JPH01321591A - Character recognizing device - Google Patents

Abstract

PURPOSE:To simply learn a new category by producing a mapping matrix and recognizing dictionary for extraction of features of a discriminating method and a feature vector of each category via the K-L evolution of a single time. CONSTITUTION:An average initial vector of the learning character categories is produced from an initial vector where the feature values detected out of the pictures of learning categories. An inter-category covariance matrix is obtained from said average initial vector and the average initial vector of the categories registered before learning. Then an inter-normalized category covariance matrix is obtained from the product obtained between a normalized mapping matrix produced previously from an intra category covariance matrix and the inter-category covariance matrix. Then the proper vectors obtained by applying the K-L evolution to the inter-normalized category covariance matrix are arranged for production of an inter-category emphasis matrix which emphasizes the difference of categories. Then a matrix obtained from the product of the normalized mapping matrix and the inter-category emphasis matrix is used as a mapping matrix for decision/analysis. While a vector obtained from the product of the average initial vector of each category and the mapping matrix is used as a feature vector of a dictionary. Thus the learning is possible in a simple process.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention provides a character & IN! related to the identification device.

(Prior art) A method of character recognition using discriminant analysis (referred to as discriminant analysis method in the following explanation) is described in Nobuyuki Otsu [Mathematical study on feature extraction in Bame 7 recognition], Denshi General Research Institute Report No. 818, July 1988, pages 188 to 191 (Reference 1), Kamiike [Hierarchical discriminant method and its character recognition system PC-OCRJ, IEICE Pattern Recognition and Understanding Study Group materials PRU86-76 (Reference 2). In the above literature, a mapping matrix used for feature extraction and a feature vector serving as a recognition dictionary are obtained in advance and used for character recognition.

According to reference document 1, the above-mentioned mapping matrix and feature vector can be obtained using a flowchart as illustrated in FIG. 2. First, an inter-category covariance matrix and an intra-category covariance matrix are determined based on a vector (initial vector) in which characteristic sounds detected from an input image are arranged (201). Next, a normalization mapping matrix for normalizing the intra-category covariance matrix is obtained from the eigenvalues and eigenvectors obtained by performing K-L expansion of the intra-category covariance matrix (202). Furthermore, the normalized inter-category covariance matrix is obtained by multiplying the normalized mapping matrix and the inter-category covariance matrix (203), and the normalized inter-category covariance matrix is −L-expanded and the resulting eigenvectors are arranged to calculate the inter-category covariance matrix. Create an emphasis matrix between categories that emphasizes differences (2
04). The matrix obtained by multiplying the normalized mapping matrix and the inter-category emphasis matrix becomes the mapping matrix for discriminant analysis (205
). Further, a feature vector that becomes a recognition dictionary in the discriminant analysis method is created by averaging each element of the vector found by multiplying the initial vector and the mapping matrix for each category (206 and 207). On the other hand, there is a principal component analysis method as a statistical-based method similar to the discriminant analysis method.
No. (Reference 3). In Reference 3, standard patterns can be learned simply by -L expansion of the covariance matrix for each category.

(Problem to be solved by the invention) In order to add, that is, learn standard patterns for new character categories using the discriminant analysis method, as can be easily understood from the explanation of Figure 2, the inter-category covariance matrix and the within-category covariance matrix It is necessary to recalculate the matrix and perform -L expansion twice and matrix multiplication five times. Therefore, compared to the above-mentioned principal component analysis method, the discriminant analysis method has a problem in that it is difficult to learn new categories.

(Means for solving the problem) In order to solve the above-mentioned problem, the device provided by the present invention transforms an initial vector, which is a vector in which characteristic sounds detected from an input image are arranged, using a mapping matrix of discriminant analysis. The character recognition device performs recognition processing by comparing the obtained feature vector with the feature vectors of registered categories, and also inputs an image of a specific category to create and register feature vectors of all categories including the specific category. and an averaging means that averages the initial vectors for each character category.

average initial vector storage means for storing an average initial vector which is an output of the averaging means; normalization mapping storage means for storing a mapping matrix for normalizing the intra-category covariance matrix of the initial vector; and the average initial vector storage means. An inter-category covariance matrix is obtained from the average initial vector of the means, and the inter-category covariance matrix is expanded into −L, and the product of the matrix whose elements are the obtained eigenvectors and the mapping matrix of the normalized mapping storage means is subjected to feature transformation. It is characterized by comprising learning processing means for outputting the product of the average initial vector of each character category from the average initial vector storage means and the feature conversion mapping as a feature vector of each character category at the same time as outputting it as a mapping. .

(Operation) FIG. 3 is a flowchart for obtaining a mapping matrix used for feature extraction and a feature vector serving as a recognition dictionary in the present invention.

From the vector (initial vector) that arranges the characteristic sounds detected from the image of learning category IJ i, learning character category IJ
An average initial vector of i is created (301).

An inter-category covariance matrix is obtained from the average initial vector of the categories registered before learning and the average initial vector of the learning category (302). Next, a normalized inter-category covariance matrix is obtained by multiplying the normalized mapping matrix created in advance from the intra-category covariance matrix of the registered categories and the inter-category covariance matrix (303), and the normalized inter-category covariance matrix is -L expansion and arrange the resulting eigenvectors to create an inter-category emphasis matrix that emphasizes the differences between categories (3
04).

A matrix obtained by multiplying the normalized mapping matrix and the inter-category emphasis matrix is used as a mapping matrix for discriminant analysis (305).

