CN101853503A - Spectral line inflexion multi-scale optimizing segmentation method and application thereof - Google Patents

Abstract

The invention belongs to the field of hyperspectral remote sensing application and in particular relates to a spectral matching and recognition method. The invention overcomes the defects that the conventional spectral matching and recognition method only considers the whole similarity measurement between spectral lines but neglects the local difference measurement between the spectral lines, and provides a spectral segmentation-based matching and recognition method. The method comprises the following steps of: firstly, performing transformation processing on the spectral lines by adopting the multi-scale wavelet transform taking a second-order Gaussian derived function as a wavelet function; secondly, extracting an optimized inflexion of a spectral curve by using the designed inflexion multi-scale optimizing algorithm; and finally, segmenting the spectral lines based on the extracted optimized inflexion information, and recognizing the spectral lines by adopting a segmentation matching method. The spectral matching and recognition method has the advantages that: wave bands with larger ground object spectrum difference and wave bands with smaller spectrum difference can be segmented into different segmentations through the inflexion segmentation so as to protrude the wave bands with the larger spectrum difference and enhance the effectiveness of spectral matching and recognition.

Description

A kind of spectral line inflexion multi-scale optimizing segmentation method and application thereof

Technical field

The invention belongs to field of hyperspectral remote sensing application, be specifically related to a kind of spectral line inflexion multi-scale optimizing segmentation method and the application in Spectral matching and identification thereof.

Background technology

In high spectrum image was handled, the Spectral matching technology was one of high spectrum terrain classification and key technique in identification.Spectral matching is judged the belonging kinds of atural object by the curve of spectrum that relatively reflects the object spectrum radiation characteristic.Spectral matching method commonly used at present generally all is based upon on the global similarity tolerance basis of the curve of spectrum, yet, in high spectrum image, their characteristic may be close between the different atural object, it is closely similar to be usually expressed as on some wave band spectral line, even the same, but there is notable difference at the other wave band.Global similarity tolerance mode has been ignored the locality difference between spectral line.Therefore, if in spectral similarity tolerance, can keep even highlight this locality difference, will help the segmentation of different atural objects.Given this, the invention provides a kind of Spectral matching and recognition methods based on the flex point segmentation, this method is carried out segmentation according to certain strategy to the curve of spectrum, the bigger wave band of different atural object SPECTRAL DIVERSITY is assigned in the different paragraphs with the less wave band of SPECTRAL DIVERSITY, with the outstanding bigger wave band of SPECTRAL DIVERSITY, the validity of enhanced spectrum coupling and identification.

Summary of the invention

The technical solution adopted for the present invention to solve the technical problems is: at first adopt multi-scale wavelet transformation that spectral line is carried out conversion process, extract the flex point of the curve of spectrum by the inflexion multi-scale optimizing algorithm, utilize flex point information that spectral line is carried out segmentation then, and take the method for segmentation coupling that spectral line is discerned.Concrete technical scheme is as follows:

A kind of spectral line inflexion multi-scale optimizing segmentation method comprises the steps:

(1) establishing optic spectrum line is R (x), x be the ripple segment number (x=1 ..., N), N is a natural number, makes that j is the yardstick variable, its initial value is 1, i.e. j=1, S ^jThe flex point number of optic spectrum line when being j for the yardstick variable; V ⁱFor making the curve of spectrum have the set of the yardstick variable of identical flex point number, i is a token variable, and initial value is 1;

(2) by following formula described optic spectrum line R (x) is carried out multi-scale wavelet transformation:

α=2j; W wherein _aF () is the multi-scale wavelet transformation function, and α is the wavelet transform dimension factor,

Be the result of optic spectrum line R (x) behind multi-scale wavelet transformation;

(3) calculating the flex point of described optic spectrum line R (x) when yardstick variable j is k, also is that j is when equaling k

Zero crossing number and zero crossing wave band position, make Z ^k={ x|W _aF (R (x))==0}, S ^k=# (Z ^k), wherein, # () is the set element counting function, i.e. statistics set Z ^kThe number of middle element, S ^kFlex point number when being k for yardstick variable j, Z ^kThe set of the wave band position of corresponding flex point when being k for yardstick variable j, if under former and later two adjacent yardstick variablees the flex point number of the curve of spectrum identical be S ^kEqual S ^K-1, V then ⁱ=V ⁱ+ { k}, otherwise i=i+1, V ⁱ=φ;

