CN1916935A - Hybrid sorting process of adjustable spectrum and space characteristics - Google Patents

Abstract

An adjustable mixed classification method of spectrum and space character includes setting a series of multimeasure windows of central pixel and using Mallat transform to set up multiresolution analysis of each window, obtaining four wavelet subbands by each time of Mallat decomposition on each window, using four said subbands to structure a rotary unchanged space character, selecting sample similar to said type and using mean value and variance measure to structure deflection vector according to spectrum and space information of sample, making character weight on different space character and selecting occupied proportion of spectrum and space character in classification.

Description

A kind of adjustable spectrum and space characteristics hybrid sorting process

Technical field

The invention belongs to computing machine remote sensing image processing and mode identification technology, is that a kind of new utilize multiple dimensioned background and multiresolution analysis extract the image space feature, and the sorting technique of spectrum and spatial information ratio in the scalable sorter.

Background technology

In recent years, the development of space technology makes us can obtain the remotely-sensed data of higher spatial resolution, and the characteristics of this image are: (1) adjacent picture elements has the correlativity of height; (2) SPECTRAL DIVERSITY in the class increases, and the SPECTRAL DIVERSITY between class reduces; (3) spatial relationship of image becomes complicated more, extensively exists with spectrum phenomenon heterogeneous, the different spectrum of homogeneity.So the spectral information of different classes of atural object exists more overlapping, traditional spectrographic technique faces difficulty in the decipher of high resolution image.Address this problem the texture and the spatial information that must utilize image, turn to based on contextual neighborhood characteristics from feature extraction and extract based on pixel.This point is reached common understanding in the world, and the whole bag of tricks of relevant spatial texture feature extraction is the focus of current research, below the more typical method of summary current research.

Utilize spatial autocorrelation can extract the repeatability textural characteristics, autocorrelation function can be used for representing the rough and smooth of texture, the space size of reflection texture primitive, and the bigger expression coarse texture of primitive, it is slower that autocorrelation function moves decline rate with window; Primitive is less, and then to move decline rate with window very fast for autocorrelation function.Pertinent literature has: Myint.S.W., N.S.N.Lam, J.Tylor.An Evaluation of four different waveletdecomposition procedures for spatial feature discrimination within and aroundurban areas.Transactions in GIS, 2002,6 (4): 403～429.

Utilize gray scale curved surface and natural shape that fractal method can token image, the most direct meaning of fractal dimension in fractal is exactly to have represented surperficial fluctuating quantity.It is fractal to present statistics owing to ground nature is fractal in certain range scale, therefore, more in the branch dimension technique of estimation application of remote sensing field.Document is as Myint.S.W.Fractal approachestexture analysis and classification of remotely sensed data.InternationalJournal of Remote Sensing, 2003,24 (9): 1925～1947.

Gray level co-occurrence matrixes is a kind of texture statistics method commonly used, it utilizes the combination condition probability density between the image greyscale level to calculate texture, by some texture statistics amounts gray scale symbiosis space is described, thereby reach the purpose of extracting space characteristics, its list of references has: Yun Zhang.Optimisation of building detection insatellite images by combining multispectral classification and texturefiltering.ISPRS Journal 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.

The Markov random field method is utilized the inner link of context pixel value, describes this relation with Markov model, and with the space characteristics of the model parameter of extracting as image.As: Y.Dong, A.K.Milne, and B.C.Forster, " Segmentation and classification of vegetated areas usingpolarimetric SAR image data; " IEEE Transactions on Geoscience and RemoteSensing, vol.39, pp.321-329,2001.

Small echo is the applied mathematics theory that the mid-80 grows up, owing to have time-frequency localization feature, dimensional variation feature and direction characteristic, makes it obtain in fields such as Flame Image Process using widely.In order to improve the decipher precision of remote sensing image, existing scholar proposes the image texture feature extraction algorithm based on wavelet transformation.Relevant document: Zhu C, Yang X.Study of remote sensing image texture analysis and classificationusing wavelet.International Journal of Remote Sensing, 1998,13 (6): 3167～3187; Zhu Changqing, Yang Xiaomei has the remote sensing image Texture classification of better resolution wavelet decomposition, Geographical Study, 1997,16 (1): 53-59; Ni Ling, Zhang Jianqing, Yao Wei, based on the Wavelet for SAR Image's Texture Analysis, Wuhan University's journal information science version, 2004,29 (4): 367-370.The characteristics of these methods are to utilize the characteristics of wavelet decomposition different frequency bands, estimate the eigenwert of extracting each subband with energy value, and the input category device is classified then.

