JPS61208186A - Correspondence processing system for on line handwriting character feature point - Google Patents

Abstract

PURPOSE:To obtain tolerance to a flicker while maintaining equal order by selecting one or plural couples of next corresponding candidata feature points in consideration of a couple of feature points which are already made to correspond and performing back track operation when correspondence processing reaches its limits. CONSTITUTION:Handwriting character data inputted on a tablet 1 is converted by a feature point extracting circuit 2 into a feature point sequence consisting of both ends of a stroke and an inflection point and sent to an input pattern storage circuit 3. A feature point correspondence circuit 6 matches the data in this input pattern storage circuit 3 with data which is read out of a standard pattern storage circuit 4 and stored in a standard pattern storage circuit 5. In this case, a calculation part 40 determines couples of feature points as objects of distance calculation successively, but couples of candidate points which are already made to correspond are held in the stack in said circuit 6; if the correspondence processing reaches its limits, the back track operation is performed by using the contents of the stack 22 and when the correspondence processing is carried out to the end an output is sent to a signal line.

Description

[Detailed Description of the Invention] [Summary] In the process of associating feature points between an input pattern of online handwritten characters and a standard pattern, a feature point matching circuit is provided, and a feature point pair that has already been matched is taken into consideration. Alternatively, you can select multiple next corresponding candidate feature point pairs, select one next corresponding candidate feature point pair, and then proceed. It is disclosed that, by configuring the characters to have the same character, the correspondence according to the order of strokes can be strictly observed, while giving some latitude to the correspondence.

[Industrial application field]

The present invention relates to an online handwritten character feature point matching processing method, and in particular, feature point matching that strictly matches feature points forming characters according to the stroke order between an input pattern and a standard pattern, regardless of the number of strokes. Regarding processing method.

[Conventional technology]

In conventional online handwritten character recognition devices, in order to improve the recognition rate by being able to deal with variations in the number of double strokes and stroke order,
Change the order of strokes.

Several methods are known in which recognition is performed by calculating the similarity between an input pattern and a dictionary pattern by expanding or contracting the length. However, since these methods calculate the distance by associating the input pattern and the standard pattern in a well-balanced manner as a whole, local differences in strokes, etc., tend to be ignored. Some kind of candidates are output even for characters that are not recognized.

The drawback was that it could not be rejected.

[Problem that the invention seeks to solve]

The present invention allows for stable recognition of characters to be recognized by strictly associating feature points that make up strokes between an input pattern and a standard pattern according to stroke order, regardless of the number of strokes.

A feature point matching processing method is provided to achieve recognition that reliably rejects misspellings and non-recognized characters.

[Means for solving problems]

FIG. 1 shows a block diagram of the principle of the present invention.

In the figure, handwritten character data input from a tablet 1 is converted by a feature point extraction circuit 2 into a feature point sequence consisting of both end points and bending points of a stroke.

It is sent to the input pattern storage circuit 3. The standard pattern storage circuit 4 stores a large number of standard pattern data for recognition expressed by similar feature point sequences, from which standard pattern data for matching are sequentially extracted and stored in the standard pattern storage circuit 5. cent. In this state, the feature point matching circuit 6 for matching feature points is activated. The circuit 6 includes a distance calculation section 40, a stack 22, and a next corresponding feature point determination circuit section 16, and performs distance calculations to sequentially determine feature point pairs. Then, the result of the correspondence is sent to the correspondence result storage circuit 7.

A signal line 8 is sent from the feature point matching circuit 6 to the inter-pattern similarity calculation circuit 9, and is a control signal indicating whether the matching has been completed to the end or has been stopped midway. Only when the signal line 8 indicates that the matching has been completed to the end, the inter-pattern similarity calculating circuit 9 is activated, and distance calculation between patterns is performed based on the matching result.

The results of distance calculation are sent to the sorting circuit 10, where the character codes sent from the standard pattern storage circuit 4 are rearranged in descending order of distance and output as a recognition result.

