CN104657995B - Utilize the Remote Sensing Image Segmentation of domain decomposition method - Google Patents

Abstract

The invention discloses a kind of Remote Sensing Image Segmentation utilizing domain decomposition method, first utilize fractional spins to carry out initialize partition to image, realize the compartmentalization of image; To the picture construction Region adjacency graph of compartmentalization; Based on Markov field modeling k average initial markers carried out to Region adjacency graph; Enter iterative part, the global iterative weights based on Markov model carry out the optimization of Gibbs sampling designations and prime area merging, simultaneously with the record of binary tree form protection merging process to the image of initial markers; Subsequently regional split is carried out to initial segmentation image, turn back to prime areaization configuration; According to the subjective scales in the nodes of Region binary-tree structure and scene, there is positive correlation, adjust the yardstick weights in each regional space context model adaptively, feasible region update mark, obtain final segmentation result.The present invention can the good adaptive updates mark of complex scene in the impact of stress release treatment and image.

Description

Utilize the Remote Sensing Image Segmentation of domain decomposition method

Technical field

The present invention relates to Remote Sensing Image Segmentation field, specifically a kind of Remote Sensing Image Segmentation utilizing domain decomposition method.

Background technology

The important component part that Iamge Segmentation is applied as image automatic interpretation, wherein, the operational use of the automatic interpretation of remote sensing image is significant in marine navigation and climatic study.It is high that remote sensing image has sharpness, and the characteristic such as contain much information, and wherein synthetic-aperture radar also has round-the-clock, round-the-clock characteristic, has nowadays been widely applied to environmental monitoring and the every field such as military.But due in the imaging process of remote sensing image, the impact of all factors can be subject to, the characteristic (wherein incident angle and coherent speckle noise affect larger) of such as imaging system, environmental factor (the wherein difference of season, weather, the Different Effects of wind field is larger) also have characteristic (the wherein surfaceness of scene self, body structure impact is larger) etc., so the inherent Discrete Field of true remote sensing image has non-stationary, especially the remote sensing image of complex scene structure, does not have overall stationarity.Therefore the impact of these factors must be considered for the segmentation of remote sensing image.

Existing image segmentation algorithm is divided into supervision and non-supervisory two large classes, but for remote sensing image, because by the training restriction of sample size and the non-stationary of imaging, usually all adopt non-supervisory method.The seventies and eighties in last century, just have Otsu and Jain etc. merely to utilize the dividing method of feature as histogram thresholding or clustering technique, but because be subject to the impact of coherent speckle noise in remote sensing image, these methods often can produce serious result of making an uproar.Afterwards, scholars gradually utilized in algorithm again and had gone up space language ambience information, wherein had the method based on region split and merge that the people such as Haris in 1998 propose, and the method can integrated multiple Regional Characteristics, reduces the impact of noise.But be difficult to find a stop criterion accurately, often there will be over-segmentation or less divided phenomenon is serious.Within 2003 afterwards, Nguyen proposes the method based on edge, and it can solve the non-stationary of image preferably, preferably the local message of Description Image, but effectively can not generate the significant result of the overall situation.

In the last few years, for the segmentation of remote sensing image, the method based on model obtains development preferably and application, and wherein the most frequently used model is MRF (markov random file).MRF method can to add up the essence that optimum way considers remote sensing image speckle noise, simultaneously for regularization provides effective space context model.Traditional MRF method make use of characteristic model and space context model simultaneously, but its model parameter is the overall situation to be estimated.This is limited for the process with non-stationary remote sensing image.For this reason, the people (2008) such as Deng (2005) and Yu, reasonably make use of edge strength, improve the adaptability that space context model is non-stationary to remote sensing images in traditional MRF model.But in the face of the remote sensing images of complex scene, this method does not consider the Scale Dependency of scene, is all controlled by global parameter to the scene of all yardsticks.And the one compromise that global parameter is all yardstick scenes is considered, the segmentation for regional area is all coarse.Therefore, in the Remote Sensing Image Segmentation of complex scene, the local scale weights adapted with complex scene must be added.

