CN1932850A - Remoto sensing image space shape characteristics extracting and sorting method - Google Patents

Abstract

One picking-up and classify method for remote sensing picture space figure character, tests the space figure structure character of pixel by a series of direction line in equal interval surrounding with center pixel. The amount of direction line is 5 to 48, and its length is controlled by Inter-item consistency referenced threshold and the biggest length referenced threshold and different with each other to embody the anisotropic of video. The square diagram of pixel's direction line reflects its structure character for picking-up space figure structure character more effectively. It reduces the dimensional number of character, and adopts length, width, pixel figure exponent, ratio of length and wide, weighted means and variance to pick-up the square diagram character of direction line to each pixel. It adopts spectrum and space structure character amalgamation classified method to select one method of many kind of neural network and machine study arithmetic to dispose high dimensional structure space.

Description

A kind of remoto sensing image space shape characteristics extracts and sorting technique

Technical field

The invention belongs to computing machine remote sensing image processing and mode identification technology, be a kind of new image shape and structure feature of utilizing remote sensing image context spectral similarity to distribute to extract, and the method for the spectrum and the space characteristics of higher-dimension being classified with neural network, support vector machine.

Background technology

High spatial resolution remote sense image can provide a large amount of topographical features, and the inner element rich details of same atural object classification information obtains characterizing, and spatial information is abundanter, and the relation of the size of atural object, shape and adjacent atural object is better reflected.Yet the spectrum statistical nature of the novel remote sensing image of this class is stable not as the low resolution image, atural object space distribution complexity, similar object presents very big spectrum heterogeneity, be embodied in the class internal variance and become big, inter-class variance reduces, the spectrum of different atural objects is overlapped, makes traditional spectral classification method can not obtain satisfied result.Therefore the remote sensing application personnel have proposed a lot of space characteristics operators in recent years, to remedy the deficiency of spectral signature.The whole bag of tricks that relevant spatial structure characteristic extracts is the focus of current research, below the more typical method of summary current research.

Gray level co-occurrence matrixes (GLCM) method is that the widely used a kind of texture in remote sensing image processing field and space characteristics extract operator, it with the spatial relationship of image greyscale value describe the pixel point between spatial structure characteristic and correlativity thereof.GLCM is general to survey pixel point on 0 °, 45 °, 90 ° and the 135 ° of directions to relation, and constitutes 4 gray level co-occurrence matrixes, adopts the stack of 4 directions to eliminate the influence of direction usually, and the space symbiotic characteristic of usefulness gray-scale value is as the tolerance of texture.The close grain gray space changes very fast, and open grain is also not obvious with the increase variation of distance, estimates carrying out filtering in the co-occurrence matrix space with different spaces, can extract the statistical attribute of a series of description texture and structural characteristics.Gray level co-occurrence matrixes statistical measurement commonly used has: average (mean), variance (variance), entropy (entropy), energy value (energy), homogeney (homogeneity), contrast (contrast) or the like.Each statistical measurement can be also referred to as auxiliary wave band as a textural characteristics image, classifies with spectral signature, and the advantage of this method is the difference that can either reflect the atural object space characteristics, again can with various categorizing system compatibilities.Choose different statistical attributes as index according to the characteristics of different atural objects, can reach the purpose of effective extraction terrestrial object information.Relevant list of references has: Yun Zhang.Optimisation of building detection in satelliteimages by combining multispectral classification and texture filtering.ISPRSJournal of Photogrammetry and Remote Sensing, 1999,54 (8): 50～60; R.M.Haralick.Statistical and structural approaches to texture.Proceeding of IEEE, 1979,67:786-804; And R.M.Haralik, K.Shanmugam, and D.Its ' hak.Textural features forimage classification, IEEE trans.Syst.Man Cybnet., vol.SMC-3, pp.610-621,1973.

