JPS5917670A - Method for retrieving information - Google Patents

Abstract

PURPOSE:To prevent busy state on the way of retrieval and to improve retrieving speed by dispersing storage information with high similarity into a storage space to store it in an information retrieving method by a matching test using storage and input information. CONSTITUTION:A characteristic vector obtained by decomposing a character to stroke and extracting from a reading character pattern is set up in a register 16. While counting up a counter 12 by a signal 18, a comparator 14 reads out dictionary feature vector successively from the leading address of a memory 10, discriminates rough matching at a [ I ] part, and when the matching is obtained, starts precise matching at a [II] part and continues the test also at the part [ I ]. If the matching is obtained by the [ I ] part before the completion of the test at the [II] part, busy state is generated. Therefore, the generation of the budy state is prevented by storing the dictionary feature vector with high similarity in separated addresses of the memory 10.

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field The present invention relates to an information retrieval technique for selecting specific stored information from pre-stored stored information.

BACKGROUND OF THE INVENTION There are various types of devices that have an information retrieval function, one example of which is a character recognition device. The conventional technology will be described below using a character recognition device as an example.

Generally, a character recognition device prepares in advance a dictionary in which feature vectors for a large number of characters are accumulated on a storage device. Then, the contents of the dictionary are sequentially searched using the feature vector (input feature vector) extracted from the read character pattern,
Recognize read characters by searching for matching feature vectors. Therefore, in order to increase the character recognition speed, it is necessary to speed up the search for feature vectors.

In order to increase the search speed of feature vectors, some conventional character recognition devices employ the following search method. That is, a rough matching test is performed on the feature vectors sequentially read out from the dictionary with the input feature vectors. Then, a certain feature vector (L
), a precise matching test is started for that feature vector (i), and in parallel with the execution, the features after the next feature vector (i+1) of that feature vector (i) are A rough matching test is performed on the vectors, and another feature vector that can be roughly matched is searched for.

Here, the feature vector matching test is generally performed as follows. Now, the feature vector of the dictionary is "(3/I, y2...., yn), and the input feature vector is ')/? (x, , x2.:...,!,,)
Usually, the first m components of the feature vectors X and H are rough classification parameters, and by comparing corresponding ones of these components, the feature vector
The rough matching of M can be checked at high speed. On the other hand, in order to check the matching (accurate matching described above) for the (-+1)th and subsequent components of the feature vectors X and Y (for example, stroke features), it is necessary to calculate the distance l according to the following formula, for example, It usually takes a considerably longer time than a rough matching test.

Here, W,, W2. ... is a weight function.

Therefore, in conventional character recognition devices that employ the above search method, while an accurate matching test involving distance calculation as shown in the above equation is being executed for a certain feature vector, the next feature vector that can be roughly matched is searched for. is searched, and the rough matching test for subsequent feature vectors in the dictionary must be interrupted at the point where the precise matching test for the previous feature vector ends (
(busy state) frequently occurs, which is a factor that slows down search speed.

OBJECT It is an object of the present invention to provide an improved information retrieval method that solves the above-mentioned problems and is effective when applied to feature vector retrieval in a character recognition device.

Overview In the prior art described above, the busy state occurs because
This is a case where feature vectors that can be roughly matched are read out from the dictionary at short time intervals. In other words, feature vectors with high mutual similarity are arranged close to each other in the dictionary and are successively read out from the dictionary (storage device) at short time intervals. Therefore, by distributing and arranging feature vectors that are highly similar to each other in the dictionary, and allowing a certain amount of time or more to be read out between feature vectors that are highly similar to each other, the chances of a busy state occurring will be reduced. and improve search speed. The present invention has been achieved by paying attention to this point.

Therefore, the present invention provides an information retrieval method in which stored information is sequentially read from a storage means, a matching test is sequentially performed between the stored information and input information from the outside, and stored information that can be matched with the input information is searched from the storage means. By distributing and storing stored information that has a high degree of similarity to each other in separate locations on the storage space of the storage means, the stored information that has a high degree of similarity to each other can be stored in the execution time of a matching test of stored information. The gist of this invention is an information retrieval method characterized in that the information is read out from the storage means at a predetermined time interval or more determined by the relationship between the following.

EMBODIMENT OF THE INVENTION An embodiment of the present invention will be described below with respect to a case where the present invention is applied to a character recognition device that performs character recognition using stroke features.

FIG. 1 is a block diagram showing one embodiment of the present invention. I
O is a memory (for example, ROM) in which a dictionary storing feature vectors for a large number of characters is stored in advance.

The format of the feature vector is such that the first two components are the major classification parameters of the character, and the remaining components are the feature parameters of each stroke of the character.

For example, the character "8" is decomposed into strokes 1 to 7 as shown in FIG. 8(b), and its feature vector B(b, ,
A2. Ag, ..., bS,) is the first two components A, , A2 is the major classification (engineering), (2), and the subsequent 7
One component 65□~bs.

are the characteristics of each of strokes 1 to 7. More specifically, for example, the component S is the perimeter of the character, the component b2 is the number of holes in the character, and the component hJ? , ~bs7 express the direction, center of gravity, length, etc. of each corresponding stroke.

Similarly, the character "IL" is decomposed into six strokes as shown in FIG. 8(C), and its feature vector C('l
+ '2 + "I + "2 +..., cs'6)
The components CI and C2 of are broadly classified (1), (2), cg, ~
C-1'a is the characteristic of each of strokes 1 to 6.

Such feature vectors of each character are continuously stored in the memory 10 along with the character code.

12 is a counter that specifies the address of the memory IO. This counter 12 is incremented by a signal 18 from a comparator 14.

