CN108596918A - A kind of remote sensing image segmentation method merging tree based on level - Google Patents
A kind of remote sensing image segmentation method merging tree based on level Download PDFInfo
- Publication number
- CN108596918A CN108596918A CN201810364791.4A CN201810364791A CN108596918A CN 108596918 A CN108596918 A CN 108596918A CN 201810364791 A CN201810364791 A CN 201810364791A CN 108596918 A CN108596918 A CN 108596918A
- Authority
- CN
- China
- Prior art keywords
- patch
- level
- node
- tree
- remote sensing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20021—Dividing image into blocks, subimages or windows
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20112—Image segmentation details
- G06T2207/20156—Automatic seed setting
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of remote sensing image segmentation methods merging tree based on level, it is related to image segmentation field.Include the following steps:Its all pixels is initialized as individual patch by the remote sensing images for 1, reading input;2, it is grown using seed region(Seeded Region Growing, SRG)It is several super-pixel patches by initialized Remote Sensing Image Segmentation;3, merge tree using level patch merging method structure level;4, tree is merged to level using top-down strategy and carries out node selection;5, for each selected node, the patch corresponding to it is exported as a result.The present invention is by using a kind of data structure for being known as level and merging tree, to store and show the implementation procedure of level patch merging method.The node for being merged tree using level proposed by the present invention is selected, and is made the atural object quilt of different scale while completely being split, to improve the segmentation precision of remote sensing images.
Description
Technical field
The present invention relates to image segmentation fields, and in particular to a kind of Remote Sensing Image Segmentation side merging tree based on level
Method.
Background technology
The remote sensing image classification of object-oriented is a kind of technology for remote sensing image information interpretation.This technology is very suitable
For the remote sensing image interpretation of high spatial resolution, because compared with Remote Image Classification traditional, pixel-based, it
Space background feature can be substantially more utilized, to improve the accuracy and reliability of classification.This technology includes two weights
Want step:(1) image segmentation, the identification of (2) patch or classification.Wherein, step (1) be remote sensing images are divided into it is several in space
The patch of upper non-overlapping copies, wherein what each patch was made of several pixels spatially abutted.The performance of step (1) exists
Largely affect the effect of step (2).Because if image segmentation the pixel for belonging to different atural objects is divided into it is same
In patch, then by inevitable mistake must occur for subsequent classification.In order to avoid such classification error, image segmentation is improved
The precision of method is very important.
Currently, the remote sensing image segmentation method of mainstream is realized based on region merging technique (or patch merging) method mostly
's.Wherein method the most famous includes dividing shape net evolution method (Fractal Net Evolution Approach, FNEA)
With Hierarchical Segmentation method (Hierarchical Segmentation, HSeg).Both methods all uses bottom-up plan
Slightly, i.e., in the starting stage of algorithm performs, each pixel is taken as independent patch (can also be carried out just using super-pixel
Beginningization, to accelerate the calculating speed of image segmentation;If it is substantially similar that super-pixel can be counted as size approximately equal, shape
Dry spot block;Common super-pixel generating algorithm includes simple linear iteration cluster (Simple Linear Iterative
Clustering, SLIC), seed region growth (Seeded Region Growing, SRG) etc.);Meet certain by merging
The patch of similarity measurement makes the shape of the gradual approaching to reality atural object of the patch after merging to (i.e. two patches);Such mistake
Journey iteration carries out, until the similarity measurement of all patches pair is below a scale threshold parameter;Scale threshold parameter needs
It is pre-set by user, the height of value determines the mean size of each patch in segmentation result.In general, scale threshold value
Higher, the merging number being allowed to is more, and the average dimension of patch is also bigger in segmentation result.Joined using different scales
Number, can split different size of atural object in remote sensing images.But in many remote sensing images, between different atural objects
Different scale it is very big, these atural objects are split simultaneously, are difficult to realize using the strategy of scale threshold parameter.
In conclusion the present invention devises a kind of remote sensing image segmentation method merging tree based on level.
Invention content
In view of the shortcomings of the prior art, purpose of the present invention is to be to provide a kind of remote sensing merging tree based on level
Image partition method, compared with existing remote sensing image segmentation method, the present invention can more effectively simultaneously by remote sensing images not
Atural object full segmentation with scale comes out, and has the characteristics that segmentation precision is high.
