Background technique
In recent years, object-based image analysis OBIA (object-based image analysis) is in GIScience
(geographic information science) and remote sensing fields (especially high-resolution remote sensing image application field)
In be just increasingly taken seriously.And image segmentation is one of the core procedure in OBIA, realizes geographic object in scene
Profile information extracts, and is basis and the premise of subsequent progress feature extraction and target identification.Compared with normal image, remote sensing image
Have cover a wide range, many characteristics such as multiband, more spatial resolutions, the ground species for including are also more abundant, therefore
Traditional dividing method is difficult to directly apply on remote sensing image.At the same time, not with remote sensing satellite spatial resolution
It is disconnected to improve, such as SPOT 5, IKONOS, QuickBird be represent meter level, the high-resolution data of sub-meter grade is widely used to
The every field of the social life such as crop yield investigation, city land plan, disaster monitoring and early warning, therefore it is directed to high-resolution
The image Segmentation Technology of rate remote sensing image has become one of research hotspot of remote sensing fields.
In, compared with low resolution remote sensing image, high-resolution remote sensing image brings spectrum and texture more abundant
Feature, the spatial details such as structure, shape of object information can be expressed clearly, previous mixed pixel problem outstanding
Effective solution is obtained, to improve the inter-class separability of adjacent atural object.On the other hand, figure is also given in the raising of spatial resolution
As segmentation causes new difficulty and challenge: " the different spectrum of jljl " phenomenon is more prominent, i.e. spectral signature between the atural object of identical type
There may be significant difference;The disturbing factors such as detailed information is increased while atural object shade, noise, cloud layer cover impact
Also more significant;Changeable ecological environment, ground abundant species and complicated artificial atural object etc. is all in City scenarios
Difficulty is caused to the accurate geographical objects that extracts.
For this purpose, scholars have proposed some solution countermeasures, most important one means first is that introducing multiple dimensioned point
Strategy is cut, to preferably disclose spatial structure characteristic of the object under different scale.For example, C.Burnet etc. proposes one kind
Based on the multi-scale division algorithm for dividing shape, by estimating the homogeneity of regional area spectral signature and heterogeneous and be iterated excellent
Change, achieves good effect[1];Well-known remote sensing business software eCognition uses fractal net work evolution algorithmic
(fractal net evolution algorithm, FNEA) carries out multi-scale division, takes full advantage of spectrum, the line of object
Information between reason, shape, level and class.It should be pointed out that multi-scale segmentation method is required by artificial in the prior art
Interpretation or trial-and-error method determine scale parameter (scale parameter, SP) that these methods cannot all be referred to as automation
Image segmentation.And the multi-scale division solution of current SP parameter adaptive is not much and sees, this, which also becomes, restricts OBIA skill
One of widely applied main bottleneck of art.
Bibliography
[1]Burnett C,Blaschke T.A multi-scale segmentation/object
relationship modelling methodology for landscape analysis[J].Ecological
modelling,2003,168(3):233-249.
[2] Shao P, Yang G, Niu X, et al.Information Extraction of High-
Resolution Remotely Sensed Image Based on Multiresolution Segmentation[J]
.Sustainability, 2014,6 (8): 5300-5310.
[3]Deng Y,Manjunath B S.Unsupervised segmentation of color-texture
regions in images and video[J].Pattern Analysis and Machine Intelligence,IEEE
Transactions on,2001,23(8):800-810.
[4]Baraldi A,Boschetti L.Operational automatic remote sensing image
understanding systems:Beyond Geographic Object-Based and Object-Oriented
Image Analysis(GEOBIA/GEOOIA).Part 2:Novel system architecture,information/
knowledge representation,algorithm design and implementation[J].Remote
Sensing,2012,4(9):2768-2817.