(4) if S ^kEqual 1, promptly optic spectrum line R (x) flex point number is 1, then makes variable je=k, skips to step (5); Otherwise k=k+1 skips to (3), obtains the flex point information of optic spectrum line R (x) under next yardstick;

(5) the maximum continuum V that the flex point number is stablized constant yardstick variable j is kept in searching ^v: $v=\underset{i}{\arg \max }\left\{\right|\max \left(V^i\right)-\min \left(V^i\right)\left|\right\},$ I=1 wherein ..., je;

(6) the more excellent yardstick J that asks for optic spectrum line R (x) is: J=min (q|q ∈ V ^v), then the flex point wave band position of more excellent yardstick is Z ^J

The application of a kind of spectral line inflexion multi-scale optimizing segmentation method in Spectral matching and identification, it specifically comprises the steps:

(1) sets atural object with reference to the M bar being arranged in the spectral line storehouse, be designated as R with reference to spectral line _p(x), (p=1 ..., M), M is a natural number, spectral line to be identified is the N bar, is designated as T _q(x), q=1 ..., N, N are natural number, x is the ripple segment number;

(2) get q bar spectral line T to be identified in the described N bar _q(x), obtain this spectral line T to be identified by described spectral line inflexion multi-scale optimizing method _q(x) yardstick segment information is designated as: { J _q, Z _q, S _q, J wherein _q, Z _qAnd S _qBe respectively this q bar spectral line T to be identified _q(x) yardstick, yardstick flex point wave band position and flex point number;

(3) with this spectral line T to be identified _q(x) yardstick is segmented into benchmark, with spectral line T _q(x) with reference to spectral line R _p(x) be divided into (S _q+ 1) individual segmentation is designated as T respectively _q ^l(x) and R _p ^l(x), (l=1 ..., S _q+ 1);

(4) be calculated as follows spectral line T to be identified _q(x) with all distances with reference to spectral line: $D(T_q\left(x\right),R_p\left(x\right))=\underset{l=1}{\overset{S_q+1}{\mathrm{\Σ}}}\mathrm{dist}(T_q^l\left(x\right),R_p^l\left(x\right)),$ Dist (T wherein _q ^l(x), R _p ^l(x)) be the distance of corresponding each segmentation of two spectral lines;

(5) by the minor increment decision rule spectral line is differentiated, if promptly

D(T _q(x)，R _p(x))＝min{D(T _q(x)，R _p(x))}，

T then _q(x) and R _p(x) genus is similar, promptly identifies this spectral line T to be identified _q(x).

Spectral line T to be identified in the described step (4) _q(x) with reference to spectral line R _p(x) distance adopts spectrum angle, Euclidean distance or other measures to calculate.

Key of the present invention is the extraction of flex point information, owing in the remote sensing survey process of object spectrum, be subjected to the influence of various factors, the object spectrum information of being obtained comprises random noise, the existence of these noises can influence the characteristic of the curve of spectrum, make the curve of spectrum have the flex point information that can not reflect the object spectrum intrinsic in a large number, therefore, before extracting flex point, need to remove in advance these and disturb flex point, to extract stable flex point information.For this reason, the present invention at first adopts the multi-scale wavelet transformation based on gaussian kernel function that the curve of spectrum is carried out level and smooth and denoising, then the curve of spectrum after the denoising is carried out flex point and extracts.The wavelet transformation of different scale has in various degree smooth effect to the curve of spectrum, and the extraction of flex point is had different influences, and on the one hand, yardstick is big more, and is level and smooth and denoising effect is strong more, and the curve of spectrum is level and smooth more, and the flex point of being extracted is insensitive more to noise; On the other hand, yardstick is big more, and the corner position deviation of being extracted is also big more, influences the precision of corner position.Therefore, the present invention has designed a kind of inflexion multi-scale optimizing method, obtaining more excellent flex point information, and based on this, carries out the segmentation coupling.

