JPH073670B2 - Stroke distance calculation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial application] The present invention relates to an inter-stroke distance calculation method capable of reflecting bending point information extracted from a stroke of a handwritten pattern in calculation of the inter-stroke distance as much as possible. .

As a method of recognizing a handwritten pattern such as a handwritten character, there is a method of recognizing the handwritten pattern by calculating the distance between the strokes forming the handwritten pattern and the strokes of the dictionary pattern.

Since the feature points included in the strokes used for the calculation include important information for pattern recognition, it is important not to lose the information in the distance calculation as described above in order to recognize the handwritten pattern with high accuracy. It becomes an important matter.

In the present specification, the term "segment" refers to a partial stroke having two arbitrary feature points in a stroke as a start point and an end point, and the term "segment group" likewise refers to a feature point associated with the start point as a start point. It is assumed that a set of a plurality of segments whose end points are rearward feature points obtained under a certain condition is pointed to and distinguished.

[Conventional technology]

As a conventional method for calculating the distance between strokes as described above, two strokes used for distance calculation are approximated by making a predetermined number of points have the same interval, The one that uses the sum of the distances between the associated point pairs (see FIG. 5) and the one that is used for distance calculation
There is a method in which the distance is calculated by applying DP matching to a generally different number of point sequences obtained by approximating one stroke at a given length as a unit at equal intervals (see FIG. 6).

[Problems to be solved by the invention]

In any of the above-mentioned methods, since the stroke shape is approximated at a certain interval without paying sufficient attention to the stroke shape, it is important to consider the stroke distance calculation originally, such as a bending point. There is a risk that the element point information will be missing from the calculation (see FIG. 7), which is one of the factors that hinder the improvement of pattern recognition accuracy.

If we simply narrow the approximate intervals in order to avoid such problems, not only will the amount of data processing increase, but also important points such as bending points and end points will be included in the values calculated by the distance calculation. Since the specific gravity occupied by the characteristic points is relatively reduced, it is difficult to say that the method is effective.

Therefore, an object of the present invention is to provide a stroke distance calculation method in which feature points occupy a high specific gravity in stroke approximation, except for the above-mentioned drawbacks of the prior art.

[Means for solving problems]

 The present invention is provided to achieve the above object, and the constitution thereof is as follows.

That is, in claim 1, when calculating the distance between the stroke expressed as time-series data of the coordinate points of the handwritten pattern and the stroke of the dictionary pattern,
A first means for extracting a bending point from a stroke and approximating the stroke by using the extracted bending point and stroke end point as the characteristic points, and determining the correspondence between the approximated characteristic points of the stroke and performing the characteristic point correspondence. In the inter-stroke calculation method including a second means for performing pattern recognition by performing distance calculation between the strokes, the second means first associates the stroke start points with each other and then associates them first. An inter-segment similarity calculation means for sequentially determining a corresponding feature point based on the feature point pair, and calculating a similarity degree between segments having an arbitrary feature point pair of a stroke as a start point and an end point. For the first and second segment groups, each of which has a corresponding feature point pair as a start point and whose end point is a backward feature point group obtained under a certain condition. Segment similarity is calculated for all combinations of pairs of a segment belonging to the segment group and a segment belonging to the second segment group, and the segment pair having the maximum similarity is determined, which corresponds to both end points of the segment pair. The feature point pair to be determined is determined as the corresponding feature point pair.

Further, in claim 2, in the association by the second means in the first aspect, two associated feature points are set as a start point and an adjacent feature point of the feature point made the start point is set as an end point. A long segment of the segments is fixed, and for the other sexpton, the sum of the distances from the starting feature point to the feature point is limited to a feature point that exceeds a fixed multiple of the length of the fixed segment for the first time. Then, the similarity between the segments is calculated.

[Action]

According to the method of the present invention, the feature points are added to the approximation of the stroke, and the feature points of the approximate stroke are associated between the strokes used for the distance calculation. A stroke, for example, determines the correspondence between the feature points forming A and B. The correspondence is made by first associating the start points of A and B, and the next corresponding feature point based on the already associated feature point pair. Determine the pair.

Assuming that the characteristic point of the stroke A, for example S, corresponds to the characteristic point of the stroke B, for example T, the inter-segment similarity is calculated in the first segment group, that is, the characteristic point S. All combinations of a segment that is an element of the segment group having the starting point and a second segment group, that is, a segment that is an element of the segment group having the characteristic point T as the starting point.

In this way, the segment pair having the highest similarity is determined, and the feature point pair corresponding to both end points of the segment pair is determined as the next corresponding feature point pair.

