CN103218613B - Handwritten Numeral Recognition Method and device - Google Patents

Abstract

The invention discloses a kind of Handwritten Numeral Recognition Method and device.This Handwritten Numeral Recognition Method, comprising: determine image to be identified; According to the gray-scale value of pixel, determine at least three kinds of pixel characteristic of the specific pixel in this image to be identified; According at least three kinds of pixel characteristic of this specific pixel, determine the corresponding covariance of this image to be identified respectively; Each covariance calculating this image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label; Corresponding to minor increment in the multiple distances each covariance for this image to be identified determined respectively, numeric class distinguishing label is defined as redundant digital class label; Numeric class distinguishing labels maximum for number in this redundant digital class label is defined as the numeric class distinguishing label to be identified in image to be identified.Visible, by utilizing this programme, the identification accuracy of handwriting digital effectively can be improved.

Description

Handwritten Numeral Recognition Method and device

Technical field

The present invention relates to Handwritten Digital Recognition technical field, particularly relate to a kind of Handwritten Numeral Recognition Method and device.

Background technology

Handwritten form as the arabic numeral of countries in the world general symbol(s) frequently appears in each fields such as mail system, cashier's check, commercial Application.And along with the develop rapidly of computer technology and digital image processing techniques, Handwritten Digital Recognition technology is widely applied, and brings great convenience to the work of people.

Because numeral often represents accurate numerical value in every field, small mistake probably brings unpredictable consequence, and therefore, the simple Handwritten Numeral Recognition Method efficiently with higher accuracy is important research direction always.

And along with the popularization and application of machine learning techniques, a lot of physicist and chemist start the data widely using Lie group theoretical research association area; Accordingly, in Handwritten Digital Recognition technical field, Lie group structured data is widely used with its good mathematic(al) structure.

Wherein, Lie group average sorter (lieMeans) is the simple and effective sorting technique of one proposed in article " AK-MeansClusteringAlgorithm " by people such as J.A.Hartigan, but it selects single covariance feature to realize classification, the solution that its gradient descent method is found is local minimum, and not necessarily global minimum, the poor performance when processing many classification problems.

Visible, the existing Manuscripted Characters Identification Method based on Lie group average sorter realizes classification by selecting single covariance feature, and it can not make full use of the spatial information of image to be identified, causes identifying that accuracy is not high.

Summary of the invention

For solving the problems of the technologies described above, embodiments provide a kind of Handwritten Numeral Recognition Method and device, to improve the identification accuracy of handwriting digital, technical scheme is as follows:

On the one hand, embodiments provide a kind of Handwritten Numeral Recognition Method, comprising:

Determine image to be identified, in described image to be identified, comprise the numeric class distinguishing label to be identified of handwritten form form;

According to the gray-scale value of pixel, determine at least three kinds of pixel characteristic of the specific pixel in described image to be identified;

According at least three kinds of pixel characteristic of described specific pixel, determine the corresponding covariance of described image to be identified respectively, wherein, the unique corresponding covariance of each pixel characteristic;

Each covariance calculating described image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label; Wherein, in described training image set, each training image comprises the numeric class distinguishing label of a hand-written bodily form formula, the numeric class distinguishing label that described training image set comprises relates to all digital classifications, and, each class numeric class distinguishing label correspondence at least three Lie group averages in described training image set, a Lie group average of each covariance each class numeric class distinguishing label corresponding of described image to be identified;

Corresponding to minor increment in the multiple distances each covariance for described image to be identified determined respectively, numeric class distinguishing label is defined as redundant digital class label;

Numeric class distinguishing labels maximum for number in described redundant digital class label is defined as described numeric class distinguishing label to be identified.

Wherein, when the computing formula of the three kinds of pixel characteristic time institute foundations determining the specific pixel in described image to be identified comprises:

${\mathrm{\φ}}_1(I,x,y)={(x,y,I(x,y),|\frac{\∂}{\∂x}I(x,y)|,|\frac{\∂}{\∂y}I(x,y)\left|\right)}^T$

${\mathrm{\φ}}_2(I,x,y)={(I(x,y),|\frac{\∂}{\∂x}I(x,y)|,|\frac{\∂}{\∂y}I(x,y)|,|\frac{{\∂}^2}{\∂x\∂x}I(x,y)|,|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\left|\right)}^T$

${\mathrm{\φ}}_3(I,x,y)={\left(\begin{array}{c}x,y,I(x,y),\left|\frac{\∂}{\∂x}I(x,y)\right|,\left|\frac{\∂}{\∂y}I(x,y)\right|,\sqrt{{\left|\frac{\∂}{\∂x}I(x,y)\right|}^2+{\left|\frac{\∂}{\∂y}I(x,y)\right|}^2},\\ \left|\frac{{\∂}^2}{\∂x\∂x}I(x,y)\right|,\left|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\right|,a\tan \left(\frac{\left|\frac{\∂}{\∂x}I(x,y)\right|}{\left|\frac{\∂}{\∂y}I(x,y)\right|}\right)\end{array}\right)}^T$

Wherein, φ _jthe jth kind pixel characteristic that (I, x, y) (j=1,2,3) are the pixel (x, y) of described image to be identified, I (x, y) represents the gray-scale value at pixel (x, y) place, for the single order local derviation on x direction, pixel (x, y) place, for the single order local derviation on y direction, pixel (x, y) place, for the second order local derviation on x direction, pixel (x, y) place, for the second order local derviation on y direction, pixel (x, y) place, 1≤x≤m, m is the row pixel value in described image to be identified, and 1≤y≤n, n is the row pixel value in described image to be identified, and T is for carrying out matrix transpose.

