CN101587189A - Texture elementary feature extraction method for synthetizing aperture radar images - Google Patents

Abstract

The invention relates to a texture elementary feature extraction method for synthetizing aperture radar images, comprising the followings: the estimated vectors corresponding to each pixel point of each image of image in a training set are obtained, and the estimated vectors corresponding to all pixel points of all images in the training set are combined; the estimated vectors are clustered to obtain texture elementary code that is texture elementary element; the estimated vectors corresponding to each pixel point in image are obtained for the feature image to be extracted, the estimated vectors corresponding to all pixel points in image are obtained after combination, and the estimated vectors are marked by using the texture elementary code; and the statistics of image marks is carried out to obtain mark histogram that is the statistics texture elementary feature of image. The invention is combined with distributional characteristics, space characteristics and dimension characteristics of image to improve the classification correctness of synthetizing aperture radar images, thereby improving the accuracy of image processing application such as classification and division based on texture elementary feature.

Description

The texture primitive feature extracting method that is used for diameter radar image

Technical field

The invention belongs to technical field of image processing, particularly a kind of texture primitive feature extracting method that is used for diameter radar image.

Background technology

In computer vision and Flame Image Process, the variation of brightness of image forms the pattern of certain specific repetition, is referred to as texture, and texture is a kind of ubiquitous visual phenomenon, and people can remove to experience texture, but is difficult to the explication of texture is formed unified understanding.Still none widely accepted texture definition at present.It is generally acknowledged that texture is gradation of image or color variation or repetition spatially.Intensity profile generally has certain regularity in the texture image, and for random grain, it also has the feature on some statistical significances.People have following common recognition to texture at present:

A) texture shows as the repetition constantly in the bigger zone of this sequence of certain local sequentiality;

B) texture exists the basic comprising unit that causes visually-perceptible, i.e. texture primitive;

C) texture can not be processed into a point process, shows as region characteristic more;

D) the texture region various piece roughly is uniform entity, and each several part has roughly the same size;

E) texture has features such as intensity, density, direction and degree of roughness.

Based on above common recognition, think that texture has two key elements: (1) texture has the elementary cell that causes visually-perceptible, i.e. texture primitive.The form of texture primitive is various, shows as some image color or grayscale mode.(2) texture primitive has certain queueing discipline, and these rules may show as certain regularity, also may show as randomness [Liu Xiaoming, texture Review Study. computer utility research, 2008, Vol.25, No.8].

Texture analysis is one of main contents of texture research, it also is important field of research in the machine vision, boundless application background is arranged, and its application comprises that remote Sensing Image Analysis (Remotely-sensedImage Analysis), medical image analysis (Medical Image Analysis), industrial surface detect fields such as (Industrial Surface Inspection), document process (Document Processing) and image retrieval.

Remote Sensing Image Analysis is a very ripe application of texture analysis, and the area of different terrain landforms can produce different textures in remote sensing images, and river, lake, field, urban area can make a distinction them by texture.People's usage space gray scales such as Haralick rely on the matrix textural characteristics and analyze remote sensing images, Rignot and Kwok analyze the SAR image with textural characteristics, Du uses the wave filter texture of deriving out from Gabor to cut apart texture, successfully is water, fresh ice, ice and ice for many years for a long time with the SAR image segmentation.2003, people such as Christodoulou developed a texture analysis system based on many features, and this system can analyze satellite cloud picture automatically, dissimilar clouds were classified, for synoptic analysis provides a useful analysis means.

A key problem of texture analysis is texture description (Texture Description), is texture feature extraction (Texture Feature Extraction) at area of pattern recognition.Many texture characteristic extracting methods have been arranged at present.Tuceryan and Jain roughly are classified as four big classes with these methods: structure analysis method, statistical analysis technique, modelling analytical approach and signal processing method.Wherein statistical analysis technique and signal processing method are being served as very important role in texture analysis.

Texture analysis research at home mainly is the concrete application of a certain method.For the method for statistics, the co-occurrence matrix method is more commonly used.In the method based on model, fractal method is used many, adopts Fractal Brown function mostly, and there also have pair fractal method to carry out to be improved; The application of Markov random field (MRF) also has, and main difficulty is determining of parameter.In the method for mathematic(al) manipulation, commonly based on method of wavelet.External mainly is that textural characteristics that several texture analysis methods are extracted is in conjunction with the general classification method, to the comparison of classifying of different images.