Summary of the invention
Goal of the invention: aiming at the problems existing in the prior art with deficiency, the present invention is by by conventional color Texture Segmentation
Multiple dimensioned J-image sequence in method JSEG introduces high-resolution remote sensing image segmentation, proposes and is referred to based on local homogeney
Object bounds constraints policy and the region merging technique based on multiple features between the adaptive SP selection strategy of target J-value, scale
Strategy realizes the multi-scale division of automation.It is carried out by the remote sensing image to different sensors type, different spaces resolution
Experiment, and dividing method (the eCognition and document 2) experimental result supervised with two kinds is compared, it was demonstrated that propose calculation
It is more accurate that method not only positions target edges, and it is more complete and be not necessarily to manual intervention to extract object outline, improves cutting procedure
The degree of automation and robustness.
Technical solution: a kind of non-supervisory dividing method of high score remote sensing image based on scale parameter Automatic Optimal, it is main to wrap
Include three steps: 1) the adaptive SP of the J-value based on local homogeney index selection, so that it is determined that best multiple dimensioned J-
Image sequence;2) image segmentation based on object bounds constraints policy between scale is realized by slightly to the multi-scale division of essence;
3) based on the region merging technique of multiple features, to cope with over-segmentation phenomenon that may be present in segmentation result.
SP is adaptively selected
It selects J-image sequence as multiscale analysis platform, and proposes the adaptively selected strategy of SP a kind of.
The calculating process of multiple dimensioned J-image is as follows: carrying out color quantizing in the space LUV to raw video first.It is measuring
Change in image, set window Z having a size of M × M (M is SP) pixel centered on pixel z, and by each pixel in window
Coordinate z (x, y) be used as its pixel value, and z (x, y) ∈ Z.The angle point in window is removed simultaneously.
If gray level sum is P in quantification image, Z is enabledpFor belong in window Z gray level p all pixels set, mp
For the corresponding pixel mean value of all pixels for belonging to gray level p, then belong in window Z the variance of same gray-level pixels and can
It indicates are as follows:
The population variance of all pixels may be expressed as: in window Z
Wherein, m is the mean value of all pixels in window Z.Then part homogeney index J-value may be defined as:
J-value=(ST-SW)/SW (3)
At this point, traversing whole picture quantification image using the corresponding J-value of pixel z as the pixel value of the pixel, SP can get
J-image when for M can get multiple dimensioned J-image sequence by changing SP.
Adaptive SP selection strategy based on J-value:
Step1: calculating SP be M (M=5,6....N) when J-image sequence, wherein M=5 be J-image allow most
Small window size, N represent the J-image of most coarse scale.
Step2: the pixel J-value mean value under all scale J-image is calculatedAnd it constructsCurve.
Step3: numerousIn point of inflexion on a curve, some inflection points the most outstanding are only selected, these inflection points should meet
The multi-scale division of constraint
In the segmentation stage, the segmentation strategy constrained based on object bounds between scale is proposed.If best J-image sequence includes
L scale, is represented by Sk(k=1,2...L), the specific implementation process is as follows:
Step1: first to most coarse scale S1It is split.According to formula (4) threshold value T1Seed region is carried out to mention
It takes, wherein μkAnd σkRespectively indicate scale SkThe J-value mean value and variance of middle all pixels.
Tk=μk-0.2σk, (k=1,2...L) (4)
In S1In, all J-value values are less than T1Pixel using four connection methods constitute connection region, as one by one
Seed region.Using seed region as starting point, area is carried out with the sequence of J-value value from small to large according to four direction up and down
Domain increases, and boundary when adjacent area crosses just constitutes S1Under segmentation result.
Step2: the object bounds of a upper scale are mapped to current scale and are modified.By current scale J-image
It is converted into a width bianry image, is only retained by mapping the object bounds of extraction between scale, and carry out morphological dilation.It is swollen
Swollen structural unit is dimensioned to M × M pixel, and M is the SP of current scale.Using the boundary after expansion by current scale J-
Image is divided into independent seed region one by one, and carries out region from small to large according to J-value value to these seed regions
Increase, boundary when adjacent area crosses is the modified result in boundary.