Summing up above these methods finds:

Though one, various space characteristics extraction algorithm principles have nothing in common with each other, but window operation all inevitably appears, also just can't avoid window effect (window-size effect) famous in the image processing, does the document that proposes relevant issues have: M.E.Hodgson.What size window for image classification? A cognitiveperspective.Photogrammetric Engineering and Remote Sensing, 1998,64 (8): 797～807.This window effect is to be caused by fixing moving window, and the window of taking is big more, and this effect is obvious more, is embodied in processing and produces band edge at the intersection of atural object later, and width is about half of window size.The window problem is the gordian technique that remote sensing image processing and space characteristics extract, and also is the major issue that needs to be resolved hurrily.Existing solution is to utilize the mode of test sample, and a plurality of different big or small windows are experimentized, and finds best size.Pertinent literature has: Zhang Jin's water, and what spring sun-light, Pan Yaozhong etc. are based on the compound high spatial resolution remote sense data qualification research of the multi-source information of SVM, remote sensing journal, 2006,10 (1): 49-57; What spring sun-light, Cao Xin, Shi Peijun, Li Jing is based on panchromatic data texture of Landsat7 ETM+ and the compound urban architecture information extraction of structural information.Must be pointed out that this window is selected just a kind of compromise proposal of actual treatment, because the multiple dimensioned characteristic that image atural object shows makes that a kind of window size is difficult to it is carried out effectively expressing and description.

Two, existing wavelet method can effectively be discerned some textural characteristics, but its weak point is: (1) does not have to solve the above window problem of narration, algorithm adopts bigger stationary window operation, often can only survey in a big way texture (as landforms on a large scale), can't be applied in the actual remote sensing image identification; (2) analysis of wavelet multiresolution rate will produce the higher dimensional space feature, and there is a large amount of information redundancies in these features, and some feature has been proved to be to be unprofitable to pattern-recognition, therefore is necessary a large amount of small echo sub-band energy values is carried out feature extraction; (3) Wavelet Texture and spectral information are not combined, especially in high-resolution remote sensing image, the compound classification for image of spectrum and texture information is vital.

Summary of the invention

At the problems referred to above, the present invention proposes a kind of adjustable spectrum and space characteristics hybrid sorting process, analyze based on multiple dimensioned background, wavelet multiresolution rate, avoided the selection problem of window, can express the feature of various different sizes and yardstick atural object more accurately, improve the precision of decipher, manual intervention is few, be applicable to the automatic classification of middle high-resolution remote sensing image, can effectively improve the nicety of grading and the efficient of such image.

Technical scheme of the present invention is: a kind of adjustable spectrum and space characteristics hybrid sorting process is characterized in that comprising the following steps:

1., with the space characteristics of describing this pixel around a series of multiple dimensioned window of center pixel, the background pyramid of structure target pixel different scale of living in;

2., with each backdrop window of Mallat wavelet transform process, construct the multiresolution pyramid of each multiple dimensioned window, the progression of wavelet decomposition is determined by window size and image resolution, each wavelet decomposition of each window all will produce 4 small echo sub-bands, be respectively low frequency, horizontal high frequency, vertical high frequency and diagonal angle high-frequency sub-band;

3., estimate the eigenwert of asking for each backdrop window multiresolution sub-image with entropy function, and with the ratio of small echo low frequency sub-image and other three director image feature value sums, construct the space invariable rotary characteristic quantity of this window size, reduce the dimension of space characteristics simultaneously;

4., according to the characteristics of wavelet transformation, the backdrop window of different sizes are adopted different wavelet transformation exponent numbers, adopt the method for root window, promptly the backdrop window of different scale is carried out the minimum window of multiresolution analysis, determines the exponent number of wavelet decomposition with this;

5., the user only need select the configuration of backdrop window voluntarily according to actual conditions, after multiple dimensioned window is selected, just calculates and generate the required space characteristics of classification automatically;

6., in the structure of sorter, except considering to treat the spectrum intervals of classification mode and sample average, also add their space length simultaneously;

7., with the irrelevance vector that the mean vector and the variance vectors of sample are tried to achieve space characteristics, weaken of the influence of the higher characteristic parameter of irrelevance to classification;

8., spectrum intervals and space length shared ratio in sorter can be regulated by the user voluntarily according to actual conditions.