Early stage texture analysis uses the method for statistics or structure to extract feature.Nearest major progress is to use the textural characteristics of multiresolution (for example Gabor conversion) and hyperchannel (several wave band binding analysis textural characteristics) to describe.Past lacks texture analysis effectively to be analyzed the texture of different scale, and the different scale here is meant same width of cloth image, carries out texture analysis on different scale.Can obtain the different texture feature of the same area like this, increase quantity of information, finally can improve the precision of classification.

Texture description based on filtering can be divided into two kinds of methods: horizontally-spliced one-tenth histogram after the filtering, as GIST method [Aude Oliva, " Gist of the Scene ", Neurobiology of Attention 2005]; Vertically conspire to create a vector after the filtering, add up again, be called texture primitive method [Manik Varma, Andrew Zisserman, " A Statistical Approach to Texture Classificationfrom Single Images ", Kluwer Academic Publishers.Printed in theNetherlands, 2004].Texture primitive is meant micromechanism basic in the natural image, and it is the fundamental element of visually-perceptible starting stage (noting the stage in advance).The research of texture primitive is to utilize the thought of sparse coding (sparse coding) to attempt the super complete image-based shading reason of study primitive s from natural image, expresses texture image with texture primitive again.

Diameter radar image (SAR image) also is a kind of texture image, and its disposal route generally is based on statistical estimate, and traditional texture analysis method texture primitive is based on filtering, therefore combines statistical estimate and texture primitive necessary.

Summary of the invention

The object of the invention is to solve the prior art deficiency, has proposed to be used for the texture primitive feature extracting method of diameter radar image, can extract the feature of the texture image with distribution character efficiently.

Technical scheme of the present invention may further comprise the steps:

Step 1 is asked for each pixel P in the image to every width of cloth image in the training set _{X, y}Corresponding estimate vector V _{X, y}, combination obtains the estimate vector { V of all pixel correspondences of all images in the training set _{X, y}} _All,

Step 2 is with the estimate vector { V of all pixel correspondences of all images in the training set _{X, y}} _AllCarry out cluster, cluster gained texture primitive code book C is texture primitive;

Step 3 is asked for each pixel P in the image to characteristic image to be extracted _{X, y}Corresponding estimate vector V _{X, y}, combination obtains the estimate vector { V of all pixel correspondences in the image _{X, y}} _One, and to this estimate vector { V _{X, y}} _OneCarry out label with the texture primitive code book C that generates in the step 2, be designated as { M _{X, y}} _One

Step 4, statistical picture label { M _{X, y}} _OneObtain label histogram, be designated as H _Mark, label histogram H _MarkBe the statistics texture primitive feature of this image;

In step 1 and step 3, described interior each the pixel P of image that asks for _{X, y}Corresponding estimate vector V _{X, y}, concrete mode is as follows,

At first, to gained gray level image T process Gabor bank of filters { F behind the quantized image ₁, F ₂..., F _n, obtain filtered image sets { I ₁, I ₂..., I _n, be designated as I;

Then, to the image I among the image sets I _p, 1≤p≤n is respectively with each pixel P _{X, y}Select m different window size for the center and obtain pane group { G _{P, 1}, G _{P, 2}..., G _{P, m}, be designated as G _pWith each pane G _{P, i}, that 1≤i≤m divides is upper left, lower-left, upper right, four region R in bottom right _{P, i, 1}, R _{P, i, 2}, R _{P, i, 3}, R _{P, i, 4}, and these zones are combined obtain zone group { R _{P, i, j}, 1≤i≤m, 1≤j≤4 are designated as R _pIf each regional view data all satisfies with a kind of statistical distribution E, this distribution comprises q parameter, to R _pIn each region R _{P, i, j}Carry out parameter estimation, obtain estimated parameter E _{P, i, j}={ E _{P, i, j, 1}, E _{P, i, j, 2}..., E _{P, i, j, q}, they are combined obtains estimated parameter group { E _{P, i, j, k}, 1≤i≤m, 1≤j≤4,1≤k≤q is designated as E _p

At last, with pixel P _{X, y}Corresponding E _{P, i, j, k}, 1≤p≤n, 1≤i≤m, 1≤j≤4,1≤k≤q conspire to create an estimate vector, are each pixel P in the image _{X, y}Corresponding estimate vector V _{X, y}

And, with pixel P _{X, y}Corresponding E _{P, i, j, k}When conspiring to create an estimate vector, successively according to k, j, the order that i, p increase progressively is connected in series.

And, described to R _pIn each region R _{P, i, j}Carry out parameter estimation, adopt the maximal possibility estimation mode to realize.