Therefore, all the feature points are considered in the calculation of the distance between strokes, and they largely contribute to the improvement of the pattern recognition accuracy.

〔Example〕

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

First, the basic concept of the present invention will be described. Handwriting pattern,
For example, the inflection points and end points extracted from the strokes displayed as time-series data of the coordinate points of the Chinese character pattern are used as feature points to approximate the strokes, and the approximate strokes are characterized by the strokes of the dictionary pattern. The points are associated with each other, and the distances between the strokes associated with the feature points are calculated and used for the pattern recognition processing.

The above-mentioned correspondence between the feature points is performed as follows. First, the start points (s ₁ and t ₁ ) of the strokes are associated with each other. And each subsequent correspondence is performed as follows. When it is assumed that both arbitrary feature points si (for example, feature points of arbitrary strokes in the dictionary pattern) and arbitrary feature points tj (for example, feature points of arbitrary strokes of the input pattern) correspond, Next feature point
sk, tl (i <k, j <l) is determined as follows.

A line segment connecting the feature points si ₊₁ and tj ₊₁ adjacent to the feature points si and tj associated as described above (this line segment is the segment described above). The lengths of sisi ₊₁ and tjtj ₊₁ . Compare the In this comparison, if | sisi ₊₁ | ≧ | tjtj ₊₁ | ・ ・ ・ (1), then sk = si ₊₁ and if | sisi ₊₁ | <| tjtj ₊₁ | ・ ・ ・ (2) , Tl = tj _{+ 1} .

In the above comparison, sk or tl can be determined, but tl or sk cannot be determined. Therefore, the characteristic point is determined as follows. For example, | sis
When i ₊₁ | ≧ | tjtj ₊₁ | (see FIG. 2), l = | sisi ₊₁ | ... (3) and Is defined (however, in the formula (4), l ₁ = | tjtj ₊₁ |, l ₂
= | Tj ₊ 1tj _{+ 2} |, ... lm = | tj ₊ m _- 1tj ₊ m |, the set of a plurality of segments as described above is the segment group described above. )
Sum Lm of the segment for the first time l = | sisi ₊₁ | more than such a point tj ₊ n (n = 1 to m) of the segments belonging to virtual segment groups connecting the feature point tj and the feature point tj ₊ n up Compute the similarity between each and the segment sisi ₊₁ . This similarity is by. However, in the above equation, and are the vectors from the start point to the end point of the two segments used for the calculation. Among the segment groups used in the calculation of equation (5), the point tj ₊ n that maximizes the similarity with the segment sisi ₊₁ is set as tl = tj ₊ n, and that point is defined as a feature point that should correspond to the feature point sk. To do.

Also, in the case of the expression (2), the same processing as above is performed to obtain the characteristic point sk that should correspond to the characteristic point tl.

With the sk and tl for which the above-described processing has been obtained as the starting point, the calculation of the characteristic points that should correspond next is sequentially performed for that stroke.

The first segment group and the second segment group may limit the range (for example, the number of segments belonging to the segment group) depending on certain conditions in order to reduce the amount of calculation in calculating the degree of similarity. For example, in this embodiment, the number of segments belonging to the first segment group is 1
I am trying. When the matching is performed, there remains feature points that cannot be matched. For example, even if it is possible to associate si and tj and si ₊₁ and tj ₊ n, t
Each feature point of j _{+ 1} , ..., tj ₊ n- ₁ remains as an unsupported feature point. Segment sisi ₊₁ for these uncorresponding feature points tj
Points (virtual feature points) to be internally divided are set by the ratio of the lengths of tj ₊₁ , tj ₊₁ tj ₊₂ , ... And these feature points are associated with uncorresponding feature points (see FIG. 3). .

In this way, all the feature points in the stroke are associated with each other.

Then, for all the feature points including the virtual feature points in the strokes associated as described above, the distances are calculated between the segments having the one-to-one corresponding feature points at both ends, and the segment length is calculated. The distance between strokes is calculated as the sum of the weighted items.

Although the above is the basis of the present invention,
If the end point of one stroke group corresponds to a feature point other than the end point of the other stroke group, and as a result, uncorresponding feature points remain (see FIG. 4), the length of the uncorresponding segment A value obtained by multiplying the sum of the values by a coefficient is added as a penalty to the distance between strokes obtained as described above.