During the above processing, candidate point pairs matched in the feature point matching circuit 6 are held in the stack 22 in the circuit 6, and the above-mentioned backtracking is performed as necessary.

[Effect]

In the process of the present invention, as feature points representing two characters, both end points of each stroke and a plurality of (0 or more) Try to use bend points. The black circles correspond to end points and the white circles correspond to bending points. Two feature point sequences to be associated are A= (a++ at,..., aN)
When + B = (bM, bz, ., bM),
From associating the first feature point aI+bM with a forced contract, so to speak, to the end point a,...,ba.

Follow the stroke order and use the feature point pairs that were matched earlier.

Pairs of feature points corresponding to the first order are determined sequentially.

In the embodiment of the present invention, in order to strictly and stably match one character as a whole regardless of the number of strokes, the following items (1) to (
It has the following characteristics shown in section 3).

(1) In order to make the correspondence strictly, the correspondence is determined based on distance calculation using multiple feature values.

Assuming that the previous corresponding feature point is (am l bJ ), the correspondence distance for the feature point (a, b,) is defined as follows, as will become clearer with reference to FIG.

D" (dB ・Wl +d, -wl+d2'
Wl )/(w, +wt+w3) However, it is the angle formed by the set of d blasphemy stones and the 5 niba.

One + (Lx 1bjbt l) It is.

Also, d3 is the angle between the cod: η and the mouth bt-+, and for 1 weight W, WZ = γ Wl = 8 ・m1n(l ak al+1 l ,
lmajl) (β and γ are certain constants).

Here, if one or both of +all+bL is the final point of the feature point arrays A and B, the distance M a 3 and the weight value W become undefined, but in that case, they are defined as follows. do.

a) If ak, b, both are the final point, d, -0°, W, = δ (constant) b) If only one of a, bt is the final point, d = 180
', w, -δ In the above distance calculation, this is referred to in section (2) described later.

If there is a matching distance within the matching range that is less than a certain threshold, it is used as a matching candidate point pair, and if there is not, backtracking is performed to return to the other previously stacked candidates. Make a correspondence. For this reason, when a plurality of corresponding point candidate pairs exist, the one with the smallest distance is set as the next corresponding feature point pair for the time being, and the others are saved in the stack together with the previous corresponding point pair.

(2) Limiting the range of correspondence to reduce the amount of calculation and to prevent unnatural correspondence.

As shown in FIG. 4, the previous corresponding point (at) is indicated by a black circle.

The primary corresponding points (ah, bt) with respect to b, ) are selected from a range that satisfies the following conditions indicated by diagonal lines in the figure.

That is, i) k≧t+1 and ρ≧j+1, and 1i) k=i+l or ρ=j+1. At that time, in order to reduce the amount of calculation and to avoid unnatural correspondence, the corresponding range is set as follows. Limit as follows.

(a) Limitation by length is fixed at bj+1 and no correspondence is made from +ai to ak.

Similarly, fix a to a i, and from bj to b[ is +
No correspondence is made with bL+1' and beyond.

佽) Restriction by non-omissible point am If (k≧i+1) is a non-omissible point, no association is made with aK41 or later. Similarly, (Q≧j+
If 1) is a non-omissible point, then No correspondence is made with 1 or later.

However, feature point ak is defined as a non-omissible point when it satisfies the following conditions.

i) aK is the start or end point of the stroke;

ii) The angle formed by am-+τ and a, ak+1 exceeds a certain dark value (acute angle).

1111--→ ・iii) If am is the starting point, the length of aK-Iak is Iak
If is the end point, a,,a? The length of I exceeds a certain darkness value (sufficiently long).

Also, if ak is a feature point constituting a standard pattern, whether or not it cannot be omitted is a question.

It may also be artificially set in advance.

(C) Restrictions based on the type of feature points i) There is no correspondence between the start point and end point of the stroke.

1i) Fix bc to b j + l and calculate b j 41 and a
, are both stroke starting points, they are not associated with a k+2 or later.