The space characteristics of image also can be effectively extracted in the mathematical morphology conversion.It obtains spatial structure characteristic by various morphological operations and conversion, and (structural element) obtains multiple dimensioned result by different operating structure elements.Morphological operation operator commonly used has: expansion, burn into watershed transform, open and close computing or the like.Pesaresi utilizes the open and close computing of different scale to construct the morphology section of image (morphological profile), and multiple dimensioned morphological feature is classified with neural network, he thinks that the open and close computing applies to remote sensing image, can detect the darker or brighter structural unit in neighbour zone, territory; Benediktsson proposes the notion of difference morphology section (derivation of morphologicalprofile) on this basis,, and with the BP neural network this feature is classified and to have obtained higher nicety of grading as new image structure feature with the difference of open and close operation result between adjacent yardstick.List of references has: M.Pesaresi, and.J.A.Benediktsson, " A new approach for the morphological segmentation of high-resolutionsatellite imagery; " IEEE Transactions on Geoscience and Remote Sensing, vol.39, no.2, pp.309-320, Feb, 2001; J.A.Benediktsson, M.Pesaresi, and K.Arnason, " Classification and feature extraction for remote sensing images from urban areasbased on morphological transformations, " IEEE Transactions on Geoscience and RemoteSensing, vol.41, no.9, pp.1940-1949, Sep, 2003; J.A.Benediktsson, J.A.Palmason, and J.R.Sveinsson, " Classification of hyperspectral data from urban areas basedon extended morphological profiles, " IEEE Transactions on Geoscience and RemoteSensing, vol.43, no.3, pp.480-491, Mar, 2005.

In addition, straight line also is a kind of important image structure feature, and different imagery zones shows different linear features, and Unsalan utilizes the line feature of image can tell city (urban), suburb (suburban) and rural area (rural area).He thinks that within the specific limits the straight line that is extracted in the imagery zone in rural area is often comparatively short and small, and distributes more at random, and this is that shortage by this zone mankind's activity causes; The city image then shows opposite feature, and straight line distributes comparatively regular, and detected straight line is also longer.Unsalan and Boyer utilize the linear feature of this feature extraction image, and utilize the probability relaxed algorithm in conjunction with normalization vegetation difference index (NDVI) remote sensing image to be classified, and have obtained effect preferably.List of references is as C.Unsalan, K.L.Boyer, " Classifying land development in high-resolutionpanchromatic satellite images using straight-line statistics; " IEEE Transactionson Geoscience and Remote Sensing, vol.42, no.4, pp.907-919, April, 2004; C.Unsalan, K.L.Boyer, " Classifying land development in high-resolution panchromaticsatellite imagery using hybrid structural-multispectral features; " IEEETransactions on Geoscience and Remote Sensing, vol.42, no.12, pp.2840-2850, December, 2004.

Summing up above these methods finds: though various space characteristics extraction algorithm principle has nothing in common with each other, still having general character to a certain extent, promptly all is to utilize the intensity profile characteristic of image in certain zone to extract feature.The high spatial resolution remote sense image sorting technique that the present invention proposes a series of space characteristics indexes (SI) and merges based on shape and spectral signature.Shape and spectrum are the concrete manifestation forms of remote sensing image texture, especially in high resolution image object detail given full expression to, the relation of adjacent picture elements and the common style characteristic that characterizes thereof become the key factor of classification.The present invention describes its space structure with the relation of pixel and neighborhood thereof, simultaneously in order more fully to utilize image feature, has proposed shape and spectrum integrated classification method based on support vector machine.

The principle of design of algorithm is: 1). utilize the spectral similarity of adjacent picture elements, purpose is to consider the spatial context feature of pixel; 2). make the pixel that is in the identical shaped zone have identical or close eigenwert, this is in order to strengthen the homogeney of high resolution image, smooth noise to a certain extent; 3). widen the eigenwert between the pixel of difformity zone, this is in order to make full use of the details characteristic of high resolution image as far as possible.

Starting point of the present invention is to utilize neighborhood gray scale similarity to measure contextual structural information, the thought of this point and gray level co-occurrence matrixes is more similar, the both carries out conversion to spectral space, GLCM transforms to the co-occurrence matrix space to spectral space, SI then transforms to the directional ray metric space to spectral space, their key distinction is: 1) .GLCM adopts the stationary window operation, and SI has cancelled the window setting, and the length of every directional ray is all different, algorithm is handled flexibly according to different structure distribution, can effectively utilize the anisotropy of image; 2) .GLCM at first reduces the gray level of image, calculates the identical pixel number of gray-scale value then, and SI has then kept the gray feature of raw video, uses the similar pixel number of homogeney threshold calculations gray-scale value then; 3) .GLCM surveys 4 directions, and SI surveys the direction more than 20.Above characteristic makes SI can obtain better effect than GLCM in remote sensing image feature extraction and classification.

Summary of the invention

The present invention proposes a kind of remote sensing image spatial structure characteristic and extracts and sorting technique, describing its contextual spatial form by the spectral similarity of pixel and neighborhood thereof distributes, then shape after the normalization and spectral signature input category device are classified, in remote sensing image feature extraction and classification, can obtain better effect than GLCM.

