CN101551864A - Image classification method based on feature correlation of frequency domain direction - Google Patents

Abstract

The invention discloses an image classification method based on the feature correlation of frequency domain direction, which mainly overcomes the shortages that the existing method has high calculation complexity, low classification precision and low robustness on image size variation. The method comprises the following steps of: (1) selecting a texture sample image and classifying the texture sample image into an image data set of training sample and an image data set of test sample; (2) implementing 2 dimensional fast Fourier transformation on the training sample image and dividing frequency domain direction subband according to the frequency and direction of the Fourier surface to obtain a frequency domain direction feature matrix; (3) calculating and obtaining a correlation pair sequence according to the correlation among subband features of the frequency domain direction feature matrix; (4) using unitary linear regression model for calculating the classification feature parameters of each correlation pair to form a classifier; (5) fitting the frequency domain direction features of the test sample image with the classifier parameters to obtain a classification label of the test sample; and (6) repeating step (5) to obtain the classification labels of all test samples. The i image classification method can be used for classifying the Brodatz texture images and the SAR images.

Description

Image classification method based on feature correlation of frequency domain direction

Technical field

The invention belongs to technical field of image processing, particularly relate to a kind of image classification method, can be used for classification texture image and synthetic aperture radar (SAR) image.

Background technology

Image classification is an important branch of pattern-recognition, is the different characteristic that is reflected in image information according to different classes of target, the image processing method that they are made a distinction.The main research contents of image classification is how image to be carried out suitable description, and extraction is the feature of presentation video attribute effectively, proposes effective classifying identification method, and on this basis image is carried out the classification of precise and high efficiency.

The application of image classification mainly contains the following aspects: analyzing image texture, Image Retrieval, target detection and identification etc.Wherein image texture analyses and classification problem are important research directions in Flame Image Process and the pattern-recognition, at image classification, cut apart, field such as computer graphics and picture coding all plays crucial effects.Produced a lot of traditional sorting techniques the eighties in 20th century, as gray level co-occurrence matrixes, and second-order statistics method, Gauss-Markov random field, local linear transformation etc.Along with going deep into that the human visual system is studied, many texture analysis models of differentiating begin to grow up more, as wavelet transformation, and Gabor conversion, Brushlet, profile ripple (Contourlet) etc.Researchers have proposed a large amount of innovations and improvement in conjunction with methods such as hyperchannel Gabor filtering, wavelet transformations to texture analysis, have improved the precision of texture analysis to a great extent.Adopt the wavelet basis that is fit to texture analysis that the textile texture is carried out damaged detection as Jasperetal, people such as Ajay Kumar and Granthan K.H Pang with Gabor filtering be used to have the damaged detection of object structures of texture phenomenon, people such as K.N.Bhanu Prakash utilize gray level co-occurrence matrixes to the lung's excusing from death image detection of fetus in the parent its whether to the maturity stage.The method that experimental results demonstrate this multiresolution analysis can obtain classifying quality preferably, has therefore obtained using widely in graphical analysis and sort research.

Traditional image classification method generally is to utilize the vector distance of textural characteristics or statistics difference to judge category attribute, image classification based on feature correlation then is based on this fact: image is that the texture information by special frequency band and direction combines, and this visually is reflected as different classes of image and has visibly different correlativity on different feature passages.Therefore, this correlativity is to distinguish a notable feature of different classes of texture.People such as Zhi-Zhong Wang and Jun-Hai Yong have carried out analyzing and proposing corresponding image search method to the correlativity of each intersubband of wavelet packet.This method at first obtains the energy feature of each subband, analyzes the interchannel correlativity of each feature then and asks and obtain sorting parameter.By the fitting degree of compare test sample and training sample correlation models, get rid of successively during test until obtaining correct class mark.Comprise that with the similar method of this thinking conversion such as utilizing small echo, profile ripple carries out feature extraction and correlation analysis, but the common drawback of this class conversion is a subband after frame fixation and the conversion divide at aspects such as frequency, directions careful inadequately.These shortcomings cause feature correlation obvious inadequately, and classification performance is limited and the size variation of image showed not enough robust, and computation complexity is higher.