To achieve the goals above, the present invention is to realize by the following technical solutions:One kind merging tree based on level
Remote sensing image segmentation method, based on level merge tree Remote Sensing Image Segmentation be for the atural object in remote sensing images to be divided into
A kind of technology of independent patch, it is the key link of remote sensing image classification or target identification.Include the following steps:
Its all pixels is initialized as individual patch by the remote sensing images for 1, reading input;
2, initialized remote sensing images are divided using seed region growth (Seeded Region Growing, SRG)
It is segmented into several super-pixel patches;
3, merge tree using level patch merging method structure level;
4, tree is merged to level using top-down strategy and carries out node selection;
5, for each selected node, the patch corresponding to it is exported as a result.
Level constructed by the step 3 merges the definition set:It is a kind of data structure being under the jurisdiction of binary tree,
Process for storing and showing the merging of level patch;Level merges tree and is made of three kinds of nodes:Root node, leaf node and in
Intermediate node;Root node indicates the patch being made of all pixels of the remote sensing images inputted;Leaf node is indicated by initializing
Some the super-pixel patch generated;Intermediate node indicates to merge generated patch by the patch of its child nodes.
Merge tree using level patch merging method structure level in the step 3, includes the following steps:Step 3.1:
A linear list list is created, is suitble to combined patch pair for storing;One patch is to including two patches;Each of list
Element includes three variables:A, the pointer variable of patch 1, b, patch 2 pointer variable, c, patch 1 and patch 2 heterogeneous degree
Magnitude Ch, it is defined as:Ch=(1-fshape)Cspec+fshapeCshape, fshapeIndicate shape similarity measurement CspecWeight;
CspecDefinition be:Wherein n1、σ1,jIndicate the spot for including seed point
The gray standard deviation of the number of pixels of block and the patch in wave band j;n2、σ2,jIt indicates (to contain the spot of seed point with patch 1
Block) spatially adjacent single pixel patch;σ1+2,jIndicate patch 1 merge with patch 2 after patch wave band j gray scale
It is accurate poor;Shape similarity measurement CshapeDefinition bep1、p2、p1+2Table respectively
Show the perimeter of the patch after patch 1, patch 2, patch 1 merge with patch 2;The C of two patcheshValue is lower, illustrates that they are more suitable
Merge;Step 3.2:The super-pixel patch pair of all suitable merging is found, and is inserted it into list;Step 3.3:With Ch
For keyword, sequence from low to high is carried out to all elements in list;Step 3.4:Level, which is carried out, using list merges tree
Structure.
The step 3.2 includes following two sub-steps:Step 3.2.1:For each super-pixel patch, office is utilized
Portion is mutually most suitable for principle, searches for and is most suitable for combined patch with it;Locally the rule of mutually most suitable principle is:For a certain
Patch (is set as patch 1), in all patches adjacent thereto connect, search and its ChIt is worth minimum patch (being set as patch 2);
In all of its neighbor patch of patch 2, search in the C with patch 2hIt is worth minimum patch (being set as patch 3);If patch 1 and patch 3
It is not the same patch, then returns to null value;Otherwise, patch 2 is returned;
Step 3.2.2:If according to the mutually most suitable principle in part, there is no combined patch is suitble to current patch, then
Continue with next patch;Otherwise, the C of combined patch and two patches is suitble to by current patch, with ithValue composition
One new element, is then inserted into list.
The step 3.4 specifically includes following four sub-step:Step 3.4.1:For all super-pixel patches create with
Corresponding level merge tree node;Each node includes 4 variables:A, it is directed toward the pointer variable of father node, b, is referred to
The pointer variable of child to the left, c, the pointer variable for being directed toward right child, d, form the patch pixel chained list head portrait element finger
Needle variable;For the node corresponding to each super-pixel patch, father node, left child and right child nodes pointer variable
It is all set to 0;
Step 3.4.2:First element in list is taken out, patch 1, the spot of the first two variable meaning in the element are merged
Block 2;The element number of list is enabled to reduce 1;A new node is created using the patch newly obtained, left and right child nodes
Pointer variable is respectively directed to the node corresponding to patch 1, patch 2;
Step 3.4.3:It is suitable with it using the mutually most suitable principle search in part for patch caused by previous step
Combined patch is closed, if there are such patch, two patches and its Ch value is formed into a new element and are inserted into list;
If being not present, carry out in next step;
Step 3.4.4:If the element number of list is less than 1, return to step 4.2;Otherwise, terminate the structure that level merges tree
It builds.