Technical scheme provided by the invention is: a kind of remote sensing image spatial structure characteristic extracts and sorting technique, it is characterized in that: by survey the spatial form architectural feature of this pixel around the extension of a series of equally spaced directional rays of center pixel, the quantity of directional ray is at 5-48, the length of directional ray is by homogeney threshold value and maximum length threshold control, unequal mutually, the anisotropy of embodiment image; Its context mechanism characteristic of directional ray histogram reflection by pixel, for more effective extraction spatial structure characteristic, reduce the dimension of feature simultaneously, adopt length, width, pixel shape index, length and width ratio, weighted mean, these 6 statistical measurements of variance to extract the directional ray histogram feature of each pixel; Adopt the method for spectrum and spatial structure characteristic integrated classification, the method for selecting in multiple neural network and machine learning algorithm is simultaneously handled high-dimensional feature space.

Aforesaid remoto sensing image space shape characteristics extracts and sorting technique, and its feature may further comprise the steps:

One, at first defining directional ray is a series of line segments that pass the center pixel, and their length and direction have nothing in common with each other, and its length is estimated with threshold value by the spectrum homogeney between adjacent picture elements and determined the radian that the angle between adjacent directional ray is set to equate;

Two, the expansion of directional ray begins to extend towards two reverse directions simultaneously from a certain center pixel, the directional ray expansion condition is: the length restriction that satisfies adjacent picture elements homogeney threshold value and directional ray, each bar directional ray is all expanded according to above condition and is extended, if one of them condition is false, then stop the expansion of this directional ray;

Three, all directional rays of this center pixel are asked in tracking, calculate the length of all directional rays, in two kinds of length computation formula, choose: city block distance (city-block distance) and Euclidean distance (Euclideandistance) according to different calculation requirements;

Four, try to achieve around the length of all directional rays of this center pixel, this series length value according to clockwise series arrangement, form the directional ray histogram of this pixel, and extract the spatial form feature of this pixel according to histogrammic distribution, adopt following six kinds of characteristic measurements: length (length), width (width), pixel shape index (pixel shape index), length and width ratio (length-width ratio), weighted mean (weighted mean), variance (variance);

Five, travel through whole image, calculate the directional ray histogram and the corresponding statistical measurement of each pixel;

Six, classify in conjunction with the spectral information of raw video and the spatial form feature of extraction, adopt diverse ways to carry out feature normalization spectral information and spatial information;

Seven, composite character input category device, there is following sorter to select: minimum distance classifier (MDC), maximum likelihood classifier (MLC), multilayer perceptron network (MLP) for the user, radial base neural net (RBF), probabilistic neural network (PNN) and support vector machine (SVM), the user can select only sorter according to different requirements;

Eight, sorter is provided with and trains, input spectrum and the conscientious classification of space composite character then, classification results to the end.

Aforesaid remoto sensing image space shape characteristics extracts and sorting technique, and it is characterized in that: alternative direction number of lines D has: 12,16,20,24, and default setting is D=20.

Principle of the present invention is:

One, at the neighborhood spectral distribution feature of certain center pixel, follow the tracks of and ask for its all directional ray, the direction number of lines of establishing a center pixel is D, and D is greater than 4; The invention provides a plurality of alternative D values, general, the D value is big more, and algorithm is stronger to the descriptive power of image space neighborhood, but D is when increasing to a certain degree, and the raising of precision is also not obvious, meanwhile but will consume the more processing time.Can have for the D value that the user selects: 12,16,20,24, default setting is D=20;

Two, the length of calculated direction line, there are two kinds of computing method available: Euclidean distance and city block distance, the difference in length of the more effective reflection directional ray of the former energy, the latter can play the effect of smothing filtering and save computing time, and the user can be provided with as required flexibly;

Three, all D bar directional ray length of computing center's pixel are stored according to clockwise direction, successively as the directional ray histogram of this pixel, for the subsequent characteristics extraction step is prepared;

Four, travel through whole image, ask for the directional ray histogram of all pixels;

Five, with six kinds of characteristic measurements that propose: length, width, pixel shape index, length and width ratio, weighted mean, variance are extracted the histogram feature of each pixel, reduce the dimension of space characteristics simultaneously;

