CN102651128B - Image set partitioning method based on sampling - Google Patents
Image set partitioning method based on sampling Download PDFInfo
- Publication number
- CN102651128B CN102651128B CN201110044790.XA CN201110044790A CN102651128B CN 102651128 B CN102651128 B CN 102651128B CN 201110044790 A CN201110044790 A CN 201110044790A CN 102651128 B CN102651128 B CN 102651128B
- Authority
- CN
- China
- Prior art keywords
- image
- prospect
- probability
- target image
- sampled images
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses an image set partitioning method based on sampling. The method comprises the following steps of: image selection: extracting characteristic vectors of color histograms of all images in an image set and performing characteristic vector clustering; extracting color distributions of the foreground and background of a sampled image; performing similarity matching on the sampled image with a target image; calculating the foreground probability of the target image based on the single sampled image; calculating the foreground probability of the target image; and constructing a Gibbs energy minimizing formula, resolving a result of a foreground or background, which corresponds to each pixel, through image partitioning, and partitioning the image according to the constructed Gibbs energy minimizing formula by using an image partitioning method. The method has the remarkable advantage that a large amount of user interaction required by foreground and background partitioning operation on each image in a large-scale image set can be reduced greatly.
Description
Technical field
The present invention relates to the method that all display foregrounds of the data set that comprises arbitrary image are cut apart, especially refer to the algorithm based on some image in image set being carried out to divided ownership display foreground on simple man-machine interactively basis.
Background technology
At present, interactively image segmentation algorithm has been simplified the task that display foreground is cut apart greatly.But interactive image segmentation algorithm in the past is all directly cut apart single image.If so the image that image set comprises is very many, it is a very thing for effort that the method that so directly before application, single image is cut apart is cut apart an image set.Associating partitioning algorithm (co-segmentation) can be cut apart all prospects of some image set simultaneously, but this class algorithm has a very strong hypothesis to image set: the prospect of image set should be the same, or has closely similar color distribution at least.And this hypothesis is not strictly followed in most of image set.
Summary of the invention
Goal of the invention: technical matters to be solved by this invention is for the deficiencies in the prior art, a kind of image set dividing method based on sampling is provided, the method is ensureing under the prerequisite of the final segmentation result quality of the each image of image set, the minimizing user interactions of maximum possible, thus facilitate user to carry out foreground segmentation to whole image set.
In order to solve the problems of the technologies described above, the invention discloses a kind of algorithm that Subgraph image set is cut apart, comprise that image cuts apart and iteration correction two parts.
Wherein, described image partitioning portion comprises the following steps:
Step 1, image is chosen: extract the proper vector of the color histogram of all images of image set, to the proper vector cluster obtaining, choose the cluster centre of K and put corresponding sampled images { I
ek| k=1 ..., K} offers user, and the value of K is random natural number; These sampled images offer user, user directly adopt single image interactive segmentation method (as, cut apart with Photoshop) be partitioned into prospect and background.
Step 2, extracts sampled images I
ek(samples) prospect and the color distribution of background: the prospect of cutting apart according to previous step user and background area, used the algorithm of classical gauss hybrid models (GMMs) to come calculating prospect and the color probability distribution situation of background in RGB color space.Gauss hybrid models is exactly respectively all prospects and background pixel point to be carried out respectively to cluster, and thinks that each classification is approximate and submit to Gaussian distribution, so each classification is calculated to corresponding color class center and variance.The ratio that in addition, also will account for prospect or the total pixel number of background according to the pixel number of each classification calculates the weights of each classification.
Step 3, sampled images is for the similar coupling of target image: for a undivided target image, the present invention uses the image retrieval algorithm based on region (integrated region matching) based on a kind of classics to find multiple picture materials sampled images similar to target image, and can calculate corresponding similarity.Described target image is the arbitrary image except sampled images in image set; The following formula of the present invention calculates similarity:
Wherein, S (I
ek, I) and be target image I and sampled images I
eksimilarity.S
g(I
ek, I) and be the image similarity based on color histogram of these two image overalls.
the image similarities of these two images based on Region Segmentation.λ
gand λ
lfor controlling weights.After the present invention is normalized all similarities that calculate, similarity be greater than 0.7 just judge two image similarities.