The invention has the beneficial effects as follows, assign in the different segmentations with the less wave band of SPECTRAL DIVERSITY by the wave band that the flex point segmentation can differ greatly object spectrum, with the outstanding bigger wave band of SPECTRAL DIVERSITY, the validity of enhanced spectrum coupling and identification.

Description of drawings

Fig. 1: spectrum flex point number is with the dimensional variation synoptic diagram;

Fig. 2: curve synoptic diagram behind the different scale wavelet transformation;

Fig. 3: spectral line coupling and identification treatment scheme synoptic diagram.

Embodiment

Below in conjunction with the drawings and specific embodiments the present invention is further described.

1, multi-scale optimizing thought.

In the remote sensing survey process of object spectrum, because the influence of various factors, the spectral information of acquisition comprises noise, and the existence of these noises can influence the characteristic of the curve of spectrum, if the curve of spectrum and Gaussian function are carried out convolution algorithm, then can reach level and smooth and denoising effect.Along with the increase of wavelet scale, curve is level and smooth more, and the extraction of flex point is insensitive more to noise, and the interference flex point that is produced by the noise introducing will be suppressed, but can bring another problem simultaneously, and promptly the position deviation of flex point also can correspondingly increase.Therefore, the size of yardstick has a direct impact the extraction of flex point, thereby influences the coupling of spectral line characteristic.

Fig. 1 has provided spectrum flex point quantity with the graph of a relation that wavelet scale changes, and can find that from figure in certain range scale, the number of flex point is constant, but the wave band position at flex point place is offset to some extent.In Gauss Wavelet Transform, yardstick is big more, and then curve is level and smooth more, just more can resist the disturbance in the virgin curve, but the flex point number of being extracted is just few more; On the other hand, yardstick is more little, and then the curve after the conversion is more near virgin curve, and the residing wave band of flex point position is just accurate more, but the interference flex point that is caused by disturbance in the flex point of extracting is also many more.Improve accuracy, the minimizing of position when therefore flex point is extracted as far as possible and disturb flex point.For object spectrum, its topmost zone of reflections and absorption band show as apparent in view crest and trough in the curve of spectrum, they are not easy to be subjected to the influence of yardstick, and when yardstick changed within the specific limits, these crests and trough can be not smoothed fall.According to this characteristic, can investigate and keep the metastable range scale of flex point number.For the flex point of the curve of spectrum that extracts, it is big more to keep the constant range scale of their numbers, and then crest between these flex points and trough are just stable more, and they just can reflect the absorption and the reflection characteristic of spectrum more exactly.If can find the range scale of keeping the constant maximum of flex point number, this range scale is exactly more excellent range scale so, and pairing flex point number is exactly more excellent flex point number.In addition, because yardstick is more little, the corner position that this yardstick extracts down is just accurate more, and therefore minimum yardstick is exactly that more excellent flex point is extracted yardstick in the more excellent range scale, and the corner position that this yardstick extracts down is exactly more excellent corner position.

Yet in actual conditions, along with the continuous increase of yardstick, curve will be more and more level and smooth, and final flex point will fade away (as shown in Figure 2).So, the range scale of keeping the constant maximum of flex point number that obtains of adding up is exactly the pairing scope of zero flex point, because it is infinitely-great.In this case, reflection and absorbing features can't embody fully.Equally, have only in one in the flex point number, reflection and absorbing features also can't embody, because any absorption band or the zone of reflections all can be positioned in the middle of two flex points.Therefore, more excellent yardstick is only sought in greater than 1 range scale keeping the flex point number.

2, inflexion multi-scale optimizing segmentation method.

The method that the present invention selects the multi-scale wavelet transformation based on the second order Gauss derived function to extract as flex point, this is mainly based on the noise inhibiting ability and the multiscale analysis ability of Gauss wavelet.Select for use Gauss's second order derived function to be as the advantage of wavelet basis function, after original signal and the Gaussian function convolution, be difficult for producing ringing effect, can avoid the influence of ringing effect to the extraction of flex point like this, the flex point that guarantees to extract is between the zone of reflections and absorption band.