Wherein, according to three kinds of pixel characteristic of described specific pixel, determine that the computing formula of the corresponding covariance institute foundation of described image to be identified comprises respectively:

$C^j=\frac{1}{\mathrm{mn}}\underset{x=1}{\overset{m}{\mathrm{\Σ}}}\underset{y=1}{\overset{n}{\mathrm{\Σ}}}{({\mathrm{\φ}}_j(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I\right))({\mathrm{\φ}}_j(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I\right))}^T,j=\mathrm{1,2,3}$

Wherein, C ^jfor the covariance corresponding to jth kind pixel characteristic, for the average of jth kind pixel characteristic in described image to be identified, T is for carrying out matrix transpose.

Wherein, each covariance calculating described image to be identified respectively and default training image set the computing formula of distance institute foundation that comprises between the corresponding Lie group average of each class numeric class distinguishing label comprise:

$d_k^j(C^j,m_k^j)=\sqrt{\underset{i=1}{\overset{d_i}{\mathrm{\Σ}}}\ln \left({\mathrm{\λ}}_i^2\right)},k=1,\·\·\·,c,j=\mathrm{1,2,3}$

Wherein, for the jth Lie group average that kth class numeric class distinguishing label is corresponding, c is the classification number of numeric class distinguishing label, λ _ic ^jwith generalized eigenvalue, d _irepresent the row or column number of covariance feature matrix.

Wherein, described default training image set comprise the Lie group average of each class numeric class distinguishing label determination mode comprise:

According to the gray-scale value of pixel, determine three kinds of pixel characteristic of each training image in described default training image set;

According to three kinds of pixel characteristic of described specific pixel, for each training image determines corresponding covariance respectively, wherein, the unique corresponding covariance of each pixel characteristic;

The covariance about same pixel characteristic corresponding for all training images is inputted corresponding Lie group average sorter, to determine the Lie group average about described pixel characteristic of each class numeric class distinguishing label.

Wherein, the specific pixel in described image to be identified comprises:

All pixels in described image to be identified;

Or,

Partial pixel point in described image to be identified, and described partial pixel point is the pixel of the handwriting area in described image to be identified, described handwriting area is a part of image-region in described image to be identified.

On the other hand, embodiments provide a kind of device for Identification of Handwritten Numerals, comprising:

Image determination module to be identified, for determining image to be identified, comprises the numeric class distinguishing label to be identified of handwritten form form in described image to be identified;

Pixel characteristic determination module, for the gray-scale value according to pixel, determines at least three kinds of pixel characteristic of the specific pixel in described image to be identified;

Covariance determination module, at least three kinds of pixel characteristic according to described specific pixel, determines the corresponding covariance of described image to be identified respectively, wherein, and the unique corresponding covariance of each pixel characteristic;

Distance determination module, for each covariance of calculating described image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label; Wherein, in described training image set, each training image comprises the numeric class distinguishing label of a hand-written bodily form formula, the numeric class distinguishing label that described training image set comprises relates to all digital classifications, and, each class numeric class distinguishing label correspondence at least three Lie group averages in described training image set, a Lie group average of each covariance each class numeric class distinguishing label corresponding of described image to be identified;

Spare labels determination module, is defined as redundant digital class label for numeric class distinguishing label corresponding to the minor increment in multiple distances of each covariance for described image to be identified being determined respectively;

Label determination module to be identified, for being defined as described numeric class distinguishing label to be identified by numeric class distinguishing labels maximum for number in described redundant digital class label.

In this programme, utilize at least three kinds of pixel characteristic of the specific pixel of image to be identified to determine at least three kinds of covariance features, and the classification utilizing at least three kinds of covariance features determined to realize handwriting digital is determined.Visible, realize with adopting single covariance feature in prior art compared with the mode of classifying, this programme makes full use of the spatial information of image to be identified, therefore, effectively can improve the identification accuracy of handwriting digital.

Accompanying drawing explanation

In order to be illustrated more clearly in the embodiment of the present invention or technical scheme of the prior art, be briefly described to the accompanying drawing used required in embodiment or description of the prior art below, apparently, accompanying drawing in the following describes is only some embodiments of the present invention, for those of ordinary skill in the art, under the prerequisite not paying creative work, other accompanying drawing can also be obtained according to these accompanying drawings.

The first process flow diagram of a kind of Handwritten Numeral Recognition Method that Fig. 1 provides for the embodiment of the present invention;

The second process flow diagram of a kind of Handwritten Numeral Recognition Method that Fig. 2 provides for the embodiment of the present invention;

The structural representation of a kind of device for Identification of Handwritten Numerals that Fig. 3 provides for the embodiment of the present invention.

Embodiment

Below in conjunction with the accompanying drawing in the embodiment of the present invention, be clearly and completely described the technical scheme in the embodiment of the present invention, obviously, described embodiment is only the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.

In order to improve the identification accuracy of handwriting digital, embodiments provide a kind of Handwritten Numeral Recognition Method and device.

First a kind of Handwritten Numeral Recognition Method that the embodiment of the present invention provides is introduced below.

As shown in Figure 1, a kind of Handwritten Numeral Recognition Method, can comprise:

S101, determines image to be identified;

When needs identify handwriting digital, first determine the image to be identified of the numeric class distinguishing label to be identified including handwritten form form, and then carry out follow-up process based on this image to be identified.

It should be noted that, numeric class distinguishing label is concrete numeral, and the classification number of its correspondence is 10 classes, and this 10 class numeric class distinguishing label is: 0-9.Wherein, this numeric class distinguishing label to be identified may be any numeral in 0-9, and the handwritten form form that this numeric class distinguishing label to be identified is corresponding can be not limited to one.

S102, according to the gray-scale value of pixel, determines at least three kinds of pixel characteristic of the specific pixel in this image to be identified;

Because stroke position contributes to discriminating digit, and the gray-scale value with the pixel of the position of stroke is different from other positions, makes the gray-scale value of pixel can as the important space information of handwriting digital to be identified.Therefore, after determining image to be identified, in order to make full use of the spatial information of image to be identified, according to the gray-scale value of pixel, at least three kinds of pixel characteristic of the specific pixel in this image to be identified can be determined.