Tree is merged to level using top-down strategy in the step 4 and carries out node selection, is included the following steps:
Step 4.1:Three linear lists are created, wherein the first two linear list is used to distinguish each layer that storage hierarchy merges tree
The two linear lists are referred to as level_table by all child nodes of node and current node layer;Third linear list is used
In the selected node of storage, referred to as node_table;Each element of these three linear lists each means to merge to level and set
The pointer variable of node;
Step 4.2:According to top-down sequence, tree successively is merged to level and carries out node selection.
The step 4.2 specifically includes following four sub-step:
Step 4.2.1:If first level_table that step 4.1 creates is table1, second level_table
For table2;
Step 4.2.2:The root node that level is merged to tree is inserted into the table1;Table1 is enabled to indicate current layer;
Step 4.2.3:For each node in table1, according to formulaIt is right to calculate node institute
The average gray standard deviation sigma of all pixels for the patch answered, wherein σjIndicate the gray standard deviation of wave band j;J is remote sensing images
Wave band number;If σ < f × σI, then present node is inserted into node_table;Otherwise, by two children of present node
Node is inserted into another table2;Wherein f is one and needs parameter set by the user;σIIndicate whole scape remote sensing images picture
The average gray standard deviation of element;
Step 4.2.4:Exchange the level_table pointed by table1 and table2;If the element number of table1 is big
In 0, then return to step 4.2.3;Otherwise, representational level merges the node selection set and has completed.
The invention has the advantages that:By using a kind of data structure for being known as level and merging tree, storing and
Show the implementation procedure of level patch merging method.The node for being merged tree using level proposed by the present invention is selected, and different rulers are made
The atural object of degree is completely split simultaneously, to improve the segmentation precision of remote sensing images.
Description of the drawings
The following describes the present invention in detail with reference to the accompanying drawings and specific embodiments;
Fig. 1 is flow chart of the method for the present invention;
Fig. 2 is the exemplary plot that the embodiment of the present invention is used to describe the middle-level structure for merging tree of the method for the present invention;
Fig. 3 is the segmentation knot of remote sensing image data and three kinds of dividing methods in contrast test used by the embodiment of the present invention
Fruit, wherein (a) is used False color image remote sensing images, color configuration is:R:Near infrared band, G:Red band,
B:Green band, (b) be the expert's manual extraction calculated for quantitative assessment patch, be (c) caused by step 2 of the present invention
Super-pixel segmentation as a result, (d) be the method for the present invention segmentation result, (e) be HRM segmentation result, (f) be HSeg segmentation knot
Fruit.
Specific implementation mode
To make the technical means, the creative features, the aims and the efficiencies achieved by the present invention be easy to understand, with reference to
Specific implementation mode, the present invention is further explained.
Referring to Fig.1, present embodiment uses following technical scheme:Merge the Remote Sensing Image Segmentation side of tree based on level
The flow chart of method, specifically includes following steps:
Step 1:Its all pixels is initialized as individual patch by the remote sensing images for reading input;
Step 2:(Seeded Region Growing, SRG) is grown by initialized remote sensing figure using seed region
As being divided into several super-pixel patches;
Step 2.1:The initialization of seed point location specifically includes following sub-step:
Step 2.1.1:Seed point is equally spaced placed, it is l to make the distance between each seed point, and l is set as 10;
Step 2.1.2:Prewitt filtering is carried out to each wave band of the remote sensing images inputted, filtering is calculated and utilized
Prewitt operators include horizontal and vertical both direction, the formula of wherein horizontal direction Prewitt operators is:
The formula of vertical direction Prewitt operators is:
For including the remote sensing images of J wave band, the filter result image of 2J wave band can be obtained;After filtering calculates,
Its boundary strength value d is calculated for each pixel:Wherein νh,jIndicate that wave band j is in the horizontal direction
Prewitt filter results, νv,jThen indicate wave band j vertical direction Prewitt filter results;J indicates wave band number;
Step 2.1.3:For each seed point pixel, search is centered on it, l is side in the square area of the length of side
The position of boundary's intensity d minimums, and it is set to the new position of the seed point;
Step 2.2:For all patches for including seed point, region growing operation is carried out one by one, specifically includes following son
Step:
Step 2.2.1:The patch (being set as patch 1) for including seed point for one, search with it is spatially adjacent,
Spectral similarity measures CspecLess than grey similarity threshold value TspecAll single pixel patches, whereinn1, σ1,jIt indicates the number of pixels of the patch comprising seed point and is somebody's turn to do
Gray standard deviation of the patch in wave band j;n2, σ2,jIt indicates spatially adjacent with patch 1 (patch for containing seed point)
Single pixel patch;σ1+2,jIndicate patch 1 merge with patch 2 after patch wave band j gray standard deviation;TspecIt is gray scale phase
Like degree threshold value, it is pre-set by user, is set to 30;If the single pixel patch for not meeting condition exists, directly
Enter step 2.2.2;For meeting all single pixel patches of conditions above, the shape similarity degree of they and patch 1 are calculated
Measure Cshape, whereinp1、p2、p1+2Respectively indicate patch 1, patch 2, patch 1 with
The perimeter of patch after the merging of patch 2;Select CshapeMinimum single pixel patch, then by itself and the current spot for including seed point
Merged block;
Step 2.2.2:Step 2.2.1 is carried out to next patch comprising seed point, it is all comprising kind until having handled
The patch of son point;
Step 2.2.3:If step 2.2.2 does not merge generation, 2.3 are entered step;Otherwise return to step 2.2.2;
Step 2.3:For the patch that scale is too small, it is adjacent the most similar patch in patch and is merged,
The operation specifically included is:Search is less than T with the presence or absence of number of pixelssizePatch;TsizeFor the scale threshold value of trifling patch,
It is set to 16;If there are such patch, it is adjacent the patch the most approximate of gray scale in patch and is merged;
If being not present, 3 are entered step;
Step 3:Merge tree using level patch merging method structure level;
In order to illustrate the structure that a level merges tree, Fig. 2 gives a simple exemplary plot:It is on the left of the figure
One scape remote sensing images include altogether five super-pixel patches, are respectively expressed as a, b, c, d, e;Right side is and left side remote sensing images
Corresponding level merges tree, and green node therein indicates leaf node, they correspond to five super-pixel patches respectively;The figure is right
The purple node of side indicates root node, it contains all super-pixel patches in remote sensing images;Blue node in the figure is
Intermediate node, it indicates to be formed by larger patch by leaf node merging.
Step 3.1:The linear list list that a length is N length is created, is suitble to combined patch pair for storing;N
Length is equal to the number of the obtained super-pixel patch of step 2;Each element of list includes three variables:A, patch 1
Pointer variable, b, patch 2 pointer variable, c, patch 1 and patch 2 heterogeneous metric Ch;Wherein, Ch=(1-fshape)
Cspec+fshapeCshape, wherein fshapeIndicate the weight of shape similarity measurement;It is set to 0.1;CspecWith CshapeDetermine
Justice is shown in step 2.2.1;The C of two patcheshValue is lower, illustrates their more suitable merging;
Step 3.2:Find the super-pixel patch pair of all suitable merging, and insert it into list, specifically include with
Lower two sub-steps:
Step 3.2.1:For each super-pixel patch, using the mutually most suitable principle in part, search is most suitable for it
Combined patch;Locally the rule of mutually most suitable principle is:For a certain patch (being set as patch 1), connect adjacent thereto
In all patches, search and its ChIt is worth minimum patch (being set as patch 2);In all of its neighbor patch of patch 2, search with
The C of patch 2hIt is worth minimum patch (being set as patch 3);If patch 1 and patch 3 are not the same patches, null value is returned;It is no
Then, patch 2 is returned;
Step 3.2.2:If according to the mutually most suitable principle in part, there is no combined patch is suitble to current patch, then
Continue with next patch;Otherwise, the C of combined patch and two patches is suitble to by current patch, with ithValue composition
One new element, is then inserted into list;
Step 3.3:With ChFor keyword, sequence from low to high is carried out to all elements in list;
Step 3.4:The structure that level merges tree is carried out using list, specifically includes following four sub-step:
Step 3.4.1:The node that corresponding level merges tree is created for all super-pixel patches;Each node
Including 4 variables:A, the pointer variable of father node, b, the pointer variable for being directed toward left child, c, the pointer for being directed toward right child are directed toward
Variable, d, form the patch pixel chained list head portrait element pointer variable;For the section corresponding to each super-pixel patch
Point, father node, left child and the pointer variable of right child nodes are all set to 0;
Step 3.4.2:First element in list is taken out, patch 1, the spot of the first two variable meaning in the element are merged
Block 2;The element number of list is enabled to reduce 1;A new node is created using the patch newly obtained, left and right child nodes
Pointer variable is respectively directed to the node corresponding to patch 1, patch 2;