The definition Gaussian function is:

$g_a\left(x\right)=2{\left(\mathrm{\πa}\right)}^{-1/2}\exp \left(\frac{{-x}^2}{4a}\right)---\left(1\right)$

Then wavelet basis function is:

${\mathrm{\ψ}}_a\left(x\right)=\frac{d^2{g}_a\left(x\right)}{d{x}^2}---\left(2\right)$

Wavelet transformation is:

W _af(x)＝f(x)*ψ _a(x) (3)

Wherein, f (x) is the curve of spectrum, and α is the wavelet scale factor.

Can prove

$W_{a}f\left(x\right)=f\left(x\right)*\[\frac{d^2{g}_a\left(x\right)}{d{x}^2}\]=\frac{d^2}{d{x}^2}(f\left(x\right)*g_a\left(x\right))---\left(4\right)$

Because basis function is Gauss's second order derived function, therefore, can think approx that the curve behind the wavelet transformation is the second derivative curve of curve before the conversion, like this, the zero crossing correspondence of curve the flex point of virgin curve after the conversion.

In above-mentioned wavelet transformation, wavelet scale is determined by scale factor α, change the wavelet transformation that scale factor can obtain different scale, utilize the different scale wavelet transformation that the curve of spectrum is handled, can obtain the knee point information under the different scale, scale factor is big more, and knee point quantity is then few more, and when yardstick increased to certain value, the knee point number reduced to 1.

In order to obtain the more excellent flex point information of curve, the present invention adopts iterative manner that this curve of spectrum is carried out multi-scale wavelet transformation and handles, wavelet scale factor-alpha value is 1 during primary iteration, later on each iteration wavelet scale factor-alpha increases by 1, up to scale factor point of inflexion on a curve quantity is dropped to till 1, and remember that the value of this out to out factor-alpha is je.Then to value 1 and the je scope in all scale factors analyze, searching makes curve have the maximum continuum of the scale factor of identical flex point number, and being more excellent scale factor with this interval smallest dimension, the pairing knee point of this scale factor is more excellent flex point.Its detailed process flow process is as follows:

(1) establishing the atural object spectral line is R (x), x be the ripple segment number (x=1 ..., N), N is a natural number, makes that j is the yardstick variable, its initial value is 1, i.e. j=1, S ^jThe flex point number of the curve of spectrum when being j for the yardstick variable; V ⁱFor making the curve of spectrum have the set of the yardstick variable of identical flex point number, i is token variable (natural number), and initial value is 1;

(2) by following formula spectral line R (x) is carried out multi-scale wavelet transformation, promptly

α=2j; W wherein _aF () is the multi-scale wavelet transformation function, and α is the wavelet transform dimension factor,

Be the result of R (x) behind multi-scale wavelet transformation;

(3) calculating the flex point of spectral line R (x) when yardstick variable j is k, also is that j is when equaling k

Zero crossing number and zero crossing wave band position, make Z ^k={ x|W _aF (R (x))==0}, S ^k=# (Z ^k), wherein, # () is the set element counting function, i.e. statistics set Z ^kThe number of middle element, S ^kFlex point number when being k for yardstick variable j, Z ^kThe set of the wave band position of corresponding flex point when being k for yardstick variable j.If under former and later two adjacent yardstick variablees the flex point number of the curve of spectrum identical (be S ^kEqual S ^K-1), V then ⁱ=V ⁱ+ { k}, otherwise i=i+1, V ⁱ=φ;

(4) if S ^kEqual 1, promptly curve of spectrum flex point number is 1, then makes variable je=k, skips to step (5); Otherwise k=k+1 skips to (3), obtains the flex point information of spectral line R (x) under next yardstick;

(5) the maximum continuum V that the flex point number is stablized constant yardstick variable j is kept in searching ^v: $v=\underset{i}{\arg \max }\left\{\right|\max \left(V^i\right)-\min \left(V^i\right)\left|\right\},$ I=1 wherein ..., je;

(6) the more excellent yardstick J that asks for spectral line R (x) is: J=min (q|q ∈ V ^v), then the flex point wave band position of more excellent yardstick is Z ^J

3, based on the spectral line matching process of flex point segmentation

The spectrum identifying can be regarded a process of spectral line being carried out match classifying according to certain measurement criterion as, in this process, calculate unknown spectral line and all classes distance at first respectively with reference to spectral line, then each distance of calculating gained is compared, find out the reference spectral line that has minor increment with unknown spectral line, and judge that unknown spectral line and this belong to same class with reference to spectral line.