Summary of the invention

The objective of the invention is to overcome the deficiency of above-mentioned prior art, a kind of image classification method based on feature correlation of frequency domain direction that is not subject to conversion such as small echo, wavelet packet and multi-scale geometric analysis tool framework is proposed, to improve, reduce computation complexity to the robustness of picture size variation and the accuracy of classification.

Technical scheme of the present invention is, subgraph in the training sample set is carried out 2 dimension fast fourier transform respectively, with direction Fourier plane is divided into different frequency domain direction subbands according to frequency then, calculate each sub belt energy feature, correlativity between analytical characteristic, and utilize one-variable linear regression to obtain sorting parameter, the fitting degree of the sorting parameter model of feature by the compare test subgraph and all kinds of images obtains classification results at last.The specific implementation process is as follows:

(1) chooses the sample image of all kinds of textures, and these sample images are divided into training sample image and two data sets of test sample image;

(2) the training sample image data set is carried out 2 dimension fast fourier transform respectively, with direction Fourier plane is divided into different frequency domain direction subbands according to the frequency of Fourier plane;

(3) calculate the energy feature of each subband, obtain the frequency domain direction eigenmatrix M of image;

(4) related coefficient between each subband feature of calculating frequency domain direction eigenmatrix M is carried out descending sort according to the related coefficient that each subband is right, obtains relevant to sequence;

(5) using Linear Regression Model in One Unknown tries to achieve relevant to the relevant right characteristic of division parameter of each feature in the sequence, composition and classification device;

(6) fitting degree with all parameters in the frequency domain direction feature of test sample image and the sorter compares, and calculates the probability that this test sample book belongs to each class image, gets the class mark of the class mark of probability maximal value correspondence as this sample;

(7) repeating step (6), the class that obtains all sample images of test sample image data centralization is marked.

The present invention has the following advantages compared with prior art:

1. because the present invention adjusts the direct control of Fourier plane and sector, overcome the framework restriction of conversion such as small echo, wavelet packet, make feature at multiband, that aspect such as multi-direction is divided is more careful, be significantly improved based on the classification performance of correlativity;

2. because frequency domain direction feature extracting method subband parameter of the present invention is adjustable, make this invention without limits, and different big or small images are had better robustness the minimum dimension of image;

3. because the present invention adopts fast fourier transform, make computation complexity reduce greatly, the feature extraction time is starkly lower than other several conversion.

Description of drawings

Fig. 1 is a realization flow synoptic diagram of the present invention;

Fig. 2 is a frequency domain direction sub-band division synoptic diagram of the present invention;

Fig. 3 be among the present invention when α gets different value, the synoptic diagram of corresponding subband center r (n);

Fig. 4 is that the subband of Brodatz texture image D6 is relevant to 28 and 29 characteristic distribution synoptic diagram;

Fig. 5 is the employed SAR data texturing of emulation experiment figure.

Embodiment

With reference to Fig. 1, specific implementation process of the present invention is as follows:

Step 1 is chosen the sample image of all kinds of textures, and these sample images are divided into training sample image and two data sets of test sample image.

The present invention uses two image data sets and carries out performance test: Brodatz texture image and SAR texture image.

1a) choosing method of Brodatz texture image sample data collection is described as follows:

77 class even grain images in the selection standard Brodatz texture storehouse are as test data, and this 77 class texture is: D1, D3, D4, D5, D6, D8, D9, D11, D14, D16, D17, D18, D19, D20, D21, D22, D23, D24, D25, D26, D27, D28, D29, D32, D33, D34, D35, D36, D37, D38, D46, D47, D48, D49, D50, D51, D52, D53, D54, D55, D56, D57, D64, D65, D66, D68, D74, D75, D76, D77, D78, D79, D80, D81, D82, D83, D84, D85, D87, D88, D92, D93, D94, D95, D96, D98, D100, D101, D102, D103, D104, D105, D106, D109, D110, D111, D112.

