CN102289797B - Method for hierarchically segmenting image based on Bhattacharyya distance - Google Patents

Abstract

The invention relates to a method for hierarchically segmenting an image based on a Bhattacharyya distance. In the process of hierarchical segmentation of an input image, the Bhattacharyya distance is adopted to measure the similarity of nodes. The method comprises the following steps of: measuring the similarity of the adjacent nodes of the current layer; selecting father nodes from the current layer; calculating the degree of joint between child nodes and the father nodes; estimating the characteristic distribution of the father nodes; searching the adjacent nodes in the layer for each father node; repeatedly performing the steps until the adjacent nodes are not found for all nodes of the current layer; and determining an area to which each node of the first layer belongs. For the Bhattacharyya distance, two nodes are regarded as two probability distributions, so that the similarity of the two nodes can be effectively measured; and because the Bhattacharyya distance has invariance in approximation scale in a hierarchical accumulation process, a fixed threshold value can be used for judging whether the two nodes are the adjacent nodes in the selection process of the adjacent nodes, so that the capacity of self-determining segmentation number and anti-interference capacity are improved.

Description

A kind of hierarchical image partition method based on Pasteur's distance

Technical field

The present invention relates to a kind of method of image segmentation, relate to specifically a kind of without the supervision formula, based on the hierarchical image partition method of Pasteur distance (Bhattacharyya Distance).

Background technology

Refer to according to the similarity degree of each pixel on feature in the image image is divided into significant several regional processes without supervision formula image segmentation.Specifically, exactly similar pixel is classified as same zone, thereby finally piece image is divided into the process in similar zone, a plurality of inside.Has very large realistic meaning without supervision formula image segmentation, because most of Visual intelligent disposal systems of using in worker, agricultural production every field all need image segmentation to process, such as target identification, intelligent medical diagnosis, intelligence mapping, intelligent remote sensing, multimedia, man-machine interaction, robot system etc.

Existing nothing supervision formula image partition method can be divided into Boundary Detection and two classifications of Region Segmentation.But because the precision of method under complex environment of Boundary Detection is lower, so Region Segmentation is the nothing supervision formula image partition method of present main flow.

Region segmentation method mainly can be divided into following three classes:

1, the method for global energy optimization.Common optimisation technique is the in twos similarity between the pixel of basis in these class methods, makes up similarity matrix, obtains optimum solution by the proper vector of finding the solution this matrix.But these class methods have two limitations: (a) the very large calculated amount of energy-optimised process need.If contain n pixel in the piece image, then this similarity matrix size is n ²* n ², the calculated amount of as seen finding the solution the proper vector of this matrix is huge.(b) in the situation of cutting apart number the unknown, segmentation precision will sharply reduce.But almost in having or not supervision formula image segmentation task, it all is unknown in advance cutting apart number.

2, regional growing method.These class methods at first (at random) choose some seeds, then round each seed, progressively merge the pixel around the seed.Although these class methods have overcome the large limitation of global energy optimization method calculated amount, it also has three limitations: (a) in the situation of cutting apart in advance the unknown of number, segmentation precision will reduce.(b) segmentation precision depends critically upon choosing of seed.If seed is chosen improper, for example a plurality of seeds have dropped in the same similar area, will reduce segmentation precision.(c) in the situation of the interference such as noise, segmentation precision can reduce.

3, the zone merges the method for division.These class methods can represent that with the method for multilayer figure the similar node of merging current layer of hierarchical perhaps divides inner dissimilar node.This class methods calculated amount is moderate.But these class methods still have following limitation: (a) in the hierarchical accumulation, each node is representing the accumulation of a plurality of bottom nodes, therefore how effectively computable between them the problem of similarity still do not solve.(b) in the situation of cutting apart number the unknown, how to determine independently in cutting procedure effectively that cutting apart number does not still solve.(c) limitation of calculating owing to similarity, in the situation of the interference such as noise, segmentation precision can reduce.

Summary of the invention

The present invention overcomes the weak point of existing all kinds of image region segmentation methods, and a kind of hierarchical image partition method based on Pasteur's distance is provided, and number is unknown and have in the situation such as noise cutting apart, and can guarantee the segmentation precision of image, and calculated amount is little.