It should be noted that, in order to ensure higher accuracy, the specific pixel in this image to be identified can comprise: all pixels in this image to be identified.Further, in order under the prerequisite ensureing higher accuracy, improve treatment effeciency, specific pixel in this image to be identified can comprise: the partial pixel point in this image to be identified, and this partial pixel point is the pixel of the handwriting area in this image to be identified, this handwriting area is a part of image-region in this image to be identified.

S103, according at least three kinds of pixel characteristic of this specific pixel, determines the corresponding covariance of this image to be identified respectively;

Wherein, the unique corresponding covariance of each pixel characteristic.

After determining at least three kinds of pixel characteristic of specific pixel point of image to be identified, the covariance of this image to be identified about each pixel characteristic can be determined, and then utilize at least three kinds of covariances determined to carry out subsequent treatment.

S104, each covariance calculating this image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label;

Wherein, in this training image set, each training image comprises the numeric class distinguishing label of a hand-written bodily form formula, the numeric class distinguishing label that this training image set comprises relates to all digital classifications, and, each class numeric class distinguishing label correspondence at least three Lie group averages in this training image set, a Lie group average of each covariance each class numeric class distinguishing label corresponding of this image to be identified.

Be understandable that, in actual applications, the numeric class distinguishing label that this training image set comprises relates to 10 numeral: 0-9; Further, the quantity of the corresponding training image of each class numeric class distinguishing label comprised in this training image set preset can be similar and different.

Further, it should be noted that, the determination mode of at least three Lie group averages that each the class numeric class distinguishing label in training image set is corresponding can comprise:

The gray-scale value of a, foundation pixel, determines at least three kinds of pixel characteristic of each training image in this training image set preset;

B, at least three kinds of pixel characteristic according to this specific pixel, for each training image determines corresponding covariance respectively, wherein, the unique corresponding covariance of each pixel characteristic;

C, the covariance about same pixel characteristic corresponding for all training images is inputted corresponding Lie group average sorter, to determine the Lie group average about this pixel characteristic of each class numeric class distinguishing label.

By the way, each class numeric class distinguishing label correspondence at least three Lie group averages in this training image set.

S105, corresponding to the minor increment in the multiple distances each covariance for this image to be identified determined respectively, numeric class distinguishing label is defined as redundant digital class label;

S106, is defined as this numeric class distinguishing label to be identified by numeric class distinguishing labels maximum for number in this redundant digital class label.

Because the covariance of this image to be identified is less to the spacing of the corresponding Lie group average of a numeric class distinguishing label, show that this numeric class distinguishing label to be identified is that the probability of this numeric class distinguishing label is larger, therefore, determine each covariance of this image to be identified and default training image set comprise between the corresponding Lie group average of each class numeric class distinguishing label distance after, corresponding to minor increment in the multiple distances that can each covariance for this image to be identified be determined respectively, numeric class distinguishing label is defined as redundant digital class label, and then numeric class distinguishing labels maximum for number in this redundant digital class label is defined as this numeric class distinguishing label to be identified.

In this programme, utilize at least three kinds of pixel characteristic of the specific pixel of image to be identified to determine at least three kinds of covariance features, and the classification utilizing at least three kinds of covariance features determined to realize handwriting digital is determined.Visible, realize with adopting single covariance feature in prior art compared with the mode of classifying, this programme makes full use of the spatial information of image to be identified, therefore, effectively can improve the identification accuracy of handwriting digital.

Below to adopt three kinds of covariance features, a kind of Handwritten Numeral Recognition Method that the embodiment of the present invention provides is introduced.

As shown in Figure 2, a kind of Handwritten Numeral Recognition Method, can comprise:

S201, determines image to be identified;

When needs identify handwriting digital, first determine the image to be identified of the numeric class distinguishing label to be identified including handwritten form form, and then carry out follow-up process based on this image to be identified.

It should be noted that, numeric class distinguishing label is concrete numeral, and the digital classification of its correspondence has 10 classes, and this 10 class numeric class distinguishing label is: 0-9.Wherein, this numeric class distinguishing label to be identified may be any numeral in 0-9, and the handwritten form form that this numeric class distinguishing label to be identified is corresponding can be not limited to one.

S202, according to the gray-scale value of pixel, determines three kinds of pixel characteristic of the specific pixel in this image to be identified;

Because stroke position contributes to discriminating digit, and the gray-scale value with the pixel of the position of stroke is different from other positions, makes the gray-scale value of pixel can as the important space information of handwriting digital to be identified.Therefore, after determining image to be identified, in order to make full use of the spatial information of image to be identified, according to the gray-scale value of pixel, three kinds of pixel characteristic of the specific pixel in this image to be identified can be determined.

It should be noted that, in order to ensure higher accuracy, the specific pixel in this image to be identified can comprise: all pixels in this image to be identified.Further, in order under the prerequisite ensureing higher accuracy, improve treatment effeciency, specific pixel in this image to be identified can comprise: the partial pixel point in this image to be identified, and this partial pixel point is the pixel of the handwriting area in this image to be identified, this handwriting area is a part of image-region in this image to be identified.