The present invention replaces the pixel point value by introducing regional estimated value, has described the distribution character of pixel and surrounding pixel point thereof, has described the feature of pixel effectively; By adding the pixel distribution parameter in different azimuth zone on every side, make feature comprise the spatial relation of image, the situation of change of textural characteristics on the different directions has been described; By introducing the distribution parameter of different scale, increased description to the different scale textural characteristics.A kind of texture primitive feature extracting method that is used for diameter radar image that the present invention proposes, the distribution character of fused images, spatial character and dimensional properties, improve the classification accuracy rate of SAR image, thereby improve classification, the accuracy rate that Flame Image Process is used such as cut apart based on the texture primitive feature.

Description of drawings

Fig. 1 process flow diagram of the present invention;

The embodiment of Fig. 2 spatial characteristics of the present invention;

The embodiment of the multiple dimensioned distribution character of Fig. 3 the present invention;

The generative process of Fig. 4 texture primitive code book;

Fig. 5 utilizes the statistics texture primitive feature of texture primitive code book computed image;

Fig. 6 embodiment of the invention utilizes the texture primitive statistical nature that image is classified;

Embodiment

Referring to Fig. 1, technical solution of the present invention may further comprise the steps:

Step 1 is asked for each pixel P in the image to every width of cloth image in the training set _{X, y}Corresponding estimate vector V _{X, y}, combination obtains the estimate vector { V of all pixel correspondences of all images in the training set _{X, y}} _All,

Step 2 is with the estimate vector { V of all pixel correspondences of all images in the training set _{X, y}} _AllCarry out cluster, cluster gained texture primitive code book C is texture primitive;

Step 3 is asked for each pixel P in the image to characteristic image to be extracted _{X, y}Corresponding estimate vector V _{X, y}, combination obtains the estimate vector { V of all pixel correspondences in the image _{X, y}} _One, and to this estimate vector { V _{X, y}} _OneCarry out label with the texture primitive code book C that generates in the step 2, be designated as { M _{X, y}} _One

Step 4, statistical picture label { M _{X, y}} _OneObtain label histogram, be designated as H _Mark, label histogram H _MarkBe the statistics texture primitive feature of this image;

In step 1 and step 3, described interior each the pixel P of image that asks for _{X, y}Corresponding estimate vector V _{X, y}, concrete mode is as follows,

At first, to gained gray level image T process Gabor bank of filters { F behind the quantized image ₁, F ₂..., F _n, obtain filtered image sets { I ₁, I ₂..., I _n, be designated as I;

Then, to the image I among the image sets I _p, 1≤p≤n is respectively with each pixel P _{X, y}Select m different window size for the center and obtain pane group { G _{P, 1}, G _{P, 2}..., G _{P, m}, be designated as G _pWith each pane G _{P, i}, that 1≤i≤m divides is upper left, lower-left, upper right, four region R in bottom right _{P, i, 1}, R _{P, i, 2}, R _{P, i, 3}, R _{P, i, 4}, and these zones are combined obtain zone group { R _{P, i, j}, 1≤i≤m, 1≤j≤4 are designated as R _pIf each regional view data all satisfies with a kind of statistical distribution E, this distribution comprises q parameter, to R _pIn each region R _{P, i, j}Carry out parameter estimation, obtain estimated parameter E _{P, i, j}={ E _{P, i, j, 1}, E _{P, i, j, 2}..., E _{P, i, j, q}, they are combined obtains estimated parameter group { E _{P, i, j, k}, 1≤i≤m, 1≤j≤4,1≤k≤q is designated as E _p

At last, with pixel P _{X, y}Corresponding E _{P, i, j, k}, 1≤p≤n, 1≤i≤m, 1≤j≤4,1≤k≤q conspire to create an estimate vector, are each pixel P in the image _{X, y}Corresponding estimate vector V _{X, y}

During concrete enforcement, characteristic image to be extracted is decided according to the image applications needs, can be the image in test pattern or the training set, or the part of image to be split.When for example cutting apart, be characteristic image to be extracted, carry out step 3, step 4 with image in zone to be split and the training set; Dividing time-like, be characteristic image to be extracted with the image in test pattern or the training set, carries out step 3, step 4.When characteristic image to be extracted belongs in the training set image, because step 1 is asked for each pixel P in the image to every width of cloth image in the training set _{X, y}Corresponding estimate vector V _{X, y}, in the step 3 characteristic image to be extracted is asked for each pixel P in the image _{X, y}Corresponding estimate vector V _{X, y}Can economize the result that asks for who slightly directly quotes step 1.Subscript x, y are respectively horizontal stroke, the ordinate of pixel, on each image all with the respective pixel point P of same coordinate _{X, y}For object is estimated.