Can extract the more excellent flex point of the curve of spectrum by the inflexion multi-scale optimizing algorithm, obtain just can carry out segmentation to spectral line after the flex point with these flex points, and take the method for segmentation coupling that spectral line is discerned, and it handles block diagram as shown in Figure 3, and concrete treatment scheme is as follows:

(1) establishes atural object with reference to the M bar being arranged in the spectral line storehouse, be designated as R with reference to spectral line _p(x), (p=1 ..., M), M is a natural number.Spectral line to be identified is the N bar, is designated as T _q(x), (q=1 ..., N), N is a natural number, x is the ripple segment number;

(2) get q bar spectral line to be identified in the described N bar, obtain the more excellent yardstick segment information of this spectral line, be designated as: { J by more excellent yardstick flex point extracting method _q, Z _q, S _q.J wherein _q, Z _qAnd S _qBe respectively this q bar spectral line T to be identified _q(x) more excellent yardstick, more excellent yardstick flex point wave band position and flex point number;

(3) with this spectral line T to be identified _q(x) more excellent yardstick is segmented into benchmark, with spectral line T _q(x) and R _p(x) be divided into (S _i+ 1) individual segmentation remembers that respectively l is segmented into T _q ^l(x) and R _p ^l(x), (l=1 ..., S _q+ 1);

(4) be calculated as follows T _q(x) with all with reference to spectral line R _p(x) distance: $D(T_q\left(x\right),R_p\left(x\right))=\underset{l=1}{\overset{S_q+1}{\mathrm{\Σ}}}\mathrm{dist}(T_q^l\left(x\right),R_p^l\left(x\right)),$ Dist (T wherein _q ^l(x), R _p ^l(x)) be the distance of corresponding each segmentation of two spectral lines, this distance can adopt spectrum angle, Euclidean distance or other measures to calculate;

(5) by the minor increment decision rule spectral line is differentiated, if promptly

D (T _q(x), R _p(x))=min{D (T _q(x), R _p(x)) }, T then _q(x) and R _p(x) genus is similar;

Repeating step (2)～(5) are discerned all spectral lines to be identified.

4, application example of the present invention is analyzed

For the validity based on the spectral line recognizer of flex point segmentation is analyzed, the inventor has extracted the sample data of 6 quasi-representative atural objects such as comprising soil, road, house, water body and vegetation from actual high spectrum image, total sample number 30814, wherein soil sample is 733,3982 in road sample, 3784 in house sample, 4410 in water body sample, 16143 in vegetation sample, 1762 in shade sample.These samples are adopted segmentation spectrum angle match classifying method and the not segmentation spectrum angle match classifying method contrast test of classifying respectively.

Specific implementation process is that 30814 samples are arranged immediately, forms spectroscopic data collection to be identified.Every class sample is got average, and with the reference spectra of average, obtain the reference spectra of 6 class atural objects, form atural object with reference to the spectral line storehouse as every class sample, then by the processing of classifying of following treatment scheme:

(1) note atural object is with reference to spectral line storehouse R _p(x), (p=1 ..., 6).Spectral line to be identified is T _q(x), (q=1 ..., 30814), x is the ripple segment number;

(2) from spectra database to be identified, choose a curve of spectrum to be identified successively, be designated as q bar spectral line to be identified, obtain the more excellent yardstick segment information of this spectral line, be designated as: { J by the inflexion multi-scale optimizing method _q, Z _q, S _q.J wherein _q, Z _qAnd S _qBe respectively this i bar spectral line T to be identified _i(x) more excellent yardstick, more excellent yardstick flex point wave band position and flex point number;

(3) with this spectral line T to be identified _q(x) more excellent yardstick is segmented into benchmark, with spectral line T _q(x) and R _p(x) be divided into (S _q+ 1) individual segmentation remembers that l is segmented into T _q ^l(x) and R _p ^l(x), (l=1 ..., S _q+ 1);