At this point, the segmentation under current scale only carries out inside the object extracted by amendment back boundary.And in order to avoid mistake
Divide phenomenon, i.e., it has been matched with practical type of ground objects is thought for the internal higher object of homogeneous degree, under current scale
No longer it is split.Judgment rule is that the J-value mean value inside the object should be less than the corresponding threshold value T of current scalek(referring to
Formula 4).On this basis, according to threshold value TkIt is split to remaining object, cutting procedure is identical as scale 1, and will segmentation
As a result it is mapped to next scale.
Step3: repeating the cutting procedure of Step2, until scale L segmentation finishes, to obtain preliminary segmentation result.
The region merging technique of multiple features
The multiple dimensioned J-image sequence of each wave band correspondence of raw video is calculated according to best SP.If raw video includes F
A wave band, for any one object q, defined feature vector Jqf=(Jq1,Jq2..., JqF), wherein each component represents pair
As J-value mean value of the q at L scale J-image of each wave band, then having for any wave band f (f=1,2 ... F)According to the definition of J-value it is found that J-value concentrated expression regional area (object)
Spectrum, texture and dimensional information, therefore by judging adjacent object qAAnd qBFeature vector between Euclidean distance judge its phase
Like degree, as shown in formula (5).
Using RAG (Region Adjacency Graphics) Lai Jinhang region merging technique, detailed process is as follows:
Step1: according to the segmentation result in scale L, the RAG of all adjacent objects is generated.
Step2: selection and any object qAAdjacent all objects calculate Euclidean distance according to formula (5).
Step3: object q if it existsBMeet D (qA,qB)≤0.1, then it is assumed that qAAnd qBBelong to same target, merges qAAnd qB
And generate new RAG.Otherwise, Step2 is returned.
Step4: repeating Step2 to Step3, traverses all objects, obtains final segmentation result.
Specific embodiment
Combined with specific embodiments below, the present invention is furture elucidated, it should be understood that these embodiments are merely to illustrate the present invention
Rather than limit the scope of the invention, after the present invention has been read, those skilled in the art are to various equivalences of the invention
The modification of form falls within the application range as defined in the appended claims.
As shown in Figure 1, the non-supervisory dividing method of high score remote sensing image based on scale parameter Automatic Optimal, mainly includes three
A step: 1) the adaptive SP selection of J-value based on local homogeney index, so that it is determined that best multiple dimensioned J-
Image sequence;2) image segmentation based on object bounds constraints policy between scale is realized by slightly to the multi-scale division of essence;
3) based on the region merging technique of multiple features, to cope with over-segmentation phenomenon that may be present in segmentation result.
SP is adaptively selected
Multi-scale division usually relies primarily on the i.e. SP of core control parameter and is independent ground for piece image division
Manage object.SP is controlled the average-size of spectrum homogeney and object inside object in segmentation result, also practical shadow
Ring the quantity of object in segmentation result.Therefore, it is of the invention first for selecting suitable multiscale analysis tool and SP how to determine
One of critical issue for first needing to solve.
Multiple dimensioned J-image sequence
The multiple dimensioned J-image sequence that JSEG is used can sufficiently reflect the homogeney of regional area spatial distribution, may be used also
It is existing only right when calculating image fusion using traditional multiscale transform analysis tool (such as wavelet transformation, contourlet transform) to avoid
The limitation of the high-frequency information sensitivity in individual directions, but similarly it is faced with the rational selection of SP.Therefore, this method selects J-
Image sequence proposes the adaptively selected strategy of SP a kind of as multiscale analysis platform.