Six, the user can select whether to need to carry out eigentransformation according to specific circumstances, and the eigentransformation method that provides comprises: decision-making Edge Gradient Feature algorithm (DBFE), principal component analysis (PCA) (PCA) and index of similarity (Similarity Index).The purpose of eigentransformation is the dimension of reduction space characteristics, increases the classification separability of feature space simultaneously;

Seven, the spatial structure characteristic and the spectral information that extract are carried out pre-service and normalization respectively, the method that spectral information adopts maximum-minimal linear to stretch, space characteristics adopts the method for histogram equalization to carry out pre-service because the numerical value span is too big.Normalized purpose is for next step effective classification;

Eight, select proper classifier, alternative comprising: minimum distance classifier, maximum likelihood classifier, multilayer perceptron network, radial base neural net, probabilistic neural network, support vector machine for composite character.Minimum distance classifier is fit to the input of 1 dimensional feature, the maximum-likelihood method fast and stable, but on the high dimensional feature processing power, be not so good as machine learning algorithm, neural net method is a research focus of handling the multidimensional remotely-sensed data in recent years, the present invention utilizes the newest fruits support vector machine of machine learning to handle spectrum and shape composite character, makes a strategic decision in the hope of utilizing these features to greatest extent;

Nine, select training sample, the parameter of sorter is set, the parameter setting of support vector machine (SVM) is robotization, adopts famous leave-one-out (LOO) method to determine penalty coefficient and the nuclear parameter of SVM;

Ten, spectrum and space structure composite character are classified, obtain classification results.

Characteristics of the present invention: having defined the notion of directional ray, by survey the spatial form architectural feature of this pixel around the extension of a series of equally spaced directional rays of center pixel, is a kind of new space characteristics extracting mode.The directional ray maximum can be surveyed 48 directions, is much higher than 4 directions of gray level co-occurrence matrixes, has stronger neighborhood descriptive power, and the length of directional ray is by homogeney threshold value and maximum length threshold control, and is unequal mutually, has embodied the anisotropy of image; Its context mechanism characteristic of directional ray histogram reflection by pixel, for more effective extraction spatial structure characteristic, reduce the dimension of feature simultaneously, the present invention proposes the directional ray histogram feature that length, width, pixel shape index, length and width ratio, weighted mean, these 6 statistical measurements of variance are extracted each pixel; Adopt the method for spectrum and spatial structure characteristic integrated classification, provide multiple neural network and machine learning algorithm to handle high-dimensional feature space simultaneously, wherein the support vector function produces the not available new feature of raw data by the nuclear space mapping, has also avoided spectrum or the spatial information decisive influence in decision-making.The present invention calculates easy, program run efficient height, manual intervention is few, is applicable to the automatic classification of high-resolution remote sensing image, can effectively improve the nicety of grading and the efficient of such image.

Description of drawings

Fig. 1 is the synoptic diagram of the directional ray of the embodiment of the invention;

Fig. 2 is the master routine operational flow diagram of the embodiment of the invention;

Fig. 3 is the directional ray expansion and the track algorithm process flow diagram of the embodiment of the invention;

Fig. 4 is the flow process of the directional ray histogram feature extraction algorithm of the embodiment of the invention;

Fig. 5 is the spectrum based on neural network and the machine learning-architectural feature sorter of the embodiment of the invention.

Embodiment

1, theoretical foundation

The basic theories that the present invention uses mainly comprises:

(1) support vector machine: it is a kind of new learning method that is based upon on the Statistical Learning Theory, the consistance and the structural risk minimization principle of learning process have been embodied, it is keeping minimizing fiducial range on the fixing basis of empiric risk, by taking all factors into consideration empiric risk and fiducial range, get according to structural risk minimization that it is compromise, thereby obtain the decision function of risk minimum.Its core concept is that the sample of the input space is mapped to the higher-dimension nuclear space by nonlinear transformation, asks for the optimum linearity decision surface with low VC dimension (complexity) at the higher-dimension nuclear space.

