CN104794726A - Parallel underwater image segmentation method and device - Google Patents

Abstract

The invention discloses a parallel underwater image segmentation method and device. The method includes the following steps: S1, dividing a group of images into different subimage groups according to similarities of image attributes; S2, allocating the different subimage groups to different computational resources respectively; S3, classifying each pixel to the corresponding clustering category via the computational resources in parallel according to the membership degree of gray, as a member of the clustering center, of each pixel of the corresponding subimage group, wherein the clustering center is a gray value. By the method and device, underground image segmentation efficiency can be improved.

Description

A kind of underwater picture Parallel segmentation method and device

[technical field]

The present invention relates to a kind of underwater picture Parallel segmentation method and device.

[background technology]

At present in oceanographic observation field, ubiquity the demand of the magnanimity underwater picture that process constantly produces, wherein especially with underwater picture segmentation problem for representative.Existing technology is mainly paid close attention to and is solved segmentation accuracy and the efficiency of single width underwater picture, and therefore, the quantity of the underwater picture gathered along with the first-class equipment of underwater camera constantly increases, existing image processing method inefficiency.

[summary of the invention]

Existing technology, the non-parallel type cluster segmentation method based on single image can not be expanded well, be applied on magnanimity underwater picture segmentation problem; Simultaneously due to only for single image, and do not consider the correlativity between multiple image, therefore the grey similarity that can not make full use of between some images is optimized.Existing method not can solve the treatment effeciency problem of magnanimity underwater picture segmentation.

In order to overcome the deficiencies in the prior art, the invention provides a kind of underwater picture Parallel segmentation method and device, to improve the efficiency of underwater picture process.

A kind of underwater picture Parallel segmentation method, comprises the steps:

S1, one group of image is divided into different subimage groups by the similarity according to image attributes;

S2, distributes to different computational resources respectively by different subimage groups;

S3, described computational resource is under the jurisdiction of the degree of membership of cluster centre concurrently according to the gray scale of each pixel of corresponding subimage group, by each classify of image element in the cluster classification of correspondence; Wherein, described cluster centre is a gray-scale value.

In one embodiment, the similarity of described image attributes is the similarity of the similarity of the content of image or the intensity profile of image.

In one embodiment, the similarity of described image attributes is the similarity of intensity profile;

Described step S1 comprises the steps:

1) grey level histogram of described one group of image is obtained;

2) peak-peak and the second largest peak value of grey level histogram is obtained;

3) distance between grey level histogram between peak-peak is calculated, and the distance between second largest peak value;

4) distance between peak-peak is less than threshold distance and two images that distance between the second largest peak value of correspondence is less than threshold distance are judged as the image that intensity profile similarity is high, and image high for similarity is put under in identical subimage group.

In one embodiment, also comprise the steps: between step S2 and step S3

S21, reads cluster centre file from distributed file system, and described cluster centre file comprises cluster centre matrix and objective function, described objective function F ' _m(U, V) is:

$F_m^{\′}(U,V)=\underset{i=0}{\overset{255}{\mathrm{\Σ}}}\underset{j=1}{\overset{2}{\mathrm{\Σ}}}{\left(u_{\mathrm{ji}}^{\′}\right)}^m\·h\left(i\right)\·{d_{\mathrm{ji}}}^2(i,v_j^{\′});$ Wherein, i is the gray-scale value of the pixel of image, and h (i) represents that in image, gray-scale value is the number of the pixel of i, u ' _jirepresent that gray-scale value i is under the jurisdiction of a jth cluster centre v ' _jdegree of membership, m is FUZZY WEIGHTED index, d _ji(i, v ' _j) be that gray-scale value i is to cluster centre v ' _jdistance;

S22, to degree of membership u ' _jiupgrade: $u_{\mathrm{ji}}^{\′}=1/\underset{k=1}{\overset{2}{\mathrm{\Σ}}}{\left(\frac{d_{\mathrm{ji}}(i,v_j^{\′})}{d_{\mathrm{ki}}(i,v_k^{\′})}\right)}^{2/(m-1)};$

