JPH03144788A - Pattern normalizing device - Google Patents

Abstract

PURPOSE:To allow the pattern normalizing device to correspond to various deformation in an input pattern by repeating geometric conversion by means of an optimized calculation method. CONSTITUTION:A control part 14 executes repeat processing based upon the optimized calculation method so that a distance value 105 between respective feature vectors obtained from an input pattern and a reference pattern is minimized and geometric conversion is repeated so that the shape of the input pattern is more approximated with the reference pattern. The input pattern is deformed by all the reference patterns and several patterns are successively selected in order from the one having the minimum distance value as upper candidates. On the other hand, a geometric conversion part 10 executes deformation so that the sum of respective picture element value in the predeformed pattern and the deformed pattern is stored. Thus, the pattern normalizing device can be allowed to correspond to various deformation of the input pattern.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to an apparatus for normalizing shapes in pattern recognition.

(Prior Art) Character recognition, which is one of the important fields in pattern recognition, is generally divided into four processes: preprocessing, feature extraction, identification, and postprocessing. Preprocessing removes noise and normalizes the shape of the input character image, and feature extraction uses image processing methods such as mask processing, contour tracing, and projection to extract feature points such as intersections and endpoints. Categories are identified by matching feature vectors composed of zero-order feature quantities as components to find feature quantities that depend on stroke arrangement, density distribution, etc., and feature vectors corresponding to each category.Finally, as post-processing, Check the identification results.

One of the factors contributing to the difficulty in character recognition is the diversity of shapes of character patterns belonging to the same category.

This problem is particularly noticeable in handwritten characters and multi-font printed characters, and several methods have already been proposed to solve this problem.

Most of these methods set a certain normalization standard for the shape, and then transform it in the preprocessing process so that optimization can be performed.
, Nagy and N. Tseng, “Normalization Technique for Handwritten Digits”
for Handprinted Numerals
) ” entitled ^5 association for
CoII of Computing Machinery
communication, Vol. 13. No, 8 (1
In the paper published on pages 475 to 481 of 970), a circumscribed rectangle is found for handwritten digits,
The shape is normalized by deforming it so that it has a constant shape. In addition, Tsugumo and Tanaka published an article entitled "Handwritten Kanji Recognition Based on Hierarchical Positional Displacement Correction Processing" at a research meeting of the Institute of Electronics, Information and Communication Engineers (material number PRU87-104.198).
In a paper published in February 2007, a method was proposed for correcting the stroke positions of handwritten Chinese characters so that the stroke intervals would be equalized.

(Problem U to be Solved by the Invention) The method described above is efficient in terms of processing amount because the optimal transformation for a certain normalization standard is determined by a certain process. However, the selection of normalization criteria is a very difficult problem, which depends not only on the feature extraction and identification algorithms but also on the recognition target. It is generally difficult to obtain an optimal deformation through constant processing for a set normalization standard, and in that case, it is considered that iterative convergence calculations are necessary.

An object of the present invention is to provide a more versatile pattern normalization device by repeating geometric transformation using an optimization calculation method.

(Means for Solving the Problems) In pattern recognition processing, the device of the present invention performs geometrical analysis on an input pattern so that the distance between feature vectors obtained by extracting features from an input pattern and a reference pattern is minimized. A device for normalizing the shape of an input pattern by performing a logical transformation, the device comprising: a multi-value processing unit that converts an input binary pattern into a shading pattern; A geometric transformation unit that performs mechanical analysis to obtain a deformation pattern; a feature extraction unit that performs feature extraction on the deformation pattern and obtains a feature vector of the deformation pattern; and a feature extraction unit that performs feature extraction on each reference pattern in advance. a dictionary for storing feature vectors of the reference pattern obtained by the above-described process; a distance calculation unit for calculating the distance between the feature vector of the deformed pattern and the feature vector of the reference pattern; a control unit that calculates the value of the deformation parameter and controls the iterative process until it is optimized; and a control unit that stores and sorts the distance values minimized for each reference pattern, and Features (Functions) The pattern normalization device of the present invention is characterized in that it is configured to include a sorter that sequentially outputs top candidates as top candidates. In this method, iterative processing is performed using an optimization calculation method so that the input pattern is minimized, and geometric transformation is repeated so that the shape of the input pattern becomes closer to the reference pattern. The input pattern is transformed with respect to all reference patterns, and some of the resulting patterns with the smallest distance values are selected as top candidates.

Further, in the geometrical transformation section, transformation is performed so that the sum of each pixel value is preserved in the pattern before transformation and the pattern after transformation.

(Example) Embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of the pattern normalization device of the present invention. The present invention will be described in detail below with reference to FIG.

The multivalue processing unit 16 receives an input pattern 109 which is a binary image.
A convolution filter process such as Gaussian is performed on the image to obtain a shading pattern 100. An example of the coefficients of a 5×5 Gaussian filter is shown below.

