Method for character recognition based on gabor filters

• US 20040017944A1
• Filed: 05/23/2003
• Published: 01/29/2004
• Est. Priority Date: 05/24/2002
• Status: Active Grant

• Current Assignee: Tsinghua University
• Original Assignee: Tsinghua University
• Inventors: Ding, Xiaoging, Liu, Changsong, Peng, Liangrui, Fang, Chi, Wang, Xuewen
• Granted Patent: US 7,174,044 B2
• Time in Patent Office: Days
• US Class Current: 382/182
• CPC Class Codes: G06V 30/10 Character recognition
• CPC Class Codes: G06V 30/18019 by matching or filtering

Abstract

In a method for character recognition based on Gabor filter group the Gabor filter'"'"'s joint spatial/spatial-frequency localization and capability to efficiently extract characters'"'"' local structural features are employed to extract, from the character image, information of the stroke direction of characters as the recognition information of characters, so as to improve the capability to resist the noises, backgrounds, brightness variances in images and the deformation of characters. Using this information, a simple and effective parameter design method is put forward to. optimally design the Gabor filter, ensuring a preferable recognition performance; a corrected Sigmoid function is used to non-linearly adaptively process the stroke direction information output from the Gabor filter group. When extracting the feature from blocks, Gaussian filter array is used to process the positive and negative values output from Gabor filter group to. enhance the discrimination ability of the extracted features.

Claims

1. A method for character recognition based on a Gabor filter group, said method comprising:

(a) pre-processing a character image, said pre-processing including receiving, pre-processing the character image of a character to be recognized, and obtaining a binary or gray image with N×

N pixels for each character to be recognized, wherein said binary or gray image for each character to be recognized is represented by a matrix [A(i, j)]<sub>N×

N</sub>;

(b) processing the matrix [A(i, j)]<sub>N×

N</sub> for the character image of each character to be recognized obtained in the step (a), said processing including;

extracting stroke direction information of the character to be recognized, said extracting including employing the Gabor filter group which is composed of K two-dimension Gabor filters to extract stroke information in K different directions from the matrix [A(i, j)]<sub>N×

N</sub> of the character image of the character to be recognized, and obtaining K matrixes [G_m(i, j)]M×

M, m=1 . . . K, of the character image of the character to be recognized, wherein each of said matrixes [G_m(i, j)]<sub>M×

M, m=</sub>1 . . . K, possesses M×

M pixels and represents the stroke information of one of the K directions;

extracting features from blocks, including extracting features from the blocks in said K matrixes [G_m(i, j)]<sub>M×

M, m=</sub>1 . . . K, of the image of the character to be recognized, and obtaining an initial recognition feature vector V with a high dimension of the character image of the character to be recognized;

compressing the features, including compressing the initial recognition feature vector V and obtaining a recognition feature vector V_c with a low dimension of the character image of the character to be recognized;

recognizing the character, including employing a specific classifier to calculate a distance between the recognition feature vector Vc and a category center vector of each character category, selecting a nearest distance from distances and the character category corresponding to the nearest distance, and calculating a character code of the character to be recognized according to a national standard code of the character category; and

(c) repeating step (b) to obtain each character code of each character image.

2. The method of claim 1, wherein processing the matrix further includes a step of optimizing parameters, said optimizing parameters including:

(1) selecting some typical character images from a training sample character image database, pre-processing the selected character images and obtaining a statistic of an average width and direction of the strokes of pre-processed images;

(2) determining an initial value of each parameter in impulse response function of the Gabor filter group according to the statistic, and determining a plurality of values of a parameter a in impulse response function for every step length within a certain value range;

(3) selecting one value from the plurality of values of a, combining said value with other determined parameters to construct the Gabor filter group, extracting the direction information of strokes from the pre-processed image of the character by the constructed Gabor filter group, and calculating a first average entropy correlation coefficient {overscore (ECC)} of the character image according to the information;

(4) repeating step (3) to obtain each first average entropy correlation coefficient {overscore (ECC)} of the image of each character corresponding to selected value of σ

, calculating an average value of all first average entropy correlation coefficients {overscore (ECC)}S of the image of each character to obtain a second average entropy correlation coefficient {overscore (ECC)} corresponding to the selected value of σ

;

(5) repeating step (4) to obtain each second average entropy correlation coefficient {overscore (ECC)} corresponding to each value of σ

; and

(6) selecting the smallest one from second average entropy correlation coefficients {overscore (ECCS)}, using the value of a corresponding to the smallest second average entropy correlation coefficient {overscore (ECC)} and other said determined parameters to optimally construct the Gabor filter group.

3. The method of claim 1, further comprising a step of adaptively processing, said adaptively processing including employing corrected Sigmoid function to adaptively process the K matrixes [G_m(i, j)]_M×

M, 1<

=m<

=K, wherein each of said matrixes represents the stroke information of one of the K directions to obtain K matrixes [F_m(i,j)]<sub>M×

M</sub>, 1<

=m<

=K, for the character image of each character to be recognized.

4. The method of claim 2, further comprising a step of adaptively processing, said adaptively processing including employing corrected Sigmoid function to adaptively process the K matrixes [G_m(i, j)]_M×

M, 1<

=m<

=K, wherein each of said matrixes represents the stroke information of one of the K directions to obtain K matrixes [F_m(i,j)]<sub>M×

M</sub>, 1<

=m<

=K, for the character image of each character to be recognized.