16 is a register. A feature vector (
input feature vector) is set in this register 16. The input feature vector has the same format as the feature vector (dictionary feature vector) in the memory (dictionary) 10.

The comparator 14 executes a matching test between the dictionary feature vector read from the memory 10 and the input feature vector in the register 16, and logically, [■] and CII
) is divided into parts. [The IL part consists of the dictionary feature vector and the first two components of the input feature vector (major classification (
], ), +21), the above-mentioned delicate matching test is performed honestly. The time required for its execution is T per feature vector. It is. The CI part of the comparator 14 performs the above-mentioned accurate matching test by performing the distance calculation as shown in the above formula for the 8th and subsequent components (stroke features) of the dictionary feature vector and the input feature vector. This is the part that does this. The time required for the execution varies depending on the number of strokes of the target character, and is generally approximately the execution time T of the matching test.

considerably longer than .

When the input feature vector is set in the register 16, the comparator 14 increments the counter 12 by the signal 18 while sequentially reading dictionary feature vectors from the first address of the memory 10. (1), +2) is determined. If matching cannot be achieved, the comparator 14 outputs the counter 12.
is counted up, the next dictionary feature vector is read from the memory 10, and a matching test of major categories (1) and (2) with the input feature vector is similarly performed.

In this way, when a major classification (]) and (2+ matching are achieved for a certain dictionary feature vector, the comparator 14
Regarding the dictionary feature vector, [At the same time as starting the matching test for the stroke feature in the 13th section, the subsequent dictionary feature vectors are sequentially read from the memory 10 while counting up the counter 1z, and the matching test in the 13th section is performed. Continue execution. Then, when a dictionary feature vector that can be matched (roughly matched) for major classifications (1) and (2) is found,
Regarding this, [in Part 11, we started a matching test (accurate matching test) regarding stroke characteristics,
Continue reading subsequent dictionary feature vectors and performing a rough matching test. However, if the [13th division] is currently executing a matching test on the previous dictionary feature vector, the counter 12 will be frozen until the matching test is completed and the next dictionary feature vector can be processed by the [13th division]. , waits for reading of subsequent dictionary feature vectors. In other words, it is in a busy state.

Such an operation causes the (T) section of the comparator 14 and the [13
This process continues until a dictionary feature vector that can be matched in the same section is found (up to the final address in the memory 10). And comparator 14. When a certain dictionary feature vector is matched in both [13 parts and [13 parts] (that is, when a dictionary feature vector that can be matched with the input feature vector is searched for), the dictionary feature vector is stored in the memory IO along with the dictionary feature vector. Outputs the character code read from.

In order to prevent the above-mentioned busy state from occurring, in this embodiment, dictionary feature vectors having a high degree of similarity such as major classification (IL) where 2+ almost match each other are distributed to distant addresses in the memory 10. is stored.

For example, the dictionary feature vector B of the character "8" in FIG. 3(b) and the dictionary feature vector C of the character rBJ in FIG. Sub B
,! : Store in C. The interval between addresses B and C is the number of dictionary feature vectors (I:"/,c) (-1: or more).If dictionary feature vectors B, C and major classification (1 ), +21 are almost the same, then similarly the dictionary feature vector [
The data is distributed and stored in discrete addresses D, E, . In this way, the conventional busy state will not occur and the search speed will increase. about this,
Let me explain with an example.

Now, the pattern of the character rBJ is read, it is decomposed into six strokes 1 to 6 as shown in Fig. 8 (α), and its feature vector A (α, , a2.α91°α'
2+..., +2Sa) is extracted and set in the register 16 as an input feature vector.

The dictionary feature vectors are sequentially read out from the first address of the memo IJIO, and matching is achieved for the first time in the [13 section of the comparator 14] for the dictionary feature vector B read out from the address B. The matching test for part 113 begins. This [13
In parallel with the execution of the matching test in the part, the dictionary feature vectors of the addresses following address B are read out one after another from the memory 10, and the matching test is performed in the [13 part of the comparator 14]. Then, when the address C is reached and the dictionary feature vector is read out, [1] of the comparator 14
Matching can be achieved in the section. However, since addresses B and C are separated by dictionary feature vectors [''/T, (The test results in this case are inconsistent.

) It does not enter a busy state and immediately starts matching test [13 copies] for dictionary feature vector C, and can continue reading dictionary feature vectors from the address next to address C. In this example, the search operation ends when the matching test of the part CD) matches the dictionary feature vector C and the character code thereof is output from the comparator 14.

Although one embodiment has been described, the present invention is not limited to this, and it goes without saying that various modified embodiments are possible. It goes without saying that the present invention is not limited to searching for dictionary feature vectors in character recognition devices, but can be widely applied to the field of information searching.

Effects As is clear from the above explanation, according to the present invention, it is possible to prevent the occurrence of a busy state during information retrieval and improve the information retrieval speed.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is a diagram schematically showing the arrangement of similar dictionary feature vectors in memory, and Fig. 8 shows character noku turns and their feature vectors. It is a diagram. 10...Memory (?), 12-...Counter, ■4
...Comparator, 16...Register. Figure 1] b Figure 2

Claims (1)

[Claims] (1) Sequentially reading stored information from the storage means and sequentially performing a matching test between it and external input information,
In an information retrieval method in which stored information that can be matched with the input information is retrieved from the storage means, storage information that is highly similar to each other is distributed and stored in separate locations on the storage space of the storage means. An information retrieval system characterized in that stored information having a high degree of similarity to each other is read out from the storage means at intervals of a predetermined time period or more determined in relation to the execution time of a matching test of stored information.