (character), and its contents are loaded from the external storage device 10 when the power is turned on (steps in FIG. 2). 22). It is preferable to use a medium exchangeable device such as a floppy disk device as the external storage device 10 so that each user can change the disk.

In order to simplify the explanation here, a case will be described in which there is only one standard pattern for each character, but a case in which there are a plurality of standard patterns for each character will be described later.

The recognition results of the recognition unit 8 are displayed on the console 11 one character at a time, or the recognition results of the entire form all at once. Although the display may be displayed in the case of unrecognition (reject), there is a drawback that the case of erroneous recognition is not displayed. In the following explanation, we will discuss the case where each character is displayed (Step 3 in Figure 2).
6). The user checks the displayed recognition results, and if the recognition is unrecognizable or incorrect, press the console 11.
Enter the code representing the correct answer (step 37 in Figure 2). This character code (i) in the control unit 4
When received, the content of the address corresponding to this character code is read from the memory 12. The memory 12 stores the learning frequency for the standard pattern P _i of each character, and like the memory 9, its contents are loaded from the external storage device 10 when the power is turned on (step 22 in FIG. 2). When the content read from the memory 12 (the standard pattern learning frequency F _i for the character code input from the console 11) is 0, the feature vector X stored in the buffer memory 7 is directly transferred to the corresponding address in the memory 9. is written as a standard pattern P _i of that character (steps 38 to 39 in FIG. 2).
At the same time, the learning frequency F _i of the standard pattern in the memory 12 is set to 1 (step 40 in FIG. 2). on the other hand,
If the learning frequency F _i is not 0, a combination of the input feature vector X and the standard pattern P _i is written into the memory 9 as a new standard pattern P _i (step 41 in FIG. 2). At the same time, F _i is incremented by 1 and written into the memory 12 (step 42 in FIG. 2).
Various synthesis methods can be considered, but for example, P _i :=(1-1/F _i +1) P _i +1/F _i +1X (1) can be used (:= indicates substitution). The recognition result or its correction result is output buffer 13.
sent to.

At the time of initial setting, the average patterns of a large number of scribes are stored in advance as standard patterns stored in the external storage device 10, and the learning frequency is set to 0 for all standard patterns. By doing this, a certain degree of recognition accuracy can be obtained even if the user has not been trained at all.
As learning progresses, adaptation to the user's glyph shapes progresses. When the power is turned off or the disk is replaced, the contents of the memories 9 and 12 are saved to the external storage device 10 by an appropriate command (steps 43 to 4 in FIG. 2).
4). An effective method is to prevent the disk from being replaced unless it is evacuated.

Next, we will discuss the case where there are multiple standard patterns for one character category. At this time, P _i , _k are stored in the memory 9 as standard patterns for the character i.
N _i (N i ; integer greater than or equal to 2) of (k=1, ..., _{N i} )
pieces are available. Learning frequencies F _i and _k are stored in the memory 12 corresponding to each. In the case of unrecognized or erroneous recognition, when the user inputs the correct character code (i) from the console 11, the feature vector X of the unknown pattern is compared with each of the standard patterns P _i and _k described above. , choose the most similar standard pattern P _i , _k0 . When the value of the learning frequency _{F i} , _k0 corresponding to this standard pattern is _{0 ,} the feature _vector
Let ₀ be 1. On the other hand, when the learning frequency F _i , _k0 is not 0, X and P _i , _k0 are combined and written into the memory 9 as a new standard pattern P _i , _k0 . at the same time,
Increase the value of F _i and _k0 by 1. As an example of the synthesis method, P _i , _k0 :=(1-1/F _i , _k0 +1) P _i , _k0 +1/F _i , _k0 +1X (2) can be used.

Next, the operation flow of the control section 4 will be explained with reference to FIG.

Figure 2 shows a schematic diagram of the operation of the control unit 4.
(Problem Analysis Diagram). (When you start (step 21),
The standard pattern and learning frequency are read as described above (step 22). Thereafter, the following steps are continued indefinitely until the termination command is input and the processing is completed (step 23). Wait until the form is set and under the reading station (Step 2).
4-25). When the form is set, it is photoelectrically converted (step 26). The following processing is performed on the entire surface of the photoelectrically converted form (step 27). The search for the region where characters exist (step 28) is normally performed in accordance with format control (not shown). Character extraction (step 29), feature extraction (step 30)
As mentioned above, steps 31 to 35 are so-called recognition processing, in which the distance D _j between the feature vector X of the input character pattern and the standard pattern P _i is calculated (step 32), and the minimum value D _J of D _j is calculated. Processing to obtain the character category number J that gives it (step 3)
3) is repeated for j=1 to N (total number of character categories) (step 31). If D is smaller than _{the J} threshold θ, the recognition result is set to J, otherwise it is set to be unrecognizable (steps 34 and 35). The input pattern and recognition result J are displayed (step 36), and when a non-blank (Null) code i is returned from the console as the correct character (step 37), the update of the standard pattern is controlled according to the value of the learning frequency. (Steps 33-42) This has been described above. As described above, after the processing has been completed for the entire surface of the form, the standard pattern and the learning frequency are saved in the external storage device by issuing the end command (step 43).