Adjust the distance d _ji(i, v ' _j) upgrade: d _ji(i, v ' _j)=| i-v ' _j|, i=0,1 ..., 255, j=1,2;

Wherein, v ' _krepresent a kth cluster centre;

S23, key-value pair in the middle of exporting, described middle key-value pair comprises: gray-scale value i, and in grey level histogram, gray-scale value is the number of pixels h (i) of i, distance metric d _ji(i, v ' _j), and degree of membership u ' _ji;

S24, judge the described objective function F of each image of described subimage group ' _mwhether (U, V) restrains, if convergence, performs step S3, if do not restrain, then performs step S35 and step S36;

S25, selection portion partial image in described subimage group, to the cluster centre v ' of described parts of images _jupgrade:

$v_j^{\′}=(\underset{i=0}{\overset{255}{\mathrm{\Σ}}}{\left(u_{\mathrm{ji}}^{\′}\right)}^m\·h\left(i\right)\·i)/\left(\underset{i=0}{\overset{255}{\mathrm{\Σ}}}{\left(u_{\mathrm{ji}}^{\′}\right)}^{m}h\left(i\right)\right);$

Then the cluster centre v ' of draw value as described subimage group of the cluster centre of described parts of images is asked for _j:

$v_j^{\′}={\stackrel{\‾}{v}}_{\mathrm{sample}}^{\′}=\frac{1}{t}\underset{i=1}{\overset{t}{\mathrm{\Σ}}}{v}_{\mathrm{ji}}^{\′},j=\mathrm{1,2};$

Wherein, v' _jirepresent the cluster centre after the renewal of i-th image in described parts of images;

S26, according to the degree of membership u ' after renewal _jiwith distance d _ji(i, v ' _j) described objective function is upgraded;

S27, exports the cluster centre after renewal and objective function to described cluster centre file;

S28, judges whether described cluster centre and described objective function restrain, if convergence, performs step S3.

In one embodiment, described cluster centre comprises the first cluster centre and the second cluster centre, the prospect of described first cluster centre and the second cluster centre difference representative image and background.

Present invention also offers a kind of underwater picture Parallel segmentation device, comprise as lower unit:

First module, is divided into different subimage groups for the similarity according to image attributes by one group of image;

Second unit, for distributing to different computational resources respectively by different subimage groups;

Unit the 3rd, is under the jurisdiction of the degree of membership of cluster centre concurrently according to the gray scale of each pixel of corresponding subimage group for described computational resource, by each classify of image element in the cluster classification of correspondence; Wherein, described cluster centre is a gray-scale value.

In one embodiment, the similarity of described image attributes is the similarity of the similarity of the content of image or the intensity profile of image.

In one embodiment, the similarity of described image attributes is the similarity of intensity profile;

Described first module also for:

Obtain the grey level histogram of described one group of image;

Obtain peak-peak and the second largest peak value of grey level histogram;

Calculate the distance between peak-peak between grey level histogram, and the distance between second largest peak value;

Distance between peak-peak is less than threshold distance and two images that distance between the second largest peak value of correspondence is less than threshold distance are judged as the image that intensity profile similarity is high, and image high for similarity is put under in identical subimage group.

In one embodiment, described cluster centre comprises the first cluster centre and the second cluster centre, the prospect of described first cluster centre and the second cluster centre difference representative image and background.

Be not optimized for problems such as the grey similarity between the extensive property of amount of images and image and improve.Therefore, the quantity of the underwater picture gathered along with the first-class equipment of underwater camera constantly increases, and needs, for the extensive property in this kind of amount of images, to solve the treatment effeciency problem of its Iamge Segmentation.

The present invention, mainly based on MapReduce multiple programming framework, proposes a kind of Parallel segmentation method of underwater picture, in order to solve the treatment effeciency problem of magnanimity underwater picture on Iamge Segmentation.

