JPS62164183A - Corresponding system for stroke string in pattern recognition - Google Patents

Abstract

PURPOSE:To attain the accurate correspondence of a stroke by providing an inter-stroke distance calculating means, a candidate stroke selecting means, an inter-stroke relative position relation deciding means, and a stroke string corresponding means, etc. CONSTITUTION:An inter-stroke distance calculating means 3 performs a distance calculation between an input pattern stroke data 1 and a reference pattern stroke data 2, and generates an inter-stroke distance matrix 4, and outputs it to a candidate stroke selecting means 5. The means 5 generates the first corresponding candidate data 6, and outputs it to an inter-stroke relative position relation deciding means 7, and the means 7 generates the second corresponding candidate data 8, and outputs it to a stroke corresponding means 9. The means 9, when one or plural number of strokes which are decided as permissible by the data 8 are present, corresponds one of them to the stroke in the present reference pattern tentatively, and saves remainders at a stack means 10. And such operations are repeated until all of the strokes in reference pattern have been corresponded, thereby the accurate correspondence can be performed.

Description

[Detailed description of the invention] [Summary] One stroke row is created between the input cover pattern and the standard pattern.
When making pair-to-one correspondences, correspondence between strokes is taken in a manner that is not affected by variations in stroke order.

[Industrial application field]

The present invention relates to a pattern recognition method, and more particularly to a method for associating an input cover pattern and two stroke strings forming a standard pattern in a one-to-one correspondence in online handwritten character recognition.

[Conventional technology]

An online system that recognizes characters input by hand using a tablet or the like must be able to correctly recognize characters without being affected by errors in stroke order.

Conventionally, a method using an inter-stroke distance matrix is well known as a method for dealing with variations in stroke order in online handwritten character recognition.

This method performs processing such as normalization on the input cover pattern, then calculates the inter-stroke distance matrix dij for all stroke pairs between it and the standard pattern, and calculates di
The correspondence between strokes is determined based on the value of j.

In this method, j that minimizes the inter-stroke distance dij is determined for the first stroke of the standard pattern, and the first stroke of the input pattern is determined as the corresponding stroke.

Figure 4 shows 1% of each stroke, indicated by 1 to 4, between the introductory pattern and the standard pattern for the character ``tree''.
tt d t j. In the figure, dI□, d 2++ d33r d44 surrounded by circles give matched stroke pairs.

By the way, if it is allowed that a certain stroke of the input pattern overlaps with multiple strokes of the standard pattern, it is sufficient to find j such that dij = min dik for each i, but it is not possible to have two overlapped correspondences. If this is not allowed, it is necessary to exclude strokes for which correspondence has already been determined. Therefore, the final result differs depending on the order of association.

For example, in the example of the character "ki" in Fig. 5, the standard pattern 1. Since the input strokes with the minimum distance to the second stroke are both the first stroke, if the strokes corresponding to the strokes of the standard pattern are determined in the order of 1, 2, and 3, the first stroke of the standard pattern has the input stroke of the input pattern. Whereas the first stroke corresponds to 3.2
．． In the order of 1, the second stroke corresponds to the first stroke.

Now, with respect to the first stroke of the standard pattern, the stroke number of the corresponding person's kataan is expressed as τ(1).

If the sum of the distances between strokes of corresponding pairs of strokes is D, then D===Σdiτ(i).However, there have been many attempts to determine the correspondence between stroke pairs so as to minimize the white. Ta.

In the example in Figure 5, the stroke pair is (1, 1) (2, 2
) When associated with (3°3), D is.

D = d I I + d 2 + dti = 4 + 7
'+ 8 = 19, whereas (L2) (2,
1) What is D when associated with (3, 3)?

D=d, □+dz++d:++=14+6+8=28, and the former association is selected.

[Problem that the invention seeks to solve]

However, generally between two stroke rows.

To determine the correspondence that minimizes D.

N! This corresponds to selecting a unique combination from among the combinations (N is the number of strokes).

As N becomes large, it becomes significantly difficult to determine such a response in real time. Another drawback is that even if such a correspondence is obtained, there is no guarantee that it is a correct correspondence.

Means for Solving Problem C] The present invention provides a stroke sequence that allows two stroke sequences of equal number to be accurately matched without duplication based on the inter-stroke distance matrix and the relative positional relationship of the strokes. It provides a mapping method, and more specifically, it uses the fact that if an incorrect mapping is performed, the relative relationship of strokes will collapse, and more accurate mapping is performed.

FIG. 1 shows the basic configuration of the present invention.

In the figure, 1 is the input pattern stroke data, 2
is standard pattern stroke data, 3 is inter-stroke distance calculation means, 4 is inter-stroke distance matrix, 5 is candidate stroke selection means, 6 is first corresponding candidate data, 7 is inter-stroke relative positional relationship determination means, 8 is the first 2 correspondence candidate data, 9 a stroke string correspondence means, and 10 a stack means.