An embodiment for implementing the inventive concept described above is shown in FIG. In FIG. 1, reference numeral 1 denotes a character input device such as a tablet, which samples coordinate points at regular time intervals and sends time-series data of the sampled coordinate points to a feature point extraction device 2. The feature point extraction device 2 performs bending point extraction processing together with noise removal and normalization processing, and sends the data to the feature point correspondence determination device 3.

On the other hand, the dictionary pattern memory 4 responds to the stroke number information (the number of pen ups and downs per character) sent from the character input device 1 and sequentially outputs dictionary pattern data of characters having the same stroke number to the feature point correspondence determining device. Send to 3. In addition,
The dictionary pattern data is stored in the dictionary pattern memory 4 after the feature point extraction processing similar to that of the input pattern data is performed in advance.

The feature point correspondence determination device 3 that receives data from the feature point extraction device 2 and the dictionary pattern memory 4 sends the segment information as described above to the inter-segment similarity degree calculation device 5 to calculate the inter-segment similarity degree, and to calculate the similarity. The feature points are associated with each other while taking into consideration the degree (the details are as described above).

After the feature point associating process is performed in this manner, the data is sent to the inter-stroke distance calculation device 6, and the process of associating the virtual feature point with the unsupported feature point is performed based on the data. (See FIG. 3), all the feature points are associated with each other, and the distance between the associated strokes is calculated.

When such stroke-to-stroke distance calculation is performed for all strokes forming a pattern, the total is sent to the candidate determination device 7 as the distance between the input character pattern and the dictionary pattern at that time used for the distance calculation.

Such a candidate distance is calculated between all the dictionary pattern data included in the range specified by the stroke number information and the input character pattern, and is sent from the inter-stroke distance calculation device 6 to the candidate determination device 7 and these candidate distances are calculated. The one with the smallest distance is selected and the character code of the corresponding dictionary pattern is output as the recognition result.

Although the example in which the inter-segment similarity is used for associating the feature points in the above-described embodiment has been described, the association may be performed by other means for obtaining the inter-segment correlation.

〔The invention's effect〕

As described above, according to the present invention, in principle, all feature points can be incorporated in the stroke distance calculation, and highly accurate pattern recognition can be directly performed.

[Brief description of drawings]

FIG. 1 is a diagram showing an embodiment of the present invention, FIG. 2 is a diagram used for explaining the correspondence of feature points, and FIG. 3 is a correspondence of feature points that have not been associated with the correspondence of FIG. FIG. 4 is a diagram used to explain processing of feature points that deviate from the correspondence principle of the present invention, FIG. 5 is a diagram used to explain distance calculation by fixed point number approximation, and FIG. 6 is FIG. 7 is a diagram used for explaining distance calculation by equidistant interval approximation, and FIG. 7 is a diagram showing a lack of bending point information in constant interval approximation. In the figure, 1 is an input character device, 2 is a feature point extraction device,
3 is a feature point correspondence determination device, 4 is a dictionary pattern memory, 5
Is an inter-segment similarity calculation device, 6 is an inter-stroke distance calculation device, and 7 is a candidate determination device.

Claims (2)

[Claims] 1. When calculating a distance between a stroke expressed as time-series data of coordinate points of a handwritten pattern and a stroke of a dictionary pattern, a bending point is extracted from the stroke, and the extracted bending point and stroke end point are extracted. A first means for approximating a stroke with the feature point as a feature point, and determining the correspondence between the feature points of the approximated stroke, calculating the distance between the strokes corresponding to the feature points, and performing the pattern recognition. In the inter-stroke calculation method including the second means, the second means first associates the start points of the strokes with each other first, and then sequentially provides the next corresponding feature point based on the previously associated feature point pairs. It has a means for calculating the similarity between segments, which is a predetermined means and calculates the similarity between segments having an arbitrary feature point pair of a stroke as a start point and an end point. For the first and second segment groups whose starting point is each pair of characteristic points and whose end points are the rearward characteristic point groups obtained under certain conditions, the segment belonging to the first segment group and the second segment group The segment-to-segment similarity is calculated for all combinations of pairs with the segment belonging to, and the segment pair having the maximum similarity is determined, and the feature point pairs corresponding to both end points of the segment pair are then associated with the corresponding feature point pair. Stroke distance calculation method characterized by determining as 2. In the associating by the second means, a long segment of two segments whose fixed feature point is a starting point and whose adjacent feature point is a starting point is fixed. , For the other segment, the similarity between segments is calculated by limiting the sum of the distances from the feature point that is the starting point to the feature point that exceeds the fixed multiple of the length of the fixed segment for the first time The inter-stroke distance calculation method according to claim 1.