The geometrical transformation unit 10 receives two inputs, a shading pattern 100 and a deformation parameter 101, performs a geometric transformation given by the deformation parameter 101 on the shading pattern 100, and outputs the result as a deformation pattern 102.

Next, the method of geometric transformation will be explained.

First, consider the case of a one-dimensional pattern for simplicity. The address of the pattern 100 before transformation is expressed as x (0≦X≦M), and the address of the pattern 102 after transformation is expressed as f (x), where the open number f (x) satisfies the following conditions.

f (0) = O (1) f (M) = M (2)
f (x) ≧O (0≦X≦M) (
3) Due to these three conditions, the pattern maintains the same size before and after deformation. In particular, according to equation (3), the fractional number f increases monotonically, so the pattern can be deformed while maintaining the order of each component of the pattern. It can be carried out.

For example, f (x), f(X)"as X'"atX'"at X+a
When using a cubic function expressed as o (4),
In addition to X, as parameters a s r & 2* a
4fll of I+ao is required. However, the formula (1
) and (2), the following equation (5) is obtained, so only two parameters are actually required. This parameter corresponds to the deformation parameter 101.

f(X)'a, X'"32X" + (i-a
s H2-ax H) x (5) If a nonlinear function is used as f (x), deformation such as expanding or contracting only a certain part of the pattern is possible.

Next, the address before transformation of the two-dimensional pattern is (x.

y), the address after transformation is expressed as (X, Y), and the two-dimensional pattern is created according to the following equations (6) and (7) using the open numbers f and g that satisfy the above conditions (1) to (3). When a geometric transformation is performed, an independent transformation is performed in each direction.

X=f (x) (6)Y
=g (y) (7) On the other hand, if we add terms for address values in other directions, as in the following equations (8) and (9), we can further transform a rectangle into a parallelogram and rotate it. It becomes possible.

X=F(x,y)=f(x)+by (8)Y=
G(x,y)=g(y)+c−x (9) In this case, the number of deformation parameters 101 is six in total, three each in the X direction and the y direction.

In the above transformation, the numerical numbers f (0), f (M)
Since the value of is constant, as shown in Figure 2(a), 0
The components at both ends of the rectangular area represented by ≦X≦M and 0≦y≦N cannot be moved, but as shown in Figure 2 (b), the background part is added outside the circumscribed square of the pattern. If this is used as the normalization processing area, components at both ends of the pattern can also be moved.

Furthermore, local deformation can be realized by using different numbers of intervals depending on the region.

Next, we will explain how to calculate pixel values.The value stored at the address (x, y) of the pattern before transformation is
(x, y), the value stored at the address (x, y) of the pattern after transformation is dat(x.

y), the pixel value is determined according to Tsuginoshiki (10).

Y-01x, yl In other words, if there are multiple (x, y) that are associated with (X, Y) by the functions F and G, in other words, the function F
Or if G has many-to-one correspondence, the corresponding src (
x, y) and stores it in dat(X, Y).

Next, the meaning of this modification will be explained using FIG. 3.

FIG. 3 shows a modified example of a one-dimensional pattern, where (a) shows the pattern before modification, and (b) shows the reference pattern.

If contraction deformation is applied to (a) from both sides, it is possible to bring the two peaks closer together, and if this deformation is performed excessively, it is expected that the positions of the two peaks will coincide. However, if the transformation is performed according to equation (10), the height of the peak obtained when the positions match will be doubled as shown in (C), so (
It can be distinguished from pattern b). In this variant,
The sum of pattern pixel values is saved.

Next, the present invention will be explained in detail using FIG. 1 again.

The feature extraction unit 11 performs feature extraction processing on the deformed pattern 102 and outputs the obtained feature vector as 103. As an example of feature extraction processing, convolution processing with a spatial differential filter as shown below is performed on a constant sampling point, and the obtained value is converted into a feature vector 103.
Output as an element.

The dictionary 12 stores feature vectors obtained by performing the processing described in the feature extraction unit 11 on all reference patterns in advance, and receives the read request signal 1 from the control unit 14.
This is read out 104 in synchronization with 07.

The distance calculation unit 13 calculates a feature vector 103 obtained from the deformed pattern and a feature vector 10 obtained from the reference pattern.
4 is input, the distance value between the two is calculated and output as 105. As a distance, for example, L 1 norm, (,=Σlal b, 1 (11) or 2 norm, CI"Σ(a+b,)" (12)
etc. can be used.

The control unit 14 inputs the deformation parameter 101 and the distance value 105, and calculates the value of the deformation parameter 101 that minimizes the distance value 105 through an iterative process using the i″iM calculation manual calculation method.Once determined, Outputting the convergence value and the corresponding reference pattern category number as a signal 106;
Furthermore, a read request signal 107 is output to the dictionary 12, and the feature vector of the next reference pattern is read.

Next, we will explain the steepest gradient method, which is one of the a optimization calculation methods.The deformation parameters 101 input to the geometric transformation unit 10 are expressed as vectors K = (k + kz , kn ,
..., kL), and when K is given, the feature extraction unit 1
If the distance value 105 obtained in 1 is D(K),
It will be as shown below.

