CN102096816B - Multi-scale multi-level image segmentation method based on minimum spanning tree - Google Patents

Abstract

The invention provides a multi-scale multi-level image segmentation method based on a minimum spanning tree. Multi-scale multi-level image segmentation is expressed and realized by a graph model. The segmentation method is suitable for initial segmentation results obtained by various images and various rules to effectively combine over-segmented regions in high level segmentation, so an over-segmentation phenomenon is avoided; moreover, multi-scale segmentation results on different levels provides characteristic description information on different levels for analysis of target structural components; and the method is very important for target recognition.

Description

Multiple dimensioned multi-level image division method based on minimum spanning tree

Technical field

This method belongs to Flame Image Process and distinguishment technical field, particularly relates to a kind of new multi-level multiple dimensioned pyramid image dividing method based on minimum spanning tree Optimum Theory and graph model characteristics.

Background technology

High spatial resolution remote sense image provides the high precision space geometry information of ground landscape, abundant texture information and multispectral information for us; Make that traditional classification of remote-sensing images method based on pixel is inapplicable; Therefore, high-resolution remote sensing image is handled the challenge face the details that image provides.For this reason; Baatz and Schape pointed out in [1] in 1999: important semantic interpretation more need be represented with the mutual relationship between object and the object in the significant image rather than with pixel one by one; Therefore, propose OO high-resolution remote sensing image Target Recognition and sorting technique, promptly at first image has been cut apart the formation object zone; And come description object with hierarchical network, be that unit carries out Target Recognition again with the object.The new sorter proof of object-based spectrum, shape, texture, spatial relationship and people's the further reasoning of knowledge is very useful in the high spatial resolution field, and it meets the rule of human recognition objective, has improved nicety of grading and details.OO image segmentation is one of main method of object acquisition; OO target identification technology is through to incompatible recognition objectives of message block such as the spectrum of object, how much, texture, spatial neighborhood relation; From theoretical and practical application, find; The quality of Object Segmentation directly influences the effect and the precision of image classification identification, because can it be directly connected to geological information and the structural information that extract middle target on the image accurately and efficiently, therefore; The object-oriented high-resolution remote sensing image is partitioned into crucial and basic in the remote sensing image processing, and the research of object-oriented dividing method also becomes one of high spatial resolution focus and difficult point.

Image segmentation is being played the part of important role in image understanding, the homogeneous region that satisfies certain criterion that obtains by it, and the combination of the built-in attribute in zone and external attribute (comprising between the zone, syntople) is significant to Target Recognition.Yet, up to the present also there is not a kind of image partition method can all targets in the image be made a distinction fully, often have over-segmentation or less divided phenomenon; The difference of parameter setting will inevitably produce different segmentation results in the dividing method; And usually, same dividing method possibly reach segmentation effect preferably to certain class targets under certain parameter, and meanwhile; Over-segmentation or less divided possibly appear in other type target; Be difficult to realize that the setting of a certain scale parameter just can reach the division fully to all targets, often can only under certain data source, select suitable partitioning parameters to certain class targets, this that is to say that details keeps and large scale is difficult to and deposits in same partitioning algorithm; Particularly for high-resolution remote sensing image; Therefore, the image segmentation of research is that a kind of can the realization divided the partitioning algorithm that comes with the particular dimensions target from image usually, selects different parameters will from image, obtain the object of different details (scale size).The fundamental purpose of image segmentation is for Target Recognition, for Target Recognition provides required characteristic.Therefore, provide that can to improve the classification and the segmentation result of accuracy of identification and help providing the required object of Target Recognition also just to become be the dividing method main task.The contradiction of cutting apart accuracy and the contradiction, over-segmentation and the less divided that are prone to the property cut apart in the image segmentation is difficult to solution with the cutting techniques of fixed size; But might solve these contradictions through the multi-scale division technology, this respect research and the main contents that have much room for improvement comprise: based on multiple dimensioned image pre-service; More suitable multiple dimensioned data structure (describing and the storage data); Cooperate better multi-scale division strategy (criterion of extracted data) with it.Present most of multi-scale division all is to combine multiscale analysis theoretical; Under a certain yardstick, show as heterogeneous textural element with things; On big one-level yardstick observation but be such scale effect of homogeneity for cutting apart criterion, different scale parameters is set obtains the segmentation result on the different scale, be i.e. large scale rough segmentation; Small scale segmentation, but do not set up contact and the spatial relationship of segmentation result between the big small scale segmentation result.Simultaneously, this image segmentation of after the multi-scale image pre-service, carrying out the inaccurate problem of obscurity boundary can occur on high-level, thereby is unfavorable for that high-level target accurately extracts.Syntoples between the attribute of atural object own, atural object composition structure and the atural object etc. are significant with classification to Target Recognition, and therefore, it is very meaningful to design the partitioning algorithm that satisfies multiple dimensioned multi-level expression.

