JPS61195478A - On-line character recognizing device - Google Patents

Abstract

PURPOSE:To reduce a calculating amount by obtaining the stroke matching by a heuristic retrieval method so that the sum of distances between matched strokes can be minimized under conditions where the stroke is matched with overlapping. CONSTITUTION:An input pattern from a tablet 1 is inputted to an inter-stroke distance calculating part 5 through an input pattern memory 2, and an inter- stroke distance between the input pattern and a reference one is calculated. A heuristic calculating part 6 obtains a heuristic function corresponding to plural recognized characters and stores it in a heuristic memory 7. A node generating part 9 obtains on a node, and a list generating part 10 adds data on a new node to a code list in a node list memory 8. When the depth of the retrieval of the node data on the node list is equal to the number of strokes of the input pattern, the number of a standard pattern in the node data is outputted as the recognized result.

Description

[Detailed Description of the Invention] (Industrial Application Field) The present invention allows characters to be extracted from the handwriting of characters input online! The present invention relates to an online character recognition device that recognizes characters.

(Prior Art and its Problems) Conventionally, there has been a method of recognizing characters using characters input online as an input pattern without using stroke order (order in which strokes are written) information.

For example, find the correspondence between the strokes that do not allow overlap between the input pattern and the standard pattern and the sum of the distances between corresponding strokes is the minimum, and then find the character for the standard pattern that minimizes the sum of the distances between the corresponding strokes. There was a method to use t-recognition acid. January 9, Showa era
Patent application (2) (hereinafter referred to as cited document (1)), dated 2007-2012, describes a method for finding the optimal correspondence between such strokes using a heuristic search method.

In this method, the distance between each standard pattern and the input pattern is determined as follows.

First, the value of the theoretical function is determined based on the cumulative value of the minimum distance when the strokes of the input pattern and the standard pattern are associated with each other with overlap allowed.

Next, the strokes of the input pattern and the standard pattern are matched without allowing overlap, so that the value of the evaluation function given as the sum of the cumulative distance value and the value of the above-mentioned force-ristic function is minimized. Prioritize the stroke correspondence and search for the optimal sequential correspondence. The cumulative value of the inter-stroke distances for this correspondence is determined and is taken as the inter-pattern distance. The result is a category t-recognition result of a standard pattern in which the distance between the input pattern and the standard pattern obtained in this way is % smaller.

The utility function used in this method is
It works effectively when the input pattern and the standard pattern are similar, and can significantly reduce the amount of calculation required for search. However, when associating dissimilar patterns such as different character patterns, the heuristic function does not work effectively, which has the disadvantage that the amount of calculation required for the search becomes large. .

For example, FIG. 2 is a diagram showing an example of a pattern for the character ``ki'', FIG. 3 is a diagram showing an example of a pattern for the character ``ki'', and FIG. 4 is a diagram showing an example of a pattern for the character ``yen''. Consider the case where the pattern in FIG. 2 is used as an input pattern, and the patterns in FIGS. 3 and 4 are used as standard patterns, and the inter-pattern distances between the input pattern and each standard pattern are determined. FIG. 5 shows the inter-stroke distances of the input pattern of FIG. 2 and the standard pattern of FIG. 3 arranged in a matrix.

FIG. 6 shows the inter-stroke distances of the input pattern of FIG. 2 and the standard pattern of FIG. 4 arranged in a matrix. The strokes of the input pattern are shown in the horizontal direction, the strokes of the standard pattern are shown in the vertical direction, and the distance between each stroke is shown in the matrix element. In Figures 5 and 6, i
is the stroke number of the standard pattern, and j is the stroke number of the input pattern. h is the value of the heuristic function, which is the sum of the minimum values of the distances between strokes in each column starting from the one with the larger stroke number of the input pattern.

Figure @7 is an example of a search tree when the optimal correspondence is determined by the heuristic search method shown in Cited Document (1) K based on the distances between strokes in Figure 5. The numbers in the circles representing the nodes are the values of the evaluation function.1 The numbers written to the right of the nodes indicate the order in which the nodes were created, and the numbers to the left of the nodes indicate the values of the evaluation function. The associated stroke numbers are shown as (i=j).

