CN103425981A - Cluster-based block mass extraction device and method - Google Patents

Abstract

The invention discloses a cluster-based block mass extraction device and method. The cluster-based block mass extraction device and method are based on existing image pyramid block mass extraction device and method, an image block mass extraction module and an image block mass merging module are improved, and models are introduced into the two modules. The module is built for image points of a downsampling module in the image block mass extraction module, and the image point gray scale and the image point block mass number are computed from the perspective of probability; the module is built for block masses in the image block mass merging module, the merging of the block masses is converted into the judgment of similarities among a plurality of modules, therefore, logical judgment is simplified, data are classified from the perspective of data distribution, the accuracy of block mass merging is ensured, and a merging process is simplified.

Description

A kind of agglomerate extraction element and method based on cluster

Technical field

The present invention relates to image processing field, in particular to a kind of agglomerate extraction element and method based on cluster.

Background technology

The agglomerate feature of image is exactly in image processing process, and according to connectedness, by pixel value in image, similar picture point is classified as an agglomerate, and this method has been simulated the mankind's visually-perceptible system.Usually, human vision is after obtaining image, and the picture point that brain can be close by color in image deteriorates to a kind of color, so image has just become the zone of several different colours, the mankind pay close attention to a kind of color according to the needs of oneself again.

At present, in target tracking algorism, adopt the agglomerate feature can effectively improve the accuracy of tracking, but to the extracting method of agglomerate feature also proposed algorithm simple, process real-time requirement.For algorithm, require briefly algorithm easy to use, the algorithm too complex can increase the complexity of target following, affects the use of tracking.For processing in real time, require algorithm to calculate Simple fast, can process video flowing, even can sacrifice certain accuracy in case of necessity.

The extracting method of existing agglomerate feature has a lot, it is generally first image to be carried out to pre-service, eliminate noise, then to image according to color classification or cluster, but the classification as a result obtained can be many, and in disorder, therefore further need to be merged these classifications, obtain important severally, finally the remaining mode such as rectangle, ellipse for classification is described.

At present, the agglomerate extracting method based on image pyramid is more typical a kind of agglomerate extracting method, and the method is calculated fairly simple, and result is more satisfactory.As shown in Figure 1, its corresponding agglomerate extraction element comprises following module: image pretreatment module, image agglomerate extraction module, image agglomerate merge module, and image agglomerate fitting module, and its agglomerate extraction step comprises:

The first step, image pretreatment module, for image is carried out to pre-service work, mainly comprise image are carried out to noise restraint and smoothing processing, make agglomerate extract the impact that less is subject to noise.

Second step, image agglomerate extraction module are numbered for each picture point to image, and wherein, described numbering represents the agglomerate under corresponding picture point.The detailed operation step of this image agglomerate extraction module is as follows:

1, the memory headroom size of the hierarchical model of computed image, to distribute enough memory headrooms, and be stored in original image the 0th layer of hierarchical model, and in initialization the 0th tomographic image, the confidence level of each picture point is 1.Wherein, hierarchical model is one group of image, as shown in Figure 2, hierarchical model at the middle and upper levels image width and highly be respectively the lower image width and the height half, the original image of input is at the 0th layer of hierarchical model, by calculating other each tomographic image of hierarchical model;

2, n tomographic image down-sampling is generated to the n+1 tomographic image.The pixel value of each picture point of n+1 tomographic image utilizes hierarchical model picture point computing method to obtain, and the picture point confidence level is set.Wherein, in the down-sampling process, the n tomographic image is reduced into to 1/4 size of n+1 tomographic image, every four picture points mean by a picture point.

Wherein, hierarchical model picture point computing method adopt formula (1-1) to be calculated, be that pixel coordinate is (i, the gray-scale value of pixel j) utilizes formula (1-1) to obtain, as shown in Figure 3, the set of 12 lower floor's pixels that obtain by the sampling template means with R for the upper strata picture point of utilizing 12 neighborhood sampling templates to obtain and the coordinate mapping relations of lower floor's picture point.The computing formula of the gray average of l pixel of n tomographic image is:

${\mathrm{gray}}_n^l=\frac{1}{{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k}{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k{\mathrm{gray}}_{n-1}^k---(1-1)$

ω wherein _kFor weight, importance corresponding to each data in 12 neighborhood sampling templates, can arrange different coefficients according to the use object.T ^kFor rejecting the factor, for example when it is 1, this picture point participates in calculating.R means the pixel confidence level, and its size is the hierarchical model size, and its value, for two-value data, when the confidence level of the some pixels of n tomographic image is 1, means that the gray scale of this pixel can represent the pixel in point set R; Be 0 o'clock, think that this picture point can not replace the pixel in the point set R of its lower floor.