The calculating process of multiple dimensioned J-image is as follows: carrying out color quantizing in the space LUV to raw video first.It is measuring
Change in image, set window Z having a size of M × M (M is SP) pixel centered on pixel z, and by each pixel in window
Coordinate z (x, y) be used as its pixel value, and z (x, y) ∈ Z.Meanwhile the consistency in order to guarantee all directions, the angle in window
Point is removed.By taking M=9 as an example, the window Z centered on pixel z is as shown in Figure 2:
If gray level sum is P in quantification image, Z is enabledpFor belong in window Z gray level p all pixels set, mp
For the corresponding pixel mean value of all pixels for belonging to gray level p, then belong in window Z the variance of same gray-level pixels and can
It indicates are as follows:
The population variance of all pixels may be expressed as: in window Z
Wherein, m is the mean value of all pixels in window Z.Then part homogeney index J-value may be defined as:
J-value=(ST-SW)/SW (3)
At this point, traversing whole picture quantification image using the corresponding J-value of pixel z as the pixel value of the pixel, SP can get
J-image when for M can get multiple dimensioned J-image sequence by changing SP.
Best scale selection based on J-value
Using J-image sequence as multiscale analysis tool, a kind of adaptive SP selection based on J-value is proposed
Strategy.
Step1: calculating SP be M (M=5,6....N) when J-image sequence, wherein M=5 be J-image allow most
Small window size, N (can carry out appropriate adjustment according to real image size, set N=30 herein) represent most coarse scale
J-image.
Step2: the pixel J-value mean value under all scale J-image is calculatedAnd it constructsCurve.
According to the definition of J-value,Point of inflexion on a curve reflects compared with the scale of front and back, the homogeneous journey of spatial distribution under current scale
Degree, which has, suddenly to be increased.So we assume that these inflection points illustrate that certain representative ground species are suitable in scene
It is split under current scale.Namely at these inflection points, the object in segmentation result is matched with practical type of ground objects just,
With same or similar spectrum homogeneous degree, and these representative atural objects can be rightCurve makes a significant impact.
Step3: in numerous inflection points, some inflection points the most outstanding are only selected, these inflection points should meetBe as far as possible effectively extracted most representational atural object type in scene.In addition,
In order to retain the detailed information of image, most fine scale (i.e. M=5) is selected always.At this point it is possible to it is comprehensive it is selected this
A little corresponding SP of inflection point and most fine dimension to determine most preferably multiple dimensioned J-image sequence jointly.
The multi-scale division of constraint
In the segmentation stage, the segmentation strategy constrained based on object bounds between scale is proposed.If best J-image sequence includes
L scale, is represented by Sk(k=1,2...L), the specific implementation process is as follows:
Step1: first to most coarse scale S1It is split.According to formula (4) threshold value T1Seed region is carried out to mention
It takes, wherein μkAnd σkRespectively indicate scale SkThe J-value mean value and variance of middle all pixels.
Tk=μk-0.2σk, (k=1,2...L) (4)
In S1In, all J-value values are less than T1Pixel using four connection methods constitute connection region, as one by one
Seed region.Using seed region as starting point, area is carried out with the sequence of J-value value from small to large according to four direction up and down
Domain increases, and boundary when adjacent area crosses just constitutes S1Under segmentation result.
Step2: the object bounds of a upper scale are mapped to current scale and are modified.According to the definition of J-image
It is found that the multiple dimensioned sequence of J-image is all of the same size with raw video.Therefore upper coarse scale segmentation result is mentioned
The object bounds taken can be mapped to the same position of current fine dimension up according to coordinate, and to point under current scale
The constraint of undercutting row.And the boundary extracted under coarse scale is difficult to standard although can determine position and its general profile of object
Determine the edge of position object, it is therefore desirable to be modified under current scale, process is as follows.
A width bianry image is converted by current scale J-image, only retains the object edges by mapping extraction between scale
Boundary, and carry out morphological dilation.Expansion structure unit size is set as M × M pixel, and M is the SP of current scale.Using swollen
Current scale J-image is divided into independent seed region one by one by the boundary after swollen, and to these seed regions according to J-
Value value carries out region growth from small to large, and boundary when adjacent area crosses is the modified result in boundary.