Consider that from the Target Recognition angle human cognitive world, recognition objective are with the different scale combination, from coarse to fine, by thin to thick cognitive process; When from a width of cloth image, judging different target, be to combine from image different details of viewed various objects and relation each other thereof, like color, shape, texture, size, adjacency, the just very high recognition accuracy of acquisition of relation such as comprise; Therefore; When computer object identification, high-resolution remote sensing image classification and Target Recognition, make full use of the target that image provides spectrum, geometry and in abutting connection with, characteristic such as comprise, will improve greatly and discern and nicety of grading; Just because of this; Set up intelligent Target Recognition and categorizing system, need obtain spectrum, shape, texture and the syntople of different scale object, set up the taxonomical hierarchy of different scale; This that is to say that we need obtain the integrated information of different scale, different levels from image, and this also just image segmentation the information that will provide.Can know by preceding surface analysis; The segmentation result that obtains through a certain group of particular dimensions parameter can not reach the purpose of describing these information fully; Therefore, the segmentation result that need the different scale parameter be obtained, from coarse to fine or by carefully describing to thick carve information; This just relates to image multiple dimensionedly cuts apart at many levels and expresses, for the Target Recognition of high-resolution remote sensing image provides favourable evidence.

External at present existing eCognition business software has realized cutting apart based on the multiple dimensioned multi-level image of region growing consideration spectrum and shape facility; The domestic software that does not also have at present a this respect is realized multiple dimensionedly cutting apart at many levels; Great majority are cut apart and all are based on the image multiscale space and select the different scale image to cut apart, and lack between the level and the contact and the description of spatial relationship between the neighboring region.

Comprise and involved relation in order to analyze better and to utilize with existing between the space syntople between one deck object, multiple dimensioned multi-level information, the levels object; Need to realize the merging of small scale target is obtained cutting apart at many levels of large scale target, this is also just meeting actual things, the atural object level is divided thought.The present invention's research has also realized cutting apart based on the multi-level multiple dimensioned image of minimum spanning tree, representes each layer data and layer interior, interlayer contact with the pyramid diagram model.The pyramid diagram model mainly contains simple graph (simple graph), dual graph (dual graph) and three kinds of pyramids of constitutional diagram (combinatorial maps) and expresses model at present.Simple graph is loop free, the figure that does not have heavy limit, its simple in structure, easy realization, and counting yield is high, and the set membership of mapping node is convenient, obtains the syntople between the summit easily, has very strong practicality, therefore, is widely used in the pyramid split graph model.

The pyramid contraction method is meant how by G _lObtain G _L+1Particularly for image segmentation; It has confirmed the character of the pyramidal every layer data that restrains height and generated; Reasonably contraction method can avoid pyramid too high and to produce storage space consumption too big, and in fact this is exactly the setting of cutting apart criterion, and the quality of criterion setting has also determined the cut zone character that generates.Haxhimusa etc. use dual graph [2-4] and constitutional diagram [5] structure pyramid respectively, with the multi-level multi-scale division that criterion has realized image of cutting apart in Boruvka minimal spanning tree algorithm and the document [6].