Step 3.4.3:It is suitable with it using the mutually most suitable principle search in part for patch caused by previous step
Combined patch is closed, if there are such patch, two patches and its Ch value is formed into a new element and are inserted into list;
If being not present, carry out in next step;
Step 3.4.4:If the element number of list is less than 1, return to step 4.2;Otherwise, terminate the structure that level merges tree
It builds;
Step 4:Tree is merged to level using top-down strategy and carries out node selection;
Step 4.1:Three linear lists are created, wherein the first two linear list is used to distinguish each layer that storage hierarchy merges tree
The two linear lists are referred to as level_table by all child nodes of node and current node layer;Third linear list is used
In the selected node of storage, referred to as node_table;Each element of these three linear lists each means to merge to level and set
The pointer variable of node;
Step 4.2:According to top-down sequence, tree successively is merged to level and carries out node selection, is specifically included following
Four sub-steps:
Step 4.2.1:If first level_table that step 4.1 creates is table1, second level_table
For table2;
Step 4.2.2:The root node that level is merged to tree is inserted into the table1;Table1 is enabled to indicate current layer;
Step 4.2.3:For each node in table1, according to formulaIt is right to calculate node institute
The average gray standard deviation sigma of all pixels for the patch answered, wherein σjIndicate the gray standard deviation of wave band j;J is remote sensing images
Wave band number;If σ < f × σI, then present node is inserted into node_table;Otherwise, by two children of present node
Node is inserted into another table2;Wherein f is one and needs parameter set by the user;σIIndicate whole scape remote sensing images picture
The average gray standard deviation of element;
Step 4.2.4:Exchange the level_table pointed by table1 and table2;If the element number of table1 is big
In 0, then return to step 4.2.3;Otherwise, representational level merges the node selection set and has completed;
Step 5:For each selected node, the patch corresponding to it is exported as a result.
The present invention can CPU be Intel (R) Core (TM) i5-6500@3.20GHz, 8.00GB memory rams,
On Windows7 flagship edition systems Remote Sensing Image Segmentation is realized using 2010 software programmings of Microsoft Visual Studio
Experiment.
Embodiment 1:The remote sensing images of No. 2 remote sensing satellites of scape high score acquisition, the scape image are used in present embodiment
See Fig. 3 (a), image size is 400 × 400 pixels, Pixel size 3.2m, and center longitude is:(E114.1547°,
N30.5842 °), it is on 2 12nd, 2015 to obtain the date, includes near-infrared, red, green, blue four wave bands altogether.
In order to embody advantage of the method for the present invention on segmentation precision, by the method for the present invention and other two kinds of remote sensing images point
The segmentation result of segmentation method is compared.Two methods for comparison are that Mixed Zone merges (Hybrid Region respectively
Merging, HRM) and Hierarchical Segmentation (Hierarchical Segmentation, HSeg), wherein HRM is a kind of classical
Remote sensing image segmentation method --- divide the basis of shape net evolution method (Fractal Net Evolution Approach, FNEA)
On advanced optimize.
In order to carry out quantitative assessment to three kinds of methods, present embodiment uses a kind of quantitative thresholding segmentation side
Method, this method are calculated by the way that the segmentation result of method to be evaluated to be compared with expert's manual extraction result of practical atural object
Go out accuracy rate, recall rate and F to measure, the range of these three evaluation criterions is (0,1), and is all met:Its value is bigger, segmentation
Precision is better.Fig. 3 (b) shows the atural object patch by expert's manual extraction in this experiment.
Table 1 shows the quantitative assessment result of three kinds of methods, it is seen then that the accurate rate, recall rate and F measurements of the method for the present invention
It is highest.Fig. 3 (b) is the super-pixel segmentation result that step 2 of the present invention generates, it is seen that all super-pixel patches therein are big
It is small uniform, and the degree that the boundary of each super-pixel patch matches with practical atural object boundary is higher.Fig. 3 (d), (e), (f)
The segmentation result for respectively illustrating the method for the present invention, HRM and HSeg is observed it can be found that HRM and HSeg fails some
Building carries out complete Ground Split.It is noted that HRM, HSeg have selected optimal scale parameter, i.e. F to measure highest
When corresponding scale parameter, value is respectively 50 and 60.In contrast, the method for the present invention is more completely extracted building
Patch, this node selecting method for being mainly attributed to merge level in step 3 of the present invention tree can be more effectively by different rulers
The atural object of degree carries out full segmentation.