Further, the feature vector that becomes a recognition dictionary in the discriminant analysis method is a vector found by multiplying the average initial vector for each category and the mapping matrix (306). Compared to FIG. 2 described above, -L expansion is performed only once, and the feature vector of the dictionary is obtained from the average initial vector, making the processing simpler. Therefore, learning can be performed with the same level of processing as the principal component analysis method.

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

FIG. 1 is a block diagram of a character recognition device with a learning function, which is an embodiment of the present invention. In the figure, 1 is an image input means, 2 is a character cutting means, 3 is a feature detection processing storage means,
4 is a feature detection means, 5 is an averaging means, 6 is an average initial vector storage means, 7 is a normalization mapping storage means, 8 is a learning processing means, 9 is a feature conversion mapping storage means, 10 is a dictionary storage means, and 11 is a 12 is a recognition processing means; and 13 is a recognition result display means.

Learning of a new category using the character recognition device shown in FIG. 1 is performed as follows.

First, an image containing characters of the learning target category is obtained from the image input means IK, and the character cutting means 2 determines the position of each character image of the learning target category from the image output from the image input means 1, and sets the character image at that position. Output a certain image sequentially.

Since the feature detection processing storage means 3 stores the processing for obtaining character features, the feature detection means 4 next extracts the characters from the character cutting means 2 according to the processing output from the feature detection processing storage means 3. Find the feature amount from the character image.

Here, a sequence of feature quantities obtained for each of a plurality of features is called an initial vector.

The averaging means 5 calculates the average value of the feature amount for each feature from the initial vector output for each character image, and outputs a set of average values corresponding to the set of features as an average initial vector. The average initial vector of the learning category output from the averaging means 5 is stored in the average initial vector storage means 6. Since the normalization mapping storage means 7 stores a mapping matrix for normalizing the intra-category covariance matrix created from the initial vectors of existing categories, the learning processing means 8 stores the mapping matrix according to the procedure shown in FIG. and a feature vector serving as a dictionary for each category. The mapping matrix output from the learning processing means 8 is stored in the feature transformation mapping storage means 9, and the other feature vector is stored in the dictionary storage means 10. Learning ends by updating the mapping matrix and feature vector. Although the above explanation deals with learning of one category, learning of a plurality of categories can be performed by operating the learning processing means 8 after the average initial vector of the plurality of categories is determined.

Further, character recognition in the character recognition device shown in FIG. 1 is performed in the following order.

The image input means 1 obtains an image containing the character to be recognized, and the character cutting means 2 determines the position of the character image to be recognized one by one from the image output from the image input means 1, and extracts the character image at that position. Output sequentially. The following processing is performed for each character image. As before, the feature detection means 4 obtains a feature amount from the character image from the character extraction means 2 according to the process output from the feature detection processing storage means 3, and the feature extraction means 11 uses the obtained feature amount as a feature. The feature vector obtained by calculating the product of the feature vector and the mapping matrix output from the feature transformation mapping storage means 9 is output. The recognition processing means 12 matches the feature vectors from the feature extraction means 11 with the feature vectors for each category stored in the dictionary storage means 10, and selects category names in descending order of distance value (or similarity value) as a result of the matching. Outputs the sign and value of. The recognition result display means 13 stores the code and value of the category name output from the recognition processing means 12, and displays an image corresponding to the code of the category name.

(Effect of the invention) As explained above, according to the present invention, - times - L
Since the mapping matrix for feature extraction using the discriminant analysis method and the feature vector for each category, which serves as a recognition dictionary, can be created simply by expansion, new categories can be easily learned.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the character recognition device with a learning function of the present invention, FIG. 2 is a diagram showing an example of the processing flow when creating a recognition dictionary in the discriminant analysis method, and FIG.
The figure is a diagram showing an example of the flow of processing for creating a recognition dictionary in the character recognition device of the present invention. In the figure, 1 is an image input means, 2 is a character cutting means, and 3
4 is a feature detection processing storage means, 4 is a feature detection means, 5 is an averaging means, 6 is an average initial vector storage means, 7 is a normalization mapping storage means, 8 is a learning processing means, 9 is a feature transformation mapping storage means, 10 11 is a dictionary storage means, 11 is a feature extraction means, and 12 is a dictionary storage means.
13 is a recognition processing means, and 13 is a recognition result display means. Agent Patent Attorney Shinsuke Honjo Figure 1 Figure 2 Figure 3

Claims (1)

[Claims] Recognition processing is performed by comparing the feature vector obtained by converting the initial vector, which is a vector of feature quantities detected from the input image, with a mapping matrix of discriminant analysis, and the feature vector of the registered category. In a character recognition device that inputs an image of a specific category and creates and registers feature vectors of all categories including the specific category, the device further comprises: averaging means for averaging initial vectors for each character category; and the averaging means. average initial vector storage means for storing an average initial vector that is an output of the average initial vector; normalization mapping storage means for storing a mapping matrix that normalizes the intra-category covariance matrix of the initial vector; An inter-category covariance matrix is obtained from the vector, and the product of the matrix whose elements are the eigenvectors obtained by performing K-L expansion of the inter-category covariance matrix and the mapping matrix of the normalized mapping storage means is output as a feature transformation mapping. At the same time, the character recognition device further comprises learning processing means for outputting the product of the average initial vector of each character category from the average initial vector storage means and the feature conversion mapping as the feature vector of each character category.