CN101038667A - Scale self-adaptive image segmentation method - Google Patents

Abstract

The invention discloses a size self-adaptive image segmentation method comprising following steps: 1) images and one or more image layers of converted results of the images determining an image gather to be segmented; 2) setting a segmentation method and a size increasing manner, and segmenting the images by a continuous variated size coefficient; 3) segmented results in different sizes forming a tree-structure image object expression based on combined relationships between image speckles; (4) defining markedness of image speckles; (5) image speckles segmented with continuous variated sizes composing an image speckle development curve; (6) forming a markedness curve and a segmentation beginning curve of the image speckles in a segmentation development process; (7) calculating extremums in the markedness curve to form a size reverse order; 8) forming an extremum size image according to the extremums in the size reverse order; 9) determining segmentation image speckles based on predominance sizes in the extremum size image. The invention defines an optimum segmentation size of the iamge object according to self-properties of the image speckles so that objects with different sizes have condign segmentation sizes, respectively.

Description

A kind of image partition method of dimension self-adaption

Technical field

The present invention relates to image segmentation and image understanding technical field, relate in particular to a kind of image partition method of dimension self-adaption.

Background technology

Image segmentation is the important algorithm of Flame Image Process, has had numerous algorithms already.Image partition method is the important means that improves in nicety of grading and the extraction terrestrial object information in remote sensing image processing, and image partition method has been realized pie graph spot object in remote sensing images.

The application image cutting techniques is the pie graph spot from image, and main image Segmentation Technology has two kinds [1,2], and is a kind of based on rim detection, follows the tracks of by the edge, and the closed curve of formation constitutes little figure spot; Another kind of based on region growing, the specific discriminant function of foundation is the figure spot with close pixel merger, as the multiscale analysis method in the remote sensing image processing etc.Based on the dividing method in zone, its principle is according to selected conformance criteria, is the mutual not process of the set of regions of crossover with image division.

The multi-scale division algorithm of using in the remote sensing image processing, its conformance criteria are also referred to as cuts apart yardstick [3], cuts apart corresponding a kind of of remote sensing images of yardstick for one and cuts apart, the different object representations [4] of cutting apart yardstick formation image tree structure.With regard to a pixel in the image, in the cutting apart of different scale, belong to different figure spot objects, form a figure spot series.Set up corresponding relation between the figure spot in actual atural object and image, just have a scale problem [5].Nature and artificial object all have its inherent yardstick that is fit to self, and yardstick is not of uniform size.Multiscale analysis method is used for image classification with the multi-scale division result, and the selection of cutting apart yardstick is lacked a quantitative criteria both at home and abroad.Thereby, need determine in the image segmentation process which cuts apart the figure spot is rational [6], be to have the figure spot that meets the inherent yardstick of atural object, this utilizes the key of image object spatial information just.Have only this figure spot that meets the inherent yardstick of atural object object, carry out spatial reasoning according to geometrical combination, spatial relationship between geometric attribute such as size that the figure spot reflected, shape and figure spot and be only significant.

1 Wang Run gives birth to, image understanding, publishing house of the National University of Defense technology, 1995, October, 399 pages.

The 2 chapters Shanxi of giving birth, image segmentation, Image Engineering (middle volume), publishing house of Tsing-Hua University, October in 2005 the 2nd edition, 477 pages.

3?Baatz?M.，Shape?A.Multiresolution?segmentation，an?optimizationapproach?for?high?quality?multi-scale?image?segmentation，in?AngewandteGeographische?informationsverarbeitung.XII.Beitrage?zum?AGIT-Symposium?Salzburg?2000，T.Strobl，T.Blaschke，and?G.Griesebner，Eds.，Karlsruhe：H.Wichmann?Verlag，2000，12-23

4?Benz?U.C.，Hofmann?P.，Willhauck?G.，et?al.Multi-resolution，object-oriented?fuzzy?analysis?of?remote?sensing?data?for?GIS-readyinformation，ISPRS?Journal?of?Photogrammetry?&?Remote?Sensing，2004，58，239-258

5?Quattrochi?D?A，Goodchild?M?F.Scale?in?remote?sensing?and?GIS.Lewis?Publishers，1997

6?Gorte?B.，Probabilistic?Segmentation?of?Remotely?Sensed?Images.In：ITC?Publication?Series，1998，63.