Step 4, the prospect probability calculation of the target image based on individual sampled images: in step 2, each width sampled images has the gauss hybrid models (GMMs) of a pair of prospect and background, so directly on target image according to this pair of gauss hybrid models, can calculate the probability that the prospect of sampled images and the color distribution of background occur on target image.Its computing method are: the probability that the classification of each pixel each cluster in the gauss hybrid models of sampled images of calculating target image occurs in obtaining Gauss model, the probability then all classifications being calculated is according to the weights weighting summation of each classification.The probability of gained, after normalization, can obtain the prospect probability of the target image based on this sampled images.The probability sum that the probability that the prospect that described normalized process is sampled images occurs at target image occurs at target image divided by prospect and the background of sampled images.The prospect probability of described target image refers to the prospect probability of each pixel.
Step 5, the prospect probability calculation of target image:: for target image similar to sampled images in step 3, use step 4 to calculate corresponding prospect probability to each target image; The final prospect probability of target image be the similarity of the sampled images that calculates using step 3 and target image as weights, weighting summation is calculated the prospect probability of the target image of the sampled images based on similar by step 4.
Step 6, minimizes according to the prospect probability structure gibbs energy of target image the label (label) that (Gibbs energy minimiza tion) and figure are cut apart (graphcut) and solve prospect corresponding to each pixel or background.Gibbs energy minimizes the method for structure and mainly considers two kinds of situations: the label (label) possible to each pixel set penalty term according to its prospect probability, and considers the penalty term of the consistance setting flatness of label between adjacent pixel (label).The gibbs energy of constructing is minimized to formula the direct application drawing of image is cut apart to (graphcut) algorithm.The process of figure automatic Segmentation image is generally by the optimization problem of organizing is above equivalent to a corresponding graph structure, then utilizes minimax flow algorithm to calculate an algorithm of minimal cut.Described gibbs energy minimizes structure and comprises two kinds of situations: the firstth, the label that each pixel is possible is set to penalty term according to its prospect probability; The secondth, consider that the consistance of label between adjacent pixel is set the penalty term of flatness; Described label refers to that pixel belongs to prospect or background, if prospect, label gets 1, otherwise gets 0.
Iteration correction comprises the following steps:
Step 7, the correction of the image that user is cut apart not too satisfied image: in all images are cut apart, user selects piece image I, by the mark of interpolation prospect and background, thereby obtains correction image.In the present invention, the unsatisfied image segmentation result of user, refers generally to foreground object that user is concerned about not by complete splitting, or the background parts be not too concerned about of user is also divided into foreground object both of these case by mistake.
Step 8, more new images: the image set T that the similar image that the image that in step of updating 7, user revises and image all and that user revises define according to step 3 forms, iteration correction flow process comprises the following steps:
Step 81, the pixel in prospect, background and neighborhood image thereof to each width similar image is sampled and is obtained sampling field, and calculates corresponding proper vector; The described method that pixel in neighborhood image is sampled comprises: centered by each pixel in prospect or the background of each width similar image, the rectangle frame of getting a n × n is sample area, rectangle frame and existing rectangle frame have at least 1/4th part not overlapping, are proper vector thereby obtain the set of a sampling neighborhood;
Step 82, on all similar images at image set T except image I, according to described proper vector, each pixel and the similarity M (v that marks image-region in computed image
i); Circular is: in proper vector, each rectangle frame and similar image do convolution, then gets the synthetic new convolved image of the locational maximal value of all convolved image respective pixel.Described new convolved image be in similar image each pixel about the prospect of correction image or the similarity of background area;
Step 83, according to prospect and the background probability of every bit on gauss hybrid models calculating similar image
Step 84, directly applies described gauss hybrid models in the prospect and the background probability that upgrade the upper every bit of image set T;
Step 85, using the similarity in step 82 as weights, is added to the prospect of image in step 84 and background probability value in step 7, to upgrade in image set T in the original prospect of every width similar image and background probability, thus the more prospect of new images and background probability value; Formula is:
If more new images is exactly the image I that user revises, M (v so
i) be that an all elements is 1 matrix entirely.If more new images is image set T other images except image I, M (v so
i) be the similarity matrix calculating in step 82,
for step 84 calculates probability;
Step 86, according to the prospect Probability p obtaining in step 85
f(v
i), the similar image after upgrading for each width, with the prospect probability after upgrading in step 85 divided by prospect probability and background probability sum the prospect probability after as normalization;
Step 87, with the prospect probability after regularization in step 86, structure gibbs energy minimizes and image corresponding to application drawing automatic Segmentation.
