CN100478992C - Vector filtering method of high spectrogram based on detection of neighborhood background - Google Patents

Abstract

A vector filter method includes presenting 3D light spectrum image to be 2D vector matrix 1 = [ Pi, j ] as 0 <i< M, 0<j<N; setting neighbourhood structure S = [ Su, v ] as Su, v = 0,1, U = -1, 0, 1, V = 1, 0, 1; sliding neighbourhood S all over 2D vector matrix; driving out background spectrum vector BG in neighbourhood range of each image element; using said vector BG to replace spectrum Pi, j in neighbourhood and using the vector to go all over the image.

Description

Vector filtering method of high spectrogram based on the neighborhood background detection

Technical field:

The present invention relates to a kind of high spectrum image Processing Algorithm, particularly a kind of general two-dimension non linearity digital filtering---based on the vector filtering method of local background's statistics.

Background technology:

The high light spectrum image-forming technology is the important component part in modern space remote sensing of the earth field.High spectrum image writes down the spectral signature information on a plurality of wave bands of each target pixel simultaneously, can be described as three-dimensional data---and therefore spectrum dimension of two space peacekeepings is also referred to as the image cube.It can obtain the type of the corresponding atural object of this pixel or the information of composition by analyzing the spectral characteristic of pixel.Particularly entered the high spectrum image of application in recent years, have wave band narrow bandwidth (being generally less than 10nm), wave band number many (tens to hundreds ofs), wave band is continuous, can obtain the object spectrum curve, both bring great potential, also bring challenges to Flame Image Process to remote sensing application.

High spectrum image is owing to the smaller bandwidth of wave band, and energy is faint, so signal to noise ratio (S/N ratio) is less, and ground footprint point is bigger, and the spectrum mixing phenomena is outstanding, and these all cause difficulty to spectral analysis, and are low as nicety of grading, and classification chart comprises a large amount of points etc. that loose.Before the high spectrum image classification data are carried out filtering, filtering ground small size spectrum singular point or noise will improve classifying quality.

Summary of the invention:

The object of the present invention is to provide a kind of vector filtering method of high spectrogram, in order to effective elimination high spectrum image noise and the diffusing point of spectrum based on local background's statistics.

Technical scheme of the present invention is as follows:

According to a kind of vector filtering method of high spectrogram that detects based on neighborhood background of the present invention, its step comprises:

A. three-dimensional spectrum picture is expressed as two-dimensional vector matrix I:

I＝[P _i，j]，0＜i＜M，0＜j＜N (1)

Wherein M is the pixel number of the every row of image, and N is a picturedeep, and M, bold-faced letter P represent the spectrum vector of an image picture elements: P=[P ₁, P ₂... P _B], subscript 1,2 ..., B is the wave band number of spectrum picture,

In spectrum picture, the spectrum of random noise or ground small size atural object is often different with the most of pixels of neighborhood, and unification is referred to as local unusual spectrum; The spectrum of probability of occurrence maximum vector is local background's spectrum in said subrange, and non-background spectrum just might be unusual spectrum or edge, replaces unusual spectrum with background spectrum, just can realize removing the purpose of random noise or ground small size atural object,

B. define structure (Structure element) S of neighborhood, the definition of neighbour structure is identical with morphologic structural element, and General Definition quadrate or circular neighbour structure with the 0 non-neighborhood pixel of expression with the effective neighborhood pixel of 1 expression, then can be expressed as:

S＝[Su，v]；Su，v＝0，1；u＝-1，0，1；v＝-1，0，1 (2)

In like manner can adopt 5 * 5 as required, neighbour structures such as 7 * 7.......

C. neighbour structure s is slided on whole two-dimensional vector matrix I;

D. in the neighborhood scope of each pixel, find the solution background spectrum vector B G;

E. the spectrum P that replaces the center pixel of neighbour structure then with the background spectrum vector B G that tries to achieve _{I, j}

F. travel through entire image at last and carry out vector filtering, above-mentioned filtering method essence is the local small probability spectrum of corrosion, the process of preserving local big probability spectrum, and therefore, the spatial area of the unusual spectrum in the image is during less than neighbour structure, and this atural object is eliminated.