Summary of the invention

The object of the present invention is to provide a kind of image partition method of dimension self-adaption.

Comprise the steps:

1) the one or more figure layers after image and principal component transform thereof and the color transformed normalization are as a result determined image set to be split;

2) set dividing method and yardstick growth pattern, and with continually varying scale coefficient split image;

3) segmentation result under the different scale is expressed with the image object that the relation of the merging between the figure spot in the cutting procedure forms tree structure;

4) definition figure spot conspicuousness;

5) the affiliated figure spot of cutting off at continuous variation yardstick branch with pixel is changed to preface composition diagram spot evolution curve;

6) the conspicuousness curve of figure spot in the computed segmentation evolution process, cut apart yardstick curve just;

7) maximum value under the different yardsticks just of calculating from the conspicuousness curve forms the yardstick inverted order;

8) each pixel is given the extreme value of first yardstick correspondence in the yardstick inverted order, forms extreme value scalogram picture;

9) determine to cut apart the figure spot with the advantage yardstick in the extreme value scalogram picture.

Described setting dividing method and yardstick growth pattern, and with continually varying scale coefficient split image: the yardstick growth pattern is that natural number increases, and cutting apart scale coefficient is natural square; The merging cost F of figure spot is calculated as follows in the dividing method:

$F=\underset{d}{\mathrm{\&Sigma;}}{w}_d(n_m{\&CenterDot;\mathrm{\&sigma;}}_d^m-({\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20071006821400101.png"\; state="\{\{state\}\}">n</figure-callout>}}_1\&CenterDot;{\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="A20071006821400101.png"\; state="\{\{state\}\}">d</figure-callout>}}^1+{\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="A20071006821400101.png"\; state="\{\{state\}\}">n</figure-callout>}}_2{\&CenterDot;\mathrm{\&sigma;}}_d^2))$

Wherein n is the pixel number before and after the merger, two figure spots before 1 and 2 representatives merge, and the figure spot after m represents to merge, σ are the mean square deviation of figure spot, d is the figure layer of image set.

Described definition figure spot conspicuousness comprises one of four kinds of modes:

D ₁＝h _d-h _c

D ₂＝h _d-(h _c+h _c′)

D ₃＝h _d/h _c

D ₄＝h _d/(h _c+h _c′)

Wherein:

$h_d=\sqrt{\underset{d}{\mathrm{\&Sigma;}}{({\mathrm{\&mu;}}_d-{\mathrm{\&mu;}}_d^{\&prime;})}^2}$

$h_c=\underset{d}{\mathrm{\&Sigma;}}{w}_d{\&CenterDot;\mathrm{\&sigma;}}_d$

The conspicuousness curve of the figure spot in the described computed segmentation evolution process, cut apart yardstick curve just: the pixel in the cutting procedure is along with the variation of cutting apart yardstick, the merger of figure spot, belong to different figure spots, the first yardstick when the figure spot forms constitutes the dimensional variation curve of pixel with cutting apart dimensional variation; The conspicuousness of figure spot constitutes the conspicuousness curve of pixel with cutting apart dimensional variation.

The described maximum value of calculating from the conspicuousness curve under the different yardsticks just forms the yardstick inverted order: with the out to out in the continuous variation yardstick is the restriction yardstick; Between smallest partition yardstick in the conspicuousness curve of pixel and the restriction yardstick, a value during peaked yardstick just sorts as yardstick on the conspicuousness curve, and as new restriction yardstick; Determine the back value in the yardstick preface successively.

Describedly determine to cut apart the figure spot with the advantage yardstick in the extreme value scalogram picture: a certain pixel is included in the figure spot that a series of extreme value yardstick branches are cut off, relatively the pixel extreme value is cut apart the ratio that the first quantity of yardstick accounts for the total pixel number of figure spot for the figure spot in each figure spot, and yardstick is as the advantage yardstick and give this pixel at the beginning of the cutting apart of the figure spot of ratio maximum; In cutting procedure, belong to the same figure of cutting apart spot and advantage yardstick and have the pixel collection that four fields are communicated with, as cutting apart figure spot result.

