JPS63282890A - Pattern identification device - Google Patents

Abstract

PURPOSE:To remove patterns other than an object to be recognized by forming a sum-set feature vector of all categories, collating the formed vector with a dictionary, separating a feature to be extracted, and identifying its corresponding pattern. CONSTITUTION:The titled device is constituted of an input pattern data base 1 for storing the data of an input pattern to be recognized, a counter 2 for counting up the number of times of processing, a reader 3 for reading out the pattern from an input pattern data base, a feature selector 4 for selecting a feature to be extracted based on an instruction from the counter 2, respective feature extracting mechanisms 7 to be started by an instruction from the selector 4, feature lists 5, 6 in the dictionary, a matching mechanism 8 for comparing the feature of the inputted pattern with that included in the dictionary and deciding a category, a feature value memory 9 for recording a feature vector for a short period, and a writer 10 for recording the recognized result in a recognized result data base 11. Consequently, patterns other than that to be recognized can be removed.

Description

[Detailed Description of the Invention] [Summary] The present invention creates a union feature vector over all categories for each feature from feature vectors for each category of unknown input patterns including various patterns, and creates a dictionary based on this. This is a pattern identification device that separates the features to be extracted and identifies the corresponding pattern. This results in
It becomes possible to exclude patterns other than those to be recognized.

[Industrial application field]

The present invention relates to an improvement in a pattern identification device that can separate and identify a target category from an input pattern containing a mixture of categories other than the target category.

[Conventional technology]

Conventionally, in order to automatically recognize pattern information such as documents and drawings and input it into an information processing device accurately, there are several major problems that cannot be considered with an OCR (optical character reader).

That is, (1) there is no restriction on the position or size of the recognition target in the document 9 drawing, and (2) patterns other than the recognition target also coexist in the document 9 drawing.

etc.

FIG. 4 is a schematic explanatory diagram of the configuration of a conventional example.

In the conventional OCR mechanism, feature quantities such as size, length, area, etc. extracted from recognition target pattern X, f2. ....

It shows the process of determining the recognition category of characters, symbols, etc. by comparing fk with corresponding features in the dictionary.

In this case, the category of the input pattern is limited by the characteristics of the dictionary, and for example, characters other than katakana must not be input to katakana recognition OCR.

[Problem that the invention seeks to solve]

In the conventional pattern recognition method described above, when characters and symbols of the target category are mixed with lines and images of other categories, it is difficult to recognize objects other than the recognition target if the mechanism shown in Figure 4 is applied as is. become. This not only causes an increase in recognition time and recognition errors, but also creates the need to newly calculate some feature quantity that distinguishes the recognition target from patterns other than the target.

SUMMARY OF THE INVENTION An object of the present invention is to provide a pattern identification device that specifically includes means for separating a target category from an input pattern in which categories other than the target category are mixed.

[Means for solving problems]

In order to achieve the above object, the present invention provides a separation procedure as shown in the principle explanatory diagram of FIG.

Document 9 Pattern data, for example raster, mechanically obtained from a drawing or the like using a photoelectric conversion device.

Let the set of vectors, contour data, etc. be 1x<1.

Unlike conventional OCR, this set also includes patterns other than characters. Each Xi is composed of n-dimensional data and a flag area.

The categories to be recognized are Sl, S2. ..., S, and @trace to Fl, F2. ..., Fk.

Conventionally, for each input pattern Xi, the special Q F
l, F2. ..., Fk was extracted and recognition was performed based on which category in the dictionary the feature vector matched with.

On the other hand, in the present invention, recognition is performed according to the procedure shown in FIG. 1, and the separation process is particularly a feature of the present application.

Learning process 1 is the same as feature selection and feature learning that have been used in the prior art.

Learning process 2 is a process of calculating a union ALut over all categories for each feature with respect to the feature vectors for each category obtained in learning process 1. Patterns belonging to either category always satisfy this union feature.