The ultimate principle of SVM is: suppose that training sample is { (x ₁, y ₁), (x ₂, y ₂) ..., (x _N, y _N), x wherein _i∈ R ^d, expression input pattern, y _i{ ± 1} represents target output to ∈.If optimizing decision face equation is: w ^Tx _i+ b=0, then weight vector w and biasing b must satisfy constraint:

y _i(w ^Tx _i+b)≥1-ξ _i

ξ wherein _iBe the slack variable under the inseparable condition of linearity, its expression pattern is to the departure degree under the ideal linearity situation.The target of SVM is to find a decision surface to make its average error error in classification minimum on training data, can derive following optimization problem:

$\mathrm{\Φ}(w,\mathrm{\ξ})=\frac{1}{2}{w}^{T}w+C\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i$

C is the positive parameter of user's appointment, and it represents the punishment degree that SVM divides sample to mistake, is the balance parameters between wrong branch sample proportion and the algorithm complex.Use the Lagrange multiplier method, finding the solution of optimizing decision face can be converted into following constrained optimization problem:

$Q\left(\mathrm{\α}\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i-\frac{1}{2}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_j{y}_i{y}_{j}K(x_i,x_j)$

{ α wherein _i} _I=1 ^NBe the Lagrange multiplier, and (5) satisfy constraint condition:

$\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_i{y}_i=\mathrm{0,0}\≤{\mathrm{\α}}_i\≤C,i=\mathrm{1,2,3}...,N$

K (x, x _i) be kernel function, satisfy the Mercer theorem, nuclear commonly used has following two kinds:

Polynomial kernel function: K=(x ^Tx _i+ 1) ^p, index p is determined by the user;

Radially basic kernel function: $K=\exp (-\frac{1}{2{\mathrm{\σ}}^2}{\left|\right|x-x_i\left|\right|}^2),$ Width cs is examined all identical to all and is specified by the user;

Select the sorter of support vector machine (SVM) as space characteristics, it is the hypothesis of considering that the characteristic of its imparametrization need not the feature space normal distribution, and the mapping of higher-dimension nuclear space is more suitable for the space characteristics input of multidimensional, because model complexity that SVM provides and input feature vector dimension are irrelevant, this makes that the feature of input pattern can diversification, kernel function is mapped to the higher-dimension nuclear space with input feature vector may produce the not available new feature of raw data, makes the inseparable pattern of script spectrum owing to the adding of space characteristics becomes and can divide.In the categorizing system, the application of SVM will be noted:

(a) being provided with of .SVM mainly is the selection of kernel function, promptly in polynomial kernel with RBF is internuclear makes one's options.Characteristics according to high resolution image, because inter-class variance is bigger, the spectral signature of its similar ground object sample is comparatively disperseed, and be not tightly round some center, the spectrum samples that is high resolution image does not have tangible center, sample there is no the weight size, and examine for RBF, it is almost nil for the output away from the input sample of node center, the distance of sample evidence decentering distance has different weights and response, yet but there is not locality in polynomial kernel, so it is more suitable for the kernel function as the high resolution image input feature vector.

(b) .C is a regularization parameter, or claims penalty coefficient, but C is controlling and treats the departure degree of merotype to decision surface in feature space, and when C increased, this departure degree increased, and when C reduces, but departure degree reduces.Its setting is with sample, support vector and treat that the distribution of merotype in feature space is relevant, consider the relation of classification time and precision, to choosing of C is significant, and suitable value can be obtained optimal results with the minimum time, improves the efficient of categorizing system.

(c) the sample district of .SVM selects also will must take into full account the different spectral signatures of atural object of the same race according to the characteristics of high resolution image.

(2) probabilistic neural network (PNN): probabilistic neural network is that Specht proposes, and its essence is the combination of Bayes decision rule and multilayer perceptron, and network divides 3 layers: input layer, mode layer and output layer.The training of PNN is very simple, and K class sample is sequentially arranged in K the pool of mode layer, to certain pool_i, and i=1,2,3 ... K, K are classification sum or pattern count, and N is all arranged _iIndividual pattern neuron, for each input vector y, then the j of pool_i neuronic activation value is:

$f(y,w_i^{\left(j\right)},\mathrm{\σ})=\frac{1}{N_i{\left(2\mathrm{\π}\right)}^{d/2}{\mathrm{\σ}}^d}\exp \left[\frac{{(y-w_i^{\left(j\right)})\·}^T\left(y-w_i^{\left(j\right)}\right)}{2{\mathrm{\σ}}^2}\right]$

W in the formula _i ^(j)J neuronic weight vector of expression pool_i determined by training sample.Output layer has K neuron, represents K pattern, and wherein the value of i output terminal is:

$o_i=\underset{j}{\mathrm{\Σ}}f(y,w_i^{\left(j\right)},\mathrm{\σ})$

The method of " victor takes entirely " is adopted in decision-making:

If: O _k＞O _i, i ≠ k, and i, k ∈ [1, K], then: y ∈ C _k

The training of PNN is very simple disposable process, to each training sample x, supposes that it is an i pattern, i.e. x ∈ C _i, its training process only adds a new neuron again in mode layer pool_i so, and its weight vector w _i ^(j)Assignment is x.