When the present invention was cut apart complicated image, if we use single yardstick split image, when the small scale object reached its inherent yardstick, the object of large scale was divided into several little figure spots by mistake; When the yardstick of selecting enough greatly the time, when the large scale object correctly was divided into a big figure spot, it was a figure spot that the object of a plurality of small scales is merged.Thereby, the present invention defines the optimum segmentation yardstick of image object to scheme the spot self attributes, by analyzing the conspicuousness change curve, from have various possible cutting object, detect most possible object diagram spot, make the object of different scale have the suitable separately yardstick of cutting apart.

Description of drawings

Fig. 1 is the image partition method procedure chart of dimension self-adaption of the present invention;

Fig. 2 is the segmentation result figure of multi-scale division method different scale;

Fig. 3 is the image segmentation result figure of dimension self-adaption of the present invention.

Embodiment

The present invention comprises the pixel value distance between the figure spot and the standard deviation of figure spot pixel value to the definition of conspicuousness, from classification angle based on pixel, can be considered as discreteness in group distance and the group, this is similar to the linear differentiation of a kind of supervised classification method-Fischer commonly used, it is group distance (average difference) maximum, discreteness (sum of squares of deviations) minimum in the group, difference is that the linear differentiation of Fischer do not have constraint on the space to the data of two classifications, and here we spatially limit the pixel collection of two classifications, promptly the figure spot is adjacent, and pixel spatially is that four fields are communicated with in the figure spot.The minor increment of figure spot has reflected the pixel value distance of figure spot and adjacent figure spot, and the standard deviation of figure spot has reflected the pixel value homogeneous degree of figure spot inside, and our the figure spot conspicuousness of definition has reflected above-mentioned characteristic simultaneously.

The image partition method of dimension self-adaption comprises the steps:

Step 1,

Image comprises remote sensing image, can be gray scale (single band) data, colored (triband) data and multi-wavelength data.The conversion of image comprises principal component transform and color transformed, and image and transformation results are through normalized, and the one or more figure layers among both are as image set to be split.

Step 2,

Setting the yardstick growth pattern is that natural number increases, and cutting apart scale coefficient is natural square; The merging cost F of figure spot is as shown in the formula calculating in the dividing method:

$F=\underset{d}{\mathrm{\&Sigma;}}{w}_d(n_m{\&CenterDot;\mathrm{\&sigma;}}_d^m-({\mathrm{<figure-callout\; id="1"\; label="n"\; filenames="A20071006821400101.png"\; state="\{\{state\}\}">n</figure-callout>}}_1\&CenterDot;{\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="A20071006821400101.png"\; state="\{\{state\}\}">d</figure-callout>}}^1+{\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="A20071006821400101.png"\; state="\{\{state\}\}">n</figure-callout>}}_2{\&CenterDot;\mathrm{\&sigma;}}_d^2))$

Wherein n is the pixel number before and after the merger, two figure spots before 1 and 2 representatives merge, and the figure spot after m represents to merge, σ are the mean square deviation of figure spot, d is the figure layer of image set.

For the specific single of cutting apart yardstick was cut apart, cutting procedure was as follows:

2.1 the neighbouring relations between figure spot and figure spot are defined as follows: the pixel set of single pixel and a plurality of pixels all can be thought the figure spot.To a figure spot, investigate its border picture dot, if the pixel of two adjacent figure spots is that four fields are adjacent, then two figure spots are that four field methods are adjacent.

2.2 in cutting apart the process of carrying out, along with the continuous merging of figure spot, the figure spot is heterogeneous constantly to be increased, when each figure spot all satisfies following condition in the image: 1. all figure spot heterogeneities all less than given threshold value; 2. the heterogeneity of the new figure spot of formation was all greater than given threshold value after any one figure spot merged with any one neighborhood figure spot again.Then think and finish once cutting apart in the cutting procedure.

2.3 the merging method in the cutting procedure is as follows: when a figure spot have conditioned disjunction more than one adjacent figure spot match-merge have repeatedly qualified figure spot to the time, just need to determine that the merger figure spot of an optimum is right, the cost minimum of its merger.To a figure spot A, investigate its neighbours territory picture dot adjacent map spot, if A and its certain adjacent map spot B satisfy following condition then claim A, B satisfies local optimum matching principle mutually: 1. the heterogeneity of the big figure spot that forms after A and the B merging is less than or equal to the heterogeneity that the adjacent figure spot with other of A merges the big figure spot of back formation; 2. be that center figure spot is sought the adjacent map spot C that satisfies heterogeneous minimum criteria after merging with B with B; 3. a plurality of figure spots that satisfy condition are arranged in A=C or (2), and A is one of them.If satisfying local optimum matching principle mutually, A, B just they are merged into a big figure spot, if satisfied then be that starting point continues to search with B.