All t when arranging, are set in initialization ^kBe 1, calculate

In rear traversal point set R

If the Euclidean distance of two gray scales is greater than the threshold value Dis of setting, its t is set ^kBe 0, adopt formula (1-1) to calculate once again, then judge whether to need to reject, if so, continue preceding step, if not, finish.Also judgement simultaneously

Value, if this value is greater than the threshold value Ran of a setting, the confidence level that this point is set is 1, otherwise the confidence level that this point is set is 0.

If 3 n+1 images are pyramidal top layer picture point, go to step 4, otherwise n add 1, and go to step 2;

4, the agglomerate that initialization arranges the top layer picture point of hierarchical model is numbered 0, and wherein the agglomerate numbering is the numbering of the agglomerate under picture point, and agglomerate 0 means that the picture point of this agglomerate is to abandon the set of no point, does not participate in follow-up calculating;

5, calculate the agglomerate numbering of each picture point of n tomographic image according to the coordinate relation of n tomographic image and n+1 tomographic image, computing method are picture point agglomerate numbering computing method, and preserve the numbering result;

Wherein, picture point agglomerate numbering computing method are determined according to the similarity of levels picture point, at first find upper strata picture point S set corresponding to lower floor's picture point, obtaining institute's agglomerate a little in set numbers, calculate the Euclidean distance of all agglomerate pixel values and this lower floor's picture point pixel value, one of the chosen distance minimum and with the threshold value Dis set relatively, if be less than setting threshold, this picture point belongs to this agglomerate, otherwise, the agglomerate of this picture point of Credibility judgement according to this picture point is numbered 0 or be a new agglomerate, if confidence level is 0, its agglomerate is numbered 0, otherwise this picture point belongs to a new agglomerate, the total number of agglomerate numbering increases by one.Wherein, the agglomerate pixel value means that first is labeled as the gray-scale value of the picture point of agglomerate, and each agglomerate only has an agglomerate pixel value, only has when the agglomerate numbering increases, and just can increase the agglomerate pixel value newly, and assignment, and change is no longer done in back.Equally, the 0th class, because there is no classification, does not therefore have the agglomerate pixel value.

If 6 n are 0, go to step 7; Otherwise n subtracts 1, go to step 5;

7, the agglomerate numbering result of output original image.

The 3rd step, image agglomerate merge module and are merged for the Output rusults to image agglomerate extraction module, to set up gray distribution model to agglomerate, according to the similarity of each agglomerate gray distribution model and the adjacency on image, two agglomerates be merged.

The 4th step, image agglomerate fitting module are carried out ellipse fitting for the agglomerate result to output, and are exported oval expression formula.

But still there is following problem in this existing agglomerate extracting method:

1, calculating parameter is too many, and described parameter does not have versatility, and generally speaking, for different images, its parameter need to arrange difference, need to, in operating process, also need parameter is carried out to adaptive adjustment to realize result optimizing;

Although in 2 agglomerate merging methods in this agglomerate extracting method, in the time of under its situation comparatively complicated at image, accuracy is higher, but its Consideration is more, calculate more complicated, extracting result for the agglomerate of the image in simple two-value situation has larger deviation.

Summary of the invention

In order to solve the above-mentioned problems in the prior art, the object of the present invention is to provide a kind of agglomerate extraction element and method based on cluster.