Similarly, ak is ai. Fixed ai to 1. If both 8 and 8 are the starting points of the stroke, no correspondence is made between bL+对 and later. These restrictions are intended to suppress unnatural correspondences such as associating actual strokes with virtual strokes (off strokes) connecting strokes.

(3) Perform backtracking in a timely manner to prevent local errors from spreading throughout the system.

If there is no feature point pair whose matching distance is within the threshold within the matching range, backtracking is performed using the feature point pairs saved in the stack.

If the stack is empty, the association is aborted.

In this case, the following measures are taken to prevent the amount of calculation due to backtracking from increasing exponentially.

(a) In general, the order in which a plurality of matching candidate point pairs are saved to the stack is set such that the one with the largest matching distance comes first. By doing this, the mapping with the smaller mapping distance, that is, the mapping with the highest probability, will be evaluated first, and if the mapping is completed successfully, the final mapping result can be obtained with less calculation. It will be done.

(b) Make the correspondence first, memorize the failed route, and prevent the same route from being taken again.

Specifically, for the previous corresponding point (ai + J).

The following processing is performed using the NXM array path (PATH) that stores the next corresponding point (ai+, bt).

i) Initialize array PATH to (0,0).

ii) For the previous corresponding point (at, bz), PATH(i
.

iii) Once the primary corresponding point (am, bL) is determined for the previous corresponding point (ai + bj), substitute (k, Q) into PATH (i, j).

iv) When back trunking is performed, in addition to the next corresponding candidate point pair (akl, bt'), the previous corresponding point pair (ai' + bj') at that time is also saved,
t and j + (7) 1? PATH(i', j')=
(k', Q') The next corresponding point pair is determined by using the zero-order feature point pair (ak'*b,'), which is exchanged with the array PATH4-IF, as the previous corresponding point pair.

(e) Add radical end information to the feature point sequence constituting the standard pattern, and if the correspondence progresses normally up to the radical end, the stack is forcibly cleared at that point.

By doing this, it is possible to suppress undesired backtracking when the correspondence has progressed sufficiently to the light.

As described above, in the present invention, if the next corresponding feature point pair cannot be found and the stack is empty and backtracking is not possible, the association is stopped at that point. For this reason, only patterns that are similar in character shape, including stroke order, are matched until the end.

〔Example〕

FIG. 5 is a functional block diagram of the feature point matching circuit shown in FIG. 1.

In the figure, when the feature point string data to be correlated is set in the feature point string data storage circuits 11 and 12, the correspondence information storage circuit 13 that stores the NXM array is first set (
Initialize to 0, O).

Next, registers 14 and 15 store the position of the previous corresponding feature point.
is initialized to (1), and then the next corresponding feature point determining circuit 16 is activated.

The circuit 16 first reads the value stored in the circuit 13 for the previous corresponding feature point pair (i, j) using the data lines 17 and 18, and if it is not (0°0), backs up. - Track processing can be performed (described later).

The next corresponding feature point is calculated, and if there is one or more corresponding candidate point pairs, the one that minimizes the matching distance is set in registers 19 and 20. If there are a plurality of corresponding candidate points, the feature point pairs other than those set in the registers 19 and 20 are saved in the stack 22 via the data line 21 together with the previous corresponding feature point pair (i, j). If there are no corresponding candidate points, back tracking processing is performed (described later).

On the other hand, when the ninth-order corresponding point pair (k, ρ) is determined, it is sent to the circuit 13 along with the previous corresponding point pair (i, j) using the data line 23, and the element (i, j) of the array is (0
, 0) to (k, ρ).

Backtrack processing is performed as follows.

When back track processing is requested from the circuit 16 through the signal line 24, the bare feature point pair most recently saved from the stack 22 (previous corresponding feature point pair (1' l
J'') and the next corresponding candidate point pair (k'Q''') are extracted and the elements of the array (19°jw) of 9 circuits 13 are converted to (k'
, Q'). Next, the values of registers 14 and 15 indicating the position of the previous corresponding point are set to ' and Q', respectively.

If the stack is empty, the mapping is aborted.