In order to reach above purpose, the technical solution adopted in the present invention is:

A kind of hierarchical image partition method based on Pasteur's distance constantly synthesizes in the hierarchical image segmentation process of accumulation at node, and the Pasteur of probability of use distance is measured the similarity between the node, and its step is as follows:

(1) initialization; Input picture as ground floor, is made up the figure of a regularization;

(2) use the similarity between the neighborhood of nodes in Pasteur's range observation current layer; Because each node is representing the accumulation of a plurality of sublayers node, the eigenwert of each node is regarded as a gaussian probability distributes, use Pasteur's range formula to calculate two distances between the gaussian probabilities distribution, obtain the similarity between interior two neighborhood of nodes of current layer;

(3) from the current layer node, choose one group of node, be used for making up father's layer; Current layer is the sublayer, and the current layer node is child node, and the node of choosing is father node;

(4) calculate in the current layer connection degree between the father node in the child node and father's layer; Calculative strategy is: the connection degree of father node and child node with they between be directly proportional in the similarity of current layer, simultaneously in father's layer in each father node and all current layers connection degree sum of child node be 1;

(5) feature of all father nodes distributes in the estimation father layer; Use the connection degree as weight, the feature of all child nodes that weighted accumulation is adjacent with this father node distributes, thereby the feature that obtains father node distributes.

(6) in father's layer, seek neighborhood of nodes for each father node;

(7) repeating step (2)～(6) until all nodes of father's layer all can not find neighborhood of nodes, have namely obtained all independently zones, and each node of father's layer has namely represented a zone;

(8) based on the connection degree of child node and father node between layers, determine which zone each node of ground floor belongs to, and finally finishes the image segmentation task.

Described hierarchical image partition method based on Pasteur's distance, in the step (3), the Selection Strategy of father node is: for each child node, all adjacent child nodes are compared with it, if this child node has maximum similarity, it namely is selected as father node, and each father node is used for representing all child nodes adjacent with it.

Described hierarchical image partition method based on Pasteur's distance, in the step (6), the process of seeking neighborhood of nodes for each father node in father's layer is as follows: for each father node, at first in the spatial neighbor zone, choose one group of candidate's neighborhood of nodes, then use the similarity between this father node of Pasteur's range observation and the candidate's neighborhood of nodes; If similarity is greater than threshold value, this candidate's node is the neighborhood of nodes of this father node; If a father node does not find any neighborhood of nodes, this father node has namely represented a new zone.

Described hierarchical image partition method based on Pasteur's distance, in the step (1), as ground floor, i.e. l=1 makes up the figure of a regularization with input picture; Each pixel is a node, for each node, and 8 neighborhood of nodes that node is this center node that the space is adjacent;

Step (2) in current layer l, is used the similarity between Pasteur range observation neighborhood of nodes i and the j

Use Pasteur's range formula to calculate two distances between the gaussian probabilities distribution, obtain the similarity between interior two neighborhood of nodes of current layer, computing formula is: $e_{\mathrm{ij}}^l=\exp \{-\frac{1}{8}{({\mathrm{\μ}}_i^l-{\mathrm{\μ}}_j^l)}^T{\mathrm{\Σ}}_{\mathrm{ij}}^l({\mathrm{\μ}}_i^l-{\mathrm{\μ}}_j^l)-\frac{1}{2}\ln [\frac{\det \left({\mathrm{\Σ}}_{\mathrm{ij}}^l\right)}{\det \left({\mathrm{\Σ}}_i^l\right)\det \left({\mathrm{\Σ}}_j^l\right)}\left]\right\},$ Wherein

With

Represent respectively the average of the feature distribution of two neighborhood of nodes,

With

Represent respectively the variance of the feature distribution of two neighborhood of nodes, ${\mathrm{\Σ}}_{\mathrm{ij}}^l=({\mathrm{\Σ}}_i^l+{\mathrm{\Σ}}_j^l)/2.$