In order to reach purpose of the present invention, the present invention realizes by the following technical solutions:

A kind of agglomerate extraction element based on cluster, comprise that image pretreatment module, image agglomerate extraction module, image agglomerate merge module and image agglomerate fitting module, wherein:

The image pretreatment module, carry out the image pre-service for the image to obtaining;

Image agglomerate extraction module, be used in hierarchical model picture point computation process, the adaptive model of a gray scale of setting up according to the gray scale to N picture point on the sampling template set in advance is processed the picture point in present image, when meeting this model, current picture point is retained, otherwise reject, when last numerical convergence, calculate the gray-scale value of upper strata picture point, preserve the adaptive model parameter of gray scale of this upper strata picture point simultaneously; Utilize the retrodict agglomerate numbering of each picture point on the orlop image of hierarchical model, obtain lower floor's picture point after the mapping relations of upper strata picture point, determining the agglomerate numbering of lower floor's picture point according to the probabilistic relation between the adaptive model of the gray scale of lower floor's picture point gray scale and upper strata picture point;

The image agglomerate merges module, for the Output rusults to image agglomerate extraction module, merged, its merging method is: agglomerate is set up to gray distribution model, according to the similarity of each agglomerate gray distribution model and the adjacency on image, two agglomerates are merged;

And image agglomerate fitting module, carry out agglomerate matching output for the Output rusults that the image agglomerate is merged to module.

Under a kind of embodiment preferably, described image agglomerate extraction module is also further numbered in computation process at the picture point agglomerate, is handled as follows:

Obtain all upper stratas picture point S set corresponding to lower image picture point according to down-sampling;

Obtain in the picture point S set of upper strata with the nearest point of the gray scale Euclidean distance of lower image picture point under agglomerate, and determine according to the pixel value of this agglomerate whether this lower image picture point belongs to this agglomerate; If the gray scale of this lower image picture point meets the adaptive model of gray scale, judge that this lower image picture point belongs to this agglomerate, otherwise, assess the confidence level of this lower image picture point and according to assessment result, this agglomerate be numbered, if confidence level is 1, this lower image picture point belongs to a new agglomerate, distributes a new agglomerate to number to this agglomerate, and preserves agglomerate pixel value and the adaptive model parameter of gray scale; If confidence level is 0, think that this lower image picture point can't sort out, its agglomerate numbering is made as to 0.

Under a kind of embodiment preferably, the image agglomerate merges the step that module merged the Output rusults of image agglomerate extraction module and comprises:

Output rusults according to image agglomerate extraction module is set up respectively the blob match model to each agglomerate;

Obtain a pair of adjacent agglomerate and compare their blob match model, if its model is not close, again obtaining a pair of adjacent agglomerate to proceed comparison; If its model approaches, merge this two agglomerates, and upgrade the blob match model parameter of this new agglomerate, again search for afterwards and find a pair of adjacent agglomerate the blob match model that upgrades the marquis according to it to compare; When completeer all adjacent agglomerates and can't continue to merge the time, finish.

Under a kind of embodiment preferably, image agglomerate extraction module adopts following mathematical expression to calculate the gray-scale value of upper strata picture point:

${\mathrm{gray}}_n^l=\frac{1}{{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k}{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k{\mathrm{gray}}_{n-1}^k;$

Wherein, ω _kFor weight, t ^kFor rejecting the factor, r means the pixel confidence level.

Under a kind of embodiment preferably, image agglomerate fitting module is carried out ellipse fitting output for the Output rusults that the image agglomerate is merged to module.

A kind of agglomerate extracting method based on cluster, its agglomerate extraction element based on cluster comprises that image pretreatment module, image agglomerate extraction module, image agglomerate merge module and image agglomerate fitting module, wherein, described method comprises:

The image pretreatment module is carried out pre-service to the image obtained;

In hierarchical model picture point computation process, the adaptive model of a gray scale that image agglomerate extraction module is set up according to the gray scale to N picture point on the sampling template set in advance is processed the picture point in present image, when meeting this model, current picture point is retained, otherwise reject, when last numerical convergence, calculate the gray-scale value of upper strata picture point, preserve the adaptive model parameter of gray scale of this upper strata picture point simultaneously; Utilize the retrodict agglomerate numbering of each picture point on the orlop image of hierarchical model, obtain lower floor's picture point after the mapping relations of upper strata picture point, determining the agglomerate numbering of lower floor's picture point according to the probabilistic relation between the adaptive model of the gray scale of lower floor's picture point gray scale and upper strata picture point;

Image agglomerate merging module is merged the Output rusults of image agglomerate extraction module, its merging method is: agglomerate is set up to gray distribution model, according to the similarity of each agglomerate gray distribution model and the adjacency on image, two agglomerates are merged;

The Output rusults that image agglomerate fitting module merges module to the image agglomerate carries out agglomerate matching output.