Although this device is controlled by the above operation flow, this is just an example, and it goes without saying that various modifications can be made.

In the above explanation, it is assumed that learning is performed one character at a time. However, if you want to check and learn the entire form at once, the capacity of the buffer memory 7 should be set aside for the number of characters on the form, and the console 11
It is sufficient to establish a relationship between the contents displayed in the buffer memory 7 and the feature vectors in the buffer memory 7, and this can be easily realized.

As a modification of this embodiment, it is easy to display the cut out character pattern together with the recognition result, or to display the target character pattern in the form image before cutting out with high brightness. realizable. Furthermore, in order to facilitate the detection of misrecognition and the selection of reject character candidates, it is also possible to grammatically process the character code string resulting from the recognition or to check it against a word dictionary.

In addition, a modification for further accelerating learning in this embodiment will be described. FIG. 3 shows a modification of a part of the flow shown in FIG. 2 (steps 38 to 42). The processing (steps 41 to 42) when the learning frequency F _i is not 0 is the same, but when F _i is 0
When , the distance D _J obtained by recognition is compared with a threshold value θ ₀ (step 51). θ ₀ takes a fairly small value. When D _J is larger than θ ₀ , it means that the input character
0). When D _J is smaller than θ ₀ , the letter i for that filler is the letter J for a general filler.
It means very close to. At this time, after transferring the standard pattern P _J to P _i , erase P _J and write the letter i.
The learning frequency F _i of is set to 1 (step 52). In step 52, φ indicates a feature vector that does not match any pattern, and substituting φ for P _J means that
This shows that _PJ has been virtually erased. Display the letters i and J with appropriate guidance sentences (step 5)
3) Instructs that the standard pattern P _J has been registered as the letter i and that the letter J should be entered and set in the form in order to re-learn the standard pattern of J if necessary. This process means the following.
For example, the characters ``WORK'' and ``仂'' are different characters,
If there is a person who has a habit of writing ``work'' as ``仂'', the standard pattern for ``work'' for this person is registered as ``仂''. This eliminates the need to learn multiple times to obtain a stable standard pattern for "work." If the standard pattern is changed in this way, there is a problem of what to do when actually inputting the character ``O'', and step 53 is where the user instructs the user in advance. In other words, if this person registers the writing habit of writing ``□仂'' when he wants to input ``仂'' at this point, there is a risk that he will forget that the standard pattern for ``仂'' has been deleted. prevent from happening.

〔Effect of the invention〕

According to the present invention, the user does not need to present a large number of learning patterns prior to initial use, and only learns erroneous recognition or unrecognizable patterns when recognizing using average standard patterns. , it is extremely easy to introduce. Furthermore, since the pattern to be learned is completely adapted to the user, there is an advantage that the recognition accuracy for the user becomes extremely high after learning.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the block configuration of an embodiment of the present invention, FIG. 2 is a flowchart of processing in the control section of the embodiment described above, and FIG. 3 is a flowchart of a modified portion of FIG. 2. 2...Photoelectric conversion unit, 4...Control unit.

Claims (1)

[Scope of Claims] 1. A first storage means for storing a plurality of standard patterns for each character category; and determining a distance between an input character pattern and each of the plurality of standard patterns stored for each character category. , the one with the smallest distance is selected, and if the selected distance is smaller than a predetermined value, the character category to which the standard pattern with the smallest distance to the input character pattern belongs is output as a recognition result, and the selected distance is In a pattern recognition device having a recognition means that determines that it is unrecognizable if it exceeds a predetermined value, and a display means that displays the input character pattern and the recognition result of the recognition means, the learning frequency of the plurality of standard patterns is stored. If the recognition result displayed on the display means is an erroneous recognition, or if the recognition means determines that it is unrecognizable, the correct character code for the input character pattern is stored. is input, selects the standard pattern with the smallest distance from the input character pattern from among multiple standard patterns belonging to the character category corresponding to the input character code, and calculates the learning frequency of the selected standard pattern. In the case of a predetermined value, the selected standard pattern is deleted from the first storage means, the input character pattern is stored in the first storage means as a standard pattern of the character category, and the stored standard pattern increasing the learning frequency of the selected standard pattern from a predetermined value and storing the increased value in the second storage means, and when the learning frequency of the selected standard pattern is greater than the predetermined value, the selected standard pattern is 1, and the selected standard pattern and the input character pattern are combined at a ratio according to the learning frequency and stored in the first storage means as a standard pattern of the character category; and A method for creating a standard pattern in a pattern recognition device, characterized in that the learning frequency of the standard pattern is further increased and the increased value is stored in the second storage means. 2. In claim 1, it is provided that standard pattern creation in a pattern recognition device includes storing the storage contents of at least one of the first storage means and the second storage means in a replaceable external storage device. Method.