Wherein, when determining three kinds of pixel characteristic of the specific pixel in this image to be identified, the computing formula of institute's foundation comprises:

${\mathrm{\φ}}_1(I,x,y)={(x,y,I(x,y),|\frac{\∂}{\∂x}I(x,y)|,|\frac{\∂}{\∂y}I(x,y)\left|\right)}^T---\left(1\right)$

${\mathrm{\φ}}_2(I,x,y)={(I(x,y),|\frac{\∂}{\∂x}I(x,y)|,|\frac{\∂}{\∂y}I(x,y)|,|\frac{{\∂}^2}{\∂x\∂x}I(x,y)|,|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\left|\right)}^T---\left(2\right)$

${\mathrm{\φ}}_3(I,x,y)={\left(\begin{array}{c}x,y,I(x,y),\left|\frac{\∂}{\∂x}I(x,y)\right|,\left|\frac{\∂}{\∂y}I(x,y)\right|,\sqrt{{\left|\frac{\∂}{\∂x}I(x,y)\right|}^2+{\left|\frac{\∂}{\∂y}I(x,y)\right|}^2},\\ \left|\frac{{\∂}^2}{\∂x\∂x}I(x,y)\right|,\left|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\right|,a\tan \left(\frac{\left|\frac{\∂}{\∂x}I(x,y)\right|}{\left|\frac{\∂}{\∂y}I(x,y)\right|}\right)\end{array}\right)}^T---\left(3\right)$

Wherein, φ _jthe jth kind pixel characteristic of the pixel (x, y) that (I, x, y) (j=1,2,3) are this image to be identified, I (x, y) represents the gray-scale value at pixel (x, y) place, for the single order local derviation on x direction, pixel (x, y) place, for the single order local derviation on y direction, pixel (x, y) place, for the second order local derviation on x direction, pixel (x, y) place, for the second order local derviation on y direction, pixel (x, y) place, 1≤x≤m, m is the row pixel value in this image to be identified, and 1≤y≤n, n is the row pixel value in this image to be identified, and T is for carrying out matrix transpose.

S203, according to three kinds of pixel characteristic of this specific pixel, determines the corresponding covariance of this image to be identified respectively;

Wherein, the unique corresponding covariance of each pixel characteristic.

After the three kinds of pixel characteristic of specific pixel point determining image to be identified, the covariance of this image to be identified about each pixel characteristic can be determined, and then utilize the three kinds of covariances determined to carry out subsequent treatment.

Wherein, according to three kinds of pixel characteristic of this specific pixel, determine that the computing formula of the corresponding covariance institute foundation of this image to be identified can comprise respectively:

$C^j=\frac{1}{\mathrm{mn}}\underset{x=1}{\overset{m}{\mathrm{\Σ}}}\underset{y=1}{\overset{n}{\mathrm{\Σ}}}{({\mathrm{\φ}}_j(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I\right))({\mathrm{\φ}}_j(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I\right))}^T,j=\mathrm{1,2,3}---\left(4\right)$

Wherein, C ^jfor the covariance corresponding to jth kind pixel characteristic, for the average of jth kind pixel characteristic in this image to be identified, T is for carrying out matrix transpose.

S204, each covariance calculating this image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label;

Wherein, in this training image set, each training image comprises the numeric class distinguishing label of a hand-written bodily form formula, the numeric class distinguishing label that this training image set comprises relates to all digital classifications, and, corresponding three the Lie group averages of each class numeric class distinguishing label in this training image set, a Lie group average of each covariance each class numeric class distinguishing label corresponding of this image to be identified.

Be understandable that, in actual applications, the numeric class distinguishing label that this training image set comprises relates to 10 numeral: 0-9; Further, the quantity of the corresponding training image of each class numeric class distinguishing label comprised in this training image set preset can be similar and different.

Wherein, each covariance calculating this image to be identified respectively and default training image set the computing formula of distance institute foundation that comprises between the corresponding Lie group average of each class numeric class distinguishing label comprise:

$d_k^j(C^j,m_k^j)=\sqrt{\underset{i=1}{\overset{d_i}{\mathrm{\Σ}}}1n\left({\mathrm{\λ}}_i^2\right)},k=1,\·\·\·,c,j=\mathrm{1,2,3}---\left(5\right)$

Wherein, for the jth Lie group average that kth class numeric class distinguishing label is corresponding, c is the classification number of numeric class distinguishing label, λ _ic ^jwith generalized eigenvalue, d _irepresent the row or column number of covariance feature matrix.

Further, preset training image set comprise the Lie group average of each class numeric class distinguishing label determination mode can comprise:

The gray-scale value of a, foundation pixel, determines three kinds of pixel characteristic of each training image in this training image set preset;

B, three kinds of pixel characteristic according to this specific pixel, for each training image determines corresponding covariance respectively, wherein, the unique corresponding covariance of each pixel characteristic;

C, the covariance about same pixel characteristic corresponding for all training images is inputted corresponding Lie group average sorter, to determine the Lie group average about described pixel characteristic of each class numeric class distinguishing label.

It should be noted that, for training image, determine that the computing formula of three kinds of pixel characteristic institute foundations of specific pixel is computing formula (1) (2) (3), determine that the computing formula of three kinds of covariance institute foundations of training image is computing formula (4).

S205, corresponding to the minor increment in the multiple distances each covariance for this image to be identified determined respectively, numeric class distinguishing label is defined as redundant digital class label;

S206, is defined as this numeric class distinguishing label to be identified by numeric class distinguishing labels maximum for number in this redundant digital class label.

Because the covariance of this image to be identified is less to the spacing of the corresponding Lie group average of a numeric class distinguishing label, show that this numeric class distinguishing label to be identified is that the probability of this numeric class distinguishing label is larger, therefore, determine each covariance of this image to be identified and default training image set comprise between the corresponding Lie group average of each class numeric class distinguishing label distance after, corresponding to minor increment in the multiple distances that can each covariance for this image to be identified be determined respectively, numeric class distinguishing label is defined as redundant digital class label, and then numeric class distinguishing labels maximum for number in this redundant digital class label is defined as this numeric class distinguishing label to be identified.