Under a kind of embodiment preferably, described image agglomerate extraction module is also further numbered in computation process at the picture point agglomerate, is handled as follows:

Obtain all upper stratas picture point S set corresponding to lower image picture point according to down-sampling;

Obtain in the picture point S set of upper strata with the nearest point of the gray scale Euclidean distance of lower image picture point under agglomerate, and determine according to the pixel value of this agglomerate whether this lower image picture point belongs to this agglomerate; If the gray scale of this lower image picture point meets the adaptive model of gray scale, judge that this lower image picture point belongs to this agglomerate, otherwise, assess the confidence level of this lower image picture point and according to assessment result, this agglomerate be numbered, if confidence level is 1, this lower image picture point belongs to a new agglomerate, distributes a new agglomerate to number to this agglomerate, and preserves agglomerate pixel value and the adaptive model parameter of gray scale; If confidence level is 0, think that this lower image picture point can't sort out, its agglomerate numbering is made as to 0.

Under a kind of embodiment preferably, the image agglomerate merges the step that module merged the Output rusults of image agglomerate extraction module and comprises:

Output rusults according to image agglomerate extraction module is set up respectively the blob match model to each agglomerate;

Obtain a pair of adjacent agglomerate and compare their blob match model, if its model is not close, again obtaining a pair of adjacent agglomerate to proceed comparison; If its model approaches, merge this two agglomerates, and upgrade the blob match model parameter of this new agglomerate, again search for afterwards and find a pair of adjacent agglomerate the blob match model that upgrades the marquis according to it to compare; When completeer all adjacent agglomerates and can't continue to merge the time, finish.

Under a kind of embodiment preferably, image agglomerate extraction module adopts following mathematical expression to calculate the gray-scale value of upper strata picture point:

${\mathrm{gray}}_n^l=\frac{1}{{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k}{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k{\mathrm{gray}}_{n-1}^k;$

Wherein, ω _kFor weight, t ^kFor rejecting the factor, r means the pixel confidence level.

Under a kind of embodiment preferably, image agglomerate fitting module is carried out ellipse fitting output for the Output rusults that the image agglomerate is merged to module.

Compared to prior art, the present invention has following beneficial effect:

1, the speed of merging agglomerate is faster, has guaranteed real-time;

2, view data is set up to model, replace the parameter of direct significance by model parameter, make the agglomerate extracting method have more versatility;

3, due to the number that has reduced parameter, make arranging of whole agglomerate extracting method convenient, for domestic consumer, can be more prone to use.

The accompanying drawing explanation

The realization of the object of the invention, functional characteristics and excellent effect, be described further below in conjunction with specific embodiment and accompanying drawing, wherein,

The structural representation that Fig. 1 is existing agglomerate extraction element;

Fig. 2 is the hierarchical model schematic diagram;

Fig. 3 is 12 neighborhood sampling template schematic diagram;

Lower floor's picture point mapping upper strata picture point schematic diagram that Fig. 4 obtains for adopting 12 neighborhood sampling templates.

Embodiment

Below in conjunction with the drawings and specific embodiments, technical scheme of the present invention is described in further detail, can be implemented so that those skilled in the art can better understand the present invention also, but illustrated embodiment is not as a limitation of the invention.

Agglomerate extracting method provided by the invention improves image agglomerate extraction module and image agglomerate merging module on the thinking of continuing to use the existing agglomerate extraction element structure based on image pyramid, in these two modules, introduces model.In image agglomerate extraction module, the picture point in the down-sampling template is set up to model, from the angle of probability, calculate picture point gray scale and picture point agglomerate numbering.Merge module at the image agglomerate agglomerate is set up to model, convert the merging of agglomerate to judgement to similarity between a plurality of models, thereby simplified the logic judgement, the angle distributed from data is sorted out data, guarantee the accuracy that agglomerate merges, simplified merging process simultaneously.