Above-mentioned every width of cloth texture image size is 640 * 640, chooses mode according to following two drawings of seeds, and 77 class even grain images are set up two databases respectively, and is as shown in table 1:

Table 1Brodatz texture dataset is provided with

Test database	The texture classes number	The subgraph size	Every class number of samples	Total sample number	Every class testing/training sample number
						Texture storehouse 128	77	128×128	25	1925	10/15
Texture storehouse 64	77	64×64	100	7700	40/60

In the table 1, texture storehouse 128 is to be nonoverlapping 25 width of cloth subgraphs with the uniform cutting of described 77 class images, the size of every width of cloth subgraph is 128 * 128, its subgraph adds up to 1925 width of cloth, in 25 width of cloth subgraphs, 10 width of cloth are as training sample, and 15 width of cloth are as test sample book, be that subgraph adds up to 770 width of cloth in the training sample image data set, subgraph adds up to 1155 width of cloth in the test sample image data set; Texture storehouse 64 is to be nonoverlapping 100 width of cloth subgraphs with the uniform cutting of described 77 class images, the size of every width of cloth subgraph is 64 * 64, its subgraph adds up to 7700 width of cloth, in 100 width of cloth subgraphs, 40 width of cloth are as training sample, 60 width of cloth are as test sample book, and promptly the training sample image data set comprises subgraph 3080 width of cloth, and the test sample image data set comprises subgraph 4620 width of cloth.

1b) choosing method of SAR image pattern data set is described as follows:

SAR image classification database is taken from the different texture zone of the true SAR image of 3 width of cloth, comprising 5 class textures, Fig. 5 has shown used this 5 class SAR data texturing figure in the experiment, from left to right is respectively the city under mountain range, waters, farmland and the high-resolution and low-resolution from top to bottom.Respectively in the uniform SAR texture region of this 5 class at random get 2000 points, slide the operation that window extracts subgraph, the subgraph size is that 128 * 128 database is designated as SAR128, the subgraph size is that 64 * 64 database is designated as SAR64, every class texture all comprises 2000 subgraphs, 500 width of cloth are as training sample, 1500 width of cloth are as test sample book, therefore the subgraph sum is 10000 width of cloth among SAR128 and the SAR64, wherein the training sample image data set comprises 2500 width of cloth subgraphs, the test sample image data set comprises 7500 width of cloth subgraphs, and is as shown in table 2:

Table 2SAR image data set is provided with

Test database	The texture classes number	The subgraph size	Every class number of samples	Total sample number	Every class testing/training sample number
						SAR128	5	128×128	2000	10000	500/1500
SAR64	5	64×64	2000	10000	500/1500

Step 2 is carried out 2 dimension fast fourier transform to given training sample image, and according to the frequency and the direction of Fourier plane Fourier plane is carried out the frequency domain direction sub-band division.

After image carried out 2 dimension fast fourier transform, obtaining Fourier plane, is the center with the DC component of Fourier plane, and the Fourier poincare half plane is obtained K frequency domain direction subband according to the method for frequency domain direction sub-band division, and its specific implementation process is as follows:

2a) Fourier plane the first half is divided into N frequency band by polar coordinates, D direction obtains K=N * D+2 sector, and different frequency bands and the directional information from the low frequency to the high frequency represented in each sector, constitutes different subbands, as shown in Figure 2.D=8 among Fig. 2, the position at sector 13 and 26 places among Fig. 2 because the information that comprises is few, is brought row into as a son and is handled, and D generally gets 16, and N generally gets 6;

2b) calculate each subband central point (r according to following formula _n, θ _d) the position:

$r_n={\left(\frac{n}{N}\right)}^{\mathrm{\α}}(R-1)+1,n=1,\·\·\·,N$

${\mathrm{\θ}}_d=\frac{\mathrm{\π}}{D}\·d,d=1,\·\·\·,D$

Wherein n represents frequency band, and N is total number of frequency band, and d represents the polar coordinates direction of Fourier poincare half plane, and D is total direction number, and R represents the maximum radius of Fourier plane, and α is adjustable parameter.When Fig. 3 represents that α gets different value, radius corresponding r _nValue, when α=1, r _nBe linear function, be illustrated on the radius and evenly get a little, when α increases, r _nTransfer nonlinear function to, it is a little more intensive to make the low frequency position get, and that HFS is got is a little sparse, meets the thought of the sub-band division in the multi-scale geometric analysis.Repeatedly experimental result shows, gets α=1.5, D=16, and the result is comparatively sane during N=6;

2c) calculate the size delta r of each subband according to following formula _n:

Δr _n＝r _n-r _n-1，n＝1，…，N

Wherein, the subband size on the identical frequency band different directions is identical.