In this programme, utilize three kinds of pixel characteristic of the specific pixel of image to be identified to determine three kinds of covariance features, and the classification utilizing the three kinds of covariance features determined to realize handwriting digital is determined.Visible, realize with adopting single covariance feature in prior art compared with the mode of classifying, this programme makes full use of the spatial information of image to be identified, therefore, effectively can improve the identification accuracy of handwriting digital.

It should be noted that, building the training image set preset can from MNIST handwriting digital data centralization, the random training image obtaining some quantity of each class numeric class distinguishing label respectively.Wherein, MNIST is the subset of famous American data set NIST, pattern-recognition common experimental data set, and this data centralization has the training set be made up of 60000 training images and the test set be made up of 10000 test patterns.

Introduce the training process corresponding based on the Handwritten Numeral Recognition Method of three kinds of covariance features that the embodiment of the present invention provides below:

(1) training image processing procedure:

1) training image set is determined wherein, I _i∈ R ^{m × n}be i-th training image, m and n represents row pixel and the row pixel value of training image, l _i∈ 1 ..., c} is I _icorresponding numeric class distinguishing label, namely represents I _ibe which numeral, N represents total number of training image, the classification number of c representative digit class label; Wherein, suppose m=n=28, N=100c, and make i=1.

2) to training image I _ithe pixel at upper point (x, y) place extracts following three kinds of pixel characteristic:

${\mathrm{\φ}}_1(I_i,x,y)={(x,y,I_i(x,y),|\frac{\∂}{\∂x}{I}_i(x,y)|,|\frac{\∂}{\∂y}{I}_i(x,y)\left|\right)}^T$

${\mathrm{\φ}}_2(I_i,x,y)={(I_i(x,y),|\frac{\∂}{\∂x}{I}_i(x,y)|,|\frac{\∂}{\∂y}{I}_i(x,y)|,|\frac{{\∂}^2}{\∂x\∂x}{I}_i(x,y)|,|\frac{{\∂}^2}{\∂y\∂y}{I}_i(x,y)\left|\right)}^T$

${\mathrm{\φ}}_3(I_i,x,y)={\left(\begin{array}{c}x,y,I_i(x,y),\left|\frac{\∂}{\∂x}{I}_i(x,y)\right|,\left|\frac{\∂}{\∂y}{I}_i(x,y)\right|,\sqrt{{\left|\frac{\∂}{\∂x}{I}_i(x,y)\right|}^2+{\left|\frac{\∂}{\∂y}{I}_i(x,y)\right|}^2},\\ \left|\frac{{\∂}^2}{\∂x\∂x}{I}_i(x,y)\right|,\left|\frac{{\∂}^2}{\∂y\∂y}{I}_i(x,y)\right|,a\tan \left(\frac{\left|\frac{\∂}{\∂x}{I}_i(x,y)\right|}{\left|\frac{\∂}{\∂y}{I}_i(x,y)\right|}\right)\end{array}\right)}^T$

Wherein, φ _j(I, x, y) (j=1,2,3) are training image I _ithe jth kind pixel characteristic of pixel (x, y), I _i(x, y) represents the gray-scale value at pixel (x, y) place, for the single order local derviation on x direction, pixel (x, y) place, for the single order local derviation on y direction, pixel (x, y) place, for the second order local derviation on x direction, pixel (x, y) place, for the second order local derviation on y direction, pixel (x, y) place, 1≤x≤28, m is this training image I _irow pixel value, 1≤y≤28, n is training image I _irow pixel value, T is for carrying out matrix transpose.

3) according to the three kinds of pixel characteristic extracted, according to following computing formula determination training image I _ithree kinds of covariances:

$C_i^j=\frac{1}{\mathrm{mn}}\underset{x=1}{\overset{m}{\mathrm{\Σ}}}\underset{y=1}{\overset{n}{\mathrm{\Σ}}}{({\mathrm{\φ}}_j(I_i,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I_i\right))({\mathrm{\φ}}_j(I_i,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I_i\right))}^T,j=\mathrm{1,2,3}$

Wherein, for training image I _ithe covariance corresponding to jth kind pixel characteristic,

for training image I _ithe average of middle jth kind pixel characteristic, T is for carrying out matrix transpose.

4) if i=N, then stop, otherwise i=i+1, repeat 2) and 3).

5) to the training image covariance obtained the Lie group average of each class numeric class distinguishing label is obtained with a jth Lie group average sorter k=1 ..., c.

(2) test pattern processing procedure:

1) test pattern I is determined, wherein, x ∈ R ^{m × n}.

2) pixel characteristic φ is extracted to the pixel at point (x, y) place on test pattern I _j(I, x, y), j=1,2,3, according to formula identical with computing formula training image being extracted to three kinds of pixel characteristic;

3) according to the three kinds of pixel characteristic extracted, the covariance C of test pattern I is determined ^j, according to formula with determine that for training image the computing formula of three kinds of covariances is identical;

(3) identifying:

1) test pattern covariance C is calculated ^jwith the Lie group average of each class numeric class distinguishing label in training image set between distance, namely

$d_k^j(C^j,m_k^j)=\sqrt{\underset{i=1}{\overset{d_i}{\mathrm{\Σ}}}1n\left({\mathrm{\λ}}_i^2\right)},k=1,\·\·\·,c,j=\mathrm{1,2,3}$

2) a jth Lie group average sorter is q to the classification results of test pattern ^j, namely

$q^j=\underset{k=1,\·\·\·,c}{\arg \min }{d}_k^j(C^j,m_k^j)$

3) according to the result of integrated three the Lie group average sorters of majority ballot criterion, the classification of test pattern is exported.

Effect of the present invention can by following experimental verification:

100 training images that Stochastic choice every class numeric class distinguishing label is corresponding from training set, Stochastic choice 200 test patterns from test set; Repeat this sampling process 10 times, the result finally exported is this average result of 10 times.The method of Experimental comparison has Lie group average sorter and the present invention.