At this point, the segmentation under current scale only carries out inside the object extracted by amendment back boundary.And in order to avoid mistake
Divide phenomenon, i.e., it has been matched with practical type of ground objects is thought for the internal higher object of homogeneous degree, under current scale
No longer it is split.Judgment rule is that the J-value mean value inside the object should be less than the corresponding threshold value T of current scalek(referring to
Formula 4).On this basis, according to threshold value TkIt is split to remaining object, cutting procedure is identical as scale 1, and will segmentation
As a result it is mapped to next scale.
Step3: repeating the cutting procedure of Step2, until scale L segmentation finishes, to obtain preliminary segmentation result.
The region merging technique of multiple features
Although being differentiated first to the internal higher object of homogeneous degree before to each multi-scale segmentation, excessively
It cuts phenomenon to be still difficult to avoid that, it is also necessary to further region merging technique processing.Due to " jljl outstanding in high-resolution remote sensing image
Different spectrum " and " same object different images " phenomenon, merely with the spectral signature inside object, there may be accidentally consolidation problems, therefore propose one
The region merging technique strategy of kind multiple features.
The multiple dimensioned J-image sequence of each wave band correspondence of raw video is calculated according to best SP.If raw video includes F
A wave band, for any one object q, defined feature vector Jqf=(Jq1,Jq2..., JqF), wherein each component represents pair
As J-value mean value of the q at L scale J-image of each wave band, then having for any wave band f (f=1,2 ... F)According to the definition of J-value it is found that the J-value concentrated expression light of regional area (object)
Spectrum, texture and dimensional information, therefore by judging adjacent object qAAnd qBFeature vector between Euclidean distance judge that its is similar
Degree, as shown in formula (5).
We use RAG (Region Adjacency Graphics) Lai Jinhang region merging technique, and detailed process is as follows:
Step1: according to the segmentation result in scale L, the RAG of all adjacent objects is generated.
Step2: selection and any object qAAdjacent all objects calculate Euclidean distance according to formula (5).
Step3: object q if it existsBMeet D (qA,qB)≤0.1, then it is assumed that qAAnd qBBelong to same target, merges qAAnd qB
And generate new RAG.Otherwise, Step2 is returned.
Step4: repeating Step2 to Step3, traverses all objects, obtains final segmentation result.
Experiment and analysis
For verify proposed method validity and reliability, to two groups of different resolutions, different sensors type it is more
Spectrum high resolution remote sensing image is tested, and with business software eCongnition (referred to as " method two ") and Shao
The high-resolution remote sensing image dividing method (referred to as " method three ") for the supervision that P et al. et al. (document 2) proposes is compared
Compared with.
Wherein, eCongnition is a internationally recognizable object-oriented of German Definiens Imaging exploitation
Classification of remote-sensing images software, the FNEA segmentation strategy used fully utilize the features such as the spectrum, texture and shape of object,
There is excellent performance in the segmentation of high-resolution remote sensing image.ECongnition is in segmentation mainly by three parameters
Control, it may be assumed that scale parameter (Scale Parameter), the main object average-size controlled in segmentation result;Form parameter
(Shape Parameter), facilitates the integrality of keeping object profile in cutting procedure;Compact degree parameter (Compactness
Parameter), the inter-class separability of object is helped to improve.The method that Shao P et al. et al. is proposed is by traditional edge
Detection is introduced into high-resolution remote sensing image segmentation, is realized the Multi resolution feature extraction of object by building object hierarchy and is divided
It cuts, achieves good effect in the segmentation of Chinese ZY-3 satellite image.The setting of relevant parameter is both needed in both the above method
It to be realized by human interpretation, the present invention determines that optimal parameter is combined by trial-and-error method in an experiment.
Test a result and analysis
Using tetra- wave band color integration data of QucikBird in 2005, location was Wuhan, China, space for experiment one
Resolution ratio is 2.4m, and picture size is 512 × 512 pixels.Image is mainly the typical urban scene areas under complex background, packet
Containing ground abundant species such as road, playground, water body and complicated man-made targets, as shown in Figure 3.