On the other hand, the next corresponding feature point pair is successfully determined without backtracking, and the value (k,
When Q) is placed in the registers 19 and 20, the end determination circuit 25 determines whether the association has reached the end point of the feature point sequence, and if so, ends the process. If not, the gate 26.27 is opened and (k, Q) is stored in the register 14.15 as the position of the new previous corresponding feature point pair.

Depending on the value of the signal line 24, the changeover switches 28 and 29 set the data from the stack 22 in the case of backtracking, and the data from the registers 19 and 20 in the case of not backtracking, in the registers 14 and 15. It is for.

FIG. 6 is a block diagram illustrating the operation of the next corresponding feature point determining circuit shown in FIG.

In the figure, when the feature point string data is set in the feature point string data storage circuit 11.12 and the position data (i, j) indicating the position of the previous corresponding point is set in the register 14.15,
The value of (i, j) is sent to the judgment circuit 34 and the value of 2 judgment circuit 3
4, the element (i, j) of the two-dimensional array stored in the correspondence information storage circuit 13 is read out, and it is
), a control signal 36 instructing to perform backtracking is sent. When the element (t,j) of the array is (0,0), the correspondence range determination circuit 38 is activated by the signal line 37, and the next corresponding point pair is determined in the manner described in the [Operation] section above. Determine the range to select.

A set of feature point pairs ((k
, Q)) are sent to the storage circuit 39, and each feature point pair (k,
ρ), the correspondence distance is calculated by the correspondence distance calculation unit 40. The matching distance is determined by comparison circuit 4.
1 and is compared with a threshold value Th stored in the register 42, and only feature point pairs smaller than the threshold value Th are sent to the next corresponding candidate point pair storage circuit 43 as ninth-order corresponding candidate point pairs. The determining circuit 44 selects the one with the minimum distance from among the feature point pairs stored in the circuit 43 and outputs it as a primary corresponding feature point pair, and rewrites the elements of the array in the three circuits 13 into the determined feature point pair. . The remaining feature point pairs that were not selected are saved to the stack (22 shown in FIG. 5).

If there is no feature point pair with a matching distance smaller than the threshold Th and the circuit 43 is empty, the decision circuit 44 sends out a control signal 45 instructing to perform backtracking.

It should be noted that the present invention provides a correspondence processing method having the following characteristics.

(1) As the distance calculated from a certain feature point pair (aklbL) and the previously corresponding feature point pair (at, bj), the vector [°1τ, which connects the feature points a and a, and the feature point S, The vector shoulder bL that connects and, the feature points a, and ak
Vector ah ak*1 connecting the next feature point a mo+
And, the feature point is.

Vector b, b, connecting the next feature point bt** of . 1 and against.

i) Difference between vector ai ak and vector b, bjO opening angle, ii) Difference in length between vector ai a, , and vectors b, bL, and iii) Vector akak+1 and vector b, bt
Calculate the distance using a weighted average of the three distances of ++ direction angle differences.

(2) As the weight of the weighted average of the above three distances.

For the distance i), vector 17x12 and vector b
The weight value is a value proportional to the average value of the length of jbL, ii
) for the distance, a constant value is used as the weighting value, and for the distance 1ii), the vector 7]0. and vector b, b,.

Multiplying each weight using a value proportional to the minimum length of 1 as a weight value.

(3) When selecting feature points, previously matched feature point pairs (
aibj) as the range specified as a reference, the next feature point al of ai. The feature points bt after 1 and bJ (
A set A= ((az++, b,
) I Q>j+1) and the next feature point of b,
. Set B = ((ak, bj, +) l k:?
:i+1), the range should be the union ALIB.

(4) As the range specified by (at, b,) above.

Set ((a=-+, bt) I Q ≧j+1
), the length of the vector sequence from the feature point I to bL is the vector ai ai. Similarly, the subset A9 is within the range specified by the length of , and the set ((am, bJ+I) l k≧i+1) is the vector bjbj. The union set A' with respect to the subset B', which is within the range specified by the length of .
The range should be UB'.