Traditional patch merging method limits union operation by the way that scale threshold parameter is arranged, to be closed as iteration
And the end condition of process;Such methods, such as FNEA, HRM, HSeg will not dependent on different scale threshold parameters is arranged
Atural object with scale is split.But in many practical remote sensing images, even for same class atural object, different scale
May be very big, it is split simultaneously, traditional patch merging method is difficult to realize.The method of the present invention utilizes layer
Secondary merging sets completely to record level patch merging process, and the top-down level by being proposed merges burl and clicks
Selection method splits the patch of different scale simultaneously, to improve the precision of Remote Sensing Image Segmentation.It is distant by a scape high score
Feel the confirmatory experiment of image, the method for the present invention can effectively split the atural object of different scale simultaneously, obtain preferable
Segmentation result.
1 quantitative assessment result of table
Method/quantitative assessment criteria | Accuracy rate | Recall rate | F is measured |
The method of the present invention | 0.9339 | 0.8128 | 0.8691 |
HRM | 0.8194 | 0.7581 | 0.7876 |
HSeg | 0.8686 | 0.7617 | 0.8117 |
The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention.The technology of the industry
Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the above embodiments and description only describe this
The principle of invention, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changes
Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and its
Equivalent thereof.
Claims (7)
1. a kind of remote sensing image segmentation method merging tree based on level, which is characterized in that include the following steps:
(1), its all pixels is initialized as individual patch by the remote sensing images for reading input;
(2), it is several super-pixel patches to be grown initialized Remote Sensing Image Segmentation using seed region;
(3), merge tree using level patch merging method structure level;
(4), tree is merged to level using top-down strategy and carries out node selection;
(5), for each selected node, the patch corresponding to it is exported as a result.
2. a kind of remote sensing image segmentation method merging tree based on level according to claim 1, which is characterized in that described
The step of (3) constructed by level merge tree definition be:It is a kind of data structure being under the jurisdiction of binary tree, for store and
Show the process that level patch merges;Level merges tree and is made of three kinds of nodes:Root node, leaf node and intermediate node;Root
Node indicates the patch being made of all pixels of the remote sensing images inputted;Leaf node indicates to be generated by initialization a certain
A super-pixel patch;Intermediate node indicates to merge generated patch by the patch of its child nodes.
3. a kind of remote sensing image segmentation method merging tree based on level according to claim 1, which is characterized in that described
The step of (3) in using level patch merging method structure level merge tree, include the following steps:
Step (3.1):A linear list list is created, is suitble to combined patch pair for storing;One patch is to including two
Patch;Each element of list includes three variables:A, the pointer variable of patch 1, b, patch 2 pointer variable, c, patch 1 with
The heterogeneous metric C of patch 2h, it is defined as:Ch=(1-fshape)Cspec+fshapeCshape, fshapeIndicate shape similarity degree
Measure CspecWeight;CspecDefinition be:Wherein n1、σ1,jIt indicates
Include the gray standard deviation of the number of pixels of the patch of seed point and the patch in wave band j;n2、σ2,jExpression (is wrapped with patch 1
Contain the patch of seed point) spatially adjacent single pixel patch;σ1+2,jIndicate the patch after patch 1 merges with patch 2
In the gray standard deviation of wave band j;Shape similarity measurement CshapeDefinition be
p1、p2、p1+2The perimeter of the patch after patch 1, patch 2, patch 1 merge with patch 2 is indicated respectively;The C of two patcheshValue is got over
It is low, illustrate their more suitable merging;
Step (3.2):The super-pixel patch pair of all suitable merging is found, and is inserted it into list;
Step (3.3):With ChFor keyword, sequence from low to high is carried out to all elements in list;
Step (3.4):The structure that level merges tree is carried out using list.