Described hierarchical image partition method based on Pasteur's distance, in the step (3), the process of choosing of father node is followed following two rules: (a) two adjacent child nodes can not be selected simultaneously; (b) each child node must will have an adjacent father node at least; This steps flow chart is as follows:

Each child node i is endowed two indexed variable a _iAnd b _i, a _i=true represents that this child node is chosen as father node; a _iIt is selected that=false represents that this child node does not have, b _i=true represents that this node does not have adjacent father node; b _i=false represents that this node has an adjacent father node at least; This step is divided into two sub-steps: first substep is to choose the child node with local extremum, and this step detects each child node once successively, therefrom selects the child node of similarity maximum as father node; It is more sparse to distribute for the child node with local extremum, is difficult to satisfy rule (b), adopts repeatedly each child node of cycle detection of the second sub-steps, select have local time extreme value child node as father node; This substep is until the indexed variable b of all nodes _iStop during=false.

Described hierarchical image partition method based on Pasteur's distance, step (4): the connection degree between the father node in child node and the father's layer in the calculating current layer

The connection degree sum of each father node and all current layer child nodes is 1 in father's layer, namely

Step (5): use the connection degree as weight, the feature of all child nodes that weighted accumulation is adjacent with this father node distributes, thereby the feature that obtains father node distributes.

Useful good effect of the present invention:

1, the present invention is based on the hierarchical image partition method of Pasteur's distance, because Pasteur's distance is regarded two nodes as two probability distribution, therefore can effectively measure the similarity degree between them, thereby improve segmentation precision and antijamming capability.

2, the present invention is based on the hierarchical image partition method of Pasteur's distance, Pasteur's distance has approximate yardstick unchangeability in the hierarchical accumulation, thereby choose in the process in the neighborhood of nodes of step 6, can judge whether neighborhood of nodes of two nodes with a fixing threshold value, thereby improve autonomous decision Segmentation Number purpose ability.

3, the present invention is based on the hierarchical image partition method of Pasteur's distance, neighborhood of nodes choosing method based on Pasteur's distance, realized limiting to determine neighborhood of nodes in conjunction with region limits and similarity, it is upper close but in the connection between the dissimilar node on the feature effectively to cut off the zone, thereby has improved segmentation precision and the autonomous Segmentation Number purpose performance that determines.

Description of drawings

Fig. 1 is the hierarchical image partition method process flow diagram based on Pasteur's distance;

Fig. 2 is for choosing the process flow diagram of one group of father node from the current layer node;

Fig. 3 is that neighborhood of nodes is chosen the process synoptic diagram.

Embodiment

Embodiment one: referring to Fig. 1.The present invention is based on the hierarchical image partition method of Pasteur's distance, constantly synthesize in the hierarchical image segmentation process of accumulation at node, the Pasteur of probability of use distance is measured the similarity between the node, and its step is as follows:

(1) initialization; Input picture as ground floor, is made up the figure of a regularization;

(2) use the similarity between the neighborhood of nodes in Pasteur's range observation current layer; Because each node is representing the accumulation of a plurality of sublayers node, the eigenwert of each node is regarded as a gaussian probability distributes, use Pasteur's range formula to calculate two distances between the gaussian probabilities distribution, obtain the similarity between interior two neighborhood of nodes of current layer;

(3) from the current layer node, choose one group of node, be used for making up father's layer; Current layer is the sublayer, and the current layer node is child node, and the node of choosing is father node; For each child node, all adjacent child nodes are compared with it, if this child node has maximum similarity, it namely is selected as father node, and each father node is used for representing all child nodes adjacent with it.

(4) calculate in the current layer connection degree between the father node in the child node and father's layer; Calculative strategy is: the connection degree of father node and child node with they between be directly proportional in the similarity of current layer, simultaneously in father's layer in each father node and all current layers connection degree sum of child node be 1;

(5) feature of all father nodes distributes in the estimation father layer; Use the connection degree as weight, the feature of all child nodes that weighted accumulation is adjacent with this father node distributes, thereby the feature that obtains father node distributes.