The image segmentation thought of structure minimum spanning tree has been saved the process that makes up the entire image minimum spanning tree, has improved efficient.But the simple threshold values criterion is to noise-sensitive, and these class methods are irrelevant, only relevant with the limit weights with raw data in division or merging process.Certainly, also can realize and criterion is considered that raw data and neighborhood relationships thereof come in being provided with, but will increase computation complexity like this at algorithm.

Citing document in the background technology:

1.Blaschke，T.and?J.Strobl，What’s?wrong?with?pixels?Some?recent?developmentsinterfacing?remote?sensing?and?GIS.GIS-Zeitschrift?fürGeoinformationssysteme，2001.14(6)：p.12-17.

2.Haxhimusa，Y.and?W.Kropatsch，Hierarchy?of?Partitions?with?Dual?GraphContraction，in?Pattern?Recognition.2003，Springer?Berlin/Heidelberg.p.338-345.

3.Haxhimusa，Y.and?W.Kropatsch，Segmentation?Graph?Hierarchies，in?Proceedingsof?Joint?Workshops?on?Structural，Syntactic，and?Statistical?PatternRecognition?S+SSPR.2004，Springer?Berlin/Heidelberg.p.343-351.

4.Kropatsch，W.G.and?Y.Haxhimusa.Grouping?and?segmentation?in?a?hierarchy?ofgraphs.in?Proceeding?of?the?16th?IS&T/SPIE?Annual?Symposium.2004.

5.Ion，A.，Y.Haxhimusa，W.G.Kropatsch，and?L.Brunz，Hierarchical?ImagePartitioning?using?Combinatorial?Maps，in?Proceeding?of?the?JointHungarian-Austrian?Conference?on?Image?Processing?and?Pattern?Recognition，D.Chetverikov，L.Czuni，and?M.Vincze，Editors.May?2005：Hungary.p.179--186.

6.Felzenszwalb，P.F.and?D.P.Huttenlocher，Efficient?Graph-Based?ImageSegmentation.International?Journal?of?Computer?Vision?200459(2)：p.167-181

Summary of the invention

Cut apart and the description problem to multiple dimensioned multi-level image; Be the inaccurate problem of avoiding occurring in the multiscale space of obscurity boundary; Relation between syntople and the levels between the ability description object the invention provides a kind of multiple dimensioned multi-level image division method based on minimum spanning tree simultaneously.

Technical scheme of the present invention comprises the steps:

Step 1, image is represented with the simple graph model that a pixel of each summit correspondence image of simple graph model connects with the limit between per two adjacent vertexes, the limit weights on every limit for two summits of this limit connection the difference between corresponding two pixels;

Step 2, step 1 gained simple graph model is carried out cutting apart based on the image of minimum spanning tree based on predefined regional merging criterion and partitioning parameters; Minimum weights limit when cutting apart from the simple graph model begins to increase up to maximum weights limit by regional merging criterion; Merge each minimum spanning tree that generates and represent the zone of a connection, obtain initial segmentation result;

Step 3, be the summit with each zone of segmentation result gained, each zone is adjacent between the zone and is connected by the limit, with the distance of the histogram between zone and neighboring region as the limit weights, the simple graph model that structure makes new advances;

Step 4, step 3 gained simple graph model is carried out cutting apart based on the image of minimum spanning tree according to the limit weights; Obtain new segmentation result; So turning back to step 3 circulation carries out the new simple graph model of structure and cuts apart; Every regional merging criterion that adopts when cutting apart in the circulation of taking turns is provided with different big or small similarity thresholds; Accomplish to obtain required multiple dimensioned multi-level image and cut apart through circulation repeatedly, whenever take turns the simple graph model formation image pyramid that simple graph model that circulation obtains and step 1 obtain.

And in step 1, the limit weights are the weight on every limit in the simple graph model, adopt the wave band weighted euclidean distance to calculate weight.

Perhaps, in step 1, the limit weights are the weight on every limit in the simple graph model, when image is the multiband image, adopt the cosine angle distance to calculate weight.

And in step 2, the comprehensive multiple region characteristic of predefined regional merging criterion comprises and considers spectrum, shape and texture.

And in step 2, partitioning parameters is set according to predefined regional merging criterion, comprises a Minimum Area area size parameter that is used to prevent to generate the zonule in the partitioning parameters.