Agglomerate extracting method provided by the invention mainly comprises following several step:

The first step, image pretreatment module are responsible for the pre-service work of image.Described pre-service can comprise carries out noise restraint and smoothing processing to image, makes agglomerate extract the impact that less is subject to noise.In addition, in pre-treatment step, can also comprise other processing modes, its pretreatment mode not done to restriction here.

Second step, image agglomerate extraction module are numbered for the agglomerate of each picture point of computed image, and wherein, the agglomerate numbering represents the agglomerate under this picture point.Its core methed comprises hierarchical model picture point computing method and picture point agglomerate numbering computing method.Introduce probability model in this image agglomerate extraction module, when guaranteeing result, can simplify computation process.

When down-sampling, a picture point on upper strata is that N the picture point (number of N is determined by the sampling template size) by lower floor calculated by hierarchical model picture point computing method, in this process, right SamplingThe gray scale of the N of a template picture point is set up a model, the picture point that meets model will be retained, otherwise reject, then utilize formula (1-2) Renewal model, same, the picture point that meets model will be retained, otherwise reject ... repeat this process, when last numerical convergence, utilize formula (1-2) can obtain the gray-scale value of upper strata picture point, preserve the model parameter of this picture point, for calculating its agglomerate numbering simultaneously.

${\mathrm{gray}}_n^l=\frac{1}{{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k}{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k{\mathrm{gray}}_{n-1}^k---(1-2)$

Wherein, ω _kFor weight, t ^kFor rejecting the factor, r means the pixel confidence level.

After hierarchical model calculating is complete, in follow-up picture point agglomerate numbering computing method, suppose that the agglomerate of top layer picture point is numbered 0, the agglomerate of lower floor's picture point numbering just can be come definite by the agglomerate numbering of upper strata picture point so.At first obtain all upper stratas picture point S set corresponding to lower floor's picture point according to the down-sampling template.Then find in S set with the nearest point of the gray scale Euclidean distance of this lower floor's picture point under agglomerate, determine according to the agglomerate pixel value whether this lower floor's picture point belongs to this agglomerate.If the gray scale of this lower floor's picture point meets the model of agglomerate pixel value, think that this lower floor's picture point belongs to this agglomerate, otherwise, assess the confidence level of this lower floor's picture point and accordingly agglomerate be numbered, if confidence level is 1, this lower floor's picture point belongs to a new agglomerate, and the agglomerate sum increases by one, preserves agglomerate pixel value and model parameter; If confidence level is 0, think that this lower floor's picture point can't sort out, its agglomerate numbering is made as to 0.

The 3rd step, agglomerate merge module to be responsible for the Output rusults of image agglomerate extraction module is merged, and smaller agglomerate or the relatively more approaching agglomerate that still is divided into two classes of agglomerate gray level model are combined.Equally, introduce model in this module and can not need complexity, discrete data heap, as long as they are set up to model, utilize their data to distribute to be merged, not only simply but also effective, for the less demanding tracking in agglomerate edge, utilizing model to merge is a reasonable disposal route, and its concrete implementation step comprises:

At first utilize the data of agglomerate to set up model to each agglomerate;

Then find a pair of adjacent agglomerate, relatively their model, if these two agglomerate models are not close, pick up a pair of adjacent agglomerate, continues relatively; If the gray distribution model of two agglomerates is more approaching, merge this two agglomerates, and upgrade the model parameter of this new agglomerate, a pair of adjacent agglomerate is found in search again.Still can't merge when having searched all adjacent agglomerates, finish.

The 4th step, image agglomerate fitting module are responsible for the agglomerate result of output is carried out to matching, and the expression formula after the output matching.

Below will be described in more detail spirit of the present invention by a specific embodiment in order better to explain.

Continuation is with reference to Fig. 1, and the agglomerate extraction element that the embodiment of the present invention provides comprises that image pretreatment module, image agglomerate extraction module, image agglomerate merge module and image agglomerate fitting module, and its step of carrying out the agglomerate extraction comprises:

Step 1, image is carried out to pre-service, for example, at the image pretreatment stage, image is carried out to gaussian filtering, eliminate the processing such as noise spot.