In the steps d, find the solution the core that the background spectrum vector is the inventive method, its success or failure will influence the accuracy of spectral information.Therefore, should clear and definite subrange in the definition of spectrum of probability of occurrence maximum, promptly the spectrum of probability of occurrence maximum refers to through cluster analysis, the central point spectrum of the class that pattern feature point is the most intensive, quantity is maximum.Can design multiple cluster method for solving according to this definition, but this algorithm adopts sliding window, too Fu Za clustering method increases calculated amount easily, because the pixel number is less in the neighbour structure, therefore recommends to adopt approximate calculation method again.

Further, the definition of background spectrum is simplified, the central point of the class of the most intensive quantity maximum of pattern feature point is reduced to the distribute central point in the most intensive zone of si unique point, thereby, be reduced to apart from si-1 unique point distance and the point of minimum, done such simplification and just can get rid of apart from the class of point that disperses and pixel negligible amounts.

More specifically establish sn nonzero element arranged in the neighbour structure, 1. the Euclidean distance of the spectrum vector of obtaining these element correspondences between in twos, form distance matrix, each vector of matrix has comprised the pixel in the neighbour structure and the distance of every other neighborhood pixel, 2. to each vectorial ascending ordering of this matrix, before 3. asking si distance and minimum value d _BG, the pixel at its place is exactly the position at background spectrum place.Represent ascending ordering with order (), || Euclidean distance between the vector is asked in expression, and I is a pixel vector, and its subscript is illustrated in the sequence number in the neighbour structure, i.e. di, j=||Ii-Ij||

$d_{\mathrm{BG}}=\min \begin{array}{}\end{array}\left[\begin{array}{c}\underset{1}{\overset{\mathrm{si}}{\mathrm{\&Sigma;}}}\mathrm{<figure-callout\; id="1"\; label="order\; d"\; filenames="C20051002631800081.png"\; state="\{\{state\}\}">order</figure-callout>}\left(\begin{array}{c}{d}_{}\end{array}\right)\left(\begin{array}{cccc}{}_{\mathrm{1,2}}& {\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="C20051002631800081.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\mathrm{1,3}}& <figure-callout\; id="1"\; label="...\; d"\; filenames="C20051002631800081.png"\; state="\{\{state\}\}">.</figure-callout>..& d_{}{}_{1,\mathrm{sn}}\end{array}\right)\\ \underset{1}{\overset{\mathrm{si}}{\mathrm{\&Sigma;}}}\mathrm{order}\left(\begin{array}{cccc}{d}_{\mathrm{2,1}}& d_{\mathrm{2,3}}& ...& d_{2,\mathrm{sn}}\end{array}\right)\\ \begin{array}{cccc}...& ...& ...& ...\end{array}\\ \underset{1}{\overset{\mathrm{si}}{\mathrm{\&Sigma;}}}\mathrm{order}\left(\begin{array}{cccc}{d}_{\mathrm{sn},1}& d_{\mathrm{sn},2}& ...& d_{\mathrm{sn},\mathrm{sn}-1}\end{array}\right)\end{array}\right]---\left(3\right)$

Wherein si generally is made as half of neighborhood pixel number, i.e. (sn-1)/2.

Advantage of the present invention is: this filtering method not only can be applicable to spectrum picture but also can be applicable to gray level image, can corrode the unusual pixel of small size, therefore can effectively eliminate the diffusing point of noise and spectrum, do not introduce non-existent new spectrum in the neighborhood scope, keep original spectrum, in addition there is certain sharpening at the edge.

Description of drawings:

Fig. 1 is a kind of 3 * 3 square neighbour structure synoptic diagram.

Fig. 2 is a background spectrum cluster synoptic diagram.

