JP3064192B2 - Feature generator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a feature value generating device used for a destination recognizing device of a package sorting device. The present invention relates to various sorting devices such as a mail address sorting device and a mail number sorting device, an OCR (optical character reading device), and a product and material management device (ID reading unit) in FA and the like. It can be used with any character or symbol recognition device.

[0002]

2. Description of the Related Art In the prior art, various feature values have been used as feature values for character and symbol image pattern recognition. However, these are roughly classified into (A) pixel values as they are and (B) ) Contour-based feature, (C) skeleton-based feature, and (D) ad-hoc feature. Table 1 shows these descriptions.

[0003]

[Table 1]

[0004] In addition, other new features include:
There is an example in which a study of pattern recognition using a frequency-based feature of the same classification as the present invention has been published, but in the conventional example, a high-dimensional coefficient obtained as a result of wavelet transform or Gabor transform is directly used as a feature. It was used.

[0005]

The conventional pattern recognition based on feature values has problems as shown in Table 2, and is not sufficient for the recognition of complex characters (for example, handwritten kanji) and complex symbols.

[0006]

[Table 2] An object of the present invention is to provide a device that can solve these problems.

[0007]

SUMMARY OF THE INVENTION A feature amount generating apparatus according to the present invention generates a feature amount for identifying a class to which a character or a symbol belongs from image data such as a character or a symbol. At the multidimensional wavelet transformation
Conversion device, wavelet transform coefficient dimension reduction device,
Normalize the size of the training teacher image and outline the character line
Teacher image preprocessing means for preparing
Each of the images processed by the stages is referred to as the multidimensional wave.
Wavelet transform by a wavelet transform unit and transform coefficients
And the transform coefficient is reduced by the wavelet transform dimension.
Depending on the device, it can be divided into three directions: vertical, horizontal, and scale.
The feature amount of the learning image whose dimension is reduced by taking the average of the partial weights
A learning feature amount generating means for obtaining a vector;
Using the feature vector obtained by the collection generator,
A dictionary that generates dictionary data that serves as an identification condition for another
And normalizing the size of the image data to be identified
Identification image pre-processing means for shaping the outline of the character line together with
The target image processed by the identification image pre-processing means is
The wave is transformed by the multi-dimensional wavelet transform device.
A transform coefficient is obtained by performing a wavelet transform, and the transform coefficient is
The vertical, horizontal, and vertical
Calculates partial weight average in three directions of Kale direction, and reduces dimension
Feature generation for obtaining the feature vector of the identified identification image
Means for generating the characteristic amount for identification.
Classification of feature vector using the dictionary data
Characterized in that it.

[0008]

[Function] By using a multidimensional wavelet transform coefficient as a feature value, a feature value that is more resistant to deformation and line thickness change than a pixel itself or a contour-based feature value is more versatile and practical than an ad hoc feature value. It can be obtained without failure like the skeleton feature.

[0009] By reducing the multidimensional wavelet transform coefficients, the restrictions on the identification method due to the feature dimension being too high are reduced, and an identification method effective for complicated identification such as Mahalanobis distance, Bayesian identification, and neural network is provided. Can be applied.

[0010]

1 to 3 show a first embodiment of the present apparatus. FIG. 1 is a diagram showing a first embodiment of the present invention, FIG. 2 is a diagram showing a handwritten kanji recognition device using the device of the present invention, and FIG. 3 is a diagram showing a dimension reduction method of a wavelet transform coefficient dimension reduction device of the device of FIG. It is a figure showing an example. I. Creation of Identification Dictionary Data by Learning Prior to character recognition, identification dictionary data is created according to the following (A) to (F).