Step 2, calculate the size of hierarchical model storage allocation according to the size of input picture.Calculate the memory headroom size of hierarchical model, and distribute enough storage spaces.At first initial pictures is copied in the 0th layer of hierarchical model.Suppose the picture traverse of input and highly be respectively W0 and H0, be i.e. the size of the 0th tomographic image.The width of the 1st tomographic image and highly be respectively [W0/2+1] and [H0/2+1] so, wherein [] means to round downwards, the width of the 2nd tomographic image and be highly [W1/2+1] and [H1/2+1] ..., the size of last layer data is 1.

Fill hierarchical model by down-sampling, with reference to figure 3, and in setting initialization down-sampling template, the weight of 12 elements is 1/12, data is set up to Gauss model simultaneously, according to formula (1-2), calculates hierarchical model.Wherein hierarchical model picture point computing method are as follows.L the picture point for the n layer

Pixel value

Computation process be:

1, the t of 12 picture point of initial n-1 layer ^kBe 1, wherein, r is known;

2, to t ^kThe pixel value that equals 1 picture point is set up Gauss model, according to formula (1-2), calculates the Gaussian distribution center

According to

Standard deviation with the calculated for pixel values Gaussian distribution of these picture points

3, judgement t ^kValue (k=1,2 ..., 12), work as t ^kBe 0 o'clock, k adds 1, if k equals 12, goes to step 5, otherwise, go to step 3.Work as t ^kBe 1 o'clock, go to step 4;

4, calculate

With

Euclidean distance, if this distance be less than

(wherein, β is input parameter, and scope is 0 ~ 4), k adds 1, goes to step 3, otherwise, reject this picture point, i.e. t ^k=0, go to step 2;

5, preserve last Calculating (for example meaning with Dir) for the agglomerate numbering.Then to t ^kSummation, if the result that summation obtains is less than a threshold value set in advance,

Be 0, otherwise

Be 1, go to step 6;

6, finish.

List file names with picture point agglomerate numbering computing method, the computation process of numbering for the agglomerate of s picture point of n-1 layer is as follows:

1, with reference to figure 4, determine picture point

The picture point S set of corresponding n layer;

2, calculation level Pixel value with the picture point S set under the Euclidean distance of agglomerate pixel value of agglomerate, if agglomerate is numbered 0, without compute euclidian distances, afterwards by the threshold value Dir of the shortest value dmin of Euclidean distance that calculates and preservation relatively, if it is less than Dir, picture point

Agglomerate be numbered and there is the corresponding agglomerate of the shortest value dmin of Euclidean distance numbering, go to step 5, otherwise go to step 3;

If 3 picture points

Confidence level be 0, the agglomerate of this picture point is numbered 0, goes to step 5, if confidence level is not 0, goes to step 4;

4, total agglomerate numbering adds 1, and the agglomerate pixel value of this newly-increased agglomerate is just picture point Pixel value, picture point

The agglomerate numbering agglomerate numbering for newly increasing just, go to step 5;

5, finish.

Step 3, agglomerate merge module to be responsible for the Output rusults of image agglomerate extraction module is merged, and smaller agglomerate or the relatively more approaching agglomerate that still is divided into two classes are combined.The calculation procedure of this module is as follows:

1, traversal agglomerate result, set up Gauss model for each agglomerate, calculates the pixel average of each agglomerate, agglomerate pixel variance, agglomerate coordinate center.Judge whether two agglomerates have intersection point, and the number of intersection point, be saved in the common factor array by result simultaneously;

2, traversal common factor array, search immediate two agglomerates of pixel average, and judge whether the Euclidean distance of the pixel average of these two agglomerates is less than β * σ, the maximal value of the standard deviation that wherein σ is two agglomerates.If be greater than, nonjoinder is rejected this from the common factor array, goes to step 2.If be greater than β * σ, merged, go to step 3.If traveled through the common factor array, go to step 4;

If exist and merged the agglomerate more than 2 times in 3 these two agglomerates, cancel current the merging, directly turn 2.Otherwise these two agglomerates are merged into to an agglomerate, change agglomerate numbering, and upgrade the every data in the common factor array, go to step 2;

4, finish to merge, recalculate agglomerate result output.