Beneficial effect: it is mutual that remarkable advantage of the present invention is that every piece image that can reduce greatly by a Large Scale Graphs image set all carries out a large number of users required in prospect and background segment operation.Particularly, the present invention only has two places to need that user's is simple mutual, (1), in image is cut apart, need user to several sampled images (samples) that choose and in order to the single image dividing method having, sampled images cut apart.(2) in iteration correction, need user to image not too satisfied in image is cut apart, simply revise, and go iteration to upgrade every other similarly dissatisfied image with these corrections.Method in the past is all cut apart single image or the image set with identical prospect, and needs a large amount of user interactions.Contrast, the present invention can carry out interactive segmentation to the data set that has arbitrarily great amount of images.
Brief description of the drawings
Below in conjunction with the drawings and specific embodiments, the present invention is done further and illustrated, above-mentioned and/or otherwise advantage of the present invention will become apparent.
Fig. 1 is process flow diagram of the present invention.
Fig. 2 is the illustration of the embodiment of the present invention one.
Fig. 3 is the image partitioning portion illustration of the embodiment of the present invention two.
Fig. 4 is the iteration retouch illustration of the embodiment of the present invention two.
Embodiment
The invention discloses a kind of algorithm that Subgraph image set is cut apart, comprise that image cuts apart and iteration correction two parts.
Wherein, image is cut apart and is comprised the following steps:
Step 1, user interactions image is chosen: extract the proper vector of the color histogram of all images of image set, to the proper vector cluster obtaining, choose the cluster centre of K and put corresponding sampled images { I
ek| k=1 ..., K} offers user.These sampled images offer user, and user directly adopts single image interactive segmentation method (as used PHOTOSHOP instrument) to be partitioned into prospect and background.Two kinds of the most classical single image dividing methods that the present invention here adopts: Lazy Snapping and Grabcut.Can reference papers about the details of these two kinds of algorithms: Yin Li, Jian Sun, Chi-Keung Tang, and Heung-Yeung Shum, " Lazy snapping, " (in Proceedings of ACM SIGG RAPH, 2004, and Carsten Rother pp.303-308), Vladimir Kolmogorov, and Andrew Blake, " Grabcut:Interactive foreground extraction using iterated graph cuts, " (ACM Transaction s on Graphics, vol.23, pp.309-314, 2004).In the image partitioning portion of step 1 corresponding diagram 1 basic procedure, sampled images is carried out to the subprocess that user interactions is cut apart.During the sampled images of Fig. 2 and Fig. 3 is capable, provide the result directly perceived of step 1.
Step 2, extract the prospect of sampled images (samples) and the color distribution of background: the prospect of cutting apart according to previous step user and background area, used the algorithm of classical gauss hybrid models (GMMs) to carry out the color distribution of calculating prospect and background.In the Grabcut list of references providing in step 1 about gauss hybrid models, be described later in detail.Summarize says, gauss hybrid models is exactly respectively all prospects and background pixel point to be carried out respectively to cluster, and think each classification be similar to submit to Gaussian distribution, so each classification is calculated to color class center and variance.The ratio that in addition, also will account for prospect or the total pixel number of background according to the pixel number of each classification calculates the weights of each classification.The foreground/background GMMs sub-process of image partitioning portion in step 2 correspondence and Fig. 1.