(5) The range specified by the above (ai, bJ) is a feature point that has been determined in advance to be non-omissible using a certain method in the above subset A9.

When a pair (ai-+, bm) of and ai, I is included, a subset A'' is obtained by deleting the pair of the feature point after b, , and a i+1 from the subset A9.

Regarding the subset B vv obtained by similar logic.

The range should be the union A “U B”.

(6) As the range specified by the above (at, b=), delete the feature point pair '(am, b,) that satisfies any of the following conditions from the union A"IJB"'. to use

+) am is the start point of the stroke, is the end point of the stroke 1i) aII is the end point of the stroke, is the start point of the stroke (7) In determining that the feature point in (5) above cannot be omitted, the feature point Determine that a is indispensable when it satisfies all of the following conditions.

i) Feature point a is the starting point or ending point of a certain stroke.

ii) For the vector amam-1 connecting the feature point a1 and the previous feature point a,-1. Feature point a.

and the next feature point a,. Vector 7 connecting 1: The angle formed by 1 is smaller than a certain threshold.

1ii) If a is the starting point, then the length of the vector mini O is r
If am is the end point, the length of the vector 8a+a+1 is longer than a certain dark value.

(8) In the order in which feature point pairs are saved to the stack, those with larger distances are saved first, and used in the back track first.

(9) What is the determination that the matching process will not proceed further?

When the matching reaches the feature point pair (ai, bJ), it is determined that the matching will not proceed further if any of the 9-order conditions is satisfied.

i) There is no feature point pair next to (am, bJ) that is located at a certain distance within the threshold.

ii) At least - Backtracking of the association has occurred for collection purposes, and (at, bj) has been associated in the past.

〔Effect of the invention〕

As explained above, according to the present invention, it is possible to perform backtracking, so that it is possible to strictly maintain the stroke order while providing some margin for positional fluctuation.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the principle of the present invention, and FIG. 2 is an explanatory diagram illustrating the relationship between input patterns and feature points. FIG. 3 is an explanatory diagram for explaining the correspondence distance, FIG. 4 is an explanatory diagram for explaining the correspondence range, and FIG. 5 is a configuration diagram of an embodiment of the feature point correspondence circuit shown in FIG. FIG. 6 shows a block diagram of an embodiment of the next corresponding feature point determination circuit shown in FIG. In the figure, 1 is a tablet and 2 is a feature point extraction circuit. 3 is an input pattern storage circuit, 4 is a standard pattern storage circuit, 5 is a standard pattern storage circuit, 6 is a feature point correspondence circuit, 7 is a correspondence result storage circuit, 9 is an inter-pattern similarity calculation circuit, and 10 is a sorting circuit. 16 is a next corresponding feature point determining circuit, and 22 is a stack. 40 represents a correspondence distance calculation unit.

Claims (1)

[Claims] One feature point sequence includes stroke end points and bending points for an input pattern input by hand, and the other includes stroke end points and bending points for a standard pattern given in advance. The feature point sequence is the other, and between the two, the first
In an online handwritten character recognition device that associates points with each other and then determines the next corresponding feature point in the order of input, the next pair of corresponding feature points to be checked for correspondence is (a_k, b_t) and the immediately preceding pair of feature points is (a_i, b_j), a distance calculation section 40 that calculates a distance based on each of the above-mentioned respective feature point pairs, and a result calculated by the distance calculation section 40. A stack 22 that holds one or more next correspondence candidate feature point pairs obtained by , and other next correspondences that were held in the stack 22 when a bottleneck occurred in the matching process that proceeded earlier. A feature point matching circuit 6 having at least a next corresponding feature point determining circuit section 16 having a function of selecting a candidate feature point pair and backtracking is provided, and the feature point matching circuit 6 is connected to the stack 22. The online method is characterized in that the matching process is proceeded based on the content, and when a blockage occurs in the matching process when the content of the stack is empty, the process is stopped. Correspondence processing method for calligraphic character feature points.