(6) in father's layer, seek neighborhood of nodes for each father node;

(7) repeating step (2)～(6) until all nodes of father's layer all can not find neighborhood of nodes, have namely obtained all independently zones, and each node of father's layer has namely represented a zone;

(8) based on the connection degree of child node and father node between layers, determine which zone each node of ground floor belongs to, and finally finishes the image segmentation task.

Embodiment two: referring to Fig. 1, Fig. 2, Fig. 3.Present embodiment is based on the hierarchical image partition method of Pasteur's distance, and the technical scheme of step (3) is described in further detail.Step (3) is used for making up father's layer for choose one group of node from the current layer node.This steps flow chart is as follows:

Each child node i is endowed two indexed variable a _iAnd b _i, a _i=true represents that this child node is chosen as father node; a _iIt is selected that=false represents that this child node does not have, b _i=true represents that this node does not have adjacent father node; b _i=false represents that this node has an adjacent father node at least; This step is divided into two sub-steps: first substep is to choose the child node with local extremum, and this step detects each child node once successively, therefrom selects the child node of similarity maximum as father node; It is more sparse to distribute for the child node with local extremum, is difficult to satisfy rule (b), adopts repeatedly each child node of cycle detection of the second sub-steps, select have local time extreme value child node as father node; This substep is until the indexed variable b of all nodes _iStop during=false.

Embodiment three: referring to Fig. 1, Fig. 2, Fig. 3.Present embodiment is based on the hierarchical image partition method of Pasteur's distance, and the technical scheme of step (6) is described in further detail.In step (6), the process of seeking neighborhood of nodes for each father node in father's layer is as follows: for each father node, at first in the spatial neighbor zone, choose one group of candidate's neighborhood of nodes, then use the similarity between this father node of Pasteur's range observation and the candidate's neighborhood of nodes; If similarity is greater than threshold value, this candidate's node is the neighborhood of nodes of this father node; If a father node does not find any neighborhood of nodes, this father node has namely represented a new zone.

Embodiment four: present embodiment is described in further detail technical scheme of the present invention in conjunction with Fig. 1～Fig. 3.

The first step: initialization.As ground floor, i.e. l=1 makes up the figure of a regularization with input picture.Each pixel is a node, for each node, and 8 neighborhood of nodes that node is this center node that the space is adjacent.

Second step: in current layer l, use the similarity between Pasteur range observation neighborhood of nodes i and the j

Because each node is representing the accumulation of a plurality of sublayers node, so can be regarded as a gaussian probability, the eigenwert of each node distributes.By using Pasteur's range formula to calculate two distances between the gaussian probabilities distribution, can obtain the similarity between interior two neighborhood of nodes of current layer.Computing formula is: $e_{\mathrm{ij}}^l=\exp \{-\frac{1}{8}{({\mathrm{\μ}}_i^l-{\mathrm{\μ}}_j^l)}^T{\mathrm{\Σ}}_{\mathrm{ij}}^l({\mathrm{\μ}}_i^l-{\mathrm{\μ}}_j^l)-\frac{1}{2}\ln [\frac{\det \left({\mathrm{\Σ}}_{\mathrm{ij}}^l\right)}{\det \left({\mathrm{\Σ}}_i^l\right)\det \left({\mathrm{\Σ}}_j^l\right)}\left]\right\},$ Wherein

With

Represent respectively the average of the feature distribution of two neighborhood of nodes,

With

Represent respectively the variance of the feature distribution of two neighborhood of nodes, ${\mathrm{\Σ}}_{\mathrm{ij}}^l=({\mathrm{\Σ}}_i^l+{\mathrm{\Σ}}_j^l)/2.$

The 3rd step: from the current layer node, choose one group of node, be used for making up father's layer.Current layer is the sublayer, and the current layer node is child node, and the node of choosing is father node.Selection Strategy is: for each child node, all adjacent child nodes are compared with it, if this child node has maximum similarity, it namely is selected as father node.Each father node is used for representing all child nodes adjacent with it.Simultaneously, choose process and also will follow following two rules: (a) two adjacent child nodes can not be selected simultaneously; (b) each child node must will have an adjacent father node at least.This steps flow chart as shown in Figure 2.