To R _pIn each region R _{P, i, j}Before carrying out parameter estimation, can be earlier to R _pIn each regional view data analyze, obtain these data and satisfy with a kind of statistical distribution E, for example Gamma distributes.Parameter estimation then can be selected existing techniques in realizing such as maximal possibility estimation mode for use, and the present invention will not give unnecessary details.Obviously, with each pixel P _{X, y}Corresponding E _{P, i, j, k}When conspiring to create an estimate vector, serial connection sequence should be consistent.The present invention advises successively according to k, j, and the order that i, p increase progressively is connected in series.

The present invention be directed to diameter radar image and extract the texture primitive feature, therefore the image in above-mentioned test pattern or the training set, the part of image to be split obviously all are diameter radar images.The gray level image T that obtains after the diameter radar image quantification is sent into Gabor bank of filters { F ₁, F ₂..., F _nGet final product.

Principle to technical solution of the present invention is illustrated below:

(1) bank of filters is selected

Textural characteristics can be made of the texture primitive of different texture primitives or various combination mode, therefore considers to assist the generation texture primitive by one group of wave filter.The standard of selective filter is in order to extract texture information better, and this texture information that requires to extract not only has a plurality of yardsticks and a plurality of directions are arranged.Wave filter commonly used has a lot, as the Gabor wave filter [Xue Yuli. based on the feature extraction and the application [D] thereof of Gabor conversion. Shandong University's master thesis, in April, 2007] etc.In the signal Processing field, the Gabor wave filter is one of best method of image identification, Gabor can be regarded as the edge that direction and yardstick all can change and the detecting device of straight line (striped), and, statistics for these microscopic features in the given area, can be used for representing basic texture information, so the present invention selects the Gabor wave filter.The two-dimensional Gabor function can be expressed as:

$g(x,y)=\left[\frac{1}{2\mathrm{\π}{\mathrm{\σ}}_x{\mathrm{\σ}}_y}\right]\exp [-\frac{1}{2}(\frac{x^2}{{{\mathrm{\σ}}_x}^2}+\frac{y^2}{{{\mathrm{\σ}}_y}^2})+2\mathrm{\πjWx}]$

Wherein, W is a Gaussian function multiple modulation frequency, and x, y are respectively the coordinate figure of two dimensions, σ _x, σ _yBandwidth, the j that is Gaussian function in spatial domain is imaginary symbols, exp exponential transform symbol.

During concrete enforcement, with n F ₁, F ₂..., F _nConstitute Gabor bank of filters { F ₁, F ₂..., F _n, 3 yardsticks are chosen in suggestion, and each yardstick has the Gabor wave filter of 8 directions, and then the suggestion value of n is 24.

(2) spatial characteristics of pixel

Do not extract the spatial characteristics of pixel when having the texture primitive feature extraction now, the present invention has considered the spatial characteristics of pixel when extracting the texture primitive feature.The texture primitive that specifically refers to pixel is not only calculated by this neighborhood of pixels, and be to calculate by the neighborhood on 4 different directions, the texture primitive of the pixel that obtains has so just comprised distribution character on the different spatial, can the distribution character of pixel and peripheral region thereof be retrained.Therefore with each pane G _{P, i}, that 1≤p≤n, 1≤i≤m divide is upper left, lower-left, upper right, four zones, bottom right, and these zones are combined, and just can obtain the center pixel spatial characteristics.M suggestion span is 1 ~ 3, G _{P, i}But pane size general recommendations value be 9,11,13 odd numbers such as grade.

(3) multiple dimensioned distribution character

Referring to accompanying drawing 2, wherein a pixel with certain gray level image I is an example, to the image I that obtains after the gray level image I filtering ₁, I ₂..., I _n, selecting pane to count m is 1, G _{P, 1}Pane size value be 9, show that to be that certain pane at center is divided upper left, lower-left, upper right, four zones, bottom right with this pixel, therefore can obtain behavior n * 4, classify 1 matrix as, be expressed as [] _(n*4,1)The multiple dimensioned distribution character of pixel

Texture exists on certain zone, the region area that some texture cell covers is big, the region area of some covering is little, original texture primitive extracting method obtains the texture primitive of different scale by selecting different bank of filters for use, and the present invention directly obtains multiple dimensioned characteristic from the neighborhood of the different sizes of pixel.Region area is too small can't to constitute texture in order to avoid, and selects different area size to estimate, to obtain multiple dimensioned distribution character.With each pixel P _{X, y}Select different window sizes for the center and obtain pane G _{P, i}, 1≤i≤m chooses the different sizes of estimating exactly, can obtain the multiple dimensioned distribution character of center pixel.