Step 3 according to the energy feature of a following formula calculating K subband, obtains proper vector V={v _k, k=1 ..., K}:

$v_k=\frac{1}{\left|A_k\right|}\underset{j\∈A_k}{\mathrm{\Σ}}\left|c_j\right|$

Wherein | c _j| the absolute value of expression sub-band coefficients, A _kBe the coefficient coordinate set of k subband, || the expression set sizes.

Step 4, repeating step 2～3 to each class texture, calculates the proper vector of all training samples, is column vector constitutive characteristic matrix M with the proper vector.

Step 5 is calculated the relevant to sequence of each class image.

5a) calculate the covariance matrix C of certain category feature matrix M, the coefficient c in the Matrix C _IjBe the related coefficient of subband i and subband j, it is right that subband i, j constitute subband;

5b) with each subband to carrying out descending sort by related coefficient ρ, select wherein ρ＞T _αBe that correlativity is relevant significantly relevant to sequence to putting into, output is relevant, considers the requirement of speed and precision here to sequence, generally gets T _α=0.4;

5c) repeating step 5a)～5b), calculate the relevant of each class image to sequence.

Step 6 is asked sorting parameter matrix X, the composition and classification device.

6a) use Linear Regression Model in One Unknown, calculate the relevant to each relevant right disaggregated model parameter in the sequence of each class texture, establishing i relevant right parameter is a _i, b _i, μ _i, σ _i, computing formula is as follows:

$a_i=\frac{m{\mathrm{\Σ}}_{h=1}^m{x}_{\mathrm{ih}}{y}_{\mathrm{ih}}-{\mathrm{\Σ}}_{h=1}^m{x}_{\mathrm{ih}}{\mathrm{\Σ}}_{h=1}^m{y}_{\mathrm{ih}}}{m{\mathrm{\Σ}}_{h=1}^m{x_{\mathrm{ih}}}^2-{\left({\mathrm{\Σ}}_{h=1}^m{x}_{\mathrm{ih}}\right)}^2}$ $b_i=\frac{1}{m}\underset{h=1}{\overset{m}{\mathrm{\Σ}}}{y}_{\mathrm{ih}}-\frac{a}{m}\underset{h=1}{\overset{m}{\mathrm{\Σ}}}{x}_{\mathrm{ih}}$

${\mathrm{\μ}}_i=\frac{{\mathrm{\Σ}}_{h=1}^m(y_{\mathrm{ih}}-{\hat{y}}_{\mathrm{ih}})}{m}$ ${\mathrm{\σ}}_i=\sqrt{\frac{{\mathrm{\Σ}}_{h=1}^m{(y_{\mathrm{ih}}-{\hat{y}}_{\mathrm{ih}})}^2}{m-2}}$

Wherein, x _Ih, y _IhRepresent i of h sample of certain class image to be correlated with respectively to two sub belt energies,

Be by equation of linear regression ${\hat{y}}_i=a_i\×x_i+b_i$ What obtain should be relevant to corresponding subband energy y _IhEstimated value, m is a sample size, when two subband correlativitys are higher, equation of linear regression can well two intersubbands of match relation, Fig. 4 be texture image D6 relevant its related coefficient is 0.9952 to 28 and 29 energy profile, Fig. 4 cathetus is that equation of linear regression obtains, can see, this straight line can well the match subband relevant to relation;

6b) according to the descending of related coefficient, with each relevant right parameter a of each class texture, b, μ, σ relevantly puts into sorting parameter matrix X to label and these parameters of related coefficient ρ, as the characteristic of division parameter of all kinds of textures, composition and classification device.

Step 7 is carried out 2 dimension fast fourier transform to given test sample image, by abovementioned steps 2～3 described frequency domain direction feature extracting methods, calculates its frequency domain direction proper vector V then;

Step 8 compares the fitting degree of all parameters in the proper vector V of given test sample book and the sorter, obtains the class mark of this test sample book.

