CN101000651B - Method for recognizing multiple texture image - Google Patents

Abstract

A method for identifying multi-source vein image includes using grey scale of image as height, calculating out angles of four surfaces, utilizing wavelet transform to separate image low frequency component from high frequency component to obtain wavelet transform coefficient and utilizing 2-D gradient formula to calculate average value of image gradient for obtaining character parameters of wavelet transform and gradient, using DS evidence theory to convert character parameters to be basic trust degrees on image and using a merge formula to obtain identification result of image to identified.

Description

A kind of recognition methods of multi-source texture image

Technical field

The present invention relates to a kind of method that is used to read or discern printing or written character or is used to discern figure, more particularly relate to a kind of recognition methods at the multi-source texture image.

Background technology

In the practical problems of texture image identification, how to select appropriate characteristic parameter to allow the people feel awkward all the time, be difficult to it be made correct identification by a kind of characteristic parameter.And because the influence of various situations, a lot of objects to be identified are more difficult to get the data set of large sample, and uncertain often, the out of true or inaccurate of the information that data set provided of small sample so just is easy to cause the incorrect of recognition result.In order to address this problem, can utilize multiple information sources that the image that will discern is extracted characteristic parameter respectively to obtain data set separately, the method for application data fusion is made identification on this basis then.

Summary of the invention

Technical matters to be solved by this invention provides a kind of recognition methods that utilizes multiple information sources at the multi-source texture image.

The technical solution used in the present invention: a kind of recognition methods of multi-source texture graphics comprises the following steps:

A. the characteristic parameter of computed image " angle ": the gray scale of image is considered as height, each pixel and its neighbours territory pixel can constitute one " pyramid ", calculate the angle of four faces, the average of these angles is expressed as " angle " characteristic parameter of this image;

B. the characteristic parameter of computed image " wavelet transformation ": utilize wavelet transformation that the low frequency component and the high fdrequency component of image are separated, can obtain the coefficient of wavelet transformation, the part of its medium-high frequency is corresponding to the details of image, calculate the mean square deviation of these presentation video detail coefficients, as the characteristic parameter of image " wavelet transformation ";

C. the characteristic parameter of computed image " gradient ": the gradient correspondence of image the localized variation of image, the average of computing formula computed image gradient of utilizing two-dimentional gradient is as the characteristic parameter of image " gradient ";

D. the characteristic parameter that utilizes the DS evidence theory that (a) and (b), (c) step are obtained is converted into the basic trust degree of each characteristic parameter to image, obtains the recognition result of image to be identified afterwards by fusion formula.

Beneficial effect of the present invention, thus the present invention can be merged the identification of finishing texture image to multiple information sources effectively.In experiment to 100 image patterns, the recognition correct rate of any method of characteristic parameter of the characteristic parameter of the characteristic parameter of employing " angle ", " wavelet transformation " or " gradient " all is lower than 85% separately, wherein utilize angle " the recognition correct rate of characteristic parameter be 82%; utilizing the characteristic parameter recognition correct rate of " wavelet transformation " is 84%, is 79% and utilize the characteristic parameter recognition correct rate of " gradient ".But having adopted after the blending algorithm based on above three kinds of methods proposed by the invention, is 93% to the recognition correct rate of same 100 image patterns, and visible this method can obtain the better recognition effect in actual applications.

Embodiment

Below by embodiment the present invention is described in further detail, can take following three kinds of diverse ways to obtain its characteristic parameter at the characteristics of texture image:

(1) " angle ": if regard the gray scale of image as height, each pixel and its neighbours territory pixel can constitute one " pyramid ".Can calculate the angle of four faces like this, the average of these angles can be represented the angle of this image.

(2) " wavelet transformation ": utilize wavelet transformation, the low frequency component and the high fdrequency component of image can be separated, mainly use the high fdrequency component after the conversion here.After image carried out wavelet transformation, can obtain the coefficient of wavelet transformation.The part of its medium-high frequency is corresponding to the details of image.For the coefficient of these presentation video details, can calculate its mean square deviation, as the second method of extracting the characteristics of image parameter.

(3) " gradient ": the gradient correspondence of image the localized variation of image, the average of computing formula computed image gradient that can utilize two-dimentional gradient here is as the 3rd characteristic parameter.

Utilize above-described method, key issue is how the result of multiple information sources gained to be merged.In the DS theory, at first according to the model of elementary probability partition function, obtain the elementary probability partition function of different aforementioned sources to different identifying objects, the rule (1) according to data fusion merges it afterwards.

$m\left(H\right)=m_1\left(H\right)\⊕...\⊕m_a\left(H\right),\∀H\∈2^{\mathrm{\Ω}}$ (1)

Here, H represents to suppose the object discerned, m ₁(H) ..., m _a(H) represent the degree of belief of different information sources for this identifying object, Ω represents the set of identifying object,

The operational symbol of then representing fusion.