Whole experimentation comprises three groups of experiments, is respectively to carry out two classification to numeral (1,9), and numeral (1,7,9) carries out three classification, and numeral (1,2,7,9) carries out four classification.

Experimental result is as shown in table 1.The result of Lie group average sorter has four, adopts the result of different covariance and the average of this three.As can be seen from Table 1, for two classification, three classification and four classification, misclassification rate of the present invention is all less than the misclassification rate of Lie group average sorter, and therefore, the present invention has better recognition effect.

Table 1

By the description of above embodiment of the method, those skilled in the art can be well understood to the mode that the present invention can add required general hardware platform by software and realize, hardware can certainly be passed through, but in a lot of situation, the former is better embodiment.Based on such understanding, technical scheme of the present invention can embody with the form of software product the part that prior art contributes in essence in other words, this computer software product is stored in a storage medium, comprising some instructions in order to make a computer equipment (can be personal computer, server, or the network equipment etc.) perform all or part of step of method described in each embodiment of the present invention.And aforesaid storage medium comprises: ROM (read-only memory) (ROM), random-access memory (ram), magnetic disc or CD etc. various can be program code stored medium.

Corresponding to embodiment of the method above, the embodiment of the present invention also provides a kind of device for Identification of Handwritten Numerals, as shown in Figure 3, can comprise:

Image determination module 110 to be identified, for determining image to be identified, comprises the numeric class distinguishing label to be identified of handwritten form form in described image to be identified;

Pixel characteristic determination module 120, for the gray-scale value according to pixel, determines at least three kinds of pixel characteristic of the specific pixel in described image to be identified;

Covariance determination module 130, at least three kinds of pixel characteristic according to described specific pixel, determines the corresponding covariance of described image to be identified respectively, wherein, and the unique corresponding covariance of each pixel characteristic;

Distance determination module 140, for each covariance of calculating described image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label; Wherein, in described training image set, each training image comprises the numeric class distinguishing label of a hand-written bodily form formula, the numeric class distinguishing label that described training image set comprises relates to all digital classifications, and, each class numeric class distinguishing label correspondence at least three Lie group averages in described training image set, a Lie group average of each covariance each class numeric class distinguishing label corresponding of described image to be identified;

Spare labels determination module 150, is defined as redundant digital class label for numeric class distinguishing label corresponding to the minor increment in multiple distances of each covariance for described image to be identified being determined respectively;

Label determination module 160 to be identified, for being defined as described numeric class distinguishing label to be identified by numeric class distinguishing labels maximum for number in described redundant digital class label.

In this programme, utilize at least three kinds of pixel characteristic of the specific pixel of image to be identified to determine at least three kinds of covariance features, and the classification utilizing at least three kinds of covariance features determined to realize handwriting digital is determined.Visible, realize with adopting single covariance feature in prior art compared with the mode of classifying, this programme makes full use of the spatial information of image to be identified, therefore, effectively can improve the identification accuracy of handwriting digital.

Wherein, described pixel characteristic determination module 120 determines that the computing formula of three kinds of pixel characteristic time institute foundations of the specific pixel in described image to be identified can comprise:

${\mathrm{\φ}}_1(I,x,y)={(x,y,I(x,y),|\frac{\∂}{\∂x}I(x,y)|,|\frac{\∂}{\∂y}I(x,y)\left|\right)}^T$

${\mathrm{\φ}}_2(I,x,y)={(I(x,y),|\frac{\∂}{\∂x}I(x,y)|,|\frac{\∂}{\∂y}I(x,y)|,|\frac{{\∂}^2}{\∂x\∂x}I(x,y)|,|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\left|\right)}^T$

${\mathrm{\φ}}_3(I,x,y)={\left(\begin{array}{c}x,y,I(x,y),\left|\frac{\∂}{\∂x}I(x,y)\right|,\left|\frac{\∂}{\∂y}I(x,y)\right|,\sqrt{{\left|\frac{\∂}{\∂x}I(x,y)\right|}^2+{\left|\frac{\∂}{\∂y}I(x,y)\right|}^2},\\ \left|\frac{{\∂}^2}{\∂x\∂x}I(x,y)\right|,\left|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\right|,a\tan \left(\frac{\left|\frac{\∂}{\∂x}I(x,y)\right|}{\left|\frac{\∂}{\∂y}I(x,y)\right|}\right)\end{array}\right)}^T$

Wherein, φ _jthe jth kind pixel characteristic that (I, x, y) (j=1,2,3) are the pixel (x, y) of described image to be identified, I (x, y) represents the gray-scale value at pixel (x, y) place, for the single order local derviation on x direction, pixel (x, y) place, for the single order local derviation on y direction, pixel (x, y) place, for the second order local derviation on x direction, pixel (x, y) place, for the second order local derviation on y direction, pixel (x, y) place, 1≤x≤m, m is the row pixel value in described image to be identified, and 1≤y≤n, n is the row pixel value in described image to be identified, and T is for carrying out matrix transpose.

Accordingly, described covariance determination module 130, according to three kinds of pixel characteristic of described specific pixel, determines that the computing formula of the corresponding covariance institute foundation of described image to be identified can comprise respectively:

$C^j=\frac{1}{\mathrm{mn}}\underset{x=1}{\overset{m}{\mathrm{\Σ}}}\underset{y=1}{\overset{n}{\mathrm{\Σ}}}{({\mathrm{\φ}}_j(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I\right))({\mathrm{\φ}}_j(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j\left(I\right))}^T,j=\mathrm{1,2,3}$

Wherein, C ^jfor the covariance corresponding to jth kind pixel characteristic, for the average of jth kind pixel characteristic in described image to be identified, T is for carrying out matrix transpose.