Summary of the invention

The object of this invention is to provide a kind of Remote Sensing Image Segmentation utilizing domain decomposition method, to solve the weak point existing for prior art, according to the yardstick of zones of different object in scene, yardstick weights in the context model of self-adaptative adjustment corresponding space, take into full account the impact of different scale region on space context model, thus improve the segmentation precision of complex scene remote sensing image.

In order to achieve the above object, the technical solution adopted in the present invention is:

Utilize the Remote Sensing Image Segmentation of domain decomposition method, detailed process is as follows:

(1), remote sensing image over-segmentation and Region adjacency graph represent: carry out watershed divide over-segmentation to the remote sensing image inputted, produce many zonules, each region has relatively consistent back scattering value; Each region R is by one group of position S _rform, these positions belong to this region; The eigenvector { Y of each position _s| s ∈ S _raverage to an eigenvector Y _r, and with specific data structure Region adjacency graph RAG Description Image, the node of adjacent map is made up of one group of region R, and its limit represents the boundary position between every a pair neighboring region;

(2), the mark optimization of Region adjacency graph and the region merging technique based on binary tree structure: k mean cluster initial markers is carried out to the image of compartmentalization, utilizes the mark x of formula (1) zoning _r, and mark x _rgive each watershed region R obtained, formula (1) is as follows:

$x_R=\arg \underset{i}{\min }\underset{s\∈S_R}{\mathrm{\Σ}}{(Y_s-{\mathrm{\μ}}_i)}^T(Y_s-{\mathrm{\μ}}_i)---\left(1\right);$

In formula (1), Y _sbe S place, position eigenvector, its element is the value of image channel, and T is transposed operator here, and this process is iteration, from Stochastic Mean-Value μ _istart, each iteration all upgrades;

Next, algorithm enters iterative part, first carries out Gibbs sampling and carries out mark optimization, and according to the global minima cost function of computing formula (2), find out optimum mark, formula (2) is as follows:

E _min＝argminE _f+E _s(2)；

In formula (2): $E_f=\frac{1}{2c}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{S_R\⊆R_i}{\mathrm{\Σ}}\underset{s\∈S_R}{\mathrm{\Σ}}\log \left(\right|{\mathrm{\Σ}}_i\left|\right)+{(y_s-{\mathrm{\μ}}_i)}^T{\mathrm{\Σ}}_i^{-1}(y_s-{\mathrm{\μ}}_i)---\left(3\right),$

In formula (2): $E_s=\mathrm{\α}\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{n}{\mathrm{\Σ}}}\underset{S\∈\∂R_i\∩\∂R_j}{\mathrm{\Σ}}g\left({\▿}_S\right)---\left(4\right),$

In formula (2)-formula (4), E _fcharacteristic model energy, E _sbe space context model energy, C is the port number in image, and n is classification number, S _rbe the position in each watershed region R, the latter is region R _ia part, R _ibe assigned class i, ∑ _ithe covariance matrix of class i, μ _ibe the average of class i, α is positive weights value, border punishment;

Next, carry out region merging technique to the compartmentalization image after optimum mark in greedy mode, merging front and after merging energy difference, merging all negative minimum Δ E according to formula (5) to the region of all same tag to calculating _minregion pair, and with the configuration that binary tree structure record merges at every turn, formula (5) is as follows:

△E _min＝E _After-E _Before(5)，

Wherein, E _afterenergy after merging, E _beforeenergy before merging.