And in step 3, the histogram between neighboring region is suitable for multiband statistical property histogram distance relatively apart from employing.

And histogram is card side's distance apart from what adopt, and the histogram between neighboring region is 0-1 apart from span.

And, simple graph model of each layer of image pyramid correspondence, the simple graph model that step 1 obtains is the bottom of image pyramid; Whole image pyramid is described with a tree construction, the corresponding relation that this tree construction reflection certain zone, upper strata is made up of the following a plurality of zones of one deck.

The present invention adopts graph model to describe image and cuts apart each layer object of being generated and the syntople between the object; Employing is based on the theoretical image segmentation of minimum spanning tree; Apart from making up limit power,, different big or small similarity thresholds realize that multiple dimensioned multi-level image cuts apart with the histogram between the zone through being set.

Description of drawings

Fig. 1 is a multi-level multi-scale division pyramid construction synoptic diagram of the present invention.

Embodiment

Below in conjunction with accompanying drawing and embodiments of the invention, technical scheme of the present invention is elaborated.The implementation procedure of embodiment is following:

Step 1, image is represented with the simple graph model that a pixel of each summit correspondence image of simple graph model connects with the limit between per two adjacent vertexes, the limit weights on every limit for two summits of this limit connection the difference between corresponding two pixels.

The limit weights are the weight on every limit in the simple graph model, and embodiment adopts the wave band weighted euclidean distance to calculate weight.Especially, when image is the multiband image, can also adopt the cosine angle distance to calculate weight.Concrete account form is a prior art, and the present invention will not give unnecessary details.

Step 2, step 1 gained simple graph model is carried out cutting apart based on the image of minimum spanning tree according to predefined regional merging criterion and partitioning parameters; Minimum weights limit when cutting apart from the simple graph model begins to increase up to maximum weights limit by regional merging criterion; Merge each minimum spanning tree that generates and represent the zone of a connection, obtain initial segmentation result.

Minimum spanning tree problem is described as seeking limit power and minimum generation tree, for image segmentation, seeks the minimum connected region problem of difference and just can be converted into minimum spanning tree problem.Therefore, between step 1 has realized the pixel of graphical representation the relation with the simple graph model representation after, can carry out image to step 1 gained simple graph model based on minimum spanning tree and cut apart.Mainly contain top-down division minimum spanning tree and bottom-up merging generates two kinds of strategies of a plurality of minimum spanning trees based on the image segmentation of minimum spanning tree.Embodiment adopts Kruskal minimal spanning tree algorithm and bottom-up consolidation strategy to realize cutting apart based on the image of minimum spanning tree, and specific algorithm is a prior art with strategy, and the present invention will not give unnecessary details.

When preestablishing regional merging criterion, advise comprehensive multiple region characteristic, comprise and consider spectrum, shape and texture etc. that practical implementation can be set as required.For the purpose of convenient enforcement reference, below provide three kinds of criterions to set schemes:

A establishes and judges whether that the object that can merge is a region S based on the criterion of cutting apart of the consideration spectrum of infima species internal variance and shape facility ₁And region S ₂, two zones merge the interior difference of class of back gained object is calculated by (2) formula, if difference is less than certain threshold value th, then region S ₁And region S ₂Can merge, if greater than then nonjoinder of this threshold value th, this criterion is expressed as (1) formula:

$P(S_1,S_2)=\left\{\begin{array}{cc}\mathrm{true}& \mathrm{if\; f}\≤\mathrm{th}\\ \mathrm{false}& \mathrm{otherwise}\end{array}\right.---\left(1\right)$

f＝w _spectral·h _spectral+(1-w _spectral)·h _shape (2)

Wherein, w _SpectralRepresent the weight of spectral signature, span is 0～1,1-w _SpectralWeight for shape facility; h _SpectralAnd h _ShapeCalculate by (3)-(6) respectively

$h_{\mathrm{spectral}}=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{w}_k{n}_{\mathrm{merge}}\·{\mathrm{\σ}}_k^{\mathrm{merge}}---\left(3\right)$

h _shape＝w _compact·h _compact+(1-w _compact)·h _smooth(4)