Each covariance that described distance determination module 140 calculates described image to be identified respectively and default training image set the computing formula of distance institute foundation that comprises between the corresponding Lie group average of each class numeric class distinguishing label can comprise:

$d_k^j(C^j,m_k^j)=\sqrt{\underset{i=1}{\overset{d_i}{\mathrm{\Σ}}}1n\left({\mathrm{\λ}}_i^2\right)},k=1,\·\·\·,c,j=\mathrm{1,2,3}$

Wherein, for the jth Lie group average that kth class numeric class distinguishing label is corresponding, c is the classification number of numeric class distinguishing label, λ _ibe Cj and generalized eigenvalue, d _irepresent the row or column number of covariance feature matrix.

For device or system embodiment, because it is substantially corresponding to embodiment of the method, so relevant part illustrates see the part of embodiment of the method.Device described above or system embodiment are only schematic, the wherein said unit illustrated as separating component or can may not be and physically separates, parts as unit display can be or may not be physical location, namely can be positioned at a place, or also can be distributed in multiple network element.Some or all of module wherein can be selected according to the actual needs to realize the object of the present embodiment scheme.Those of ordinary skill in the art, when not paying creative work, are namely appreciated that and implement.

In several embodiment provided by the present invention, should be understood that, disclosed system, apparatus and method, not exceeding in the spirit and scope of the application, can realize in other way.Current embodiment is a kind of exemplary example, should as restriction, and given particular content should in no way limit the object of the application.Such as, the division of described unit or subelement, is only a kind of logic function and divides, and actual can have other dividing mode when realizing, and such as multiple unit or multiple subelement combine.In addition, multiple unit can or assembly can in conjunction with or another system can be integrated into, or some features can be ignored, or do not perform.

In addition, described system, the schematic diagram of apparatus and method and different embodiment, not exceeding in the scope of the application, can with other system, module, technology or methods combining or integrated.Another point, shown or discussed coupling each other or direct-coupling or communication connection can be by some interfaces, and the indirect coupling of device or unit or communication connection can be electrical, machinery or other form.

The above is only the specific embodiment of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (5)

1. a Handwritten Numeral Recognition Method, is characterized in that, comprising:

Determine image to be identified, in described image to be identified, comprise the numeric class distinguishing label to be identified of handwritten form form;

According to the gray-scale value of pixel, determine at least three kinds of pixel characteristic of the specific pixel in described image to be identified;

According at least three kinds of pixel characteristic of described specific pixel, determine the corresponding covariance of described image to be identified respectively, wherein, the unique corresponding covariance of each pixel characteristic;

Each covariance calculating described image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label; Wherein, in described training image set, each training image comprises the numeric class distinguishing label of a hand-written bodily form formula, the numeric class distinguishing label that described training image set comprises relates to all digital classifications, and, each class numeric class distinguishing label correspondence at least three Lie group averages in described training image set, a Lie group average of each covariance each class numeric class distinguishing label corresponding of described image to be identified;

Corresponding to minor increment in the multiple distances each covariance for described image to be identified determined respectively, numeric class distinguishing label is defined as redundant digital class label;

Numeric class distinguishing labels maximum for number in described redundant digital class label is defined as described numeric class distinguishing label to be identified;

When the computing formula of the three kinds of pixel characteristic time institute foundations determining the specific pixel in described image to be identified comprises:

${\mathrm{\φ}}_1(I,x,y)={(x,y,I(x,y),|\frac{\∂}{\∂x}I\left(x,y\right)|,|\frac{\∂}{\∂y}I\left(x,y\right)\left|\right)}^T$

${\mathrm{\φ}}_2(I,x,y)={\left(I\right(x,y),|\frac{\∂}{\∂x}I\left(x,y\right)|,|\frac{\∂}{\∂y}I\left(x,y\right)|,|\frac{{\∂}^2}{\∂x\∂x}I\left(x,y\right)|,|\frac{{\∂}^2}{\∂y\∂y}I\left(x,y\right)\left|\right)}^T$

${\mathrm{\φ}}_3(I,x,y)={\left(\begin{array}{c}x,y,I(x,y),\left|\frac{\∂}{\∂x}I(x,y)\right|,\left|\frac{\∂}{\∂y}I(x,y)\right|,\sqrt{|\frac{\∂}{\∂x}I(x,y){|}^2+|\frac{\∂}{\∂y}I(x,y){|}^2},\\ \left|\frac{{\∂}^2}{\∂x\∂x}I(x,y)\right|,\left|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\right|,a\tan \left(\frac{\left|\frac{\∂}{\∂x}I(x,y)\right|}{\left|\frac{\∂}{\∂y}I(x,y)\right|}\right)\end{array}\right)}^T$

Wherein, φ _jthe jth kind pixel characteristic that (I, x, y) (j=1,2,3) are the pixel (x, y) of described image to be identified, I (x, y) represents the gray-scale value at pixel (x, y) place, for the single order local derviation on x direction, pixel (x, y) place, for the single order local derviation on y direction, pixel (x, y) place, for the second order local derviation on x direction, pixel (x, y) place, for the second order local derviation on y direction, pixel (x, y) place, 1≤x≤m, m is the row pixel value in described image to be identified, and 1≤y≤n, n is the row pixel value in described image to be identified, and T is for carrying out matrix transpose;

According to three kinds of pixel characteristic of described specific pixel, determine that the computing formula of the corresponding covariance institute foundation of described image to be identified comprises respectively:

$C^j=\frac{1}{mn}\underset{x=1}{\overset{m}{\Σ}}\underset{y=1}{\overset{n}{\Σ}}\left({\mathrm{\φ}}_j\right(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j(I\left)\right){\left({\mathrm{\φ}}_j\right(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j(I\left)\right)}^T,j=1,2,3$

Wherein, C ^jfor the covariance corresponding to jth kind pixel characteristic, for the average of jth kind pixel characteristic in described image to be identified, T is for carrying out matrix transpose.