Step 3, segmentation result data organization

Cut apart the corresponding segmentation result of yardstick,, form a series of segmentation result for one with continually varying yardstick split image.The figure spot of cutting apart under the maximum fractionation yardstick is as root node, and all figure spots that are merged into this figure spot in cutting procedure are as child node, and the figure spot on the child node is again the parent node of figure spot before all merge, and expresses with this segmentation result of forming tree structure.

Step 4, figure spot conspicuousness define,

Pixel mean value is different between figure spot and adjacent figure spot, forms the distance of two figure spots, in all adjacent figure spot set, has the mean distance value of a minimum.The spectrum mean distance value of figure spot can be defined as follows:

$h_d=\sqrt{\underset{d}{\mathrm{\&Sigma;}}{({\mathrm{\&mu;}}_d-{\mathrm{\&mu;}}_d^{\&prime;})}^2}$

$h_c=\underset{d}{\mathrm{\&Sigma;}}{w}_d{\&CenterDot;\mathrm{\&sigma;}}_d$

Standard deviation h _cThe dispersion degree of pixel value in the representative graph spot, w in the formula _dThe weights of representing the d wave band, σ _dThe mean square deviation of representing the d wave band.Figure spot conspicuousness is defined by following four kinds of modes:

D ₁＝h _d-h _c

D ₂＝h _d-(h _c+h _c′)

D ₃＝h _d/h _c

D ₄＝h _d/(h _c+h _c′)

The foundation of step 5, figure spot evolution curve

Along with the growth of cutting apart yardstick, for a pixel, along with the variation of yardstick, little figure spot is merged into big figure spot, and the succession of figure spot merger can uniquely be determined, forms the evolution of figure spot.

Step 6, is cut apart yardstick curve just at the conspicuousness curve of figure spot;

Pixel in the cutting procedure belongs to different figure spots along with the variation of cutting apart yardstick, and the first yardstick when the figure spot forms constitutes the dimensional variation curve of pixel with cutting apart dimensional variation; The conspicuousness of figure spot constitutes the conspicuousness curve of pixel with cutting apart dimensional variation.Obviously, according to the definition of figure spot conspicuousness, when target figure spot kept stablizing, the variation of adjacent figure spot also can cause the variation of this figure spot conspicuousness.

Step 7, the yardstick inverted order that forms

Pixel at first defines out to out and is the restriction yardstick along with a series of figure spots of the variation formation of cutting apart yardstick and the conspicuousness of figure spot.The conspicuousness that between smallest dimension and restriction yardstick, has a maximum.Maximum saliency object cut apart just yardstick as first value in sorting for yardstick; This value conduct restriction simultaneously yardstick, and the like, the yardstick inverted order formed.

Step 8, extreme value scalogram picture

Each pixel is given the extreme value of first yardstick correspondence in the yardstick inverted order, forms extreme value scalogram picture.In image segmentation result, for each adjacent picture elements with identical maximum conspicuousness, and have an identical yardstick just of cutting apart, can be communicated with the figure spot that constitutes by the neighbours territory, be referred to as maximum conspicuousness set, this cuts apart the figure spot that forms under the first yardstick is maximum conspicuousness figure spot, should, maximum conspicuousness figure spot comprises maximum conspicuousness set.

Step 9, determine to cut apart the figure spot with the advantage yardstick in the extreme value scalogram picture:

For a pixel, may exist a plurality of maximum conspicuousness figure spots to comprise this pixel, if above-mentioned a plurality of maximum conspicuousness figure spots are by cutting apart scale size arrangement just, obviously, according to the process that the figure spot merges, the figure spot with maximum first yardstick comprises above-mentioned all figure spots.

A certain pixel is included in the figure spot that a series of extreme value yardstick branches are cut off, relatively the pixel extreme value is cut apart the ratio that the first quantity of yardstick accounts for the total pixel number of figure spot for the figure spot in each figure spot, and yardstick is as the advantage yardstick and give this pixel at the beginning of the cutting apart of the figure spot of ratio maximum; In cutting procedure, belong to the same figure of cutting apart spot and advantage yardstick and have the pixel collection that four fields are communicated with, as cutting apart figure spot result.