The present invention considers and is incorporated in underwater picture segmentation problem by parallel processing thought, consider the phenomenon of the intensity profile that ubiquity is similar between underwater picture simultaneously, give a kind of underwater picture Parallel segmentation method based on MapReduce, while the accuracy of guarantee Iamge Segmentation, the treatment effeciency of dividing method on extensive underwater picture data set can be improved.Owing to have employed parallel calculating method, multiple processing unit can simultaneously image data processing, therefore the dividing method being compared to single processing unit has higher treatment effeciency on Large Scale Graphs image set, and efficiency can improve further with the expansion of process number of unit simultaneously; Also can avoid the double counting to similar image to the consideration of intensity profile similarity simultaneously, and then avoid unnecessary calculating consumption.

[accompanying drawing explanation]

Fig. 1 is the process flow diagram of the underwater picture Parallel segmentation method of an embodiment of the present invention.

[embodiment]

Below the preferred embodiment of invention is described in further detail.

As shown in Figure 1, a kind of underwater picture Parallel segmentation method of embodiment, comprises the steps:

S1, one group of image is divided into different subimage groups by the similarity according to image attributes.

The similarity of described image attributes can be the similarity etc. of the similarity of the content of image or the intensity profile of image, and the similarity based on image attributes is divided into groups to large-scale image collection, obtains different subimage groups.Such as, if image itself is with content tab, mark the content contained by image, then can divide into groups according to content tab, otherwise adopt intensity profile similarity.

The concrete grammar adopting the similarity of intensity profile to carry out dividing into groups can be as follows:

1) grey level histogram of described one group of image is obtained;

2) peak-peak and the second largest peak value of all grey level histograms is obtained;

3) distance between grey level histogram between peak-peak is calculated, and the distance between second largest peak value;

4) distance between peak-peak is less than threshold distance and two images that distance between the second largest peak value of correspondence is less than threshold distance are judged as the image that intensity profile similarity is high, and image high for similarity is put under in identical subimage group.

Such as, for the image that the different frame extracted by same video scene is formed, if several frame obviously has the similarity in content and intensity profile, then can put same group under.

For the image of same subimage group, based on the partitioning algorithm of cluster by having the parameters such as similar grey level histogram and cluster centre, therefore, can handle together, thus achieve optimization on the treatment effeciency of dividing method.

S2, distributes to different computational resources respectively by different subimage groups, and the image of same group can be packaged as a compressed file, and is distributed on same computational resource.Here computational resource can be computing unit or computing cluster, and same computing cluster can be the part computational resource that virtual grate goes out from whole cluster.

Then, same computational resource carries out the cluster segmentation of parallelization to same subimage group.

S3, the first clustering parameter such as Initialize installation cluster centre number, FUZZY WEIGHTED exponential sum maximum iteration time.These clustering parameters, owing to being for same subimage group, therefore can be set to same or analogous numerical value to the different images of same subimage group.

S4, reads in cluster centre file from distributed file system.Contain the initial numerical value such as cluster centre matrix, objective function in cluster centre file, store on a distributed, and be thus continually updated in follow-up MapReduce iteration.

Cluster centre a: gray-scale value of representative image prospect or background, such as, the cluster centre of prospect may be a darker gray-scale value, and the cluster centre of background may be a brighter gray-scale value; Here the cluster centre entry of a matrix element in cluster centre file starts can be initialized, a random numerical value.Cluster centre is embodied in grey level histogram, is exactly two peak values representing prospect or background respectively.A cluster centre is corresponding with a cluster classification.

Objective function F ' _m(U, V): the objective function in algorithm iteration optimizing process.Along with each iteration of algorithm, objective function will reduce.Objective function expression formula is: $F_m^{\′}(U,V)=\underset{i=0}{\overset{255}{\mathrm{\Σ}}}\underset{j=1}{\overset{2}{\mathrm{\Σ}}}{\left(u_{\mathrm{ji}}^{\′}\right)}^m\·h\left(i\right)\·{d_{\mathrm{ji}}}^2(i,v_j^{\′});$ Wherein i is the gray-scale value of the pixel of image, and value can between 0-255, under representing eight gray levels, and the gray-scale value span of each pixel of image.U ' _jibe in certain grey level histogram, gray-scale value i is under the jurisdiction of a jth cluster centre v ' _jdegree of membership, m is FUZZY WEIGHTED index, the power of blur level in method for expressing, and h (i) is in the grey level histogram of this image, and gray-scale value is the number of the pixel of i, d _ji(i, v ' _j) be that gray-scale value i is to cluster centre v ' _jdistance.