Fig. 3 is the vector filtering program flow diagram that background of the present invention detects.

Fig. 4 is the process flow diagram of simplifying procedures that background spectrum of the present invention is found the solution.

Fig. 5 is the nearly true color composite diagram of original high spectrum.

Fig. 6 is the filtered images of of the present invention 3 * 3 square neighbour structures.

Fig. 7 is the filtered images of of the present invention 5 * 5 square neighbour structures.

Fig. 8-the 1st is without the original high-spectrum of neighbour structure filtering of the present invention.

Fig. 8-the 2nd is through of the present invention 5 * 5 square filtered images of neighbour structure that form.

Fig. 9-the 1st, high spectrum image is through the filtered outline map of vector median.

Fig. 9-the 2nd, the outline map of high spectrum image behind neighbour structure vector filtering of the present invention.

Figure 10-the 1st, the filtered outline map of colored aviation image vector median.

Figure 10-the 2nd, colored aviation image outline map after via neighbour structure vector filtering of the present invention.

Embodiment:

Provide better embodiment of the present invention according to Fig. 1～Figure 10-2 below, and described in detail, so that method characteristics of the present invention and function are described better, enable to be easier to understand the present invention, rather than be used for limiting scope of the present invention.

Please consult Fig. 1 earlier, Fig. 2 and Fig. 3, as shown in the figure, the vector filtering method of high spectrum image of the present invention is realized by the vector filtering program 100 that background shown in Figure 3 detects.

S1001 is expressed as two-dimensional vector matrix I to three-dimensional spectrum picture:

I＝[P _i，j]，0＜i＜M，0＜j＜N (1)

Wherein M is the pixel number of the every row of image, and N is a picturedeep, and bold-faced letter P represents the spectrum vector of an image picture elements: P=[P ₁, P ₂... P _B], subscript 1,2 ..., B is the wave band number of spectrum picture,

S1002, structure (Structure element) S of definition neighborhood, General Definition quadrate or circle, as shown in Figure 1, present embodiment adopts 3 * 3 square or circular neighbour structure, the non-neighborhood pixel of 0 expression, the effective neighborhood pixel of 1 expression can be expressed as:

S＝[Su，v]；Su，v＝0，1；u＝-1，0，1；v＝-1，0，1 (2)

In like manner can adopt 5 * 5 as required, neighbour structures such as 7 * 7......,

S1003, neighbour structure S slides on whole two-dimensional vector matrix I,

S1004 asks the background spectrum vector B G of the pixel vector of I under the neighbour structure,

S1005, with the spectrum of the image I under the pixel in the middle of the structure of background spectrum vector B G replacement neighborhood,

S1006, is the traversal image I through with? if do not finish, just return and carry out S1003; If finish, then carry out S1007,

S1007, filtering finishes.

It is the core of the inventive method that described S1004 finds the solution background spectrum vector B G, and its success or failure will influence the accuracy of spectral information.Therefore, should first clear and definite subrange in the definition of spectrum of probability of occurrence maximum, as shown in Figure 2, the spectrum of probability of occurrence maximum refers to through cluster analysis, the central point spectrum of the class of the most intensive quantity maximum of pattern feature point.Can design multiple cluster method for solving according to this definition, but this algorithm adopts sliding window, too Fu Za clustering method increases calculated amount easily, again because the pixel number is less in the neighbour structure, therefore, in the present embodiment, adopt approximate calculation method as shown in Figure 4, promptly by 200 realizations of simplifying procedures of background spectrum, its operating procedure is as follows successively:

S2001, the spectrum vector of obtaining all elements correspondence under the neighbour structure distance between in twos,

S2002, to the ascending ordering of distance value of each pixel and all other neighborhood pixels,

S2003, ask each pixel and all other neighborhood pixel distances preceding Si minimum value with,

S2004, before asking Si apart from minimum value and the pixel position of minimum,

S2005 confirms that the pixel spectra at place is exactly a background spectrum.