In conjunction with Fig. 2, this step is described in detail as follows.Each child node i is endowed two indexed variable a _iAnd b _ia _i=true represents that this child node is chosen as father node; a _iIt is selected that=false represents that this child node does not have.b _i=true represents that this node does not have adjacent father node; b _i=false represents that this node has an adjacent father node at least.This step is divided into two sub-steps.First substep is to choose the child node with local extremum.This step detects each child node once successively, therefrom selects the child node of similarity maximum as father node.Because it is more sparse having that the child node of local extremum distributes, be difficult to satisfy above-mentioned (b) rule, so the second sub-steps each child node of cycle detection repeatedly, select have local time extreme value child node as father node.This substep is until the indexed variable b of all nodes _iStop during=false.

The 4th step: the connection degree between the father node in child node and the father's layer in the calculating current layer

Calculative strategy is: the connection degree of father node and child node with they between be directly proportional in the similarity of current layer, the connection degree sum of each father node and all current layer child nodes is 1 in father's layer simultaneously, namely

The 5th step: the feature of estimating all father nodes in father's layer distributes.This estimation procedure is as follows: use the connection degree as weight, the feature of all child nodes that weighted accumulation is adjacent with this father node distributes, thereby the feature that obtains father node distributes (comprising average and variance), and for example mean value computation is as follows:

The 6th step: for each father node is sought neighborhood of nodes in father's layer.This process is as follows: for each father node, at first choose one group of candidate's neighborhood of nodes in the spatial neighbor zone, then use the similarity between this father node of Pasteur's range observation and the candidate's neighborhood of nodes.If similarity is greater than given threshold value φ _e, this candidate's node is the neighborhood of nodes of this father node.If a father node does not find any neighborhood of nodes, this father node has namely represented a new zone.So this step can independently determine a new zone, thereby realized autonomous decision Segmentation Number purpose function.Fig. 3 has provided the synoptic diagram of this step.

The 7th step: repeating step 2～6, until all nodes of father's layer all can not find neighborhood of nodes, namely obtained all independently zones, each node of father's layer has namely represented a zone.

The 8th step: based on the connection degree of child node and father node between layers, determine which zone each node of ground floor belongs to, and finally finishes the image segmentation task.

Claims (7)

1. hierarchical image partition method based on Pasteur's distance, each pixel in each tomographic image is a node, constantly synthesize in the hierarchical image segmentation process of accumulation at node, the Pasteur of probability of use distance is measured the similarity between the node, and its step is as follows:

(1) initialization; Input picture as ground floor, is made up the figure of a regularization;

(2) use the similarity between the neighborhood of nodes in Pasteur's range observation current layer; Because each node is representing the accumulation of a plurality of sublayers node, the eigenwert of each node is regarded as a gaussian probability distributes, use Pasteur's range formula to calculate two distances between the gaussian probabilities distribution, obtain the similarity between interior two neighborhood of nodes of current layer;

(3) from the current layer node, choose one group of node, be used for making up father's layer; Current layer is the sublayer, and the current layer node is child node, and the node of choosing is father node;

(4) calculate in the current layer connection degree between the father node in the child node and father's layer; Calculative strategy is: the connection degree of father node and child node with they between be directly proportional in the similarity of current layer, simultaneously in father's layer in each father node and all current layers connection degree sum of child node be 1;

(5) feature of all father nodes distributes in the estimation father layer: use the connection degree as weight, the feature of all child nodes that weighted accumulation is adjacent with this father node distributes, thereby the feature that obtains father node distributes;

(6) in father's layer, seek neighborhood of nodes for each father node;

(7) repeating step (2)～(6) until all nodes of father's layer all can not find neighborhood of nodes, have namely obtained all independently zones, and each node of father's layer has namely represented a zone;

(8) based on the connection degree of child node and father node between layers, determine which zone each node of ground floor belongs to, and finally finishes the image segmentation task.

2. the hierarchical image partition method based on Pasteur's distance according to claim 1, it is characterized in that: in the step (3), the Selection Strategy of father node is: for each child node, all adjacent child nodes are compared with it, if this child node has the similarity of the maximum between interior two neighborhood of nodes of current layer, it namely is selected as father node, and each father node is used for representing all child nodes adjacent with it.