Aforesaid adjustable spectrum and space characteristics hybrid sorting process is characterized in that: root window gets 2 * 2 or 4 * 4.

Ultimate principle of the present invention is:

One, a series of multiple dimensioned window of center pixel is set with user's needs according to specific circumstances, and this is to observe the multiple dimensioned focus characteristics of surveying target for the anthropomorphic dummy;

Two, set up the multiresolution analysis of each window with the Mallat conversion, the progression of wavelet decomposition is determined by window size and image resolution;

Three, the Mallat each time to each window decomposes, can obtain 4 small echo sub-bands, ask for the information entropy of this 4 sub-frequency bands and estimate, and construct the space characteristics of an invariable rotary with these 4 values, eliminate the influence of direction, reduce the dimension of spatial information simultaneously;

Four, through after the above step, the employed space characteristics of sorter also just determines thereupon, and whole process only needs the user selecting the configuration of multiwindow at the beginning as the case may be;

Five, on remote sensing images, select the sample of respective classes according to priori, spectrum and spatial information according to sample, estimate structure irrelevance vector with average and variance, different space characteristics is carried out characteristic weighing, limit the weight of unsettled feature in sorter;

Six, the user can select spectrum and space characteristics shared ratio in classification, can prejudge according to priori and imaging characteristic, also can adjust according to sorted result.

Beneficial effect of the present invention

Extract multiple dimensioned spatial information by a series of backdrop window around the center pixel, avoided the limitation of single window operation, ascending window has been avoided the selection problem of window, because atural object is multiple dimensioned system, single window setting may be suitable for some atural object, but can not find a yardstick can be applicable to the characteristics of all atural objects, a kind of effective method is the multiple dimensioned type of focusing of simulation eye-observation object, can express the feature of various different sizes and yardstick atural object so more accurately, improve the precision of decipher; Can effectively cooperate multiple dimensioned windowhood method by the Mallat multiresolution analysis, the multiresolution analysis of different scale can further embody the principle that layering focuses on, the directivity textural characteristics that extracts can effectively be described the spatial characteristics of atural object, the suitable space characteristics of structure on the different resolution grade of each background; By the structure of invariable rotary space characteristics, avoided the directional problems of Wavelet Texture, simultaneously also reduced the dimension of space characteristics because be not the sub-channel of each small echo all be the feature that helps classifying; Constructed the irrelevance vector with sample data statistics, limited the little and unsettled space characteristics shared proportion in sorter that distributes the classification contribution; The ratio of spectral information and space characteristics makes the selection of characteristic of division more flexible in the employing scalable factor control sorter, has 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 middle high-resolution remote sensing image, can effectively improve the nicety of grading and the efficient of such image.

Description of drawings

Fig. 1 is a master routine operational flow diagram of the present invention.

Fig. 2 is multiple dimensioned multiresolution analysis process flow diagram.

Fig. 3 is the flow process of spatial texture feature extraction algorithm.

Fig. 4 is adjustable spectrum-spatial classification method.

Embodiment

1, theoretical foundation

The basic theories that the present invention uses mainly comprises:

(1) utilize wavelet transformation to construct the spatial parameter of pixel.(x is y) in resolution for 2 dimension image f _{2 '}Under wavelet coefficient can be calculated as follows:

$c_{m,n}^{\left(j\right)}=\underset{l,k}{\mathrm{\Σ}}{\mathrm{\φ}}_{k-2m}{\mathrm{\φ}}_{l-2n}{f}_{k,l}^{j-1},w_{m,n}^{(j,h)}=\underset{l,k}{\mathrm{\Σ}}{\mathrm{\φ}}_{k-2m}{\mathrm{\ψ}}_{l-2n}{f}_{k,l}^{j-1}$

$w_{m,n}^{(j,v)}=\underset{l,k}{\mathrm{\Σ}}{\mathrm{\φ}}_{k-2m}{\mathrm{\φ}}_{l-2n}{f}_{k,l}^{j-1},w_{m,n}^{(j,d)}=\underset{l,k}{\mathrm{\Σ}}{\mathrm{\ψ}}_{k-2m}{\mathrm{\ψ}}_{l-2n}{f}_{k,l}^{j-1}$

c _{M, n} ^(j), w _{M, n} ^{(j, h)}, w _{M, n} ^{(j, v)}And w _{M, n} ^{(j, d)}The wavelet coefficient of representing low frequency, horizontal high frequency, vertical high frequency and diagonal angle high frequency sub-image respectively, j is for decomposing exponent number, and ψ (x) is the one dimension wavelet function, has the effect of high-pass filtering; φ (x) is the unidimensional scale function, the effect with low-pass filtering.Select suitable wavelet basis for use, obtain the wavelet coefficient of each sub-image, just can obtain each number of sub images of wavelet decomposition through linear mapping.