(4) calculate T respectively by following formula _q(x) with 6 with reference to spectral line R _p(x) distance: $D(T_q\left(x\right),R_p\left(x\right))=\underset{l=1}{\overset{S_q+1}{\mathrm{\Σ}}}\mathrm{dist}(T_q^l\left(x\right),R_p^l\left(x\right)),$ Dist (T wherein _q ^l(x), R _p ^l(x)) expression formula is: $\mathrm{dist}(T_q^l\left(x\right),R_p^l\left(x\right))=\frac{\underset{x=B_l}{\overset{E_l}{\mathrm{\Σ}}}{T}_q^l\left(x\right)R_p^l\left(x\right)}{\sqrt{\underset{x=B_l}{\overset{E_l}{\mathrm{\Σ}}}{T}^k\left(x\right)T_q^l\left(x\right)\sqrt{\underset{x=B_l}{\overset{E_l}{\mathrm{\Σ}}}{R}_p^l\left(x\right)R_p^l\left(x\right)}}}$

B _lBe the initial ripple segment number of l segmentation of spectral line, E _lThe termination ripple segment number of l segmentation of spectral line.

(5) by the minor increment decision rule spectral line is differentiated, if promptly

D (T _q(x), R _p(x))=min{D (T _q(x), R _p(x)) }, T then _i(x) and R _j(x) genus is similar;

(6) repeating step (2)～(5) are discerned all spectral lines to be identified.

Experimental result is shown in table 1 and 2, and experimental result shows that the inventive method can effectively be improved the Spectral matching effect, improves the Target Recognition precision, and the Target Recognition accuracy is improved significantly.

Table 1 is based on the spectrum angle recognizer of flex point segmentation

Table 2 direct sunshine spectral corner recognizer

The evaluation index explanation:

Application example analysis of the present invention adopts confusion matrix to estimate, and confusion matrix is a kind of expression-form commonly used in the area of pattern recognition, and it describes the true type of sample and the relation between the recognition result type, is a kind of common method of estimating classification performance.Confusion matrix is defined as follows:

In the formula, m _IjThe pixel that expression test site planted agent belongs to the i class is assigned to the sum of all pixels that goes in the j class, and n is the classification number.If the element value of obscuring on the diagonal line is big more, then presentation class result's reliability is high more, as if the element value on the off-diagonal is big more in the confusion matrix, represents that then the phenomenon of mis-classification is serious more.

The nicety of grading evaluation index has multiple, adopts producer's precision (Producer Accuracy), overall accuracy (Overall Accuracy) and three kinds of indexs of Kappa coefficient here.

(1) producer's precision (PA)

Refer to that the correct number of categories of a certain classification accounts for the ratio of this classification pixel sum in the reference data, the production precision is embodied in the confusion matrix and then is:

$\mathrm{PA}=\frac{m_{\mathrm{ii}}}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}}$

(2) overall accuracy (OA)

Refer to that total correct number of categories accounts for the ratio of total sampling number, it has reflected the correct degree that classification results is total.Utilize confusion matrix to be expressed as:

$\mathrm{OA}=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{m}_{\mathrm{ii}}}{\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{m}_{\mathrm{ij}}}$

(3) Kappa coefficient

Because total classification precision has only been utilized the element on the confusion matrix diagonal line, and do not utilize the information of whole confusion matrix, still owe not enough as comprehensive measurement of error in classification, the Kappa coefficient can utilize the information of confusion matrix all sidedly, can be used as the overall target that nicety of grading is estimated, the Kappa coefficient can calculate with following formula:

$K=\frac{N\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{m}_{\mathrm{ii}}-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left(m_{i+}{m}_{+i}\right)}{N^2-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\left(m_{i+}{m}_{+i}\right)}$

In the formula, n is classification matrix ranks numbers, m _IjBe the element value of the capable j row of i in the confusion matrix, m _I+And m _{+ i}, the capable summation and the row summation of difference presentation class confusion matrix, N is the total inspection value, i.e. all elements sum in the hybrid matrix.