S5, starts mapping (Map) stage (the Map stage is the Data dissemination stage in MapReduce framework) mapping reduction (MapReduce) program, obtains the grey level histogram information of each image.

Owing to having obtained the grey level histogram of each image in step 1, therefore directly call the grey level histogram calculated here.

S6, according to obtained grey level histogram information, upgrades the distance d of every width image _ji(i, v ' _j) and subordinated-degree matrix, distance d here _ji(i, v ' _j) be the Euclidean distance of each gray level to cluster centre, and the subordinated-degree matrix gray level identified belonging to pixel is under the jurisdiction of the relation of each cluster centre.

The calculating of degree of membership and more new formula are $u_{\mathrm{ji}}^{\′}=1/\underset{k=1}{\overset{2}{\mathrm{\Σ}}}{\left(\frac{d_{\mathrm{ji}}(i,v_j^{\′})}{d_{\mathrm{ki}}(i,v_k^{\′})}\right)}^{2/(m-1)};$

The more new formula of distance is: d _ji(i, v ' _j)=| i-v ' _j|, i=0,1 ..., 255, j=1,2.

S7, completes after upgrading and calculating, arranges middle key-value pair, and outputs to next stage and process.Here middle key-value pair needs to preserve the information such as grey level histogram, and the distance d calculated in the Map stage _ji(i, v ' _j), the information such as subordinated-degree matrix, these information will be pushed to next stage by cluster classification and process.

Here the information that middle key assignments centering comprises has: gray value information i, and in grey level histogram, gray-scale value is the number of pixels h (i) of i, distance metric d _ji(i, v ' _j), and degree of membership u ' _ji.These information (namely key-value pair), will such as, according to its cluster classification, d _ji(i, v ' _j) or u ' _jiin a jth cluster centre if 1, then represent it is prospect, 2 represent backgrounds, therefore belong to 1 will be passed to a reduce, and belong to another reduce that will be passed to of 2, two reduce by while, process concurrently.

S8, judges whether objective function restrains, if convergence, then no longer enter reduction (Reduce) stage (the Reduce stage is Data Collection in MapReduce framework and merging phase) of this iteration, so far, the Map stage terminates, and performs step S13; Otherwise perform step S9, carry out the process operation in Reduce stage.

Judge whether objective function restrains, if convergence, then can start a function in map stage, this function will export the image after segmentation, and no longer the execution reduce stage just exits whole algorithm, represent that algorithm terminates.If objective function is not restrained, illustrate and need further to upgrade, therefore this time iteration also needs the process operation carrying out the reduce stage.On the whole, one time iteration comprises two parts: map stage and reduce stage, and the map stage terminates the rear just execution reduce stage, can certainly select not perform.Whether so-called objective function restrains, and refers to, the change of the target function value that the current target function value that calculates and cluster centre file store is less than a threshold value.

S9, starts the Reduce stage of MapReduce program, resolves the middle key-value pair delivered to from the Map stage, by the information analysis such as distance metric, degree of membership out, for subsequent treatment.In the Reduce stage, the image of same group carries out iteration by using the mean value of common cluster centre or multiple cluster centre, this be by image between grey similarity cause.

Here for every width image, be according to formula originally upgrade each cluster centre, the implication of each symbol in formula as previously mentioned.Because image has grey similarity, therefore no longer need in algorithm to calculate all like this every width image or upgrade, but in each iteration, utilize the parts of images therefrom extracted, such as 3-5 width, after above-mentioned formulae discovery cluster centre, asks for average and the result that remaining image just will upgrade using this average as this, and do not need to calculate again, thus decrease calculated amount.