In the search tree shown in FIG. 7, the input pattern and the standard pattern both have the letter K "tree" and are similar patterns to each other. Therefore, here the number of nodes created is 10
The optimal correspondence can be found using the minimum necessary number of nodes.

This is an example of a search tree. The numbers in the figure are the same as in FIG. Here, the input pattern is the character "tree" and the standard pattern is the character "yen", and the two patterns are not similar. Therefore, the number of created nodes is 21, which is larger than in the case of FIG. Since the amount of calculation required for the search is approximately #t proportional to the number of nodes to be created, in this case, a large amount of calculation is required for the search.

Furthermore, in the conventional method as shown in cited document (1), such a search needs to be performed for all standard patterns. Generally 4Il! Since there are many patterns in the quasi-patterns that are not similar to the input pattern, character recognition using conventional methods requires a large amount of calculation and requires a lot of searching to match dissimilar patterns. There is a need to do.

For this reason, the conventional method has the drawback of requiring a large amount of calculation.

(Objective of the Invention) The object of the present invention is to reduce the amount of calculation required in a method for realizing high-performance online characters glI! using the distance between turns determined based on the optimal correspondence between the strokes of an input pattern and a standard pattern. It is said that it can significantly reduce the

(Structure of the Invention) The online character recognition device according to the present invention includes a handwriting detection unit that detects the handwriting marks when writing characters and reads them as an input pattern, and 1&! a standard pattern memory that stores standard patterns for a plurality of characters to be recognized; an inter-stroke distance calculation unit that calculates the distance between strokes between the input pattern and the standard pattern; a heuristic calculation unit that calculates a heuristic function for the plurality of recognized characters, with a minimum distance of 4 when the strokes of the input pattern and the standard pattern are matched without allowing overlap; The evaluation value obtained as the sum of the cumulative distance and the heuristic function when It is composed of a search control section and k, which output a character corresponding to the standard pattern when all the strokes of the standard pattern and the input pattern are associated with each other.

(Principle of the invention) Now, the principle of the present invention will be explained.

The online character recognition device according to the present invention uses a hierarchy search method (artificial intelligence principle) known in the field of artificial intelligence to efficiently find the optimal stroke correspondence between an input pattern and a standard pattern. Japanese combination tool
Yukyokai (1983), pp. 63-81). In the method shown in cited document (1) K, the correspondence between the strokes and the input pattern is determined for each individual standard pattern, but in the present invention, the correspondence with all the standard patterns is calculated in parallel. Okona Maro prioritizes the best node among all these search trees and proceeds with the search.

First, define the distance between the two points. Let N be the number of strokes in the input pattern. When the number of strokes is H standard patterns, the inter-stroke distance between the j-th stroke of the input pattern and the first stroke of the k-th standard pattern is defined as 'iid (km it j).

However, 1≦i≦N, 1≦j≦t, and 1≦≦. The distance between strokes is assumed to be the sum of Euclidean distances between the coordinate values of writing points forming strokes, as described in Cited Document (1). In addition to this, various definitions of the distance between strokes shown in the cited document (1) can be used.

Next, strokes are associated based on this distance between strokes. In the present invention, under the condition that strokes are matched without allowing overlap, strokes are matched such that the sum of the distances between corresponding strokes is the minimum. Stroke correspondence is performed using a heuristic search method.

The branches of the search tree are extended by giving priority to nodes with small evaluation values f (k, p). A search tree is created in parallel for each LIS pattern. These plurality of search trees are distinguished by the standard pattern number k (1≦to≦K). Evaluation value f (k,
p) is.

f (k, p) = h (k, p)
: The value p=Q. Here, p indicates the search depth (0≦p≦N), and T (k, j) indicates the stroke number of the standard pattern corresponding to the j-th stroke of the input pattern.

However, T (k, j) all take different values in 1≦j≦p. Here, h (k, p) is a Hi-17 stake function. Each node in the search tree has an evaluation value f (k
, p) and the stroke numbers of the standard patterns that have been associated so far (referred to as correspondence history). This correspondence history allows strokes to be associated without duplication.