8a) take out i relevant right Linear Regression Model in One Unknown sorting parameter a of j class image _i, b _i, μ _i, σ _iAnd should relevant subband feature (x in the test sample book feature to correspondence _i, y _i);

8b) according to the one-variable linear regression equation $\hat{y}=a\×x+b$ Try to achieve subband feature y _iEstimated value The error of calculation

8c) obtain test sample book and meet each relevant Probability p of j class image the sorting parameter model according to following formula _Ij:

$p_{\mathrm{ij}}=\left\{\begin{array}{c}1,\mathrm{if}|y_i-{\hat{y}}_i-{\mathrm{\&mu;}}_i|<{3\mathrm{\&sigma;}}_i\\ 0,\mathrm{if}|y_i-{\hat{y}}_i-{\mathrm{\&mu;}}_i|>3{\mathrm{\&sigma;}}_i\end{array}\right.,i=1,\&CenterDot;\&CenterDot;\&CenterDot;,L,$

Wherein L is that j class image is relevant to relevant right total number in the sequence;

8d) calculate the probability P that test sample book belongs to j class image _j:

$P_j=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}};$

8e) repeating step 8a)～8d), calculate the probability that this test sample book belongs to each class image, obtain probability set P={P _j, j=1 ..., S}, wherein S is total classification number of image set;

8f) with step 8e) the probability set P that obtains is by descending sort, if most probable value P _MaxUnique, then test sample book is classified as the class of maximum probability correspondence; If P _MaxNot unique, then this test sample book is included into the region of rejection, show that the class mark of this sample is uncertain.

Step 9, repeating step 8 is asked the class mark and the output of all test sample image of test sample image data centralization.

Effect of the present invention can further specify by following emulation experiment.

1. emulation content: the Brodatz texture image of the present invention and small echo, wavelet packet, profile ripple and the experiment of SAR image classification.

2. simulated conditions: Intel (R) Pentium (R) 4CPU, 3.00GHz, Windows XP system, Matlab7.4.0 operation platform.

3. The simulation experiment result:

Respectively texture image and SAR image are classified in the experiment.

The A.Brodatz texture image classification

Experiment parameter is provided with as shown in table 3.Consider the robustness of small echo, wavelet package transforms, its subband minimum dimension is 16 * 16, so the decomposition number of plies of database 128 and database 64 is different, gets 3 layers and 2 layers of conversion respectively.The direction number that the profile wavelength-division is separated and the number of plies are according to experimental result, and what get is optimized parameter.Decompose for frequency domain direction of the present invention, the repeatedly experiment of a plurality of databases shows, gets parameter alpha=1.5, D=16, and classification results is more sane during N=6.

Table 3Brodatz Texture classification experiment parameter is provided with

Picked at random 10 groups of training, test sample books experimentize, and ask the average result of 10 subseries, and following experimental result is arranged:

1) method of using the correlativity classification is classified to Brodatz texture storehouse 128 and texture storehouse 64, obtains the experimental result shown in following table 4,5.

The correlativity classification results in table 4Brodatz texture storehouse 128

Feature	Correctly	Refusal	Mistake	Standard deviation
						3 layers of decomposition of small echo	85.82	5.95	8.23	1.74
3 layers of decomposition of wavelet packet	87.64	0	12.36	2.00
					2 layers of decomposition of profile ripple, direction number 4,8	83.51	5.31	11.19	3.15
The present invention (α=1.5, D=16, N=6)	96.04	0	3.96	0.87

The correlativity classification results in table 5Brodatz texture storehouse 64

Feature	Correctly	Refusal	Mistake	Standard deviation
						2 layers of decomposition of small echo	23.95	73.79	2.26	5.04
2 layers of decomposition of wavelet packet	52.72	38.72	8.56	1.26
					2 layers of decomposition of profile ripple, direction number 4,8	46.00	48.68	5.32	1.45
The present invention (α=1.5, D=16, N=6)	83.43	0	16.57	0.74

From table 4, table 5 as can be known, frequency domain direction feature application of the present invention in the correlativity sorting technique, can be obtained good classifying quality.With widely used based on wavelet transformation sorting technique and multi-scale geometric analysis method profile Bob, the classification accuracy rate that the present invention obtains improves significantly, for texture storehouse 128, classification accuracy rate of the present invention improves nearly 9 percentage points at least, for texture storehouse 64, advantage is more obvious, and accuracy improves 24% at least, this result fully verified feature of the present invention on frequency and direction, divide more careful, the advantage that classification performance is good.