(3), based on the regional split of binary tree structure and the adaptive updates of hierarchy: return original configuration by binary tree structure and carry out regional split, adaptively cost function is upgraded according to formula (6) to the region of each yardstick, finally carry out renewal to split image by step (2) to merge, formula (6) is as follows:

$E_{\min }=\arg \min (E_f+{\hat{E}}_s)---\left(6\right),$

In formula (6):

${\hat{E}}_s=\underset{i=1}{\overset{m-1}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})\underset{S\∈\∂R_i\∩\∂R_j}{\mathrm{\Σ}}g\left({\▿}_S\right)---\left(7\right),$

In formula (7):

$({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})=\frac{N_i+N_j-2N_{\min }}{N_{\max }-N_{\min }}+d---\left(8\right),$

Wherein m represents number of regions, N _iexpressive notation is the binary tree nodes that the section object of i is corresponding, N _jexpressive notation is the binary tree nodes that the section object of j is corresponding, N _maxrepresent the maximum node number in binary tree set, N _minrepresent the minimum node number in binary tree set.

Compared with prior art, beneficial effect of the present invention is embodied in:

1, the present invention is before image tagged, utilizes watershed divide over-segmentation that image is carried out compartmentalization expression, utilizes this dividing method based on region, significantly reduces the impact of speckle noise, and improve a lot compared with the method based on pixel in counting yield.

2, have employed binary tree structure when the optimization mark of the present invention in region and region merging technique to have recorded to merge and configure, for follow-up regional split provides configuration information, save assessing the cost of regional split, and utilize binary tree nodes to be linked together cleverly by the yardstick weights in complexity and space context model, compared with the scheme based on many, improve the efficiency of algorithm.

3, the present invention utilizes the positive correlation that the subjective scales in binary tree nodes and scene has; via linear transformation, nodes is converted to the local scale factor, improves space context model; while the level and smooth large-scale structure of protection, Mesoscale and microscale structure obtains further refinement.Compared with existing single markov random file method, improve the segmentation precision of remote sensing image, obtain better visual segments effect.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of the inventive method.

Fig. 2 is the detail flowchart of the inventive method.

The original image that Fig. 3 (a) is remote sensing image.

The ground truth figure that Fig. 3 (b) is original image.

Fig. 3 (c) is for utilizing the segmentation result of traditional MRF method.

The segmentation result that Fig. 3 (d) is the inventive method.

Embodiment

Utilize the Remote Sensing Image Segmentation of domain decomposition method, comprise the following steps:

(1), remote sensing image over-segmentation and Region adjacency graph represent: carry out watershed divide over-segmentation to the remote sensing image inputted, produce many zonules, each region has relatively consistent back scattering value; Each region R is by one group of position S _rform, these positions belong to this region; The eigenvector { Y of each position _s| s ∈ S _raverage to an eigenvector Y _r, and with specific data structure Region adjacency graph RAG Description Image, the node of adjacent map is made up of one group of region R, and its limit represents the boundary position between every a pair neighboring region;

(2), the mark optimization of Region adjacency graph and the region merging technique based on binary tree structure: k mean cluster initial markers is carried out to the image of compartmentalization, utilizes the mark x of formula (1) zoning _r, and mark x _rgive each watershed region R obtained, formula (1) is as follows:

$x_R=\arg \underset{i}{\min }\underset{s\∈S_R}{\mathrm{\Σ}}{(Y_s-{\mathrm{\μ}}_i)}^T(Y_s-{\mathrm{\μ}}_i)---\left(1\right);$

In formula (1), Y _sbe S place, position eigenvector, its element is the value of image channel, and T is transposed operator here, and this process is iteration, from Stochastic Mean-Value μ _istart, each iteration all upgrades;

Next, algorithm enters iterative part, first carries out Gibbs sampling and carries out mark optimization, and according to the global minima cost function of computing formula (2), find out optimum mark, formula (2) is as follows:

E _min＝argminE _f+E _s(2)；

In formula (2): $E_f=\frac{1}{2c}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{S_R\⊆R_i}{\mathrm{\Σ}}\underset{s\∈S_R}{\mathrm{\Σ}}\log \left(\right|{\mathrm{\Σ}}_i\left|\right)+{(y_s-{\mathrm{\μ}}_i)}^T{\mathrm{\Σ}}_i^{-1}(y_s-{\mathrm{\μ}}_i)---\left(3\right),$