Step 3, sampled images is for the similar coupling of target image: for a undivided target image, image retrieval algorithm based on region (the integrated region matching of the present invention based on a kind of classics, IRM) find multiple picture materials sampled images similar to target image, and can calculate corresponding similarity.The details that realizes about IRM can articles of reference James Z.Wang, Jia Li, and Gio Wiederhold, " Simplicity:Se mantics-sensitive integrated matching for picture libraries; " (IEEE Transactions on Patte rn Analysis and Machine Intelligence, vol.23, pp.947-963,2001).Sampling in this step corresponding diagram 1, the similar coupling sub-process of target image.The present invention calculates measurement similarity with following formula:
Wherein, S (I
ek, I) and be target image I and sampled images I
eksimilarity.S
g(I
ek, I) and be the image similarity based on color histogram of these two image overalls.
the image similarities of these two images based on Region Segmentation.λ
gand λ
lfor controlling weights.After the present invention is normalized all similarities that calculate, similarity be greater than 0.7 just think two image similarities.
Step 4, based on individual sampled images I
ekeach pixel x of target image
iprospect probability
calculate:
In step 2, each width sampled images has the gauss hybrid models (GMMs) of a pair of prospect and background, so directly on target image according to this pair of gauss hybrid models, calculate the probability that the prospect of sampled images and the color distribution of background occur on target image
with
the detailed description of the Grabcut list of references that the details of computing method can refer step 1 provides.Following formula:
Circular may be summarized to be following steps: calculate each pixel probability that the classification of each cluster obtains occurring in Gauss model in gauss hybrid models, the probability then all classifications being calculated is according to the weights weighting summation of each classification.The probability of gained, after normalization, can obtain the prospect probability of the target image based on this sampled images.The probability sum that this normalized process probability that to be prospect occur at target image occurs at target image divided by prospect and background.(note: the prospect probability of the target image here refers to the prospect probability of each pixel.)
Step 5, the prospect Probability p of target image
f(x
i) calculating: several sampled images similar to target image that step 3 finds, each calculates corresponding prospect probability by step 4
so, the final prospect probability of target image is the sampled images that calculates with step 3 and the similarity S (I of target image
ek, I) and as weights, weighting summation is calculated the prospect probability of the target image of the sampled images based on similar by step 4.Capable having provided every image according to step 4,5 results of calculating of initial probability of Fig. 2 and Fig. 3.Be following formula:
Step 6, minimizes according to the prospect probability structure gibbs energy of target image the label (label) that (Gibbs energy minimiz ation) and figure are cut apart (graphcut) and solve prospect corresponding to each pixel or background.Gibbs energy minimizes the method for structure and mainly considers two kinds of situations: the label (label) possible to each pixel set penalty term according to its prospect probability, and considers the penalty term of the consistance setting flatness of label between adjacent pixel (label).The gibbs energy of constructing is minimized to formula the direct application drawing of image is cut apart to (graphcut) algorithm.About constructing, gibbs energy minimizes and the details reference document Yuri Y.Boykov and Marie-Pie rre Jolly of figure partitioning algorithm, " Interactive graph cuts for optimal boundary & region segmentation of object s in n-d images; " in Proceedings of IEEE International Conference on Computer Visio n, 2001, pp.105-112.The image of this step corresponding diagram 1 is cut apart.Fig. 2,3 the capable result directly perceived that has provided this step of image segmentation result.
The minimized formula of gibbs energy is:
Wherein, X
tall pixels of target image,
e
1(x
i) be pixel x
ithe single order energy term at place, E (x
i, x
j) be adjacent pixel x
i, x
jsecond order energy term.Neigh (i) represents the neighborhood territory pixel of pixel i.Figure partitioning algorithm is exactly to minimize above-mentioned formula to obtain satisfied X value.Above-mentioned formula comprises two, i.e. E
1(x
i) and E
2(x
i, x
j).Minimized two kinds of situation: the E of the gibbs energy of the 3rd article in these two corresponding claims books
1(x
i) refer to pixel x
iset penalty term according to its prospect probability; E
2(x
i, x
j) be to consider adjacent pixel x
i, x
jbetween the consistance of label set the penalty term of flatness.These two have ensured respectively the rationality of every image segmentation result from two angles: (1) E
1(x
i) ensureing the larger pixel of prospect probability, the possibility that is finally calculated to be prospect by the minimized formula of gibbs energy is larger, (2) E
2(x
i, x
j) ensureing that between neighbor pixel, label has certain consistance, the prospect of image segmentation result is level and smooth.In general this flatness comprises two aspects: the edge of prospect the boundary part of background (with) and the interior zone of prospect on image are all level and smooth.Particular content can refer step 1 in listed two sections of documents about Lazy Snapping and Graphcut dividing method.It is to these two summations of all image slices vegetarian refreshments that gibbs energy minimizes formula.