(2) traditional minimum distance classifier: $d_{\mathrm{wi}}=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}\left|\right|\mathrm{ij}-l_{w,j}\left|\right|,$ Wherein, d _WiRefer to the spectrum intervals of pixel i to the w class, ij represents the spectral value of pixel i at wave band j, and p is the wave band number of sensor, l _w=[l _{W, 1}, l _{W, 2}, l _{W, 3}... l _{W, p}], l wherein _{W, j}Be the averaged spectrum value of w class ground object sample on wave band j, j=1,2,3 ... p.The classification of pixel i is so: $w^{*}=\underset{w}{\min }\left(d_{\mathrm{wi}}\right).$

2, the structure of space characteristics

Extracted the space characteristics of pixel in the past with wavelet transformation, and often be confined to window and the yardstick fixed, what the researchist was concerned about is difference and the comparisons of various yardsticks to feature extraction.In fact, people's brain and vision are based on continuous yardstick to the observation and the understanding of object, determine the locus and the relation of observed object with background continuous and that constantly enlarge.For the characteristic of simulating human recognition objective, the present invention carries out wavelet transformation to a series of different centers window, and the space characteristic parameter of structure pixel expands to multiwindow and multiple dimensioned to single window and yardstick, extracts the multiscale space information of pixel.Accompanying drawing 1 expression spatial information of the present invention extracts characteristics, and with the space characteristics of two kinds of pyramid evaluating objects pixels, a kind of is the background of the residing different scale of target pixel, as multiwindow 64 * 64,32 * 32,16 * 16 and 8 * 8; Another kind is that each backdrop window is adopted the Mallat pyramid.If original window is I, establishes I and obtain low frequency, vertical, level and diagonal angle director image through i rank wavelet transformation and be respectively AP_i, VD_i, HD_i, DD_i, i=1,2,3.As the eigenwert of each sub-image, formula is as follows with entropy Entropy (brief note is ENT):

ENT＝-∑∑Q(i，j)*log(Q(i，j))；

$Q(i,j)=\frac{{\left|P(i,j)\right|}^2}{\sqrt{\underset{i,j}{\mathrm{\Σ}}{\left|P(i,j)\right|}^2}}$

Wherein, (i j) is wavelet coefficient to P.If sub-image AP_i, VD_i, HD_i, the eigenwert of DD_i is respectively: E (AP_i), E (VD_i), E (HD_i) and E (DD_i).If be combined into the proper vector (n=3i+1) of a n dimension with the eigenwert of each sub-image, so too much spatial parameter can reduce the spectral content in the characteristic of division, and the opinion scale of space characteristics and spectral signature is different, can cause the distortion of hybrid classification feature.So the present invention will make up with the sub-image eigenwert of single order wavelet decomposition, constructs new space characteristic parameter, C1, C2, C3, C4 ... Cm, wherein m is the dimension of space characteristics.For certain window, whenever carry out a wavelet decomposition and will produce 4 new eigenwerts, this paper constructs 1 new normalization eigenwert Cj with these 4 values, like this, this window whenever carries out a wavelet decomposition and just obtains 1 new spatial parameter Cj, and its computing method are as follows, j=1 in the formula, 2,3 ... m, i are the wavelet decomposition exponent number:

$\mathrm{Cj}=\frac{E(\mathrm{AP}\_i)}{E(\mathrm{VD}\_i)+E(\mathrm{HD}\_i)+E(\mathrm{DD}\_i)}$

3, implementation procedure

(1), as the case may be need select the configuration of multiple dimensioned backdrop window with the user, alternative backdrop window has: 64 * 64,32 * 32,16 * 16 and 8 * 8, and the foundation of selection is the characteristics and the image spatial resolution of atural object.The principle of this multiwindow configuration is based on multiple dimensioned many background focus characteristic of human eye observed object, and meets the multiple dimensioned unified characteristic distributions of image atural object.When concrete operations, can select one of them window (be exactly traditional one-window operation this moment), also can select the combination of a plurality of windows simultaneously.