The search starts from p=o and proceeds to the node with the smallest evaluation value f (k, p). That is, find the node with the minimum evaluation value f (k, p) among the nodes created so far, and calculate the evaluation value f (k , p) is repeated. When the search depth p=N and f(k, FJ) becomes the minimum among the evaluation values of the nodes to be searched, the strokes are completely matched and the search ends. .

The standard pattern number k that gives this node becomes the recognition result.

The heuristic function h (k, p) is the cumulative value of the minimum distance when the strokes of the input pattern and the standard pattern are associated with each other with overlap allowed.

That is, h (k, p) = 0: p
=FJ(2). If the hierarchical function is defined in this way, the search can be performed more efficiently than K.

An example of character recognition according to the present invention will be shown.

The pattern of the character ``tree'' in Figure 2 is the input pattern, the pattern of the character ``tree'' in Figure 3 is the standard pattern number = the standard pattern of l, the pattern of the character ``circle'' in Figure 4. Consider a case where characters are recognized using the standard pattern number =J as the standard pattern and the above two patterns as standard patterns. As mentioned above, with the input pattern shown in Figure 2! The distance between strokes of the standard pattern shown in FIG. 3 is shown in FIG. 5, and the distance between strokes of the input pattern shown in FIG. 2 and the standard pattern shown in FIG. 4 is shown in FIG.

This toki, -ristic function h (1, 0) = 12
．． Since h (2,0) = 24, first the evaluation value f
(1.0) = Search starts from Minoru's node.

The search proceeds according to the search tree shown in FIG. On the other hand, the evaluation value f (2,0)=u. This evaluation value is larger than the minimum evaluation value at any point in time when searching the search tree in FIG. Therefore, a search for stroke correspondence is not performed for the standard pattern whose standard pattern number is 2.

In other words, the search ends without creating the search tree shown in FIG. The recognition result is the minimum evaluation value f (1, 4) = 1
The standard pattern number that gives the node 2 is the character of the standard pattern =l, that is, the character ``tree''.

As described above, according to the present invention, the amount of calculation required for IR search can be significantly reduced. In the conventional method as described in the cited document (1), it was necessary to also search the search tree shown in FIG. 8 in the above example.

In the above example, the number of standard patterns was two, but since the number of standard patterns is generally larger, the effect of reducing the amount of calculation by the present invention is even greater.

(Example) An example of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the present invention.

The character pattern input from the tufflet 1 is held in the input pattern memory 2. Standard pattern memory 3Kt
[A standard pattern of recognized characters is maintained.] From the input pattern memory 2, the first stroke of the input pattern SO) is sent to the control signal from the control unit 4, and from the standard pattern memory 3, the first stroke of the standard pattern is sent to the standard pattern number. pattern M
(k.

i) is read out. The inter-stroke distance calculation unit 5 reads pattern 5(i) and pattern M(k.

The interstroke distance d(k, i, j) between i) is calculated. Assume that this distance between strokes d(k, i, j) uses the distance between the strokes and the distance shown in the cited document (1)K.

First, find the value of the heuristic function h (k, p). The determined distance between strokes d (k.

i, j) t, the heuristic calculation unit 6 calculates the heuristic function h(k, p
) is calculated. This Hieriste.

The value of /l1lah (k, p) is held in the heuristic memory 7.

The node list memory 8 contains data Nd of nodes of the search tree.
are stored in a list format in descending order of evaluation value f (k, p). The node data Nd has an evaluation value f (k,
p), cumulative distance δd (k, p) (value of evaluation value not including heuristic), standard pattern number k
, search depth p, and correspondence history. Compatible shoes! holds the stroke number of the next standard pattern that has been associated so far. It is empty if there is no correspondence at all.

In the initial state, the evaluation value f is stored in the node list memory 8.
(k, J)) is sufficiently large, and data Nd of one node is held.

Next, a node at %I)=OK is created by the node creation section 9. The evaluation value f (k・p) when p=OK is the heuristic function h (
k, p). From the control unit 4, the control signal is p.
takes the value p=0・l≦≦ and is sequentially output. In response to this control signal, the value of the heuristic function h (k, p) is output from the hierarchic memory 7 to the node creation unit 9 in accordance with pK. The node creation unit 9 creates new node data Ndnew. The data Ndnew of the new node created here is: Search depth: p = 0 Standard pattern memory k (heuristic function h I same as k of (k, 0) Cumulative distance: sd (ke p) =.