2, the structure of shape and structure feature

The principle of design of PSI is: 1). utilize the spectral similarity of adjacent picture elements, purpose is to consider the spatial context feature of pixel; 2). make the pixel that is in the identical shaped zone have identical or close eigenwert, this is in order to strengthen the homogeney of high resolution image, smooth noise to a certain extent; 3). widen the eigenwert between the pixel of difformity zone, this is in order to make full use of the details characteristic of high resolution image as far as possible.

At first defining directional ray is a series of line segments that pass the center pixel, and as shown in Figure 1, their length has nothing in common with each other, and its length is estimated with threshold value by the spectrum homogeney between adjacent picture elements and determined.The tracking and the calculation procedure of directional ray are as follows:

1). homogeney is estimated:

${\mathrm{PH}}_i(x,y)={\left(\underset{s=1}{\overset{n}{\mathrm{\Σ}}}{(p_s^{\mathrm{sur}}(x,y)-p_s^{\mathrm{cen}})}^2\right)}^{1/2}$

Wherein, PH _i(x, y) current neighborhood pixel (x, y) the homogeney measure value on i bar directional ray, the p of expression _s ^CenThe spectral value of expression center pixel on wave band s, p _s ^SurRepresent the spectral value of current neighborhood pixel on wave band s, n represents the wave band number.

2). the expansion of directional ray: every directional ray all according to specific rule therefrom imago unit set out and expand simultaneously towards both sides, the condition of i bar directional ray expansion is: (a). the PH of current pixel _i(x is y) less than threshold value T1; (b). the total length of this directional ray is less than threshold value T2.The theoretical value of T1 should be to get the mean value of the interior mean square deviation of class of all kinds of samples, can regulate according to specific circumstances in experiment; T2 is a scale factor, should determine according to the size of atural object interested, also can utilize the variation of T2 to extract multi-scale information.

3). establishing D is the directional ray sum of a pixel, travels through whole image, according to 1), 2) two the step can follow the tracks of all D bar directional rays that obtain each pixel respectively.

4). calculate the length of i bar directional ray:

$d_i={({(m^{e1}-m^{e2})}^2+{(n^{e1}-n^{e2})}^2)}^{\frac{1}{2}}$

Or d _i=max{|m ^E1-m ^E2|, | n ^E1-n ^E2|

(m wherein ^E1, n ^E1) represent the cell coordinate ranks number of this directional ray one end, (m ^E2, n ^E2) represent the ranks number of another end points.Therefore obtain any pixel (i, directional ray length sequences j): d (i, j)=[d ₁, d ₂..., d _D].

3, implementation procedure

(1), parameter T1, T2 and the D of feature extraction algorithm are set.Under the default situations, T1 gets the mean value of the interior mean square deviation of class of all kinds of samples, and T2 gets image length or width 0.35 times, D=20.All D bar directional rays according to threshold value T1, a certain center of T2 following calculation pixel.In concrete operation, can adjust threshold size flexibly according to operation result.

(2), the computing formula of choice direction line length, Euclidean distance or city block distance, the former can effectively embody the difference between the directional ray length, the latter can reduce computing time in level and smooth space characteristics.Calculate the directional ray length of certain center pixel subsequently according to selected range formula.

(3), in the direction of the clock store all directional ray length of this center pixel successively, form the directional ray histogram of D dimension.Travel through whole image, follow the tracks of the directional ray of asking for all pixels, store the directional ray histogram of each pixel, so that carry out feature extraction.

(4), six kinds of characteristic measurements that propose with the present invention: length, width, pixel shape index, length and width ratio, weighted mean, variance are extracted the histogrammic statistical attribute of each pixel directional ray, and like this, each pixel just forms 6 dimension space architectural features.

The computing method of these 6 kinds of statistical natures are as follows:

(a). length (length):

$\mathrm{length}=\underset{i=1}{\overset{D}{\max }}\left(H(i,j)\right)$

Wherein (i j) represents pixel (i, directional ray histogram j) to H.