$h_{\mathrm{smooth}}=n_{\mathrm{merge}}\·\frac{l_{\mathrm{merge}}}{b_{\mathrm{merge}}}---\left(5\right)$

$h_{\mathrm{compact}}=n_{\mathrm{merge}}\·\frac{l_{\mathrm{merge}}}{\sqrt{n_{\mathrm{merge}}}}---\left(6\right)$

Wherein, k representes the wave band numbering, and K representes total wave band number, w _kRepresent the K-band weight,

Expression merges the standard deviation of rear region, n _MergeExpression merges rear region size, l _MergeExpression merges the girth of rear region, b _MergeExpression merges the minimum boundary rectangle girth of rear region, w _CompactRepresent complex-shaped degree, span is 0～1,1-w _CompactRepresent the shape, flat slippery.

B is based on the merging criterion of only considering spectral characteristic of infima species internal variance: establish and judge whether that the object that can merge is a region S ₁And region S ₂, two zones merge the interior difference of class of back gained object is calculated by (8) formula, if difference is less than certain threshold value th, then region S ₁And region S ₂Can merge, if greater than then nonjoinder of this threshold value th, then merging criterion is expressed as (7) formula:

$P(S_1,S_2)=\left\{\begin{array}{cc}\mathrm{true}& \mathrm{if\; f}\≤\mathrm{th}\\ \mathrm{false}& \mathrm{otherwise}\end{array}\right.---\left(7\right)$

$f=\underset{k=1}{\overset{K}{\mathrm{\Σ}}}{w}_k{\mathrm{\σ}}_k^{\mathrm{merge}}---\left(8\right)$

Wherein, k representes the wave band numbering, and K representes total wave band number, w _kRepresent the K-band weight, Expression merges the standard deviation of rear region.

C only considers the merging criterion of spectral characteristic based on stable and consistent learning algorithm in the Statistical Learning Theory: establish and judge whether that the object that can merge is a region S ₁And region S ₂, if satisfy condition then region S ₁And region S ₂Can merge, otherwise nonjoinder.

Wherein, n ₁And n ₂The expression region S ₁And S ₂In pixel count (number of vertex), n representes to work as front, w _nBe to work as the front in the merging process from small to large by the limit weights, it connects S ₁And S ₂Two zones of expression; δ is the probability number of value between 0 and 1, and its variation is little to threshold affects, has the fine setting effect, and according to probability theory, establishing its value during enforcement usually is 0.1; M is the loss upper bound, and it is corresponding to the supremum of limit power; && representes to satisfy simultaneously.If P is (S ₁, S ₂) then merge region S for TRUE ₁And S ₂, if P (S ₁, S ₂) be then nonjoinder of FALSE.

Can also adopt other criterions as required during practical implementation.

Partitioning parameters is set according to predefined regional merging criterion, generates the zonule when cutting apart, and the present invention proposes to comprise in the partitioning parameters one and is used for Minimum Area area size parameter.When certain region area when cutting apart, occurring less than Minimum Area area size parameter; That can carry out is treated to: begin in proper order from small to large by the limit weights again; If when two zones that the front connected do not belong to same zone; And certain region area in two zones that this limit connected then merges these two zones less than specifying the Minimum Area area.Begin to increase up to maximum weights limit through the minimum weights limit from the simple graph model again, guaranteed that like this similarity is maximum between the merging zone by regional merging criterion.

Step 3, be the summit with each zone of segmentation result gained, each zone is adjacent between the zone and is connected by the limit, with the distance of the histogram between zone and neighboring region as the limit weights, the simple graph model that structure makes new advances.

The invention proposition as the limit weights, can utilize the statistical distribution characteristic in zone with the distance of the histogram between zone and neighboring region like this.Existing histogram distance measure has a variety of, like heuristic histogram distance, weighted mean variance distance, nonparametric detection statistics histogram distance (Kolmogorov-Smirnov distance, Cramer/von Mieses type statistics, c ²Distance), information theory divergence histogram distance, geodesic distance etc.The present invention need adopt and be suitable for multiband statistical property histogram distance relatively.Consulting document and experiment back discovery c ²Distance (card side's distance) is better than other distance in the performance of image segmentation, so embodiment selects histogram c for use ²Distance is as calculating the limit weight function that connects two neighboring regions, and its span is 0-1.Distance concrete account form in card side's is a prior art, and the present invention will not give unnecessary details.