The distance between stroke calculation means 3 calculates the distance between each stroke data 1 of the input pattern and all the stroke data 2 of the standard pattern, and calculates the distance between each stroke data 1 of the input pattern and all the stroke data 2 of the standard pattern, and calculates the distance between each stroke data 1 of the input pattern and all the stroke data 2 of the standard pattern, and
Create.

The candidate stroke selection means 5 identifies a plurality of strokes of the input cover turn with a small distance between strokes as corresponding candidates for each stroke of either the input cover turn or the standard pattern, for example, each stroke of the standard pattern. In this way, the first correspondence candidate data 6 is created.

The inter-stroke relative positional relationship determining means 7 uses the first correspondence candidate data 6 to determine each stroke of the input pattern from which strokes for which correspondence with other strokes has already been determined are removed from among correspondence candidates for a certain stroke in the standard pattern. , determine whether the relative positional relationship between that stroke and the previous stroke of the incoming pattern is permissible based on the relative positional relationship between the current stroke of the standard pattern and the previous stroke. death.

The second correspondence candidate data 8 is created using only the permissible data.

The stroke string matching means 9 uses the second matching candidate data 8
If there is one or more strokes determined to be permissible, one of them is temporarily associated with the stroke of the current standard pattern, and the rest are saved in the stack means 10 and used as the secondary standard. Proceed to match the strokes of the pattern. If there is no stroke determined to be permissible, the correspondence candidates saved in the past are retrieved from the stacking means 10.

Undo the mapping operation. , and repeat these operations until all the standard pattern strokes are matched.

Note that if the stack means 10 becomes empty during the middle of the process and the mapping cannot proceed further, this means that the 9 is originally trying to map different 9 characters, so the mapping is stopped at that point. .

[Effect]

According to the present invention, the two stroke sequences of the input cover pattern and the standard pattern are basically associated with the strokes of the standard pattern in the order in which the strokes of the input cover pattern are input.

Multiple candidate strokes for correspondence between both patterns are selected with small distances between strokes, and among them, combinations that satisfy the correspondence of all strokes are searched for strokes with reasonable relative positions. be done.

This search uses the so-called backtrunk method, which when stroke matching fails midway through, replaces the previously matched strokes with other matching candidates and redoes the matching, so even if the stroke order changes, It is not essentially affected.

〔Example〕

Next, details of the present invention will be explained according to the embodiment shown in FIG.

In the figure, 11 and 12 are stroke string storage units for storing each stroke string of two patterns such as an input cover pattern and a standard pattern, 13 is a stroke number register in which the number of strokes of an input cover pattern is set, and 14 is an interval between strokes. a distance calculation unit; 15 is an inter-stroke distance matrix storage unit;
16 is a stroke number register, 17 is a stroke string correspondence result storage section, 18 is a correspondence candidate selection section, 19 is a first correspondence candidate storage section, and 20 is a relative positional relationship determination section.

21 represents a second correspondence candidate storage unit, 22 represents a control unit, and 23 represents a stack circuit.

The stroke string storage unit 11.12 contains two stroke strings A=(a, a, ++aH) having the same number of strokes.

B=(b+, bz, . . . bN) is input. however.

N is the number of strokes ai, bj is the stroke, and each has a certain number of feature point sequences ai = ((xi+, yi+), ... (xi +
yt, 4) ) bj= ((xL, yL), ++(Xj
x + yjs )).

In FIG. 2, when stroke strings A and B are set in the stroke string storage sections 11 and 12 and the number of strokes N is set in the one-stroke number register 13, the inter-stroke distance calculation section 14 is first activated.

Stroke distance d between each stroke ai and bj of B
ij is stored in the inter-stroke distance matrix storage section 15.

Here, the distance between strokes dij for strokes at and bj is defined as linear.

(M is the number of feature points constituting a stroke) When the calculation of inter-stroke distances for all stroke pairs A and B is completed, the stroke number register 16 is set to 1''.

The stroke number register 16 contains the number i of the stroke at which the correspondence is being considered at that point in the stroke sequence A.
is stored.

The matching is performed in order from i=l until the number of strokes reaches N.

The results of the association are stored in the association result storage section 17.

Currently, the stroke number register 16 has the value i
It is assumed that (1≦i≦N) is set. The correspondence candidate selection unit 18 selects the stroke at as a candidate in the stroke sequence B.

(1) Stroke bj also corresponds to or is not attached to any stroke. (2) On the condition that the distance between strokes dij is smaller than the threshold value dmax, the pair of stroke numbers j and i of bj (
1+j) is set in the first correspondence candidate storage section 19.