Claims (3)

1. a spectral line inflexion multi-scale optimizing segmentation method comprises the steps:

(1) establishing optic spectrum line is R (x), x be the ripple segment number (x=1 ..., N), N is a natural number, makes that j is the yardstick variable, its initial value is 1, i.e. j=1, S ^jThe flex point number of optic spectrum line when being j for the yardstick variable; V ⁱFor making the curve of spectrum have the set of the yardstick variable of identical flex point number, i is a token variable, and initial value is 1;

(2) by following formula described optic spectrum line R (x) is carried out multi-scale wavelet transformation:

α=2j; W wherein _aF () is the multi-scale wavelet transformation function, and α is the wavelet transform dimension factor, R ^%jBe the result of optic spectrum line R (x) behind multi-scale wavelet transformation;

(3) calculating the flex point of described optic spectrum line R (x) when yardstick variable j is k, also is j R when equaling k ^%jZero crossing number and zero crossing wave band position, make Z ^k={ x|W _aF (R (x))==0}, S ^k=# (Z ^k), wherein, # () is the set element counting function, i.e. statistics set Z ^kThe number of middle element, S ^kFlex point number when being k for yardstick variable j, Z ^kThe set of the wave band position of corresponding flex point when being k for yardstick variable j, if under former and later two adjacent yardstick variablees the flex point number of the curve of spectrum identical be S ^kEqual S ^K-1, V then ⁱ=V ⁱ+ { k}, otherwise i=i+1, V ⁱ=φ;

(4) if S ^kEqual 1, promptly optic spectrum line R (x) flex point number is 1, then makes variable je=k, skips to step (5); Otherwise k=k+1 skips to (3), obtains the flex point information of optic spectrum line R (x) under next yardstick;

(5) the maximum continuum V that the flex point number is stablized constant yardstick variable j is kept in searching ^v:

$v=\underset{i}{\arg \max }\left\{\right|\max \left(V^i\right)-\min \left(V^i\right)\left|\right\},$ I=1 wherein ..., je;

(6) the more excellent yardstick J that asks for optic spectrum line R (x) is: J=min (q|q ∈ V ^v), then the flex point wave band position of more excellent yardstick is Z ^J

2. the application of the described a kind of spectral line inflexion multi-scale optimizing segmentation method of claim 1 in Spectral matching and identification, it specifically comprises the steps:

(1) sets atural object with reference to the M bar being arranged in the spectral line storehouse, be designated as R with reference to spectral line _p(x), (p=1 ..., M), M is a natural number, spectral line to be identified is the N bar, is designated as T _q(x), q=1 ..., N, N are natural number, x is the ripple segment number;

(2) get q bar spectral line T to be identified in the described N bar _q(x), obtain this spectral line T to be identified by the described spectral line inflexion multi-scale optimizing segmentation method of claim 1 _q(x) yardstick segment information is designated as: { J _q, Z _q, S _q, J wherein _q, Z _qAnd S _qBe respectively this q bar spectral line T to be identified _q(x) yardstick, yardstick flex point wave band position and flex point number;

(3) with this spectral line T to be identified _q(x) yardstick is segmented into benchmark, with spectral line T _q(x) with reference to spectral line R _p(x) be divided into (S _q+ 1) individual segmentation remembers that respectively l is segmented into T _q ^l(x) and R _p ^l(x), (l=1 ..., S _q+ 1);

(4) be calculated as follows spectral line T to be identified _q(x) with all distances with reference to spectral line: Dist (T wherein _q ^l(x), R _p ^l(x)) be the distance of corresponding each segmentation of two spectral lines;

(5) by the minor increment decision rule spectral line is differentiated, if promptly

D(T _q(x)，R _p(x))＝min{D(T _q(x)，R _p(x))}，

T then _q(x) and R _p(x) genus is similar, promptly identifies this spectral line T to be identified _q(x).

3. the application of spectral line inflexion extracting method according to claim 2 in Spectral matching and identification is characterized in that spectral line T to be identified in the described step (4) _q(x) with reference to spectral line R _p(x) distance adopts spectrum angle, Euclidean distance or other measures to calculate.

Images