3. the hierarchical image partition method based on Pasteur's distance according to claim 1, it is characterized in that: in the step (6), the process of seeking neighborhood of nodes for each father node in father's layer is as follows: for each father node, at first in the spatial neighbor zone, choose one group of candidate's neighborhood of nodes, then use the similarity between this father node of Pasteur's range observation and the candidate's neighborhood of nodes; If similarity is greater than threshold value, this candidate's node is the neighborhood of nodes of this father node; If a father node does not find any neighborhood of nodes, this father node has namely represented a new zone.

4. the hierarchical image partition method based on Pasteur's distance according to claim 2, it is characterized in that: in the step (6), the process of seeking neighborhood of nodes for each father node in father's layer is as follows: for each father node, at first in the spatial neighbor zone, choose one group of candidate's neighborhood of nodes, then use the similarity between this father node of Pasteur's range observation and the candidate's neighborhood of nodes; If similarity is greater than threshold value, this candidate's node is the neighborhood of nodes of this father node; If a father node does not find any neighborhood of nodes, this father node has namely represented a new zone.

5. each described hierarchical image partition method based on Pasteur's distance according to claim 1～4 is characterized in that: in the step (1), as ground floor, i.e. l=1 makes up the figure of a regularization with input picture; Each pixel is a node, for each node, and 8 neighborhood of nodes that node is this node that the space is adjacent;

In the step (2), in current layer l, use the similarity between Pasteur range observation neighborhood of nodes i and the j

Use Pasteur's range formula to calculate two distances between the gaussian probabilities distribution, obtain the similarity between interior two neighborhood of nodes of current layer, computing formula is: $e_{\mathrm{ij}}^l=\exp \{-\frac{1}{8}{({\mathrm{\μ}}_i^l-{\mathrm{\μ}}_j^l)}^T{\mathrm{\Σ}}_{\mathrm{ij}}^l({\mathrm{\μ}}_i^l-{\mathrm{\μ}}_j^l)-\frac{1}{2}\ln [\frac{\det \left({\mathrm{\Σ}}_{\mathrm{ij}}^l\right)}{\det \left({\mathrm{\Σ}}_i^l\right)\det \left({\mathrm{\Σ}}_j^l\right)}\left]\right\},$ Wherein

With

Represent respectively the average of the feature distribution of two neighborhood of nodes,

With

Represent respectively the variance of the feature distribution of two neighborhood of nodes, ${\mathrm{\Σ}}_{\mathrm{ij}}^l=({\mathrm{\Σ}}_i^l+{\mathrm{\Σ}}_j^l)/2.$

6. the hierarchical image partition method based on Pasteur's distance according to claim 5, it is characterized in that: in the step (3), the process of choosing of father node is followed following two rules: (a) two adjacent child nodes can not be selected simultaneously; (b) each child node must will have an adjacent father node at least; This steps flow chart is as follows:

Each child node i is endowed two indexed variable a _iAnd b _i, a _i=true represents that this child node is chosen as father node; a _iIt is selected that=false represents that this child node does not have, b _i=true represents that this node does not have adjacent father node; b _i=false represents that this node has an adjacent father node at least; This step is divided into two sub-steps: first substep is to choose the child node with local extremum, and this step detects each child node once successively, therefrom selects the child node of the similarity maximum between interior two neighborhood of nodes of current layer as father node; It is more sparse to distribute for the child node with local extremum, is difficult to satisfy rule (b), adopts repeatedly each child node of cycle detection of the second sub-steps, select have local time extreme value child node as father node; This substep is until the indexed variable b of all nodes _iStop during=false.

7. the hierarchical image partition method based on Pasteur's distance according to claim 6 is characterized in that: step (4): the connection degree between the father node in child node and the father's layer in the calculating current layer

, the connection degree sum of each father node and all current layer child nodes is 1 in father's layer, namely

Step (5): use the connection degree as weight, the feature of all child nodes that weighted accumulation is adjacent with this father node distributes, thereby the feature that obtains father node distributes.

Images