S10, utilizes the key-value pair after resolving, jointly upgrades cluster centre and objective function to same group image.Owing to processing for a cluster in a Reducer (Reducer:Reduce phase process has a computing unit of same keys), what therefore upgrade is only the part of objective function, just can carry out the cumulative summation of objective function after the Reduce stage terminates.

S11, the new cluster centre calculated and target function value will be written in output file, and this output file, as the result of cluster centre file current iteration and backup, is kept in distributed file system equally.So far, the Reduce stage terminates, and a MapReduce iteration terminates simultaneously.

S12, judges whether objective function and cluster centre restrain.If the two is all restrained, then the iterative optimization procedure of clustering method terminates, and obtains final degree of membership result, and performs step S13; Otherwise after covering old cluster centre file with output file, delete output file and start next round iteration.

Here cluster centre convergence refers to the numerical value that each iteration all needs to upgrade cluster centre, when in twice iterative process, this numerical value change is little, such as represent kth time iteration, represent kth-1 iteration, twice iterative numerical change is less than a threshold value, thinks that cluster centre is restrained.

S13, according to the subordinated-degree matrix finally obtained, being subordinate to angle value according to each pixel and each cluster classification, classifying, as detected target and image background two class to each pixel of same group image; Thus generate the image after splitting and be saved in distributed file system.

In each iterative process, if through judging, objective function and cluster centre are all restrained, and the subordinated-degree matrix that so current iterative computation obtains is exactly final result; Otherwise carry out next iteration, again upgrade subordinated-degree matrix according to given formula above.

Above content is in conjunction with concrete preferred implementation further description made for the present invention, can not assert that specific embodiment of the invention is confined to these explanations.For general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, some simple deduction or replace can also be made, all should be considered as belonging to the scope of patent protection that the present invention is determined by submitted to claims.

Claims (9)

1. a underwater picture Parallel segmentation method, is characterized in that, comprises the steps:

S1, one group of image is divided into different subimage groups by the similarity according to image attributes;

S2, distributes to different computational resources respectively by different subimage groups;

S3, described computational resource is under the jurisdiction of the degree of membership of cluster centre concurrently according to the gray scale of each pixel of corresponding subimage group, by each classify of image element in the cluster classification of correspondence; Wherein, described cluster centre is a gray-scale value.

2. underwater picture Parallel segmentation method as claimed in claim 1, it is characterized in that, the similarity of described image attributes is the similarity of the similarity of the content of image or the intensity profile of image.

3. underwater picture Parallel segmentation method as claimed in claim 1, it is characterized in that, the similarity of described image attributes is the similarity of intensity profile;

Described step S1 comprises the steps:

1) grey level histogram of described one group of image is obtained;

2) peak-peak and the second largest peak value of grey level histogram is obtained;

3) distance between grey level histogram between peak-peak is calculated, and the distance between second largest peak value;

4) distance between peak-peak is less than threshold distance and two images that distance between the second largest peak value of correspondence is less than threshold distance are judged as the image that intensity profile similarity is high, and image high for similarity is put under in identical subimage group.

4. underwater picture Parallel segmentation method as claimed in claim 1, is characterized in that, also comprise the steps: between step S2 and step S3

S21, reads cluster centre file from distributed file system, and described cluster centre file comprises cluster centre matrix and objective function, described objective function F ' _m(U, V) is:

$F_m^{\′}(U,V)=\underset{i=0}{\overset{255}{\mathrm{\Σ}}}\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{\left(u_{\mathrm{ji}}^{\′}\right)}^m\·h\left(i\right)\·{d_{\mathrm{ji}}}^2(i,v_j^{\′});$ Wherein, i is the gray-scale value of the pixel of image, and h (i) represents that in image, gray-scale value is the number of the pixel of i, u ' _jirepresent that gray-scale value i is under the jurisdiction of a jth cluster centre v ' _jdegree of membership, m is FUZZY WEIGHTED index, d _ji(i, v ' _j) be that gray-scale value i is to cluster centre v ' _jdistance;