Step 4, image agglomerate fitting module are responsible for the agglomerate result of output is carried out to process of fitting treatment, and the output fitting result.For example, in the present embodiment, described image agglomerate fitting module is ellipse fitting agglomerate module, and it selects the ellipse fitting method based on least square method commonly used to calculate the ellipse of each agglomerate, and this paper does not do and is described in detail these detailed computing method.

The foregoing is only the preferred embodiments of the present invention; not thereby limit the scope of the claims of the present invention; every equivalent structure or conversion of equivalent flow process that utilizes instructions of the present invention and accompanying drawing content to do; or directly or indirectly be used in other relevant technical fields, all in like manner be included in scope of patent protection of the present invention.

Claims (10)

1. the agglomerate extraction element based on cluster, comprise image pretreatment module, image agglomerate extraction module, image agglomerate merging module and image agglomerate fitting module, it is characterized in that:

The image pretreatment module, carry out the image pre-service for the image to obtaining;

Image agglomerate extraction module, be used in hierarchical model picture point computation process, the adaptive model of a gray scale of setting up according to the gray scale to N picture point on the sampling template set in advance is processed the picture point in present image, when meeting this model, current picture point is retained, otherwise reject, when last numerical convergence, calculate the gray-scale value of upper strata picture point, preserve the adaptive model parameter of gray scale of this upper strata picture point simultaneously; Utilize the retrodict agglomerate numbering of each picture point on the orlop image of hierarchical model, obtain lower floor's picture point after the mapping relations of upper strata picture point, determining the agglomerate numbering of lower floor's picture point according to the probabilistic relation between the adaptive model of the gray scale of lower floor's picture point gray scale and upper strata picture point;

The image agglomerate merges module, for the Output rusults to image agglomerate extraction module, merged, its merging method is: agglomerate is set up to gray distribution model, according to the similarity of each agglomerate gray distribution model and the adjacency on image, two agglomerates are merged;

And image agglomerate fitting module, carry out agglomerate matching output for the Output rusults that the image agglomerate is merged to module.

2. according to the agglomerate extraction element based on cluster claimed in claim 1, it is characterized in that, described image agglomerate extraction module is also further numbered in computation process at the picture point agglomerate, is handled as follows:

Obtain all upper stratas picture point S set corresponding to lower image picture point according to down-sampling;

Obtain in the picture point S set of upper strata with the nearest point of the gray scale Euclidean distance of lower image picture point under agglomerate, and determine according to the pixel value of this agglomerate whether this lower image picture point belongs to this agglomerate; If the gray scale of this lower image picture point meets the adaptive model of gray scale, judge that this lower image picture point belongs to this agglomerate, otherwise, assess the confidence level of this lower image picture point and according to assessment result, this agglomerate be numbered, if confidence level is 1, this lower image picture point belongs to a new agglomerate, distributes a new agglomerate to number to this agglomerate, and preserves agglomerate pixel value and the adaptive model parameter of gray scale; If confidence level is 0, think that this lower image picture point can't sort out, its agglomerate numbering is made as to 0.

3. according to the agglomerate extraction element based on cluster claimed in claim 1, it is characterized in that, the image agglomerate merges the step that module merged the Output rusults of image agglomerate extraction module and comprises:

Output rusults according to image agglomerate extraction module is set up respectively the blob match model to each agglomerate;

Obtain a pair of adjacent agglomerate and compare their blob match model, if its model is not close, again obtaining a pair of adjacent agglomerate to proceed comparison; If its model approaches, merge this two agglomerates, and upgrade the blob match model parameter of this new agglomerate, again search for afterwards and find a pair of adjacent agglomerate the blob match model that upgrades the marquis according to it to compare; When completeer all adjacent agglomerates and can't continue to merge the time, finish.

4. according to the agglomerate extraction element based on cluster claimed in claim 1, it is characterized in that, image agglomerate extraction module adopts following mathematical expression to calculate the gray-scale value of upper strata picture point:

${\mathrm{gray}}_n^l=\frac{1}{{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k}{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k{\mathrm{gray}}_{n-1}^k;$

Wherein, ω _kFor weight, t ^kFor rejecting the factor, r means the pixel confidence level.