Evaluation value: f (k, p) = h (k,
0) Correspondence history: It consists of empty data.

In the list creation part 1α, the new node data Ndnew
A process of adding to the node list in the t-node list memory 8 is performed. That is, the new node data Ndn
Evaluation value f (k, p)'' in ew from the beginning of the node list to evaluation value f (k') of node data Nd in the node list
, b compared to p).

New node data Ndnew is added before node data Nd that satisfies f (k, p) < f (k', p).

The node creation unit 9 first reads data Ndi of the node at the head of the node list in the node list memory 8 (the one with the smallest evaluation value f(k, p)).

Based on the data Nd of this node, i, j
.

Find the value of p. That is, K is the standard pattern number in the node data Nd, and i is the number of strokes that are not retained at tζ in the correspondence history, and j = p.
'+l, P=p'+1. (7t, nodal data N
These control signals are sent to the control section 4, and further sent to the input pattern memory 2, the standard pattern memory 3, and the heuristic memory 7. According to these control signals, the input pattern memory 2 outputs the pattern S (j), the standard pattern memory 3 outputs the haha turn M (k, i), and the inter-stroke distance calculation unit 5 outputs the inter-stroke distance glld (k, j ,
j) is required. Further, the heuristic function h (k, p) is output from the heuristic memory 7.

The node creation unit 9 obtains new node data Ndnew based on these values.

That is, search depth: p = = p' + 1 m Quasi-pattern number: ni = ni' Cumulative distance: Sd (g), p) ± 9d (k', I
)') +d(g)slej) Evaluation value: f(k,
p) Tsukasa d (k, p) 10h (k, p) correspondence wear! M
: Add i to Nd's correspondence history.

It is. The data Ndnew of these new nodes is.

It is output to the list creation section 10. In the list creation unit 10, new node data N is created by the process described above.
The mark is added to the node list in the dnew1z node list memory 8.

The search depth t of the node data Nd at the head of the node list is N
When t is equal to , the search ends. This node data N
The signature pattern number k in d is output as a recognition result.

The embodiments of the present invention have been described above, but for example, the second embodiment
If you want to find candidates, enter the search depth y at the beginning of the node list.
=N and the search can be continued until a standard pattern number k different from the first candidate is found. It is clear that candidate X can be found in the same way.

The node list memory 8 holds data for all nodes, but to reduce the amount of memory and calculation, the evaluation value f
Only the data of the top n nodes of the value of (k, p) may be retained. At this time, the value of n is determined appropriately so that the two-weave performance does not deteriorate.

(Effects of the Invention) According to the present invention, the amount of calculation can be significantly reduced without deteriorating recognition performance compared to conventional methods.

[Brief explanation of drawings]

Figure 1 is a configuration diagram showing one embodiment of the present invention, Figures 2 to 8
The figure is a diagram for explaining the principle of the present invention and problems of the prior art. Figure Kj? 1...Tablet% 2...Input pattern memory. 3...6m pattern memory, 4...control unit, 5...stroke distance calculation unit, 6...
・・・・Hitame-Ristic Calculation Department,
- 7... Heuristic memory% 8...
...memory at node νλ, 9...node placement section,
lO...List made-!

Claims (1)

[Claims] A handwriting detection unit that detects handwriting when writing characters and reads it as an input pattern, a standard pattern memory that holds standard patterns for multiple characters to be recognized, and an interstroke distance that calculates the distance between strokes between the input pattern and the standard pattern. a heuristic calculation unit that calculates a heuristic function for the plurality of characters to be recognized based on the distance between the strokes, and a minimum distance when the strokes of the input pattern and the standard pattern are correlated with overlap allowed; Then, the cumulative distance is calculated by prioritizing the stroke matching that minimizes the evaluation value obtained as the sum of the cumulative distance and the heuristic function when the strokes of the input pattern and the standard pattern are matched without allowing overlap. and a search control unit that outputs a character corresponding to the standard pattern when all the strokes of the input pattern are associated with any standard pattern of the plurality of characters to be recognized. .