2. method according to claim 1, it is characterized in that, each covariance calculating described image to be identified respectively and default training image set the computing formula of distance institute foundation that comprises between the corresponding Lie group average of each class numeric class distinguishing label comprise:

$d_k^j(C^j,m_k^j)=\sqrt{\underset{i=1}{\overset{d_i}{\Σ}}ln\left({\mathrm{\λ}}_i^2\right)},k=1,...,c,j=1,2,3$

Wherein, for the jth Lie group average that kth class numeric class distinguishing label is corresponding, c is the classification number of numeric class distinguishing label, λ _ic ^jwith generalized eigenvalue, d _irepresent the row or column number of covariance feature matrix.

3. method according to claim 2, is characterized in that, described default training image set comprise the Lie group average of each class numeric class distinguishing label determination mode comprise:

According to the gray-scale value of pixel, determine three kinds of pixel characteristic of each training image in described default training image set;

According to three kinds of pixel characteristic of described specific pixel, for each training image determines corresponding covariance respectively, wherein, the unique corresponding covariance of each pixel characteristic;

The covariance about same pixel characteristic corresponding for all training images is inputted corresponding Lie group average sorter, to determine the Lie group average about described pixel characteristic of each class numeric class distinguishing label.

4. method according to claim 1, is characterized in that, the specific pixel in described image to be identified comprises:

All pixels in described image to be identified;

Or,

Partial pixel point in described image to be identified, and described partial pixel point is the pixel of the handwriting area in described image to be identified, described handwriting area is a part of image-region in described image to be identified.

5. a device for Identification of Handwritten Numerals, is characterized in that, comprising:

Image determination module to be identified, for determining image to be identified, comprises the numeric class distinguishing label to be identified of handwritten form form in described image to be identified;

Pixel characteristic determination module, for the gray-scale value according to pixel, determines at least three kinds of pixel characteristic of the specific pixel in described image to be identified;

Covariance determination module, at least three kinds of pixel characteristic according to described specific pixel, determines the corresponding covariance of described image to be identified respectively, wherein, and the unique corresponding covariance of each pixel characteristic;

Distance determination module, for each covariance of calculating described image to be identified respectively and default training image set comprise distance between the corresponding Lie group average of each class numeric class distinguishing label; Wherein, in described training image set, each training image comprises the numeric class distinguishing label of a hand-written bodily form formula, the numeric class distinguishing label that described training image set comprises relates to all digital classifications, and, each class numeric class distinguishing label correspondence at least three Lie group averages in described training image set, a Lie group average of each covariance each class numeric class distinguishing label corresponding of described image to be identified;

Spare labels determination module, is defined as redundant digital class label for numeric class distinguishing label corresponding to the minor increment in multiple distances of each covariance for described image to be identified being determined respectively;

Label determination module to be identified, for being defined as described numeric class distinguishing label to be identified by numeric class distinguishing labels maximum for number in described redundant digital class label;

When the computing formula of the three kinds of pixel characteristic time institute foundations determining the specific pixel in described image to be identified comprises:

${\mathrm{\φ}}_1(I,x,y)={(x,y,I(x,y),|\frac{\∂}{\∂x}I\left(x,y\right)|,|\frac{\∂}{\∂y}I\left(x,y\right)\left|\right)}^T$

${\mathrm{\φ}}_2(I,x,y)={\left(I\right(x,y),|\frac{\∂}{\∂x}I\left(x,y\right)|,|\frac{\∂}{\∂y}I\left(x,y\right)|,|\frac{{\∂}^2}{\∂x\∂x}I\left(x,y\right)|,|\frac{{\∂}^2}{\∂y\∂y}I\left(x,y\right)\left|\right)}^T$

${\mathrm{\φ}}_3(I,x,y)={\left(\begin{array}{c}x,y,I(x,y),\left|\frac{\∂}{\∂x}I(x,y)\right|,\left|\frac{\∂}{\∂y}I(x,y)\right|,\sqrt{|\frac{\∂}{\∂x}I(x,y){|}^2+|\frac{\∂}{\∂y}I(x,y){|}^2},\\ \left|\frac{{\∂}^2}{\∂x\∂x}I(x,y)\right|,\left|\frac{{\∂}^2}{\∂y\∂y}I(x,y)\right|,a\tan \left(\frac{\left|\frac{\∂}{\∂x}I(x,y)\right|}{\left|\frac{\∂}{\∂y}I(x,y)\right|}\right)\end{array}\right)}^T$

Wherein, φ _jthe jth kind pixel characteristic that (I, x, y) (j=1,2,3) are the pixel (x, y) of described image to be identified, I (x, y) represents the gray-scale value at pixel (x, y) place, for the single order local derviation on x direction, pixel (x, y) place, for the single order local derviation on y direction, pixel (x, y) place, for the second order local derviation on x direction, pixel (x, y) place, for the second order local derviation on y direction, pixel (x, y) place, 1≤x≤m, m is the row pixel value in described image to be identified, and 1≤y≤n, n is the row pixel value in described image to be identified, and T is for carrying out matrix transpose;

According to three kinds of pixel characteristic of described specific pixel, determine that the computing formula of the corresponding covariance institute foundation of described image to be identified comprises respectively:

$C^j=\frac{1}{mn}\underset{x=1}{\overset{m}{\Σ}}\underset{y=1}{\overset{n}{\Σ}}\left({\mathrm{\φ}}_j\right(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j(I\left)\right){\left({\mathrm{\φ}}_j\right(I,x,y)-{\stackrel{\‾}{\mathrm{\φ}}}_j(I\left)\right)}^T,j=1,2,3$

Wherein, C ^jfor the covariance corresponding to jth kind pixel characteristic, for the average of jth kind pixel characteristic in described image to be identified, T is for carrying out matrix transpose.