In the method for the present inventionCurve is as shown in figure 4, describe as scale parameter M constantly increases, indexVariation
Situation.Wherein vertical dotted line has corresponded to best SP, these scales and most fine dimension together form best multiple dimensioned J-image
Sequence, corresponding scale parameter are M ∈ [5,13,18,28], and final segmentation result is as shown in Figure 5.
In method two, set Scale Parameter as 77, Shape Parameter be 50, Compactness
Parameter is 40.It is identical that each wave band proportion is set in document 2, parameter " Scale Parameter " is 30, parameter
" Shape Heterogeneous Degree " is 0.4, parameter " Compactness Parameter " and " Smoothness
Parameter " is 0.5.Two methods experimental result difference is as shown in Figure 6, Figure 7.
For the ease of being compared to distinct methods experimental result, the present invention is to the typical subject or position progress in scene
Mark.Wherein, position A, B is sports ground, and position C, D, F are building, position E is road, and position G is man-made lake.Pass through
Visual analysis can be seen that three kinds of methods and effectively be extracted the playground areas of position A, and wherein this method and method two position
The obvious ratio method three in the edge on lawn is more accurate, and there are certain over-segmentation phenomenons for method three;For the playground area of position B
Domain, method two do not extract the profile on lawn, and method three then obscures part runway zone with lawn;The building of position C
Structure is complicated, and only this method not only maintains the integrality of object outline, while target edges have been accurately positioned, method two,
Three are respectively present accidentally segmentation and over-segmentation phenomenon;The building shape rule of position D, F, this method and method two position roof
Edge it is more accurate, but there are less divided phenomenons in position F method two;Only method three is effectively extracted the artificial of position G
The profile information in lake.In terms of comprehensive, the method for the present invention and method two position the edge detail information ability side of being substantially better than of object
Method three, and the method for the present invention keeps more complete for the profile of bulk homogenous area.Method three, which can be identified effectively, to be belonged to not
With type but the similar adjacent atural object of spectral signature, but there is also positioning accuracy outstanding is low and over-segmentation problem.
Test two results and analysis
Two three wave band high-resolution air remote sensing DOM (Digital Orthophoto Map) images of selection of experiment, data
Acquisition time is in March, 2009, and spatial resolution 0.6m, having a size of 512 × 512 pixels, location is Nanjing of China, is such as schemed
Shown in 8.It is decreased by being compared with Fig. 3 as can be seen that testing the data background complexity that two use, but variety classes
The minutias such as spectrum, texture, the edge of object are more significant, and object includes that the large area homogenous area of regular shape and structure are answered
Miscellaneous, textural characteristics man-made structures abundant, therefore the complete of target edges and keeping object profile is accurately positioned to partitioning algorithm
Property made higher requirement.
In this methodCurve is as shown in Figure 9.According toCurve is it is found that the corresponding scale parameter of best multiple dimensioned sequence is
M ∈ [5,10,12,21,25,27], segmentation result is as shown in Figure 10.
In method two, set Scale Parameter as 100, Shape Parameter be 50, Compactness
Parameter is 50.It is identical that each wave band proportion is set in document 14, Scale Parameter is 50, Shape
Heterogeneous Degree is that 0.3, Compactness Parameter and Smoothness Parameter are 0.5.
Two methods experimental result is respectively as shown in Figure 11, Figure 12.