4. a kind of remote sensing image segmentation method merging tree based on level according to claim 3, which is characterized in that described
The step of (3.2) include following two sub-steps:
Step (3.2.1):For each super-pixel patch, using the mutually most suitable principle in part, search is most suitable for closing with it
And patch;Locally the rule of mutually most suitable principle is:For a certain patch, be set as patch 1, it is adjacent thereto connect it is all
In patch, search and its ChIt is worth minimum patch, is set as patch 2;In all of its neighbor patch of patch 2, search with patch 2
ChIt is worth minimum patch, is set as patch 3;If patch 1 and patch 3 are not the same patches, null value is returned;Otherwise, it returns
Patch 2;
Step (3.2.2):If according to the mutually most suitable principle in part, there is no combined patch is suitble to current patch, then after
The continuous next patch of processing;Otherwise, the C of combined patch and two patches is suitble to by current patch, with ithValue composition one
A new element, is then inserted into list.
5. a kind of remote sensing image segmentation method merging tree based on level according to claim 3, which is characterized in that described
The step of (3.4 specifically include following four sub-step:
Step (3.4.1):The node that corresponding level merges tree is created for all super-pixel patches;Each node packet
Containing 4 variables:A, it is directed toward the pointer variable of father node, b, the pointer variable for being directed toward left child, c, the pointer for being directed toward right child become
Amount, d, form the patch pixel chained list head portrait element pointer variable;For the node corresponding to each super-pixel patch,
Its father node, left child and the pointer variable of right child nodes are all set to 0;
Step (3.4.2):First element in list is taken out, patch 1, the patch of the first two variable meaning in the element are merged
2;The element number of list is enabled to reduce 1;A new node, the finger of left and right child nodes are created using the patch newly obtained
Needle variable is respectively directed to the node corresponding to patch 1, patch 2;
Step (3.4.3):For patch caused by previous step, it is suitble to it using the mutually most suitable principle search in part
If two patches and its Ch value are formed a new element and are inserted into list by combined patch there are such patch;If
It is not present, then carries out in next step;
Step (3.4.4):If the element number of list is less than 1, return to step 4.2;Otherwise, terminate the structure that level merges tree
It builds.
6. a kind of remote sensing image segmentation method merging tree based on level according to claim 1, which is characterized in that described
Tree is merged to level using top-down strategy in step (4) and carries out node selection, is included the following steps:
Step (4.1):Three linear lists are created, wherein the first two linear list is used to distinguish each layer of section that storage hierarchy merges tree
The two linear lists are referred to as level_table by all child nodes of point and current node layer;Third linear list is used for
Store selected node, referred to as node_table;Each element of these three linear lists, which is each meant to level, merges burl
The pointer variable of point;
Step (4.2):According to top-down sequence, tree successively is merged to level and carries out node selection.
7. a kind of remote sensing image segmentation method merging tree based on level according to claim 6, which is characterized in that described
The step of (4.2) specifically include following four sub-step:
Step (4.2.1):If first level_table that step (4.1) creates is table1, second level_table
For table2;
Step (4.2.2):The root node that level is merged to tree is inserted into the table1;Table1 is enabled to indicate current layer;
Step (4.2.3):For each node in table1, according to formulaIt calculates corresponding to the node
Patch all pixels average gray standard deviation sigma, wherein σjIndicate the gray standard deviation of wave band j;J is the wave of remote sensing images
Hop count mesh;If σ < f × σI, then present node is inserted into node_table;Otherwise, two children of present node are saved
Point is inserted into another table2;Wherein f is one and needs parameter set by the user;σIIndicate whole scape remote sensing images pixel
Average gray standard deviation;
Step (4.2.4):Exchange the level_table pointed by table1 and table2;If the element number of table1 is more than
0, then return to step 4.2.3;Otherwise, representational level merges the node selection set and has completed.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810364791.4A CN108596918A (en) | 2018-04-23 | 2018-04-23 | A kind of remote sensing image segmentation method merging tree based on level |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810364791.4A CN108596918A (en) | 2018-04-23 | 2018-04-23 | A kind of remote sensing image segmentation method merging tree based on level |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108596918A true CN108596918A (en) | 2018-09-28 |