5. according to the agglomerate extraction element based on cluster claimed in claim 1, it is characterized in that, image agglomerate fitting module is carried out ellipse fitting output for the Output rusults that the image agglomerate is merged to module.

6. the agglomerate extracting method based on cluster, its agglomerate extraction element based on cluster comprises that image pretreatment module, image agglomerate extraction module, image agglomerate merge module and image agglomerate fitting module, it is characterized in that, described method comprises:

The image pretreatment module is carried out pre-service to the image obtained;

In hierarchical model picture point computation process, the adaptive model of a gray scale that image agglomerate extraction module is set up according to the gray scale to N picture point on the sampling template set in advance is processed the picture point in present image, when meeting this model, current picture point is retained, otherwise reject, when last numerical convergence, calculate the gray-scale value of upper strata picture point, preserve the adaptive model parameter of gray scale of this upper strata picture point simultaneously; Utilize the retrodict agglomerate numbering of each picture point on the orlop image of hierarchical model, obtain lower floor's picture point after the mapping relations of upper strata picture point, determining the agglomerate numbering of lower floor's picture point according to the probabilistic relation between the adaptive model of the gray scale of lower floor's picture point gray scale and upper strata picture point;

Image agglomerate merging module is merged the Output rusults of image agglomerate extraction module, its merging method is: agglomerate is set up to gray distribution model, according to the similarity of each agglomerate gray distribution model and the adjacency on image, two agglomerates are merged;

The Output rusults that image agglomerate fitting module merges module to the image agglomerate carries out agglomerate matching output.

7. the agglomerate extracting method based on cluster as claimed in claim 6, is characterized in that, described image agglomerate extraction module is also further numbered in computation process at the picture point agglomerate, is handled as follows:

Obtain all upper stratas picture point S set corresponding to lower image picture point according to down-sampling;

Obtain in the picture point S set of upper strata with the nearest point of the gray scale Euclidean distance of lower image picture point under agglomerate, and determine according to the pixel value of this agglomerate whether this lower image picture point belongs to this agglomerate; If the gray scale of this lower image picture point meets the adaptive model of gray scale, judge that this lower image picture point belongs to this agglomerate, otherwise, assess the confidence level of this lower image picture point and according to assessment result, this agglomerate be numbered, if confidence level is 1, this lower image picture point belongs to a new agglomerate, distributes a new agglomerate to number to this agglomerate, and preserves agglomerate pixel value and the adaptive model parameter of gray scale; If confidence level is 0, think that this lower image picture point can't sort out, its agglomerate numbering is made as to 0.

8. the agglomerate extracting method based on cluster as claimed in claim 6, is characterized in that, the image agglomerate merges the step that module merged the Output rusults of image agglomerate extraction module and comprises:

Output rusults according to image agglomerate extraction module is set up respectively the blob match model to each agglomerate;

Obtain a pair of adjacent agglomerate and compare their blob match model, if its model is not close, again obtaining a pair of adjacent agglomerate to proceed comparison; If its model approaches, merge this two agglomerates, and upgrade the blob match model parameter of this new agglomerate, again search for afterwards and find a pair of adjacent agglomerate the blob match model that upgrades the marquis according to it to compare; When completeer all adjacent agglomerates and can't continue to merge the time, finish.

9. the agglomerate extracting method based on cluster as claimed in claim 6, is characterized in that, image agglomerate extraction module adopts following mathematical expression to calculate the gray-scale value of upper strata picture point:

${\mathrm{gray}}_n^l=\frac{1}{{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k}{\∑}_{R\∈k}{\mathrm{\ω}}_k{t}_{n-1}^k{r}_{n-1}^k{\mathrm{gray}}_{n-1}^k;$

Wherein, ω _kFor weight, t ^kFor rejecting the factor, r means the pixel confidence level.

10. the agglomerate extracting method based on cluster as claimed in claim 6, is characterized in that, image agglomerate fitting module is carried out ellipse fitting output for the Output rusults that the image agglomerate is merged to module.

Images