In formula (2): $E_s=\mathrm{\α}\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{n}{\mathrm{\Σ}}}\underset{S\∈\∂R_i\∩\∂R_j}{\mathrm{\Σ}}g\left({\▿}_S\right)---\left(4\right),$

In formula (2)-formula (4), E _fcharacteristic model energy, E _sbe space context model energy, C is the port number in image, and n is classification number, S _rbe the position in each watershed region R, the latter is region R _ia part, R _ibe assigned class i, ∑ _ithe covariance matrix of class i, μ _ibe the average of class i, α is positive weights value, border punishment;

Next, carry out region merging technique to the compartmentalization image after optimum mark in greedy mode, merging front and after merging energy difference, merging all negative minimum Δ E according to formula (5) to the region of all same tag to calculating _minregion pair, and with the configuration that binary tree structure record merges at every turn, formula (5) is as follows:

△E _min＝E _After-E _Before(5)，

Wherein, E _afterenergy after merging, E _beforeenergy before merging.

(3), based on the regional split of binary tree structure and the adaptive updates of hierarchy: return original configuration by binary tree structure and carry out regional split, adaptively cost function is upgraded according to formula (6) to the region of each yardstick, finally carry out renewal to split image by step b to merge, formula (6) is as follows:

$E_{\min }=\arg \min (E_f+{\hat{E}}_s)---\left(6\right),$

In formula (6):

${\hat{E}}_s=\underset{i=1}{\overset{m-1}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})\underset{S\∈\∂R_i\∩\∂R_j}{\mathrm{\Σ}}g\left({\▿}_S\right)---\left(7\right),$

In formula (7):

$({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})=\frac{N_i+N_j-2N_{\min }}{N_{\max }-N_{\min }}+d---\left(8\right),$

Wherein m represents number of regions, N _iexpressive notation is the binary tree nodes that the section object of i is corresponding, N _jexpressive notation is the binary tree nodes that the section object of j is corresponding, N _maxrepresent the maximum node number in binary tree set, N _minrepresent the minimum node number in binary tree set.

Specific embodiment:

The present embodiment Applied Digital image processing techniques carries out Accurate Segmentation to the remote sensing image of complex scene.Its invention process is as follows:

Reference Fig. 1 first carries out remote sensing image over-segmentation and Region adjacency graph represents.Carry out watershed divide over-segmentation with reference to the remote sensing image of Fig. 2 to input, produce many zonules, each region has relatively consistent back scattering value.Each region R is by one group of position S _rform, these positions belong to this region.The eigenvector { Y of each position _s| s ∈ S _raverage to an eigenvector Y _r.And with specific data structure Region adjacency graph (RAG) Description Image, the node of adjacent map is made up of one group of region R, and its limit represents the boundary position between every a pair neighboring region.

The mark optimization of Region adjacency graph and the region merging technique based on binary tree structure is carried out with reference to Fig. 1.With reference to Fig. 2, k mean cluster initial markers is carried out to the image of compartmentalization, utilize the mark x of formula (1) zoning _r, and mark x _rgive each watershed region R obtained; Next, algorithm enters iterative part, first carries out Gibbs sampling and carries out mark optimization, according to the global minima cost function of computing formula (2), finds out optimum mark; Next, carry out region merging technique to the compartmentalization image after optimum mark in greedy mode, merging front and after merging energy difference, merging all negative minimum Δ E according to formula (5) to the region of all same tag to calculating _minregion pair.And with the configuration that binary tree structure record merges at every turn.

$x_R=\arg \underset{i}{\min }\underset{s\∈S_R}{\mathrm{\Σ}}{(Y_s-{\mathrm{\μ}}_i)}^T(Y_s-{\mathrm{\μ}}_i)---\left(1\right);$

Wherein Y _sbe S place, position eigenvector, its element is the value of image channel, and T is transposed operator here.This process is iteration, from Stochastic Mean-Value μ _istart, each iteration all upgrades.