Referring to accompanying drawing 3, wherein a pixel with certain gray level image I is an example, and showing with this pixel is that 3 different size panes are got at the center, i.e. m=3, and get G _{P, 1}Pane be of a size of 5, G _{P, 2}Pane be of a size of 9, G _{P, 3}Pane be of a size of 13 and can obtain behavior n * 3 * 4, classify 1 matrix as, be expressed as [] _(n*3*4,1)

(4) the texture primitive code book chooses

Use the method for cluster that texture primitives all in the training set is carried out cluster, obtain cluster centre, as the typical texture primitive.The vector that is about to all images in the training set carries out cluster, obtains the representative of cluster centre as texture primitive.So greatly reduce the quantity of texture primitive, thereby reduced calculated amount.

Referring to accompanying drawing 4,, the generative process of texture primitive code book is described to comprise in the training set that N width of cloth image is an example: to gained gray level image T1 behind the x width of cloth image quantization, T2 ..., TN sends into Gabor bank of filters { F respectively ₀, F ₁..., F _n, filtering obtains image sets { I1 respectively ₁, I1 ₂..., I1 _n, { I2 ₁, I2 ₂..., I2 _n... { IN ₁, IN ₂..., IN _n.To the gained image sets, estimate to obtain estimated parameter group { E1 respectively then ₁, E1 ₂..., E1 _n, { E2 ₁, E2 ₂..., E2 _n... { EN ₁, EN ₂..., EN _n, wherein comprising m pane, the sub-piece of 4 directions and q estimated parameter are example to scheme IN, with pixel P _{X, y}Corresponding EN _{P, i, j, k}, 1≤p≤n, 1≤i≤m, 1≤j≤4,1≤k≤q be successively according to p, i, and the order that j, k increase progressively conspires to create estimate vector VN _{X, y}, promptly be followed successively by EN _1,1,1,1, EN _1,1,1,2... EN _{1,1,1, q}, EN _1,1,2,1... EN _{1, m, 4, q}, EN _2,1,1,1... EN _{N, m, 4, q}.

Each image I 1 ₁, I1 ₂..., I1 _n..., I2 ₁, I2 ₂..., I2 _n..., IN ₁, IN ₂..., IN _nAll pixel P in the corresponding image _{X, y}Estimate vector set on image can be designated as V={V _{X, y}| P _{X, y}∈ (I1, I2 ..., IN) }.

At last the estimate vector among the V is carried out cluster, gather into the k class, generate the texture primitive code book.

(5) label histogram of image

At first with the texture primitive number consecutively in the texture primitive code book, texture primitive with all pixels of piece image carries out label according to the numbering of texture primitive code book then, again label is added up by the numerical values recited of label, obtained label histogram, as the feature of this image.

Referring to accompanying drawing 5, for the corresponding grey scale image I of certain characteristic image to be extracted, through Gabor bank of filters { F ₁, F ₂..., F _n, obtain filtered image sets I ₁, I ₂..., I _n, estimate to obtain E ₁, E ₂..., E _n, be concatenated into pixel P _{X, y}Corresponding estimate vector V _{X, y}To each estimate vector V _{X, y}C carries out label by the texture primitive code book, adds up obtaining label histogram again according to the label size.

The invention provides following examples, realize the SAR image classification method based on the statistics texture primitive.

The synthetic aperture radar (SAR) system has the round-the-clock ability of earth surface being carried out high-resolution imaging of round-the-clock, be similar to other coherent measurement systems, the imaging process of synthetic-aperture radar can be subjected to the influence of the coherent spot (Speckle) of similar noise, coherent spot is an inherent characteristic of radar image, simultaneously also make the SAR image table reveal tangible textural characteristics, the employing textural characteristics is classified to the SAR image and is cut apart a focus that also becomes present research.Step based on the SAR image classification method of adding up texture primitive is as follows:

At first N width of cloth training sample image in the training set is handled, obtained the vector set V={V of training set correspondence _{X, y}| P _{X, y}∈ (I1, I2 ..., IN) };