(b). width (width):

$\mathrm{width}=\underset{i=1}{\overset{D}{\min }}\left(H(i,j)\right)$

(c). pixel shape index (mean):

$\mathrm{mean}=\underset{i=1}{\overset{D}{\mathrm{\Σ}}}{d}_i/D$

(d). weighted mean (w-mean):

$w-\mathrm{mean}=\underset{i=1}{\overset{D}{\mathrm{\Σ}}}\frac{a\·(k_i-1)}{{\mathrm{st}}_i}{d}_i/D$

K wherein _iThe length of representing i bar directional ray, α is the ratio regulatory factor, st _iVariance for the grey scale pixel value of forming i bar directional ray is used for limiting the weight of unsane directional ray in characteristic statistics.

(e). Aspect Ratio (ratio):

$\mathrm{ratio}=\arctan \frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\underset{\min }{\mathrm{sort}}}^i\left(H(i,j)\right)}{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\underset{\max }{\mathrm{sort}}}^i\left(H(i,j)\right)}$

Wherein ${\underset{\max }{\mathrm{sort}}}^n\left(H(i,j)\right)$ With ${\underset{\min }{\mathrm{sort}}}^n\left(H(i,j)\right)$ Represent pixel (i, j) n in the directional ray line histogram minimum and maximum value respectively.

(d). standard deviation (SD):

$\mathrm{SD}=\underset{i=1}{\overset{D}{\mathrm{SD}}}\left(H(i,j)\right)$

$=\frac{1}{D-1}\underset{i=1}{\overset{D}{\mathrm{\Σ}}}\sqrt{{(d_i-\mathrm{mean})}^2}$

(5) the spatial structure characteristic dimension is more if spectral band is less, can select to carry out the feature selecting operation.This categorizing system provides 3 kinds of dimensions to reduce and feature selecting algorithm: independent component analysis (ICA), decision-making Edge Gradient Feature (DBFE) and index of similarity algorithm (Similarity Index).Because it is easy that the index of similarity method is calculated, required CPU time is minimum, also can guarantee computational accuracy simultaneously, so acquiescence is used this method.

(a). independent component analysis:

The cardinal rule of ICA is: a given set of eigenvectors x, but the task of algorithm is exactly to determine the inverse matrix W of a N * N, this vector is gathered carried out linear transformation:

y＝Wx

Make result vector y (i), i=1,2 ... N is separate.The key of ICA algorithm is the method for discrimination of independence,

Here adopt the interactive information between minimizer to estimate the W matrix.Interactive information between the y component is defined as:

$I\left(y\right)=-H\left(y\right)+\underset{i=1}{\overset{N}{\mathrm{\Σ}}}H\left(y\left(i\right)\right)$

Wherein H (y (i)) is the combination entropy of y (i).It is 0 that statistics between the y (i) independently is equivalent to I (y) because this moment joint probability density and corresponding the long-pending of marginal probability density equate, formula (3) minimize the maximization that is equivalent to following formula:

$J\left(W\right)=\ln \left|\det \left(W\right)\right|+E\left[\underset{i=1}{\overset{S}{\mathrm{\Σ}}}\ln p_i\left(y\left(i\right)\right)\right]$

The following formula both sides are to the W differentiate, and arrangement can get the iterative formula of gradient descent method:

W(t)＝W(t-1)+μ(t)(I-E[φ(y)y ^T])W ^-T(t-1)

(b). decision-making Edge Gradient Feature (DBFE):

This algorithm can make full use of the characteristics of sorter, selects needed feature from decision boundary.The theoretical foundation of DBFE is to utilize the position at each classification decision-making edge to reject unnecessary characteristic information.

(c). index of similarity (Similarity Index);

Adopt the spectral band after characteristic similarity between variable screens conversion, establish p and be the signal dimension before the feature selecting, q be after the feature selecting the signal dimension, the task of algorithm is exactly deletion (p-q) dimensional signal from feature set.Algorithm employing maximum information compression index (Maximal Information Compression Index, MICI) the p dimensional feature is got rid of:

$\mathrm{MICI}(x,y)=\mathrm{var}\left(x\right)+\mathrm{var}\left(y\right)-\sqrt{{(\mathrm{var}\left(x\right)+\mathrm{var}\left(y\right))}^2-4\mathrm{var}\left(x\right)\mathrm{var}\left(y\right)\left({1-p(x,y)}^2\right)}$

$\mathrm{\ρ}(x,y)=\mathrm{cov}(x,y)/\sqrt{\mathrm{var}\left(x\right)\mathrm{var}\left(y\right)}$

(x y) is at 0 o'clock, represents two feature linear dependences, and this moment, the error of feature selecting was 0 as MICI; (x, when y) increasing, the correlativity of two features reduces, the error increase of feature selecting as MICI.(x is that two features are to (x, the y) eigenwert of projection on its major component direction, the error of representation feature compression y) to MICI.The feature selecting algorithm of this paper is as follows:

1). the p dimensional feature is normalized to [0,1].