Step 4, step 3 gained simple graph model is carried out cutting apart based on the image of minimum spanning tree according to the limit weights; Obtain new segmentation result; Turning back to step 3 circulation then carries out the new simple graph model of structure and cuts apart; Every regional merging criterion that adopts when cutting apart in the circulation of taking turns is provided with different big or small similarity thresholds; Accomplish to obtain required multiple dimensioned multi-level image and cut apart through circulation repeatedly, whenever take turns the simple graph model formation image pyramid that simple graph model that circulation obtains and step 1 obtain.

The image of this step minimum spanning tree is cut apart also can adopt Kruskal minimal spanning tree algorithm and the realization of bottom-up consolidation strategy.Just to be used for obtaining initial segmentation result different with the criterion of step 2, and the regional merging criterion of step 4 is used for more high-rise cutting apart, so employing range statistics characteristic designs and gets final product during practical implementation.The regional merging criterion that embodiment sets in this step is following:

$P(v_i,v_j)=\left\{\begin{array}{cc}\mathrm{true}& \mathrm{ifw}\left((v_i,v_j)\right)\≤\mathrm{th}\\ \mathrm{false}& \mathrm{otherwise}\end{array}\right.$

Wherein, w ((v _i, v _j)) two neighboring region v of expression current layer connection _iAnd v _jThe limit weights, calculate gained

Histogram distribution is approaching more, w ((v _i, v _j)) value more little, it has reflected the similarity degree in two zones, when limit weights that connect two zones during less than certain similarity threshold; Think the same area, th is given threshold value, and threshold value is big more; Allow interregional distributional difference big more, the number of regions that merges in the layer is many more, accelerates pyramid and shrinks; Form easily large scale (area) zone, threshold size will directly influence cuts apart for what of the number of plies, determines the segmentation precision of each layer.

Whether constantly step 3 and step 4 are carried out in circulation, and each circulation obtains a simple graph model, stop when required multiple dimensioned multi-level image is cut apart up to obtaining, so can set control voluntarily as required by the user during practical implementation and circulate.Every regional merging criterion that adopts when cutting apart in the circulation of taking turns is identical, but different big or small similarity thresholds are set in the partitioning parameters, just can generate the segmentation result of different levels and different scale.According to interpretation; Usually threshold value is arranged between the 0.1-0.3 and can obtains better segmentation effect; Suggestion is set similarity threshold and is increased progressively at interval by 0.05 in multi-level cutting procedure; The user can set the maximum similarity threshold value according to different situations and come Control Circulation to stop, and for example can set when similarity threshold is increased to 0.8 to stop.

Like Fig. 1 finding, after many wheel circulations obtain multiple dimensioned multi-level image segmenting structure, obtain an image pyramid.Simple graph model of each layer of image pyramid correspondence, each simple graph model have reflected the syntople between this yardstick inner region, and the simple graph model that step 1 obtains is the bottom of image pyramid; Whole image pyramid is described with a tree construction, the corresponding relation that this tree construction reflection certain zone, upper strata is made up of the following a plurality of zones of one deck.For the purpose of convenient understanding, Fig. 1 adopts simple three layer image pyramids to give an example, and in step 1 raw video is made up the simple graph model that obtains the bottom by eight neighborhood syntoples, is designated as G ₀((V ₀, E ₀)), G wherein ₀, V ₀, E ₀Represent respectively the bottom simple graph model, the summit in this layer simple graph model, connect the limit on summit; Obtain the simple graph model of the second layer when running to step 3 for the first time, the initial segmentation result of performance step 2 gained is designated as G ₁((V ₁, E ₁)), G wherein ₁, V ₁, E ₁Represent respectively the second layer simple graph model, the cut zone in this layer simple graph model, connect the limit of neighboring region; When circulation runs to step 3, obtain the 3rd layer simple graph model for the second time, the gained segmentation result was designated as G when performance ran to step 4 for the first time ₂(V ₂, E ₂)), G wherein ₂, V ₂, E ₂Represent the 3rd layer simple graph model, the cut zone in this layer simple graph model respectively, connect the limit of neighboring region.Possibly have more multilayer during practical implementation, principle is consistent with this example.