E _min＝argminE _f+E _s(2)，

In formula (2): $E_f=\frac{1}{2c}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{S_R\⊆R_i}{\mathrm{\Σ}}\underset{s\∈S_R}{\mathrm{\Σ}}\log \left(\right|{\mathrm{\Σ}}_i\left|\right)+{(y_s-{\mathrm{\μ}}_i)}^T{\mathrm{\Σ}}_i^{-1}(y_s-{\mathrm{\μ}}_i)---\left(3\right),$

In formula (2): $E_s=\mathrm{\α}\underset{i=1}{\overset{n-1}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{n}{\mathrm{\Σ}}}\underset{S\∈\∂R_i\∩\∂R_j}{\mathrm{\Σ}}g\left({\▿}_S\right)---\left(4\right),$

Wherein, E _fcharacteristic model energy, E _sit is space context model energy.C is the port number in image, and n is classification number, S _rbe the position in each watershed region R, the latter is region R _ia part, R _ibe assigned class i, ∑ _ithe covariance matrix of class i, μ _iit is the average of class i.α is positive weights value, border punishment.

△E _min＝E _After-E _Before(5)，

Wherein, E _afterenergy after merging, E _beforeenergy before merging.

With reference to Fig. 1, the image after initial segmentation is carried out to the adaptive updates of regional split based on binary tree structure and hierarchy.Return original configuration with reference to Fig. 2 by binary tree structure and carry out regional split, adaptively cost function is upgraded according to formula (6) to the region of each yardstick, finally by step b, renewal is carried out to split image and merge.

$E_{\min }=\arg \min (E_f+{\hat{E}}_s)---\left(6\right),$

In formula (6):

${\hat{E}}_s=\underset{i=1}{\overset{m-1}{\mathrm{\Σ}}}\underset{j=i+1}{\overset{m}{\mathrm{\Σ}}}({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})\underset{S\∈\∂R_i\∩\∂R_j}{\mathrm{\Σ}}g\left({\▿}_S\right)---\left(7\right),$

In formula (7): $({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})=\frac{N_i+N_j-2N_{\min }}{N_{\max }-N_{\min }}+d---\left(8\right),$

Wherein m represents number of regions, N _iexpressive notation is the binary tree nodes that the section object of i is corresponding, N _jexpressive notation is the binary tree nodes that the section object of j is corresponding.N _maxrepresent the maximum node number in binary tree set, N _minrepresent the minimum node number in binary tree set.

First image is carried out watershed divide over-segmentation in the present embodiment, by image-region, to reduce the impact of coherent speckle noise.In the present invention, controlled the degree of over-segmentation by the variance adjusting Gaussian filter, in the present embodiment, the variance of Gaussian filter is set to 1.0, obtains the compartmentalization image of height over-segmentation.

Due to the remote sensing image of complex scene, the region affecting different scale of space context model has different effects.So in the present embodiment, utilize the configuration that binary tree structure storage area merges, carry out regional split fast.And utilize the positive correlation that the subjective scales in binary tree nodes and scene has; the yardstick weights of Serial regulation space context model; nodes is converted to the local scale factor; improve space context model; while the level and smooth large-scale structure of protection, Mesoscale and microscale structure obtains further refinement.The value of the b in the present embodiment is set to 1, to obtain best segmentation result.

See figures.1.and.2 carry out image all process after, at the final mark of the last preservation image of program, export final segmentation result figure, as shown in Fig. 3 (d).

As shown in Figure 3, method of the present invention is to the segmentation result of remote sensing image, and while the level and smooth large-scale structure of protection, Mesoscale and microscale structure obtains further refinement.Compared with existing single markov random file method, improve the segmentation precision of remote sensing image, obtain better visual segments effect.