By comparing the classification results in texture storehouse 128 and texture storehouse 64, we find that along with reducing of subgraph size, classification performance of the present invention descends well below other several conversion.Trace it to its cause, be since the frequency domain direction feature extracting method at the frequency domain direct control, and parameter is adjustable, the subband variable size, therefore the variation of picture size is little to this method affect, and other several conversion are subjected to the influence of transformation framework, consider robustness, general antithetical phrase mark is very little has relatively high expectations, therefore restricted to the minimum dimension of image.Experimental result has proved absolutely that classification performance of the present invention is better than other several methods, and picture size is had robustness preferably.

2) the feature extraction time, the time of frequency domain direction feature extraction of the present invention is starkly lower than other several conversion, and its computation complexity is low.Following table is that the texture image to one 1024 * 1024 size carries out conversion, extracts the used time of its energy feature.

The feature extraction time of the various conversion of table 6, the image size is 1024*1024

Conversion	3 layers of decomposition of small echo	3 layers of decomposition of wavelet packet	2 layers of decomposition of profile ripple, direction number 4,8	The present invention (α=1.5, D=16, N=6)
					The feature extraction time	1.07s	4.38s	5.76s	0.73s

Be the feature extraction time of several conversion shown in the table 5, experimental result proves absolutely because the present invention uses fast fourier transform and realizes, so the feature extraction time be starkly lower than other several conversion, computation complexity reduces greatly.

The B.SAR image classification

Experiment parameter is provided with as shown in table 7.The parameter setting and the Brodatz texture image of the experiment of SAR image classification are similar.

Table 7SAR classification experiments parameter is provided with

Two database SAR128 and SAR64 to the SAR image classify, and obtain being depicted as average classification results 10 times as table 8, table 9.

The correlativity classification results of table 8 database SAR128

Feature	Correctly	Refusal	Mistake	Standard deviation
						3 layers of decomposition of small echo	95.3	2.94	1.76	0.41
3 layers of decomposition of wavelet packet	79.18	8.32	12.5	4.50
					2 layers of decomposition of profile ripple, direction number 4,8	84.00	12.66	3.34	8.60
The present invention (α=1.5, D=16, N=6)	96.43	0	3.57	0.34

The correlativity classification results of table 9 database SAR64

Feature	Correctly	Refusal	Mistake	Standard deviation
						2 layers of decomposition of small echo	58.87	38.20	2.93	7.81
2 layers of decomposition of wavelet packet	55.48	39.57	4.95	1.24
					2 layers of decomposition of profile ripple, direction number 4,8	58.02	36.64	5.34	1.50
The present invention (α=1.5, D=16, N=6)	92.49	0	7.51	0.70

Result shown in table 8, the table 9 shows that the classifying quality that frequency domain direction of the present invention decomposes is best.For database SAR128, classification accuracy rate of the present invention is higher by 1.13% than wavelet transformation, and is higher by 12.43% than profile wave convert, higher by 17.25% than wavelet package transforms; Database SAR64, classification results advantage of the present invention is more remarkable, exceeds more than 30 percentage point than other three kinds of conversion, has fully verified validity of the present invention.

In addition, because the texture information that the subgraph of 64 * 64 sizes comprises is abundant not as 128 * 128 subgraphs, more be difficult to distinguish the classification of texture relatively, so classification accuracy rate will descend.But the present invention has only descended 4%, and additive method descends significantly, reduces by 24% at least.This experimental result has verified that further the present invention has robustness preferably for the variation of picture size.