2). calculate the compression index of every pair of feature one by one, and obtain maximum MICI, establish it corresponding to wave band a, b.

3). calculate $S_a=\underset{i=1}{\overset{q}{\mathrm{\Σ}}}\mathrm{MICI}(a,i)$ With $S_b=\underset{i=1}{\overset{q}{\mathrm{\Σ}}}\mathrm{MICI}(b,i),$ Compare S _aAnd S _bSize, if S _a＞S _b, then wave band a deletion; Otherwise then wave band b deletion.

4). make p=p-1, if p=q, then algorithm stops; If not, then changing (2) over to continues to carry out.

(6), store the spatial structure characteristic of each pixel, as auxiliary wave band and the participative decision making classification together of original spectrum wave band.

(7), spectral signature and planform feature are carried out pre-service and normalization, so that be input to sorter.Because the difference of spectral information and space characteristics adopts diverse ways that both are carried out normalization here respectively:

(8), select proper classifier, the default categories device is support vector machine (SVM), this is because its computing velocity is fast, and has significant advantage handling on the higher-dimension composite character.

(9), select training sample, selected classification risen trains and learn, then image is classified according to priori, classification chart to the end.

Claims (3)

1, a kind of remoto sensing image space shape characteristics extracts and sorting technique, it is characterized in that: by survey the spatial form architectural feature of this pixel around the extension of a series of equally spaced directional rays of center pixel, the quantity of directional ray is at 5-48, the length of directional ray is by homogeney threshold value and maximum length threshold control, unequal mutually, the anisotropy of embodiment image; Its context mechanism characteristic of directional ray histogram reflection by pixel, for more effective extraction spatial structure characteristic, reduce the dimension of feature simultaneously, adopt length, width, pixel shape index, length and width ratio, weighted mean, these 6 statistical measurements of variance to extract the directional ray histogram feature of each pixel; Adopt the method for spectrum and spatial structure characteristic integrated classification, the method for selecting in multiple neural network and machine learning algorithm is simultaneously handled high-dimensional feature space.

2, remoto sensing image space shape characteristics as claimed in claim 1 extracts and sorting technique, and its feature may further comprise the steps:

1., at first defining directional ray is a series of line segments that pass the center pixel, and their length and direction have nothing in common with each other, and its length is estimated with threshold value by the spectrum homogeney between adjacent picture elements and determined the radian that the angle between adjacent directional ray is set to equate;

2., the expansion of directional ray begins to extend towards two reverse directions simultaneously from a certain center pixel, the directional ray expansion condition is: the length restriction that satisfies adjacent picture elements homogeney threshold value and directional ray, each bar directional ray is all expanded according to above condition and is extended, if one of them condition is false, then stop the expansion of this directional ray;

3., all directional rays of this center pixel are asked in tracking, calculate the length of all directional rays, in two kinds of length computation formula, choose: city block distance (city-block distance) and Euclidean distance (Euclideandistance) according to different calculation requirements;

4., try to achieve around the length of all directional rays of this center pixel, this series length value according to clockwise series arrangement, form the directional ray histogram of this pixel, and extract the spatial form feature of this pixel according to histogrammic distribution, adopt following six kinds of characteristic measurements: length (length), width (width), pixel shape index (pixel shape index), length and width ratio (length-width ratio), weighted mean (weighted mean), variance (variance);

5., travel through whole image, calculate the directional ray histogram and the corresponding statistical measurement of each pixel;

6., classify, adopt diverse ways to carry out feature normalization to spectral information and spatial information in conjunction with the spectral information of raw video and the spatial form feature of extraction;

7., composite character input category device, there is following sorter to select: minimum distance classifier (MDC), maximum likelihood classifier (MLC), multilayer perceptron network (MLP) for the user, radial base neural net (RBF), probabilistic neural network (PNN) and support vector machine (SVM), the user can select only sorter according to different requirements;

8., sorter is provided with and trains, input spectrum and the conscientious classification of space composite character then, classification results to the end.

3, remoto sensing image space shape characteristics as claimed in claim 1 or 2 extracts and sorting technique, and it is characterized in that: alternative direction number of lines D has: 12,16,20,24, and default setting is D=20.

Images