Claims (1)

1. utilize the Remote Sensing Image Segmentation of domain decomposition method, it is characterized in that: comprise the following steps:

(1), remote sensing image over-segmentation and Region adjacency graph represent: carry out watershed divide over-segmentation to the remote sensing image inputted, produce many zonules, each region has relatively consistent back scattering value; Each region R is by one group of position S _rform, these positions belong to this region; The eigenvector { Y of each position _s| s ∈ S _raverage to an eigenvector Y _r, and with specific data structure Region adjacency graph RAG Description Image, the node of adjacent map is made up of one group of region R, and its limit represents the boundary position between every a pair neighboring region;

(2), the mark optimization of Region adjacency graph and the region merging technique based on binary tree structure: k mean cluster initial markers is carried out to the image of compartmentalization, utilizes the mark x of formula (1) zoning _r, and mark x _rgive each watershed region R obtained, formula (1) is as follows:

$x_R={\mathrm{argmin}}_i\underset{s\∈S_R}{\Σ}{(Y_s-{\mathrm{\μ}}_i)}^T(Y_s-{\mathrm{\μ}}_i)---\left(1\right):$

In formula (1), Y _sbe S place, position eigenvector, its element is the value of image channel, and T is transposed operator here, and this process is iteration, from the average μ of random class i _istart, each iteration all upgrades;

Next, algorithm enters iterative part, first carries out Gibbs sampling and carries out mark optimization, and according to the global minima cost function of computing formula (2), find out optimum mark, formula (2) is as follows:

E _min＝argminE _f+E _s(2)；

In formula (2): $E_f=\frac{1}{2c}\underset{i=1}{\overset{n}{\Σ}}\underset{S_R\⊆R_i}{\Σ}\underset{s\∈S_R}{\Σ}log\left(\right|{\mathrm{\Σ}}_i\left|\right)+{(y_s-{\mathrm{\μ}}_i)}^T{\mathrm{\Σ}}_i^{-1}(y_s-{\mathrm{\μ}}_i)---\left(3\right),$

In formula (2): $E_s=\mathrm{\α}\underset{i=1}{\overset{n-1}{\Σ}}\underset{j=i+1}{\overset{n}{\Σ}}\underset{S\∈\∂R_j\∩\∂R_j}{\Σ}g\left({\▿}_S\right)---\left(4\right),$

In formula (2)-formula (4), E _fcharacteristic model energy, E _sbe space context model energy, C is the port number in image, and n is classification number, S _rbe the position in each watershed region R, the latter is region R _ia part, R _ibe assigned class i, Σ _ithe covariance matrix of class i, μ _ibe the average of class i, α is positive weights value, g (▽ _s) be border punishment;

Next, carry out region merging technique to the compartmentalization image after optimum mark in greedy mode, merging front and after merging energy difference, merging all negative minimum Δ E according to formula (5) to the region of all same tag to calculating _minregion pair, and with the configuration that binary tree structure record merges at every turn, formula (5) is as follows:

ΔE _min＝E _After-E _Before(5)，

Wherein, E _afterenergy after merging, E _beforeenergy before merging.

(3), based on the regional split of binary tree structure and the adaptive updates of hierarchy: return original configuration by binary tree structure and carry out regional split, adaptively cost function is upgraded according to formula (6) to the region of each yardstick, finally carry out renewal to split image by step (2) to merge, formula (6) is as follows:

$E_{\min }=\arg min(E_f+{\hat{E}}_s)---\left(6\right),$

In formula (6):

${\hat{E}}_s=\underset{i=1}{\overset{m-1}{\Σ}}\underset{j=i+1}{\overset{m}{\Σ}}({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})\underset{S\∈\∂R_i\∩\∂R_j}{\Σ}g\left({\▿}_S\right)---\left(7\right),$

In formula (7):

$({\mathrm{\β}}_{R_i}+{\mathrm{\β}}_{R_j})=\frac{N_i+N_j-2N_{min}}{N_{max}-N_{min}}+d---\left(8\right),$

Wherein m represents number of regions, N _iexpressive notation is the binary tree nodes that the section object of i is corresponding, N _jexpressive notation is the binary tree nodes that the section object of j is corresponding, N _maxrepresent the maximum node number in binary tree set, N _minrepresent the minimum node number in binary tree set.