Iteration correction comprises the following steps:
Step 7, the correction of the image that user is cut apart not too satisfied image: in all images are cut apart, user selects piece image I, by the mark of interpolation prospect and background, (as user interactions correction in Fig. 4).Every decision block whether image user is satisfied with in this step corresponding diagram 1.In the present invention, the unsatisfied image segmentation result of user, refers generally to foreground object that user is concerned about not by complete splitting, or the background parts be not too concerned about of user is also divided into foreground object both of these case by mistake.
Step 8, more new images: the image set T that the similar image that the image that in step of updating 7, user revises and image all and that user revises define according to step 3 forms, the iteration correction flow process of the iteration retouch of this step corresponding diagram 1, it comprises the following steps:
Step 81, samples and obtains sampling field the pixel in prospect, background and neighborhood image thereof, and calculate corresponding proper vector; Described sampling neighborhood computing method are as follows: centered by the pixel in each prospect or background, get the rectangle frame of a n × n, rectangle frame and existing rectangle frame have at least 1/4th part not overlapping, are proper vector thereby obtain the set of a sampling neighborhood;
Step 82, on all images at image set T except image I, according to described proper vector, each pixel and the similarity that marks image-region in computed image; Circular is: each rectangle frame and similar image do convolution, then gets the synthetic new convolved image of the locational maximal value of all convolved image respective pixel.This new convolved image with regard to each pixel in similar image about the prospect of user annotation or the similarity of background area;
Step 83, the gauss hybrid models of the prospect of extraction mark and the pixel of background;
Step 84, directly applies described gauss hybrid models in the prospect and the background probability that upgrade the upper every bit of image set T;
Step 85, using the similarity in step 82 as weights, is added to the prospect of image in step 84 and background probability value to upgrade on the original prospect of image set T and background probability, thus the more prospect of new images and background probability value; Can sketch as following formula:
Here, if more new images is exactly the image I that user revises, M (v so
i) be that an all elements is 1 matrix entirely.If more new images is image set T other images except image I, M (v so
i) be the similarity matrix calculating in step 82,
for step 84 calculates probability;
Step 86, according to the prospect Probability p obtaining in step 85
f(v
i), for more new images of each width, use the prospect probability upgrading in step 85 divided by prospect probability and background probability sum;
Step 87, with the prospect probability after regularization in step 86, structure gibbs energy minimizes and image corresponding to application drawing automatic Segmentation.
Embodiment 1
For clear, show intermediate result intuitively, the present invention the corresponding centre of each target image of all figure and net result according to column alignment.