According to the method that generates the statistics texture primitive among the present invention, select a kind of clustering method that V is carried out cluster, obtain one group of typical texture primitive as the texture primitive code book, the embodiment of the invention selects random forest as clustering method, no matter select which kind of clustering method all in protection domain of the present invention when being noted that concrete enforcement, this process as shown in Figure 4;

Then with the V of all images in training set and the test set _{X, y}Make the label statistics by the texture primitive code book, obtain the corresponding label histogram of each image, as the statistics texture primitive feature of training set and test set image, as shown in Figure 5;

Utilize the KNN method to classify at last, the KNN method is the K nearest neighbor method, propose in nineteen sixty-eight at first by Cover and Hart, and be a mature methods in theory.The thinking of this method is simple, intuitive very: if the great majority in the sample of the Y of sample in feature space (being the most contiguous in the feature space) the most similar belong to some classifications, then this sample also belongs to this classification.This method is only deciding the classification for the treatment of under the branch sample according to the classification of one or several the most contiguous samples in the class decision-making.Setting the Y value among the embodiment is 1, and the KNN method deteriorates to nearest neighbor method, and promptly which kind of sample to be tested image the most similar sample on feature space belongs in the test set, just judges which kind of sample to be tested is.As Fig. 6, image in certain width of cloth test set among the embodiment, extract its label histogram after because be in the K class, just this image is assigned to the K class with the statistics texture primitive feature samples of its training set image that closes on most.

Claims (3)

1. a texture primitive feature extracting method that is used for diameter radar image is characterized in that comprising following

Step:

Step 1 is asked for each pixel P in the image to every width of cloth image in the training set _{X, y}Corresponding estimate vector V _{X, y}, combination obtains the estimate vector { { V of all pixel correspondences of all images in the training set _{X, y}} _All,

Step 2 is with the estimate vector { V of all pixel correspondences of all images in the training set _{X, y}} _AllCarry out cluster, cluster gained texture primitive code book C is texture primitive;

Step 3 is asked for each pixel P in the image to characteristic image to be extracted _{X, y}Corresponding estimate vector V _{X, y}, combination obtains the estimate vector { V of all pixel correspondences in the image _{X, y}} _One, and to this estimate vector { V _{X, y}} _OneCarry out label with the texture primitive code book C that generates in the step 2, be designated as { M _{X, y}} _One

Step 4, statistical picture label { M _{X, y}} _OneObtain label histogram, be designated as H _Mark, label histogram H _MarkBe the statistics texture primitive feature of this image;

In step 1 and step 3, described interior each the pixel P of image that asks for _{X, y}Corresponding estimate vector V _{X, y}, concrete mode is as follows,

At first, to gained gray level image T process Gabor bank of filters { F behind the quantized image ₁, F ₂..., F _n, obtain filtered image sets { I ₁, I ₂..., I _n, be designated as I;

Then, to the image I among the image sets I _p, 1≤p≤n is respectively with each pixel P _{X, y}Select m different window size for the center and obtain pane group { G _{P, 1}, G _{P, 2}..., G _{P, m}, be designated as G _pWith each pane G _{P, 1}, that 1≤i≤m divides is upper left, lower-left, upper right, four region R in bottom right _{P, i, 1}, R _{P, i, 2}, R _{P, i, 3}, R _{P, i, 4}, and these zones are combined obtain zone group { R _{P, i, 1}, 1≤i≤m, 1≤j≤4 are designated as R _pIf each regional view data all satisfies with a kind of statistical distribution E, this distribution comprises q parameter, to R _pIn each region R _{P, i, j}Carry out parameter estimation, obtain estimated parameter E _{P, i, j}={ E _{P, i, j, 1}, E _{P, i, j, 2}..., E _{P, i, j, q}, they are combined obtains estimated parameter group { E _{P, i, j, k}, 1≤i≤m, 1≤j≤4,1≤k≤q is designated as E _p

At last, with pixel P _{X, y}Corresponding E _{P, i, j, k}, 1≤p≤n, 1≤i≤m, 1≤j≤4,1≤k≤q conspire to create an estimate vector, are each pixel P in the image _{X, y}Corresponding estimate vector V _{X, y}

2. texture primitive feature extracting method as claimed in claim 1 is characterized in that: with pixel P _{X, y}Corresponding E _{P, i, j, k}When conspiring to create an estimate vector, successively according to k, j, the order that i, p increase progressively is connected in series.

3. texture primitive feature extracting method as claimed in claim 1 or 2 is characterized in that: described to R _pIn each region R _{P, i, j}Carry out parameter estimation, adopt the maximal possibility estimation mode to realize.

Images