S22, to degree of membership u ' _jiupgrade: $u_{\mathrm{ji}}^{\′}=1/\underset{k=1}{\overset{2}{\mathrm{\Σ}}}{\left(\frac{d_{\mathrm{ji}}(i,v_j^{\′})}{d_{\mathrm{ki}}(i,v_k^{\′})}\right)}^{2/(m-1)};$

Adjust the distance d _ji(i, v ' _j) upgrade: d _ji(i, v ' _j)=| i-v ' _j|, i=0,1 ..., 255, j=1,2;

Wherein, v ' _krepresent a kth cluster centre;

S23, key-value pair in the middle of exporting, described middle key-value pair comprises: gray-scale value i, and in grey level histogram, gray-scale value is the number of pixels h (i) of i, distance metric d _ji(i, v ' _j), and degree of membership u ' _ji;

S24, judge the described objective function F of each image of described subimage group ' _mwhether (U, V) restrains, if convergence, performs step S3, if do not restrain, then performs step S35 and step S36;

S25, selection portion partial image in described subimage group, to the cluster centre v of described parts of images _j' upgrade:

$v_j^{\′}=(\underset{i=0}{\overset{255}{\mathrm{\Σ}}}{\left(u_{\mathrm{ji}}^{\′}\right)}^m\·h\left(i\right)\·i)/\left(\underset{i=0}{\overset{255}{\mathrm{\Σ}}}{\left(u_{\mathrm{ji}}^{\′}\right)}^{m}h\left(i\right)\right);$

Then the cluster centre v ' of draw value as described subimage group of the cluster centre of described parts of images is asked for _j:

$v_j^{\′}={\stackrel{\‾}{v}}_{\mathrm{sample}}^{\′}=\frac{1}{t}\underset{i=1}{\overset{t}{\mathrm{\Σ}}}{v}_{\mathrm{ji}}^{\′},j=\mathrm{1,2};$

Wherein, v ' _jirepresent the cluster centre after the renewal of i-th image in described parts of images;

S26, according to the degree of membership u ' after renewal _jiwith distance d _ji(i, v ' _j) described objective function is upgraded;

S27, exports the cluster centre after renewal and objective function to described cluster centre file;

S28, judges whether described cluster centre and described objective function restrain, if convergence, performs step S3.

5. underwater picture Parallel segmentation method as claimed in claim 1, it is characterized in that, described cluster centre comprises the first cluster centre and the second cluster centre, the prospect of described first cluster centre and the second cluster centre difference representative image and background.

6. a underwater picture Parallel segmentation device, is characterized in that, comprises as lower unit:

First module, is divided into different subimage groups for the similarity according to image attributes by one group of image;

Second unit, for distributing to different computational resources respectively by different subimage groups;

Unit the 3rd, is under the jurisdiction of the degree of membership of cluster centre concurrently according to the gray scale of each pixel of corresponding subimage group for described computational resource, by each classify of image element in the cluster classification of correspondence; Wherein, described cluster centre is a gray-scale value.

7. underwater picture Parallel segmentation device as claimed in claim 6, it is characterized in that, the similarity of described image attributes is the similarity of the similarity of the content of image or the intensity profile of image.

8. underwater picture Parallel segmentation device as claimed in claim 6, it is characterized in that, the similarity of described image attributes is the similarity of intensity profile;

Described first module also for:

Obtain the grey level histogram of described one group of image;

Obtain peak-peak and the second largest peak value of grey level histogram;

Calculate the distance between peak-peak between grey level histogram, and the distance between second largest peak value;

Distance between peak-peak is less than threshold distance and two images that distance between the second largest peak value of correspondence is less than threshold distance are judged as the image that intensity profile similarity is high, and image high for similarity is put under in identical subimage group.

9. underwater picture Parallel segmentation device as claimed in claim 6, it is characterized in that, described cluster centre comprises the first cluster centre and the second cluster centre, the prospect of described first cluster centre and the second cluster centre difference representative image and background.