Fig. 2 has provided the example of cutting apart an image set.This image set is totally 5 images, comprises 1 sampled images (the first row the 1st width figure) and 4 target images (the second row 4 width figure).First, user carries out craft with Lazysnapping algorithm to sampled images to be cut apart, and is partitioned into prospect and background, and gained is the first row the 2nd width figure, and this image set has only carried out image cuts apart the result that can obtain satisfaction.Choose and the user of sampled images are cut apart the step 1 corresponding to image partitioning portion in concrete implementation step by hand.According to the result of cutting apart, the present invention can automatically be cut apart and not need other mutual of user other 4.In concrete implementation step, the step 2,3,4 of image partitioning portion, 5 execution result can obtain initial probability.Wherein step 3,4 target images and the sampled images similarity of calculating are respectively 0.95,0.85,0.87,0.92. step 4 and 5 according to putting in order, and the initial probability of every the target image calculating is as shown in the 3rd row in Fig. 2.For the ease of initial probability results of observing all pixels directly perceived, each pixel is mapped as different colors according to probability numbers, and white color represents that corresponding initial probability is higher.The corresponding initial probability of white point is 1, and the corresponding initial probability of black color dots is 0.The gibbs energy minimization problem form that initial probability provides according to the 6th step of image partitioning portion in concrete implementation step, with the result that this gibbs energy minimization problem of classical figure partitioning algorithm solution obtains be exactly final segmentation result.In step 6, gibbs energy minimization problem form mainly comprises two.Wherein Section 1 ensures the consistance of final segmentation result and initial probability, and the initial high final segmentation result of pixel of probability is more likely prospect, and the initial low final segmentation result of pixel of probability is more likely background; Section 2 ensures to cut apart the flatness of object edge, i.e. the smooth of the edge in the segmentation result of final image object.So the 4th row is the segmentation result of this image set target image in Fig. 2.
Embodiment 2
The image set of Fig. 3 has 9 images, and wherein 2 is sampled images (the first row the 1st width and the 3rd width figure in Fig. 3), and remaining 7 is target image (the second row in Fig. 3).Sampled images is cut apart by hand for user, segmentation result is the first row the 2nd width and the 4th width figure in Fig. 3, calculate according to similarity, the sampled images of Fig. 3 the 1st row the 1st width figure is similar with 1st～3 width target images of the 2nd row, and the sampled images of Fig. 3 the 1st row the 3rd width figure is similar with the target image of 4th～7 width of the 2nd row.For this image set, in image cutting procedure of the present invention and Fig. 2, image set is similar.Every target image all can calculate sampled images similarly, calculate the initial probability (as shown in result is as capable in the initial probability of the 3rd row of Fig. 3) of each target image according to similarity, then go out image segmentation result according to initial probability calculation, shown in the 4th row.In this figure, there is the unsatisfied image of some image segmentation results, image the present invention for needs correction marks in the drawings with red frame, the the 1st, 2,4,5 width figure in the 4th row, and provided corresponding correction segmentation result at next line, wherein the 3rd, 6,7 width figure are and satisfied result.The centre that each width target image is corresponding and net result are according to row alignment.The present invention taking the target image of the 1st, 3 width figure as example, elaborates the iteration makeover process that how carries out segmentation result according to user's sign correction in Fig. 4.
Fig. 4 has provided the whole process flow diagram of cutting apart data set in Fig. 3.This figure has provided the example of an iteration correction.In the rightmost sampled images row of this figure, provide the prospect (background is to be replaced by black) of sampled images (step 1) and the manual sampled images being partitioned into of user.Target image in the 1st this image set of behavior of this figure, the present invention is here representational has chosen 3.Corresponding every centre and the net result that image is cut apart of image of each row.This cutting procedure is as follows: (1) to sampled images according to step 2 extraction prospect and background the color probability distribution in RGB color space; (2) target image according to sampled images in step 3 the similar coupling for target image, the sampled images prospect that utilization calculates and the background color probability distribution in RGB color space, the prospect probability calculation of calculating the target image based on individual sampled images according to step 4; (3) target image, according to step 5, calculates the prospect probability of target image, and this prospect probability is as shown in the 2nd row of Fig. 4: each pixel is used according to color quantizing and shown, color is whiter, represents that this pixel prospect probability more approaches 1; (4) target image, according to step 6, calculates initial segmentation result.This result is as Fig. 4 the 3rd row demonstration.From result, the segmentation result of the first width target image is gratifying.But in target image Fig. 4 in the 1st row the 3rd width figure, Fig. 4 the segmentation