■Give an initial value to.

■i=1.2.3. ..., for g+ = D (
Find Kt '') D (Kt -> (13). However, in the vector Kl+, - is a vector in which only the first component of K is changed by Δ(〉O) in the positive direction and the negative direction, respectively. , and can be expressed as in the following equation (11).

Kt ” = (kr ” l k+ +
・kr(j+i).

k+"= + + Δ), and D' = D (K-α・G/IIGI+> (
Find a positive scalar value α that minimizes 15).

■If α becomes small enough and the value of K converges, the process ends; otherwise, K- becomes −α・G/llG11 (16)
As ■.

The control unit 14, in which the vector G represents the steepest gradient direction and the scalar α represents the update amount, uses the deformation parameter 101 and the distance value 105 to repeat the above-mentioned steps 1 to 2 until convergence.

In this method, all components of the vector are updated in step (3), so that horizontal and vertical transformations are performed simultaneously. However, in ■, the gradient is obtained by varying the parameters using the variational method without deforming in other directions.

In general, the optimal combination of deformations for each direction is for both directions! Therefore, the control unit 14 divides the iterative processing of ■ to ■ into two phases in the X direction and the y direction, updates parameters related to only one direction of deformation in each phase, and Execute the phases one time at a time
Repeated times.

The sorter 15 is configured to include a memory that stores the signal 106 output from the control section 14, a comparator, and the like. Among the data of the signal 106, the distance convergence values are sorted on the memory, and the upper candidate data is outputted as the signal 108 together with the corresponding category number.

Although the description has been made here as a device that minimizes the distance between two feature vectors, it is also possible to use an optimization criterion that maximizes the degree of similarity rather than the distance, for example.

In this case, if the degree of similarity is expressed as D(K), then the control unit 1
The optimization iterative process in step 4 is as follows.

■Give an initial value to.

■i=1.2.3. ..., for L, g+ = D
(Kl ”)-D (Kl-)
Find (13).

■Vector G= (g+ , g21gs ,...gL
), and D' = D (K-α・G/IIGI+> (
Find a positive scalar value α that maximizes 15).

■If α becomes small enough and the value of K converges, the process ends; otherwise, K- is converted to −α・G/IIGII (te>
As ■.

Furthermore, by sorting the convergence values in order from the maximum value in the sorter 15, a transformation that maximizes the similarity is realized.

(Effects of the Invention) As described above, by using the pattern normalization device of the present invention, it is possible to deal with more diverse deformations of input patterns than in the past. Moreover, by setting the optimization criteria not only based on minimizing the distance between feature vectors but also according to the recognition target, highly versatile normalization processing can be realized.

In addition, the geometric transformation unit performs transformation so that the sum of pixel values is preserved. As a result, for example, the input pattern is the kanji numeral ``2'', and the reference pattern is the kanji numeral ``2''.
When using "-", even if the positions of both strokes of "2" are matched by contraction deformation in the vertical direction, the height of the obtained stroke will be doubled, so the distance between the feature vectors will become large, and "-" The present invention has such excellent effects.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the pattern normalization device of the present invention, FIG. 2 is a diagram showing an example of geometric transformation within the normalization area, and FIG. It is a figure which shows the example of calculation of a pixel value. 10...Geometric conversion unit, 11...Feature extraction unit, 12
... Dictionary, 13... Distance calculation section, 14... Control section, 15... Sorter, 16... Multivalue processing section. Figure 1 - Movement/ (b) (a) Pattern before transformation (C) Pattern after transformation (b) Comparison pattern

Claims (2)

[Claims] (1) In pattern recognition processing, geometric transformation is performed on the input pattern so that the distance between the feature vectors obtained by extracting features from the input pattern and the reference pattern is minimized, and the shape of the input pattern is A normalization device that includes a multi-value processing unit that converts an input binary pattern into a shading pattern, and a geometry unit that executes mechanics on the shading pattern to obtain a deformed pattern according to given deformation parameters. a feature extraction unit that performs feature extraction on the deformed pattern to obtain a feature vector of the deformed pattern; a distance calculation unit that calculates a distance between the feature vector of the deformation pattern and the feature vector of the reference pattern; a distance calculation unit that calculates the value of the deformation parameter that minimizes the value of the distance, and a sorter that stores the distance values minimized for each of the reference patterns, performs sorting, and outputs the distance values as top candidates in order from the one with the smallest value. A pattern normalization device comprising: (2) In the pattern normalization device according to claim 1, the geometric transformation unit may
If a function representing the post-transformation address is determined by the transformation parameter, and there are a plurality of pre-transformation addresses corresponding to the post-transformation address, the sum of the values stored in each of the pre-transformation addresses is calculated, and the post-transformation address is A pattern normalization device characterized in that a pattern normalization device stores data in a pattern normalization device.