The identification dictionary data need only be generated once at the beginning, and the created identification dictionary data is used to generate the identification dictionary data even if it is difficult. Can be recognized. (A) A sufficient number of kanji images are prepared as learning teacher images. The sufficient number means at least a plurality of numbers for each class (character type), and the greater the number, the higher the pattern recognition ability. (B) In the character size normalizing device 5 of the preprocessing device 4,
The vertical and horizontal sizes of the circumscribed rectangle of the image prepared in (A)
Normalize to a predetermined size that is determined in advance (make the size of the circumscribed rectangles of all character images uniform). Here, the normalization is performed by linearly enlarging / reducing the image. However, the normalization method may be various nonlinear normalizations, and is not limited to linear enlargement / reduction. (C) In the smoothing device 6 of the pre-processing device 4, the outer shape of the character line that has been distorted by the enlargement operation or the like in the character size normalizing device 5 is adjusted by a smoothing operation such as a media filter. (D) In the multidimensional wavelet transform device 2 of the feature amount generating device 1, the preprocessing device 4 performs a two-dimensional wavelet transform on each of the preprocessed images, and obtains the transform result (wavelet transform coefficient). (E) The wavelet transform coefficient dimension reduction device 3 of the feature amount generation device 1 converts the coefficients obtained by the multidimensional wavelet transform device 2 into a vertical direction (y direction), a horizontal direction (x direction),
By blurring in three directions (taking a partial weighted average) in the scale direction, a feature amount vector with reduced dimensions is obtained. FIG. 4 shows an example of a dimension reduction due to blurring. In the example of FIG. 4, specific numerical values (the size and weight of the blur filter, the jump width of the subscript of the feature amount, etc.) are merely examples, and are not limited to these numerical values. (F) The dictionary creation device 7 uses the feature amount vector of the learning image obtained by the feature amount generation device 1 to generate dictionary data serving as an identification condition for class (type) identification.

For example, when the pseudo Mahalanobis distance is used as an identification method, when the feature amount vector of the learning image is
(A) average vector for each character class (type), (b)
The eigenvalues and eigenvectors of the covariance matrix with respect to the average vector for each character class (type) are obtained for all classes (types) and used as dictionary data.

However, the dictionary data differs depending on the identification method. For example, when class (type) identification is performed by a neural network, the weighting factor of the neural network is dictionary data. II. Character identification I. After the identification condition data is created, the following (G) to
Characters are identified by (K). (G) The character size normalizing device 5 of the preprocessing device 4
Normalize the character image to be identified. The normalization method and the normalization size are exactly the same as I (B). (H) The smoothing device 6 of the preprocessing device 4 performs smoothing of the normalized image. The smoothing method is exactly the same as I (C). (I) The multi-dimensional wavelet transform device 2 of the feature amount generating device 1 generates wavelet transform coefficients from the pre-processed image. The wavelet transform method (base, range of target scale, quantization method, etc.) is exactly the same as I (D). (J) The feature amount in which the two-dimensional wavelet transform coefficient is reduced in dimension by the wavelet transform coefficient dimension reducing device 3 of the feature amount generating device 1 is obtained. The lower dimensional method is I (E)
And exactly the same. (K) The identification device 8 identifies to which class (kind) the feature vector determined by II (G) to (J) belongs using the identification dictionary data determined by I.

For example, in the case of performing identification based on pseudo Mahalanobis distance, (a) using an average vector for each character class (type) obtained as identification dictionary data, and an eigenvalue and an eigenvector of a covariance matrix for the average vector, A pseudo Mahalanobis distance of a feature vector (features obtained by II (G) to (J)) for the class is obtained.

(B) By comparing the pseudo Mahalanobis distances for all the classes, it is recognized that they belong to the class (kind) having the smallest value. Note that the identification method given in the example is merely an example, and the present invention is not limited to this. III. Wavelet Transform (Wavelet Transform) An outline of the wavelet transform is shown in FIGS.

As shown in FIG. 4B, the shape of the wavelet transform base changes depending on the value of a (this a is called a scale). FIG. 4 (C) shows si of three types of cycles.
An example in which an n signal and two triggers overlap is shown.

By using the wavelet transform,
The following features can be exhibited. (1) Frequency analysis in the time / frequency space can be performed.

The position (time) of a unique signal can be detected. Time change of the frequency distribution can be grasped. (2) Multi-resolution analysis (MRA)
lysis).

Various signal components can be extracted with a time accuracy corresponding to each frequency. (Higher frequency components can be extracted with finer time accuracy.) Information in a wide frequency range can be handled at once. (3) Various basis functions can be used. (4) Inverse transformation is possible.