The effect that present embodiment adopts two width of cloth remote sensing images to come verification algorithm, a width of cloth are colorful digital aeroplane photography photograph, and a width of cloth is the high spectrum image of 124 wave bands.

Fig. 5 is that Fig. 6, Fig. 7 are respectively the result of 3 * 3 and 5 * 5 square neighbour structure filtering through 124 wave band high spectrum images behind the radiant correction, and the small size atural object in the filtered image is corroded, and spectral distribution is more concentrated, and the details of different scale is eliminated.

Fig. 8-1 and Fig. 8-2 is that the unsupervised classification before and after the filtering of colorful digital aviation image compares, and clearly, a large amount of points that loose in the raw video (Fig. 8-1) are by filtering.

Median vector filtering is a kind of filtering method commonly used, difference for the edge performance of verifying median filter method and filtering method of the present invention, to having carried out rim detection behind two kinds of image filterings, the algorithm of spectrum picture rim detection has adopted the sobel operator based on Euclidean distance:

Gx＝||P[i+1，j+1]+2*P[i+1，j]+P[i+1，j-1]-(P[i-1，j+1]+2*P[i-1，j]+P[i-1，j-1])||

Gy＝||P[i+1，j-1]+2*P[i，j-1]+P[i-1，j-1]-(P[i+1，j+1]+2*P[i，j+1]+P[i-1，j+1])||

$G=\sqrt{{\mathrm{Gx}}^2+{\mathrm{Gy}}^2}$

Fig. 9-1 and Fig. 9-2 and Figure 10-1 and Figure 10-2 are respectively that the rim detection effect of high spectrum image and coloured image compares, Fig. 9-1 and Figure 10-1 image border is a median vector filtering gradient map, Fig. 9-2 and Figure 10-2 is filtering gradient map of the present invention, both have passed through same linear stretch, the latter's edge is more precipitous, dynamic range is bigger, and this situation is more obvious in high spectrum image.

Claims (3)

1, a kind of vector filtering method of high spectrogram that detects based on neighborhood background, its step comprises:

A. three-dimensional spectrum picture is expressed as two-dimensional vector matrix I:

I＝[P _i，j]，0＜i＜M，0＜j＜N

Wherein M is the pixel number of the every row of image, and N is a picturedeep, and bold-faced letter P represents the spectrum vector of an image picture elements: P=[P ₁, P ₂... P _B], subscript 1,2 ..., B is the wave band number of spectrum picture,

B. define neighbour structure S,

C. neighbour structure S is slided in the entire image scope,

D. in the neighborhood scope of each pixel, find the solution background spectrum vector B G,

E. the spectrum P that replaces the center pixel of neighbour structure then with the background spectrum vector B G that tries to achieve _{I, j},

F. last, go through all over entire image and carry out vector filtering.

2, the vector filtering method of the high-spectrum image that detects based on neighborhood background according to claim 1, it is characterized in that, neighbour structure S in the step is 3 * 3 or 5 * 5 or 7 * 7 square or circular neighbour structure, and, can be expressed as with the 0 non-neighborhood pixel of expression with the effective neighborhood pixel of 1 expression:

S＝[Su，v]；Su，v＝0，1；u＝-1，0，1；v＝-1，0，1。

3, the vector filtering method of the high spectrum image that detects based on neighborhood background according to claim 1, it is characterized in that, the background spectrum vector B G that finds the solution in the described steps d is the most intensive pattern feature point successively, the central point of the class of quantity maximum is reduced to the distribute central point in the most intensive zone of Si unique point, obtains the distance between the spectrum vector of all elements correspondence under the field structure, the ascending ordering of distance value to each pixel and all other neighborhood pixels, ask each pixel and all other field pixel distances Si minimum value with, before asking Si apart from minimum value and the pixel position of minimum, the pixel spectra of confirming the place at last is exactly a background spectrum.

Images