The separation process is a process of sequentially applying the union features created in the learning process 2 to the input pattern set to separate only input patterns that have a possibility of belonging to any category. That is, it is determined that an input pattern that does not satisfy any one of the union features is not a recognition target. By performing this processing, it is possible to select a target pattern to be recognized from a large set of input patterns at an early stage and reduce unnecessary calculations for patterns other than the target patterns.

The recognition process uses the candidate pattern (
This process performs recognition using feature vectors in the dictionary (including some patterns other than those to be recognized). This process is the same as what has been done conventionally.

[For production]

g in Fig. 1, the embedded explanatory drawing, and Fig. 2 (which is the main part of the invention)
The operation will be explained in detail using the dictionary α) and the features shown in (6) and (C) in the same figure.

Step 1 (Learning process 1) Categories to be recognized: English, numbers, kanji, line 9 images, etc. 5j (j
= 1 to M) and the feature size, length, area, etc. FkCk = 1 to K) to be extracted, and set the feature f for each category.
Find fk fjk. Here, fjk is expressed in the range of real numbers as shown in FIG. 2(b). Sj and Fk are shown in Figure 2 (α
) and store them in a dictionary.

fjk = [ajk, bjk], ajk≦bjk This expression allows fjk to take a real value, as well as setting logical values (0, 1) and real number ranges (expressions such as greater than or equal to, less than or equal to...) is also possible.

Note that the processing in this step is the same as that used in the prior art.

ΔNEW (Learning process 2) The feature quantities fjk are unionized for each feature Fk. That is, the feature Qk of the union of the features Fk is determined using the following formula. (See Figure i2 (C)) gk=(MIN(to al, to a2, -..., aM
k).

MAX(bl, b2, . . . , 6Mk)) Find this feature σk for all k and record it in the dictionary table.

Step 3 (Separation process 1) All patterns Xi in the input pattern set (Xi)
Find the feature F1 for. If the value of Fl is not included in the union feature g1, 0 is set in the flag area of the pattern Xi. If it is included in gl, 1 is set in the flag area.

Step 4 (separation process k) Processing similar to step 3 is performed for feature F2.

However, the patterns Xi targeted for feature calculation are limited to those with step arms and 7 lags.

This process is repeated for all features Fk (, , = 2 to k).

Step 5 (Recognition processing) When step 4 is completed, input pattern set (Xi)
Among them, the pattern with flag=1 is the pattern to be recognized. For this target only, as before,
The category is determined by matching with the feature vector. This completes the process.

〔Example〕

FIG. 3 is an explanatory diagram showing the configuration of an embodiment of the present invention.

This mechanism consists of an input pattern database 1. which stores input pattern data to be recognized. Counter that counts the number of times of processing 2. A book reader 6 reads out patterns from the input pattern database, and a feature selector 4 selects features to be extracted according to instructions from a counter. Each feature extraction mechanism activated by the instruction of the feature selector7. Features in the dictionary 5
,6. Matching mechanism that compares the features of the input pattern with features in the dictionary to determine the category 8. and a feature amount memory for short-term recording of special a amount vectors9. Writer 1 that records recognition results in recognition result database 11
Consists of 0.

The operation of this mechanism is as follows.

A flag=1 is initially set for each turn stored in the initial setting input pattern database 1.

Further, it is assumed that learning has already been completed and the feature amount fjk for each category and the feature Qk of the union are recorded in the dictionary.

Separation Process First, the operation starts by setting an initial trigger in the counter 2. The counter increments the phase count by 1 upon a trigger. Counter 2 holds the same value until the phase count is read out for each turn.

The counter 2 instructs the reader 6 to read the pattern. The reader 6 has a function of sequentially reading out only those with flag=1 from the input pattern database 1. After reading one pattern data, the reader 6 selects the percentage selector 4.
send a signal to.

The feature selector 4 determines the feature to be extracted based on the phase count signal from the counter 2, and then activates the selected feature extraction mechanism 7 based on the signal from the reader 6.

Feature extraction mechanism 7 receives input pattern data from reader 3 and calculates features. After the calculation is completed, the feature data is transferred to the matching mechanism 8.