The speed can be easily increased by using the orthogonal wavelet. IV. Example of Generation of Low-Dimensional Feature Amount by Blur FIG. 4 shows an example of a reduction in dimension due to blur according to the first embodiment.

In FIG. 4, s is the scale, x is the horizontal length of the preprocessed image, y is the vertical length of the preprocessed image, and f (s, x, y) is the normalized value. Unfiltered blur filter function, f (-1, x, y) has a scale of 1
The weight applied to the small wavelet transform coefficients, f (0,
x, y) is the weight applied to the wavelet transform coefficient of the scale, f (1, x, y) is the weight applied to the wavelet transform coefficient having the larger scale by 1, i is the blur filter range in the x direction (horizontal direction of the image), j indicates the blur filter range in the y direction (vertical direction of the image), k indicates the blur filter range in the scale direction, and S indicates the number of types of scale.

The number of wavelet transform coefficients of one image at a certain scale differs depending on the value of the scale. Since the shift interval of the wavelet basis differs depending on the scale, the number of shifts in the x direction at a certain shift S is represented by X (S) and the number of shifts in the y direction is represented by Y (S).

[0023]

Since the apparatus of the present invention is configured as described above, the following effects are obtained. (1) By using a multidimensional wavelet transform coefficient as a feature quantity, a feature quantity resistant to deformation and a change in line thickness can be universally and reliably obtained, and a kanji character which is a very complicated character Can also be recognized. (2) By reducing the dimensions of the multidimensional wavelet transform coefficients, it is possible to apply an advanced identification method such as a pseudo Mahalanobis distance, and to improve the pattern recognition ability.

[Brief description of the drawings]

FIG. 1 is a diagram showing a first embodiment of the present invention.

FIG. 2 is an overall configuration diagram of the first embodiment (a diagram showing a handwritten kanji recognition device using the device of FIG. 1).

FIG. 3 is a diagram showing an outline of a wavelet transform.

4 is a diagram illustrating an example of a dimension reduction method of a wavelet transform coefficient dimension reduction device of the feature value generation device of FIG. 1 (an example of generation of a low-dimensional feature value by blurring).

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Feature generation apparatus, 2 ... Multidimensional wavelet transform apparatus, 3 ... Wavelet transform coefficient dimension reduction apparatus, 4 ... Preprocessing apparatus, 5 ... Character size normalization apparatus, 6 ... Smoothing apparatus, 7 ... Dictionary creation apparatus, 8 ... identification device 9 ... pattern recognition device

Continuation of the front page (56) References JP-A-6-251200 (JP, A) JP-A-6-28523 (JP, A) JP-A-59-177684 (JP, A) "Vol. 93 No. 228 (PRU93-53) p. 81-88 "Machine Vision and Applications", Vol. 6 No. 2/3 p. 110-123 (1993) (58) Field surveyed (Int. Cl. ⁷ , DB name) G06K 9/46 G06K 9/62 JICST file (JOIS)

Claims (1)

(57) [Claims] From the image data, such as 1. A character or symbol, the feature amount generating unit for generating a feature amount for identifying the class belongs, such as the character or symbol, and multi-dimensional wavelet transformation device, lowering the wavelet transform coefficients dimension outside the character line as well as normalization and apparatus, the size of the learning tutor image
A teacher image preprocessing means for shaping, and an image processed by the teacher image preprocessing means.
Wavelet by the multi-dimensional wavelet transformer.
To obtain a conversion coefficient, and the conversion coefficient is
Vertical, horizontal and scale
Average the partial weights in three directions,
Feature generation for obtaining feature vector of learned learning image
Means and a feature amount vector obtained by the learning feature amount generation means.
And generate dictionary data as identification conditions for class identification.
Dictionary creation means, which normalizes the size of the image data to be identified and
Identification image preprocessing means for adjusting the outline of the line; and a target image processed by the identification image preprocessing means.
Each of the multidimensional wavelet transformers has
The transform coefficient is obtained by bullet transform, and the transform coefficient
Vertical, horizontal,
Average partial weights in three directions of the scale direction
The discriminating feature that obtains the discriminating image feature vector
Generation means, and the identification feature quantity generation means
The class of the feature quantity vector is identified using the dictionary data.
A feature amount generation device characterized by differentiating .