Claims (8)

1. multiple dimensioned multi-level image division method based on minimum spanning tree is characterized in that: may further comprise the steps,

Step 1, image is represented with the simple graph model that a pixel of each summit correspondence image of simple graph model connects with the limit between per two adjacent vertexes, the limit weights on every limit for two summits of this limit connection the difference between corresponding two pixels;

Step 2, step 1 gained simple graph model is carried out cutting apart based on the image of minimum spanning tree according to predefined regional merging criterion and partitioning parameters; Minimum weights limit when cutting apart from the simple graph model begins to increase up to maximum weights limit by regional merging criterion; Merge each minimum spanning tree that generates and represent the zone of a connection, obtain initial segmentation result; The region characteristic that predefined regional merging criterion is considered comprises spectrum;

Step 3, be the summit with each zone of segmentation result gained, each zone is adjacent between the zone and is connected by the limit, with the distance of the histogram between zone and neighboring region as the limit weights, the simple graph model that structure makes new advances;

Step 4, step 3 gained simple graph model is carried out cutting apart based on the image of minimum spanning tree according to the limit weights; Obtain new segmentation result; Turning back to step 3 circulation then carries out the new simple graph model of structure and cuts apart; Every regional merging criterion that adopts when cutting apart in the circulation of taking turns is provided with different big or small similarity threshold th; Accomplish to obtain required multiple dimensioned multi-level image and cut apart through circulation repeatedly, whenever take turns the simple graph model formation image pyramid that simple graph model that circulation obtains and step 1 obtain; Regional merging criterion in that this step is set is following:

$P(v_i,v_j)=\left\{\begin{array}{cc}\mathrm{true}& \mathrm{if\; w}\left((v_i,v_j)\right)\≤\mathrm{th}\\ \mathrm{false}& \mathrm{otherwise}\end{array}\right.$

Wherein, w ((v _i, v _j)) two neighboring region v of expression current layer connection _iAnd v _jThe limit weights, calculate gained w ((v _i, v _j)) ∈ [0,1].

2. multiple dimensioned multi-level image division method according to claim 1 is characterized in that: in step 1, the limit weights are the weight on every limit in the simple graph model, adopt the wave band weighted euclidean distance to calculate weight.

3. multiple dimensioned multi-level image division method according to claim 1 is characterized in that: in step 1, the limit weights are the weight on every limit in the simple graph model, when image is the multiband image, adopt the cosine angle distance to calculate weight.

4. multiple dimensioned multi-level image division method according to claim 1 is characterized in that: in step 2, the comprehensive multiple region characteristic of predefined regional merging criterion comprises and considers spectrum, shape and texture.

5. multiple dimensioned multi-level image division method according to claim 1; It is characterized in that: in step 2; Partitioning parameters is set according to predefined regional merging criterion, comprises a Minimum Area area size parameter that is used to prevent to generate the zonule in the partitioning parameters.

6. multiple dimensioned multi-level image division method according to claim 1 is characterized in that: in step 3, the histogram between neighboring region is suitable for multiband statistical property histogram distance relatively apart from employing.

7. multiple dimensioned multi-level image division method according to claim 6 is characterized in that: histogram is card side's distance apart from what adopt, and the histogram between neighboring region is 0-1 apart from span.

8. multiple dimensioned multi-level image division method according to claim 1 is characterized in that: simple graph model of each layer of image pyramid correspondence, and the simple graph model that step 1 obtains is the bottom of image pyramid; Whole image pyramid is described with a tree construction, the corresponding relation that this tree construction reflection certain zone, upper strata is made up of the following a plurality of zones of one deck.

Images