It is one identical as experiment, the present invention in scene typical subject or position marked.By to not Tongfang
The visual analysis of method experimental result can be seen that three kinds of methods and accurately be partitioned into the playground lawn of position A and the behaviour of position B
Runway, but only the method for the present invention integrality for effectively maintaining playground profile, method two two larger cut zone it
Between there are some long and narrow false units, such as region adjacent with lawn on the outside of runway;For the fritter lawn of position C, method
Two there are less divided phenomenons;For the building roof of position D, three kinds of method segmentation effects are close, and method two, three further mentions
The texture information inside roof is taken, but there are over-segmentation phenomenon and edge position inaccurates for method three;The playground of position E, F are seen
Structure is complicated for platform, and texture information is abundant, and only the method for the present invention maintains the complete of the ceiling area of grandstand and is effectively extracted
The minutia of two sides auxiliary building;The method of the present invention and method two are accurately extracted the road area positioned at position G, H, method
Three have accidentally segmentation problem;For the large area vegetation region of the court of position I, the tennis court of position J and position K
Domain, three kinds of methods are more accurate to the edge positioning of object, but occur again in the segmentation in tennis court in method two long and narrow
False unit problem.In summary analysis further demonstrates proposed algorithm it can be concluded that with conclusion as experiment one kind
Reliability.
Precision evaluation
It mainly analyzes by visual observation above and the segmentation effect of distinct methods is evaluated, precision index will be used herein
Further quantitative analysis is done to experimental result.In the method that Deng et al. is proposed (document 3), index J-value is not only
It is defined as follows for calculating multiple dimensioned J-image sequence also as the evaluation index of segmentation precision:
Wherein, U is the sum of all pixels in image, and R is the region sum in segmentation result, WrAnd JrRespectively r-th of region
Internal sum of all pixels J-value corresponding with its.When the corresponding precision index of segmentation resultMore hour illustrates segmentation knot
Averaged spectrum homogeney in fruit inside object is higher, then segmentation effect is better.Due to atural object huge number in remote sensing image,
It is also not quite similar to the evaluation angle of segmentation precision with standard in different applications, and what Deng et al. et al. was proposedSpatial distribution situation in segmentation result is evaluated on the whole, there is good versatility[4].Therefore, the present invention adopts
WithPrecision evaluation is carried out to the experimental result of three kinds of methods.
To J in segmentation resultrDistribution situation is analyzed, by the corresponding J-value value of all objects in [0,1] section
Uniform quantization is 20 units, then JrDistribution curve is as shown in figs. 13 and 14:
The object for belonging to the different sections J-value in three kinds of methods is respectively represented in figure with the solid line of different colours respectively
The shared specific gravity in segmentation result, and then to represent three kinds of methods corresponding for dotted lineIndex.By comparing Figure 13,14 can
To find out: in two groups of experiments, the present invention proposes method precision indexOther two methods are significantly better than that, with visual point
It is consistent to solve result.J in two groups of experimentsrCurvilinear trend is roughly the same, and the typical feature of difference major embodiment in the scene is concentrated
Section, such as [0.2, the 0.55] section in experiment one and [0.1,0.4] section in experiment two.In addition in experiment two, three
The relatively experiment one of the segmentation precision of kind algorithm is significantly increased, main reason is that the remote sensing image spatial discriminations that experiment two uses
Rate is higher, and background is relatively easy and atural object minutia is more prominent, therefore the object of experiment extraction and its boundary are more nearly
Type of ground objects in actual scene.
For the automatic segmentation of high-resolution remote sensing image, the invention proposes a kind of novel non-supervisory multiple dimensioned point
Segmentation method.This method fully utilizes the spectrum and textural characteristics of object, proposes based on biography based on local homogeney index
Adaptive scale parameter (SP) selection strategy of J-value, enable typical feature type in scene with its
It is split in the best scale J-image matched.On this basis, the multi-scale division strategy proposed make region segmentation by
A upper scale extracts the constraint of institute's object bounds, and is modified under current scale to these boundaries, avoids between scale
The accumulation of error.And the region merging technique strategy based on multiple features is with then can effectively distinguishing the variety classes with similar spectral feature
Object avoids and accidentally merges phenomenon.Experiment shows that proposed method is fixed compared with eCognition and traditional supervised segmentation method
Position target edges are accurate and extraction object outline is more complete, have higher segmentation precision, while can be realized automation
High-resolution remote sensing image segmentation, whole process are not necessarily to manual intervention, are a kind of generic and effective non-supervisory solution.