The matching mechanism 8 learns the features to be matched from the phase count signal of the counter 2, and compares the features from the feature extraction mechanism 7 with the features gk transferred from the dictionary. The input pattern is characterized by the writer 10 along with the flag.
will be forwarded to. The flag is set to 1 if it matches, and is set to 1 if it does not match.

The writer 10 writes the obtained special di, A, and flag into the original 6th place of the input pattern database.

Counter 2 is read by reader 3, which is the last data in the database, and after the matching process is completed, 7
Increment aids count. The counter 2 considers the phase count to be between 1 and 1 to be a separation process, and in addition to transferring the union feature gk in the dictionary to the matching mechanism 8, it also sends a signal to the feature selector 7 and the writer 1 (9) and outputs the data. Control the flow. The separation process is completed when the phase count exceeds K, and then the process moves to the g-identification process.

Recognition processing At this time, the counter 2 not only transfers the characteristic light fjk for each category from within the dictionary to the matching device 11 but also directly activates the matching mechanism 8. In the recognition process, all necessary features are recorded in the input pattern database, so there is no longer a need to extract features.

The matching mechanism 8 records the winning cover pattern characteristic vector transferred from the reader 6 in a characteristic memory 9 for a short period of time. Matching mechanism 8 uses feature light fjk and feature amount memory for all categories? Identification is performed by comparing the special at vectors within.

The identified category name and input pattern number are writer 0.
The recognition results are written into the recognition result database 11. The input pattern that does not match any category is assumed to be the input pattern.

After the reader 6 has read out and recognized all the resonating input patterns, the counter 2 checks that the incremented 7 aids count has exceeded (K+1) and stops the mechanism.

〔Effect of the invention〕

The present invention brings about the following effects.

(1) Patterns other than those to be recognized can be excluded.

Moreover, no special discrimination logic is required for this purpose, and learning is easy because features that can be easily obtained from each characterization of the recognition target category are used.

(2) Patterns other than those to be recognized are excluded early in the process, so processing time is not wasted. Generally, the number of input patterns is N (0), the number of recognized t·1 categories is M, the number of feature calculations is KN(0) times, and the number of matchings with features is KMN(0) times. Moreover, this includes patterns that do not need to be recognized. In contrast, in the present invention, the number of data is reduced to N(1) by the separation process for the first feature, and the separation process for the next feature is performed for N(1). Therefore, the total number of feature calculations is N(0) ten N(1) + N(2)+...+NK
≦KN(0)(N(D)≧N(1)≧・・・・・・≧N
(K)), and the number of feature matching is also N(0) + N(1) + N(2)+...
10N (K) + KIliiN prisoner 111 separation process-
---1L recognition processing9 is expected to decrease significantly.

[Brief explanation of drawings]

FIG. 1 is an explanatory diagram of the principle of the present invention, FIG. 2 is a detailed explanatory diagram of the main parts of the present invention, FIG. 6 is an explanatory diagram of the configuration of an embodiment of the present invention, and FIG. 4 is an explanatory diagram of a conventional example. In the figure, 1 is an input pattern data memory, 2 is a counter, 5 is a reader, 4 is a feature selector, 5 and 6 are feature lights, 7 is a feature extraction mechanism, 8 is a matching mechanism, 9 is a memory with features, 10 11 indicates a writer, and 11 indicates a recognition result data memory.

Claims (1)

[Scope of Claims] A learning processing unit that calculates a feature vector for each category of figures and a union feature over the categories of each feature; a dictionary that stores each of the feature vectors and the union feature; , an input pattern data storage unit that records the original data of the input pattern, feature vectors, and flags indicating the features to be extracted; a feature selector that selects a feature to be extracted by the input pattern designation means; and a feature selector that is activated by an instruction from the feature selector to read an input pattern from the input pattern data storage unit and calculate each feature amount. a feature extraction mechanism section, which compares the features from the feature extraction mechanism section with feature amounts for each category and union feature amounts in the dictionary, outputs a feature to be extracted based on the match, and outputs a feature to be extracted based on the match; A pattern identification device comprising: a matching processing section that rewrites a flag in a storage section; and a recognition result data memory that records a pattern corresponding to the feature to be extracted.