Claims (4)

1, based on the image classification method of feature correlation of frequency domain direction, the specific implementation process is as follows:

(1) chooses the sample image of all kinds of textures, and these sample images are divided into training sample image and two data sets of test sample image;

(2) the training sample image data set is carried out 2 dimension fast fourier transform respectively, with direction Fourier plane is divided into different frequency domain direction subbands according to the frequency of Fourier plane;

(3) calculate the energy feature of each subband, obtain the frequency domain direction eigenmatrix M of image;

(4) related coefficient between each subband feature of calculating frequency domain direction eigenmatrix M is carried out descending sort according to the related coefficient that each subband is right, obtains relevant to sequence;

(5) using Linear Regression Model in One Unknown tries to achieve relevant to the relevant right characteristic of division parameter of each feature in the sequence, composition and classification device;

(6) fitting degree with all parameters in the frequency domain direction feature of test sample image and the sorter compares, and calculates the probability that this test sample book belongs to each class image, gets the class mark of the class mark of probability maximal value correspondence as this sample;

(7) repeating step (6), the class that obtains all sample images of test sample image data centralization is marked.

2, image classification method according to claim 1, wherein the described frequency according to Fourier plane of step (2) is divided into different frequency domain direction subbands with direction with Fourier plane, carries out as follows:

2a) DC component with Fourier plane is the center, the Fourier poincare half plane is divided into N frequency band and D direction by polar radius and direction, obtain K=N * D+2 sector, different frequency bands and the directional information from the low frequency to the high frequency represented in each sector, constitutes different subbands;

2b) utilize following formula to calculate each subband central point (r _n, θ _d) the position:

$r_n={\left(\frac{n}{N}\right)}^{\mathrm{\&alpha;}}(R-1)+1,$ n＝1，…，N

${\mathrm{\&theta;}}_d=\frac{\mathrm{\&pi;}}{D}\&CenterDot;d,$ d＝1，…，D

Wherein n represents frequency band, and N is total number of frequency band, and d represents the polar coordinates direction of Fourier poincare half plane, and D is total direction number, and R represents the maximum radius of Fourier plane, and α is adjustable parameter, and D generally gets 16, and N generally gets 6, and α generally gets 1.5;

2c) calculate subband size delta r on each frequency band according to following formula _n:

Δr _n＝r _n-r _n-1，n＝1，…，N，

Wherein, the subband size on the identical frequency band different directions is identical;

2d) with (r _n, θ _d) be the subband center, Δ r _nDraw window for the subband size, obtain K frequency domain direction subband.

3, image classification method according to claim 1, the probability that this test sample book of the described calculating of step wherein (6) belongs to each class image, carry out according to the following procedure:

3a) take out certain class image i relevant right Linear Regression Model in One Unknown sorting parameter a _i, b _i, μ _i, σ _iAnd should relevant subband feature (x in the test sample book feature to correspondence _i, y _i);

3b) according to the one-variable linear regression equation $\hat{y}=a\&times;x+b$ Try to achieve subband feature y _iEstimated value The error of calculation

3c) obtain test sample book and meet such each relevant probability P the sorting parameter model according to following formula _i:

$P_i=\left\{\begin{array}{c}1,\mathrm{if}|y_i-{\hat{y}}_i-{\mathrm{\&mu;}}_i|<{3\mathrm{\&sigma;}}_i\\ 0,\mathrm{if}|y_i-{\hat{y}}_i-{\mathrm{\&mu;}}_i|>{3\mathrm{\&sigma;}}_i\end{array}\right.,$ i＝1，…，L，

Wherein L is that j class image is relevant to relevant right total number in the sequence;

3d) calculate the probability P that test sample book belongs to such image:

$P=\frac{1}{L}\underset{i=1}{\overset{L}{\mathrm{\&Sigma;}}}{P}_i;$

3e) repeating step 3a)～3d), calculate the probability that this test sample book belongs to each class image.

4, image classification method according to claim 1, wherein the described class mark of getting the class mark of probability maximal value correspondence as sample of step (6) is by descending sort, if most probable value P with all probability that obtain _MaxUnique, then test sample book is classified as P _MaxCorresponding class; If P _MaxNot unique, then this test sample book is included into the region of rejection.

Images