Family
ID=63614552
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810364791.4A Pending CN108596918A (en) | 2018-04-23 | 2018-04-23 | A kind of remote sensing image segmentation method merging tree based on level |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108596918A (en) |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102915541A (en) * | 2012-10-31 | 2013-02-06 | 上海大学 | Multi-scale image segmenting method |
CN103500344A (en) * | 2013-09-02 | 2014-01-08 | 中国测绘科学研究院 | Method and module for extracting and interpreting information of remote-sensing image |
CN104112007A (en) * | 2014-07-16 | 2014-10-22 | 深圳大学 | Data storage, organization and retrieval methods of image gradation segmentation result |
CN104657995A (en) * | 2015-02-12 | 2015-05-27 | 合肥工业大学 | Remote sensing image segmentation method adopting region splitting technology |
WO2018037853A1 (en) * | 2016-08-26 | 2018-03-01 | シャープ株式会社 | Image decoding apparatus and image coding apparatus |
-
2018
- 2018-04-23 CN CN201810364791.4A patent/CN108596918A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102915541A (en) * | 2012-10-31 | 2013-02-06 | 上海大学 | Multi-scale image segmenting method |
CN103500344A (en) * | 2013-09-02 | 2014-01-08 | 中国测绘科学研究院 | Method and module for extracting and interpreting information of remote-sensing image |
CN104112007A (en) * | 2014-07-16 | 2014-10-22 | 深圳大学 | Data storage, organization and retrieval methods of image gradation segmentation result |
CN104657995A (en) * | 2015-02-12 | 2015-05-27 | 合肥工业大学 | Remote sensing image segmentation method adopting region splitting technology |
WO2018037853A1 (en) * | 2016-08-26 | 2018-03-01 | シャープ株式会社 | Image decoding apparatus and image coding apparatus |
Non-Patent Citations (4)
Title |
---|
方勇 等: "《办公自动化技术 原理、使用、维护》", 30 April 1997, 电子科技大学出版社 * |
苏腾飞 等: "基于优化合并的高分辨率遥感影像分割算法", 《地球信息科学》 * |
苏腾飞 等: "基于纹理特征与区域生长的高分辨率遥感影像分割算法", 《国土资源遥感》 * |
陆泉: "《图像语义信息可视化交互研究》", 31 July 2015, 国防图书馆出版社 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105608474B (en) | Region adaptivity plant extraction method based on high resolution image | |
CN108052966B (en) | Remote sensing image scene automatic extraction and classification method based on convolutional neural network | |
Chan et al. | Evaluation of Random Forest and Adaboost tree-based ensemble classification and spectral band selection for ecotope mapping using airborne hyperspectral imagery | |
CN110414577A (en) | A kind of laser radar point cloud multiple target Objects recognition method based on deep learning | |
CN107292339A (en) | The unmanned plane low altitude remote sensing image high score Geomorphological Classification method of feature based fusion | |
CN111199214A (en) | Residual error network multispectral image ground feature classification method | |
CN101847263B (en) | Unsupervised image division method based on multi-target immune cluster integration | |
CN106611421B (en) | The SAR image segmentation method constrained based on feature learning and sketch line segment | |
CN106683102B (en) | SAR image segmentation method based on ridge ripple filter and convolutional coding structure learning model | |
CN103888541B (en) | Method and system for discovering cells fused with topology potential and spectral clustering | |
CN109829425B (en) | Farmland landscape small-scale ground feature classification method and system | |
CN109145965A (en) | Cell recognition method and device based on random forest disaggregated model | |
Su et al. | Machine learning-assisted region merging for remote sensing image segmentation | |
CN109785329A (en) | Based on the purple soil image segmentation extracting method for improving SLIC algorithm | |
CN111598045A (en) | Remote sensing farmland change detection method based on object spectrum and mixed spectrum | |
CN108615229B (en) | Collision detection optimization method based on curvature point clustering and decision tree | |
CN109815357A (en) | A kind of remote sensing image retrieval method based on Nonlinear Dimension Reduction and rarefaction representation | |
CN101980251A (en) | Remote sensing classification method for binary tree multi-category support vector machines | |
CN107833224A (en) | A kind of image partition method based on multi-level region synthesis | |
Xiao et al. | Citrus greening disease recognition algorithm based on classification network using TRL-GAN | |
CN107992856A (en) | High score remote sensing building effects detection method under City scenarios | |
CN116129262A (en) | Cultivated land suitability evaluation method and system for suitable mechanized transformation | |
Radhika et al. | Ensemble subspace discriminant classification of satellite images | |
Vasudevan et al. | A Hybrid Approach for Plant Disease Detection Using E-GAN and CapsNet. | |
CN103164487B (en) | A kind of data clustering method based on density and geological information |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180928 |