Concrete fusion formula is as follows:

$m\left(H\right)=\frac{1}{1-k}\underset{H^{\′}\∩H^{\′\′}=H}{\mathrm{\Σ}}{m}_1\left(H^{\′}\right)m_2\left(H^{\′\′}\right)$ (2)

Wherein

$k=\underset{H^{\′}\∩H^{\′\′}=\mathrm{\φ}}{\mathrm{\Σ}}{m}_1\left(H^{\′}\right)m_2\left(H^{\′\′}\right)$ (3)

So just can discern texture image based on a plurality of characteristic parameters.In fusion process, at first can obtain the basic trust degree of three information sources for these two identifying objects.Next, can obtain the degree of belief of different aforementioned sources according to (2) and (3) for each identifying object.In order to increase the order of accuarcy of identification, introducing can't be distinguished this situation, represents with Θ.

At last, according to corresponding criterion, can obtain the affiliated image category of image to be identified.

If X is an image to be identified, extract its characteristic parameter with above three kinds of diverse ways respectively, obtain following result:

x ₁＝0.36520466；x ₂＝11.7546527；x ₃＝2.51799266。

These three kinds of methods can be regarded three information sources as, and they belong to H respectively to X ₁, H ₂Probability as follows, table 1.

Table 1

P(xj/Hn)	Crit1	Crit2	Crit3
				Hypothese 1	P(x1/H1)＝ 0.0224791	P(x2/H1)＝ 0.08988746	P(x3/H1)＝ 0.64044944
Hypothese 2	P(x1/H2)＝ 0.14606742	P(x2/H2)＝ 0.06741573	P(x3/H2)＝ 0.38202247

They belong to H to X ₁, H ₂With the degree of belief of Θ be respectively as follows, table 2.

Table 2

mj(.)	Crit1	Crit2	Crit3
				Hypothese 1	m1(H1)＝ 0	m2(H1)＝ 0.20682763	m3(H1)＝ 0.08391414
Hypothese 2	m1(H2)＝ 0.58308047	m2(H2)＝ 0	m3(H2)＝ 0
				Ensemble	m1(Θ)＝ 0.41691953	m2(Θ)＝ 0.79317237	m3(Θ)＝ 0.91608586

Can get following result of calculation according to formula (2) and (3):

m(H ₁)＝0.13559，m(H ₂)＝0.50402，m(Θ)＝0.36039.

And then the image category that can be drawn under the X by decision rule is the second class image.

By above example as can be known, thereby this method can merge the identification of finishing texture image to multiple information sources effectively. through experiment to 100 image patterns, adopt the recognition correct rate of three kinds of methods all to be lower than 85% separately, the recognition correct rate that wherein utilizes method 1 is 82%, utilizing the recognition correct rate of method 2 is 84%, is 79% and utilize the recognition correct rate of method 3.Having adopted after the blending algorithm based on above three kinds of methods that this patent proposed, is 93% to the recognition correct rate of same 100 image patterns, and visible this method can obtain the better recognition effect in actual applications.

Above said content only is the basic explanation of the present invention under conceiving, and according to any equivalent transformation that technical scheme of the present invention is done, all should belong to protection scope of the present invention.

Claims (1)

1. the recognition methods of a multi-source texture graphics comprises the following steps:

A. the characteristic parameter of computed image " angle ": the gray scale of image is considered as height, each pixel and its neighbours territory pixel can constitute one " pyramid ", calculate the angle of four faces, the average of these angles is expressed as " angle " characteristic parameter of this image;

B. the characteristic parameter of computed image " wavelet transformation ": utilize wavelet transformation that the low frequency component and the high fdrequency component of image are separated, can obtain the coefficient of wavelet transformation, the part of its medium-high frequency is corresponding to the details of image, calculate the mean square deviation of these presentation video detail coefficients, as the characteristic parameter of image " wavelet transformation ";

C. the characteristic parameter of computed image " gradient ": the gradient correspondence of image the localized variation of image, the average of computing formula computed image gradient of utilizing two-dimentional gradient is as the characteristic parameter of image " gradient ";

D. by calculate above (a) and (b) and (c) in characteristic parameter, multiple information sources during promptly as image recognition, adopt the DS evidence theory that the result of multiple information sources gained is merged, at first according to the model of elementary probability partition function, obtain the elementary probability partition function of different aforementioned sources to different identifying objects, the rule (1) according to data fusion merges it afterwards:

$m\left(H\right)=m_1\left(H\right)\⊕...\⊕m_a\left(H\right),\∀H\∈2^{\mathrm{\Ω}}---\left(1\right)$

Here, H represents to suppose the object discerned, m ₁(H) ..., m _a(H) represent the degree of belief of different information sources for this identifying object, Ω represents the set of identifying object,

The operational symbol of then representing fusion,

Concrete fusion formula is as follows:

$m\left(H\right)=\frac{1}{1-k}\underset{H^{\′}\∩H^{\′\′}=H}{\mathrm{\Σ}}{m}_1\left(H^{\′}\right)m_2\left(H^{\′\′}\right)---\left(2\right)$

Wherein,

$k=\underset{H^{\′}\∩H^{\′\′}=\mathrm{\φ}}{\mathrm{\Σ}}{m}_1\left(H^{\′}\right)m_2\left(H^{\′\′}\right)---\left(3\right)$

In fusion process, obtain the basic trust degree of a plurality of information sources for identifying object, next, obtain the degree of belief of different aforementioned sources, in order to increase the order of accuarcy of identification for each identifying object according to (2) and (3), introducing can't be distinguished this situation, represents with Θ;

At last, according to corresponding criterion, obtain the affiliated image category of image to be identified.