Here, condition (1) can be easily verified by referring to the contents of the association result storage section 17.

A plurality of correspondence candidates ((i.

L), (i, jz) ++ (i, j+s) )
When (n≧0) is set, the relative positional relationship determination unit 20 is activated in the ninth order.

In this part, from the matching result storage unit 17.

Stroke b corresponding to the previous stroke ai-1
j' and store the actual stroke data for each correspondence candidate (t+JK) in the stroke string storage unit 11゜12.
After extracting the strokes, it is determined whether the relative relationship between the strokes tljx and bj' is acceptable by comparing it with the relative relationship between the strokes ai and ai-1, and only acceptable correspondence candidates are stored in the second correspondence candidate storage section 21. send to

When a plurality of (0 or more) correspondence candidates are set in the second correspondence candidate storage section 21, the control section 22 performs the following operations.

will be activated.

In the control unit 22, the number of correspondence candidates set in the second correspondence candidate storage unit 21 is (1) 1 (i, j).

To the association result storage unit 17 (centre i, D).

If i=N, output the correspondence result of correspondence result storage section ■7,
Finish the process.

If i<N, 1+1 is set in the stroke number register 16 and the matching process is repeated.

(2) 2 or more ((i, L), ... (x+jm)
)in the case of.

(i, jl) is set in the association result storage unit 17.

The remaining correspondence candidates ((i, jz), ... (ilJII
)) is sent to the stack circuit 23.

Set i+1 in the stroke number register 16 and repeat the process.

(3) In the case of 0 pieces.

The most recently stored correspondence candidate (i', j') is retrieved from the stack circuit 23.

Store (i', j') in the association result storage unit 17.

The value of the stroke number register 16 is set to i'+1 and the process is repeated.

If the stack circuit 23 is empty, the process ends.

Corresponding results are not output.

The determination of the two stroke pairs ai and a, -1 and bj and bj' in the relative positional relationship determining section 20 is performed as follows.

Pectroux dimension connecting the start and end points of the stroke at-+ P.

The vector connecting the end point of al-1 and the start point of at is P2at
The vector connecting the starting point and ending point of is P.

The vector connecting the start point and end point of bj' is q1, the vector connecting the end point of bj' and the start point of bj is q2, the vector connecting the start point and end point of bj is q3. (Figure 3) Here, R = 1 (angle between n and n) - (angle between measure and equation) 1+1
(Angle between P2 and P3) (Angle between qz and 93) If defined as 1, the smaller R is, the more similar the relative relationships between at and ai-1 and between bj and bj' become.

Therefore, a certain threshold value Rmax is determined, and the calculated R is Rm
Only if it is smaller than ax, the corresponding (ai-+
, bj'), it is determined that (at, bj) is acceptable as a correspondence candidate.

〔Effect of the invention〕

According to the present invention, since the correspondence is determined by taking into consideration not only the absolute position coordinates of each stroke, but also the relative positional relationship between the strokes, there is an effect that more accurate stroke correspondence is possible.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the basic configuration of the present invention, FIG. 2 is a diagram showing the configuration of one embodiment of the invention, FIG. 3 is an explanatory diagram of the relative positional relationship of strokes, and FIG. 4 is an explanatory diagram of the distance matrix between strokes. , FIG. 5 is an explanatory diagram of an example of correspondence between strokes that do not allow overlap. In Figure 1. 1: Stroke data of input pattern 2: Stroke data of standard pattern 3: Interstroke distance calculation means 4: Interstroke distance matrix 5: Candidate stroke selection means 6: First corresponding candidate data 7: Interstroke relative positional relationship determination means 8 : Second correspondence candidate data 9: Stroke string correspondence means 10 Ni stack means 2nd figure Stroke 2 Nememura I standing position 1%. $3 b

Claims (1)

[Scope of Claims] In pattern recognition for associating two stroke sequences of the same number one-to-one with no overlap, a means for calculating inter-stroke distances for all stroke pairs (
3), and b. means for selecting a plurality of corresponding candidate strokes from the other stroke sequence for a certain stroke in one stroke sequence based on the distance between strokes calculated above (5);
), and c, compare the relative positional relationship between any two strokes in one stroke sequence and the relative positional relationship between any two strokes in the other stroke sequence, and determine a stroke pair that is an acceptable correspondence candidate. d, a stack means (7) for storing stroke pairs of corresponding candidates;
10) and e, sequentially associate two stroke sequences based on the corresponding candidate stroke pairs whose relative positional relationship of the strokes is acceptable, and transfer the corresponding candidate stroke pairs remaining from the correspondence to the stacking means. storage means (9), and if the correspondence between the two stroke sequences fails midway, the correspondence candidate stroke pairs stored in the stacking means are used to associate the stroke sequences. A stroke sequence mapping method characterized by redoing.