(2), each pixel in the image is all carried out the analysis of Mallat wavelet multiresolution rate according to the determined multiple dimensioned window of the first step, the present invention uses the tight support orthogonal wavelet of Daubechies, and alternative wavelet basis has db1, db2, db3, db4, and db8.The notion of root window is proposed in the analytic process, the minimum window that is multiresolution analysis (is defaulted as 4 * 4 windows, in order to improve practicality, 2 * 2 root window is available in addition), promptly 64 * 64 windows carry out 4 rank wavelet transformations, 32 * 32 windows carry out 3 rank wavelet transformations, and 16 * 16 windows carry out 2 rank conversion, and 8 * 8 carry out the single order conversion.Usually adopt stationary window to compare with existing algorithm, 4 * 4 root window as conversion have reduced window effect (window-size effect) to a great extent, make Wavelet Texture can be applied to the classification of actual remote sensing image, solved the practicality problem of algorithm.Obtain by big window transform because of root window simultaneously, thus the advantage of large scale also had simultaneously, when improving large tracts of land texture recognition precision, also can be with respect to the expression of minutia.

(3), be foundation with the Mallat transformation results of many backdrop window, the Wavelet Texture sequence of structure center pixel, the wavelet transformation each time of each backdrop window all will increase by 1 dimension space feature, the building method of space characteristics sees a joint for details.The advantage of this spatial parameter has: 1) reduced the dimension of spatial signature vectors, kept the balance of spectral content and space content in the classified information, solved the normalization problem of space characteristic parameter; 2) effectively replenished the defective of spectral information in remote Sensing Interpretation; 3), avoided the drawback of traditional single window wavelet method with the backdrop window anthropomorphic dummy's of pyramid vision; 4) each backdrop window is adopted multiresolution analysis, extract the multiscale space feature of atural object.

(4), after above step finishes, the spectrum of each pixel and space characteristics are all determined in the image, whole process only needs the user to select the configuration of backdrop window, so this method possesses good practicality and operability.

(5), select training sample, and according to the average and the variance statistical computation irrelevance vector of sample, the space characteristics in the sorter is weighted: traditional minimum distance classifier is according to the atural object characteristics of image and priori:

$d_{\mathrm{wi}}=\underset{j=1}{\overset{p}{\mathrm{\Σ}}}\left|\right|\mathrm{ij}-l_{w,j}\left|\right|$

Wherein, d _WiRefer to the spectrum intervals of pixel i to the w class, ij represents the spectral value of pixel i at wave band j, and p is the wave band number of sensor, l _w=[l _{W, 1}, l _{W, 2}, l _{W, 3}... l _{W, p}], l wherein _{W, j}Be the averaged spectrum value of w class ground object sample on wave band j, j=1,2,3 ... p.The classification of pixel i is:

$w^{*}=\underset{w}{\min }\left(d_{\mathrm{wi}}\right)$

In sorter, add space characteristic parameter, can get:

$d_{\mathrm{wi}}^s=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{h}_{w,i,k}\×\left|\right|s_{i,k}-s_{w,k}\left|\right|$

Wherein, d _Wi ^sExpression pixel i is to the space length of w class atural object, h _{W, i, k}Be the weight of pixel i with respect to k spatial parameter of w class atural object, promptly adaptive weighted to each space characteristics component, establish pixel i and the corresponding weight vector of w class atural object is h _{W, i}=[h _{W, i, 1}, h _{W, i, 2}, h _{W, i, 3}H _{W, i, m}], we are with two the most basic statistics of sample---variance and average are determined the power h of each characteristic component _{W, i, k}, its calculating relates to 5 vectors: 1) the spatial signature vectors s of pixel i _i=[s _{I, 1}, s _{I, 2}, s _{I, 3}S _{I, m}], wherein m is the space characteristics dimension; 2) suppose that w class atural object has n training sample, q=1,2,3 ... n, the spatial signature vectors of its q sample is: s _{W, q}=[s _{W, q, 1}, s _{W, q, 2}, s _{W, q, 3}S _{I, q, m}]; 3) can get the mean vector s of such spatial signature vectors by the training sample of w class atural object _w=[s _{W, 1}, s _{W, 2}, s _{W, 3}S _{W, m}]; 4) variance vectors that can get such spatial signature vectors by w class ground object sample is v _w=[v _{W, 1}, v _{W, 2}, v _{W, 3}V _{W, m}]; 5) obtain x by above 4 vector calculation _iWith respect to w class atural object depart from vector d _{W, i}=[d _{W, i, 1}, d _{W, i, 2}, d _{W, i, 3}D _{W, i, m}], wherein:

$s_{w,k}=\frac{1}{n}\underset{q=1}{\overset{n}{\mathrm{\Σ}}}{s}_{w,q,j}$

$v_{w,k}=\frac{1}{n-1}\underset{q=1}{\overset{n}{\mathrm{\Σ}}}{(s_{w,q,k}-s_{w,k})}^2$

N is the sample number of w class, k=1, and 2,3 ... m, and:

$d_{w,i,k}=\frac{1}{v_{w,k}\×|s_{i,k}-s_{w,k}|}$

D _{W, i, k}Normalization obtains h _{W, i, k}:

$h_{w,i,k}=\frac{d_{w,i,k}}{\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{d}_{w,j,k}}$

h _{W, i, k}Setting be in order to weaken of the influence of the higher characteristic parameter of irrelevance, to mean this eigenwert the object space feature description is indifferent over the ground, cause the instability of spatial parameter, therefore need using h because irrelevance is higher to classification _{W, i, k}Weaken of the influence of this parameter to classification.

(6), select and after the space characteristics weighting finishes, just can be input to composite character in the sorter at training sample.The design of hybrid classification device need be considered two problems, the one, sorter must be taken spectral signature and spatial information simultaneously into account in decision process, the 2nd, the linear module of spectrum and space characteristics is different, normalization mode difference is major issues so how to control both proportions in sorter.The present invention has designed a kind of fusion spectrum of adjustable ratio joint and the sorter of space characteristics, to address the above problem, in decision process, not only consider to treat the spectrum intervals of merotype and sample average, the space length that adds pattern and sample simultaneously, and scale-up factor is set regulates both roles in sorter, its formula is as follows:

$w^{*}=\underset{w}{\min }(d_{\mathrm{wi}}\×(1-a)+d_{\mathrm{wi}}^s\×a\×b)$

A regulates parameter, the ratio in the control sorter between space parameter and the spectral information, and b is a displacement parameter, and being provided with of it can influence a value with optimal classification effect, makes a value with best classifying quality produce skew, and this point will describe in the accompanying drawings.Like this, the setting of passing ratio coefficient has solved spectrum and the skimble-scamble problem of space characteristics dimension on the one hand, and on the other hand, both can carry out artificial adjustment according to priori and classifying quality to the contribution of classification.

Claims (2)

1. adjustable spectrum and space characteristics hybrid sorting process, its feature may further comprise the steps:

1., with the space characteristics of describing this pixel around a series of multiple dimensioned window of center pixel, the background pyramid of structure target pixel different scale of living in;

2., with each backdrop window of Mallat wavelet transform process, construct the multiresolution pyramid of each multiple dimensioned window, the progression of wavelet decomposition is determined by window size and image resolution, each wavelet decomposition of each window all will produce 4 small echo sub-bands, be respectively low frequency, horizontal high frequency, vertical high frequency and diagonal angle high-frequency sub-band;

3., estimate the eigenwert of asking for each backdrop window multiresolution sub-image with entropy function, and with the ratio of small echo low frequency sub-image and other three director image feature value sums, construct the space invariable rotary characteristic quantity of this window size, reduce the dimension of space characteristics simultaneously;

4., according to the characteristics of wavelet transformation, the backdrop window of different sizes are adopted different wavelet transformation exponent numbers, adopt the method for root window, promptly the backdrop window of different scale is carried out the minimum window of multiresolution analysis, determines the exponent number of wavelet decomposition with this;

5., the user only need select the configuration of backdrop window voluntarily according to actual conditions, after multiple dimensioned window is selected, just calculates and generate the required space characteristics of classification automatically;

6., in the structure of sorter, except considering to treat the spectrum intervals of classification mode and sample average, also add their space length simultaneously;

7., with the irrelevance vector that the mean vector and the variance vectors of sample are tried to achieve space characteristics, weaken of the influence of the higher characteristic parameter of irrelevance to classification;

8., spectrum intervals and space length shared ratio in sorter can be regulated by the user voluntarily according to actual conditions.

2, adjustable spectrum as claimed in claim 1 and space characteristics hybrid sorting process, it is characterized in that: root window gets 2 * 2 or 4 * 4.

Images