5. The method of claim 3, wherein the step of adaptively processing further comprises:

(1) normalizing said K matrixes [G_m(i, j)]<sub>M×

M</sub>, 1<

=m<

=K, to obtain K matrixes {[G′

_m(i,j)]<sub>M×

M</sub>}, 1≦

m≦

K;

(2) determining values of parameters of corrected Sigmoid function ψ

_m(t) according to the requirements of recognition capacity and experiences; and

(3) non-linear adaptively processing said K matrixes {[G′

_m(i,j)]<sub>M×

M</sub>}, 1≦

m≦

K, using the determined corrected Sigmoid function ψ

_m(t) to obtain K matrixes [F_m(i,j)]<sub>M×

M</sub>.

6. The method of claim 4, wherein the step of adaptively processing further comprises:

(1) normalizing said K matrixes [G_m(i, j)]<sub>M×

M</sub>, 1<

=m<

=K, to obtain K matrixes {[G′

_m(i,j)]<sub>M×

M</sub>}, 1≦

m≦

K ;

(2) determining values of parameters of corrected Sigmoid function ψ

_m(t) according to the requirements of recognition capacity and experiences; and

(3) non-linear adaptively processing said K matrixes {[G′

_m(i,j)]<sub>M×

M</sub>}, 1≦

m≦

K, using the determined corrected Sigmoid function ψ

_m(t) to obtain K matrixes [F_m(i,j)]<sub>M×

M</sub>, 1≦

m≦

K.

7. The method of any of claims 1-6, wherein the step of extracting features from blocks further including:

(1) evenly dividing, each of said K matrixes [G_m(i, j)]<sub>M×

M</sub> or each of said K matrixes {[F_m(i,j)]<sub>M×

M</sub>} into P×

P rectangular areas which are overlapped with each other and have a length L for each side of each rectangular;

(2) respectively calculating a first weighted sum of positive values and a second weighted sum of negative values of all pixels within the area, at the center of each rectangular area;

(3) forming a first feature vector S_m⁺ of positive values and a second feature vector S_m⁻

of negative values according to the first weighted sum and the second weighted sum of each rectangular area of each matrix, wherein the dimensions of S_m⁺ and S_m⁻

are both P²; and

(4) merging the first feature vector S_m⁺ and the second feature vector S_m⁻

for each [G_m(i, j)]<sub>M×

M</sub> or each {[F_m(i,j)]<sub>M×

M</sub>} as an initial recognition feature vector V=[S⁺₁ S⁻

₁ S⁺₂ S⁻

₂ . . . S⁺_K S⁻

_K] with a dimension of 2KP².

8. The method of any of claims 1-6, wherein said step of compressing features further includes:

(1) calculating a transformation matrix φ

used in the step of compressing features; and

(2) transposing the transformation matrix φ

, multiplying the transposed transformation matrix φ

by the initial recognition feature vector V of the character image of the character to be recognized thus compressing said initial recognition feature vector V and obtaining the recognition feature vector V_c with a low dimension for the character image of the character to be recognized.

9. The method of claim 7, wherein said step of compressing features further includes:

(1) calculating a transformation matrix φ

used in the step of compressing features; and

(2) transposing the transformation matrix φ

, multiplying the transposed transformation matrix φ

by the initial recognition feature vector V of the character image of the character to be recognized thus compressing said initial recognition feature vector V and obtaining the recognition feature vector V_c with a low dimension for the character image of the character to be recognized.

10. The method of claim 8, wherein said calculating a transformation matrix φ

includes;

(1) pre-processing the character image, extracting information of stroke'"'"'s direction of the character, adaptively processing the images and extracting features from blocks regarding all training sample characters of each character category, and obtaining a recognition feature vector sequence for each character category;

(2) calculating a first category center of each character category according to each recognition feature vector sequence of each character category;

(3) calculating a second category center of all character categories, according to each first category center of each character category;

(4) calculating a between-class scatter matrix Σ

b according to each first category center and the second category center;

(5) calculating a within-class scatter matrix Σ

w according to each first category center; and

(6) calculating first n largest, non-zero eigenvalues of matrix Σ

<sup>−

1</sup>wΣ

b, and arranging first n eigenvectors corresponding to said first n largest non-zero eigenvalues column by column to form the transformation matrixφ

.

11. The method of claim 9, wherein said calculating a transformation matrix φ

includes;

(1) pre-processing the character image, extracting information of stroke'"'"'s direction of the character, adaptively processing the images and extracting features from blocks regarding all training sample characters of each character category, and obtaining a recognition feature vector sequence for each character category;

(2) calculating a first category center of each character category according to each recognition feature vector sequence of each character category;

(3) calculating a second category center of all character categories, according to each first category center of each character category;

(4) calculating a between-class scatter matrix Σ

b according to each first category center and the second category center;

(5) calculating a within-class scatter matrix Σ

w according to each first category center; and

(6) calculating first n largest, non-zero eigenvalues of matrix Σ

<sup>−

1</sup>wΣ

b, and arranging first n eigenvectors corresponding to said first n largest non-zero eigenvalues column by column to form the transformation matrix φ

.