Claims (6)

1. the image partition method of a dimension self-adaption is characterized in that comprising the steps:

1) the one or more figure layers after image and principal component transform thereof and the color transformed normalization are as a result determined image set to be split;

2) set dividing method and yardstick growth pattern, and with continually varying scale coefficient split image;

3) segmentation result under the different scale is expressed with the image object that the relation of the merging between the figure spot in the cutting procedure forms tree structure;

4) definition figure spot conspicuousness;

5) the affiliated figure spot of cutting off at continuous variation yardstick branch with pixel is changed to preface composition diagram spot evolution curve;

6) the conspicuousness curve of figure spot in the computed segmentation evolution process, cut apart yardstick curve just;

7) maximum value under the different yardsticks just of calculating from the conspicuousness curve forms the yardstick inverted order;

8) each pixel is given the extreme value of first yardstick correspondence in the yardstick inverted order, forms extreme value scalogram picture;

9) determine to cut apart the figure spot with the advantage yardstick in the extreme value scalogram picture.

2. the image partition method of a kind of dimension self-adaption according to claim 1, it is characterized in that described setting dividing method and yardstick growth pattern, and with continually varying scale coefficient split image: the yardstick growth pattern is that natural number increases, and cutting apart scale coefficient is natural square; The merging cost F of figure spot is calculated as follows in the dividing method:

$F=\underset{d}{\mathrm{\&Sigma;}}{w}_d(n_m\&CenterDot;{\mathrm{\&sigma;}}_d^m-(n_1\&CenterDot;{\mathrm{\&sigma;}}_d^1+n_2\&CenterDot;{\mathrm{\&sigma;}}_d^2))$

Wherein n is the pixel number before and after the merger, two figure spots before 1 and 2 representatives merge, and the figure spot after m represents to merge, σ are the mean square deviation of figure spot, d is the figure layer of image set.

3. the image partition method of a kind of dimension self-adaption according to claim 1 is characterized in that described definition figure spot conspicuousness comprises one of four kinds of modes:

D ₁＝h _d-h _c

D ₂＝h _d-(h _c+h _c′)

D ₃＝h _d/h _c

D ₄＝h _d/(h _c+h _c′)

Wherein:

$h_d=\sqrt{\underset{d}{\mathrm{\&Sigma;}}{({\mathrm{\&mu;}}_d-{\mathrm{\&mu;}}_d^{\&prime;})}^2}$

$h_c=\underset{d}{\mathrm{\&Sigma;}}{w}_d\&CenterDot;{\mathrm{\&sigma;}}_d$

4. the image partition method of a kind of dimension self-adaption according to claim 1, it is characterized in that the figure spot in the described computed segmentation evolution process the conspicuousness curve, cut apart yardstick curve just: the pixel in the cutting procedure is along with the variation of cutting apart yardstick, the merger of figure spot, belong to different figure spots, first yardstick when the figure spot forms constitutes the dimensional variation curve of pixel with cutting apart dimensional variation; The conspicuousness of figure spot constitutes the conspicuousness curve of pixel with cutting apart dimensional variation.

5. the image partition method of a kind of dimension self-adaption according to claim 1 is characterized in that the described maximum value of calculating under the different yardsticks just from the conspicuousness curve, and form the yardstick inverted order: with the out to out in the continuous variation yardstick is the restriction yardstick; Between smallest partition yardstick in the conspicuousness curve of pixel and the restriction yardstick, a value during peaked yardstick just sorts as yardstick on the conspicuousness curve, and as new restriction yardstick; Determine the back value in the yardstick preface successively.

6. the image partition method of a kind of dimension self-adaption according to claim 1, it is characterized in that describedly determining to cut apart the figure spot with the advantage yardstick in the extreme value scalogram picture: a certain pixel is included in the figure spot that a series of extreme value yardstick branches are cut off, relatively the pixel extreme value is cut apart the ratio that the first quantity of yardstick accounts for the total pixel number of figure spot for the figure spot in each figure spot, and yardstick is as the advantage yardstick and give this pixel at the beginning of the cutting apart of the figure spot of ratio maximum; In cutting procedure, belong to the same figure of cutting apart spot and advantage yardstick and have the pixel collection that four fields are communicated with, as cutting apart figure spot result.

Images