result mistake of the 1st row the 2nd width figure foreground object has been treated as in a part of meadow, such result be need revise.For unsatisfied image segmentation result, the manual mark of user does not have the image-region (step 7) of distinguishing.Be 0.85 because of the similarity that in the 1st row the 3rd width figure, Fig. 4, the 1st row the 2nd width figure calculates according to step 3 in Fig. 4 again, i.e. two image iteration corrections together.According to step 8, in Fig. 4 in the 1st row the 2nd width figure by user interactions correction, mark meadow is background, sees step 81.The present invention calculates the gauss hybrid models of mark modification region according to user's mark, i.e. then step 83 calculates the initial probability after upgrading according to step 84～86.Finally constructing gibbs energy minimizes equation and cuts apart to solve with figure and obtain revising segmentation result, i.e. step 87.In image 4b because user is directly to this image correction, so do not need calculation procedure 82 mentioned with the similarity of other images.So M (v in the time of new images 4b more in step 85
i) be that an all elements is 1 matrix entirely.The present invention is according to calculating other dissatisfied images, as the correlation map M (v between the 1st row the 2nd width figure in unsatisfied Fig. 4 of the 1st row the 3rd width figure and user annotation in Fig. 4
i).About the demonstration directly perceived of the correlation map of the 1st row the 2nd width figure in the 1st row the 3rd width figure and Fig. 4 in Fig. 4 as shown in the 6th row the 2nd width figure in Fig. 4.M (v
i) each element of matrix region between 0 to 1, this figure is according to 1 corresponding white, and the principle of 0 corresponding black quantizes to show this matrix.In the time upgrading in Fig. 4 the 1st row the 3rd width figure, the present invention only needs calculation procedure 84～86 and using correlation map as weights, directly remove to revise the probability of this figure with the region of user interactions correction, obtain revised probability, then construct gibbs energy according to step 87 and minimize equation and cut apart to solve with figure and obtain revising segmentation result.The present invention is according to user's all image segmentation results similar to user's correction image of mark iteration correction.Each image of final image collection can be partitioned into satisfied result.
The invention provides a kind of thinking and method of the image set dividing method based on sampling; method and the approach of this technical scheme of specific implementation are a lot; the above is only the preferred embodiment of the present invention; should be understood that; for those skilled in the art; under the premise without departing from the principles of the invention, can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.In the present embodiment not clear and definite each ingredient all available prior art realized.
Claims (2)
1. the image set dividing method based on sampling, is characterized in that, comprises image partitioning portion;
Described image partitioning portion comprises the following steps:
Step 1, image is chosen: extract the proper vector of the color histogram of all images of image set, carry out proper vector cluster, choose the sampled images I corresponding to central point of K cluster
ek, k=1 ..., K, adopts single image automatic Segmentation to go out prospect and the background of sampled images;
Step 2, extracts sampled images I
ekprospect and the color distribution of background: use gauss hybrid models calculating sampling image I
ekprospect and the color probability distribution of background in RGB color space;
Step 3, sampled images is for the similar coupling of target image: for target image, use the image retrieval algorithm based on region to find out the picture material sampled images similar to target image, and calculating corresponding similarity, described target image is the arbitrary image except sampled images in image set;
Step 4, the prospect probability calculation of the target image based on individual sampled images: calculate the probability that the prospect of sampled images and the color distribution of background occur on target image on target image; Computing method are: the probability that the classification of each pixel each cluster in the gauss hybrid models of sampled images of calculating target image occurs in obtaining Gauss model, then the probability all classifications being calculated is according to the weights weighting summation of each classification, and the ratio that accounts for prospect or the total pixel number of background according to the pixel number of each classification calculates the weights of each classification; The probability of gained, after normalization, is the prospect probability of the target image based on this sampled images; The probability sum that the probability that the prospect that described normalized process is sampled images occurs at target image occurs at target image divided by prospect and the background of sampled images;
Step 5, the prospect probability of calculating target image: for target image similar to sampled images in step 3, use step 4 to calculate corresponding prospect probability to each target image; The final prospect probability of target image be the similarity of the sampled images that calculates using step 3 and target image as weights, weighting summation is calculated the prospect probability of the target image of the sampled images based on similar by step 4;
Step 6, minimizes formula and cuts apart the result that solves prospect corresponding to each pixel or background by figure according to the prospect probability structure gibbs energy of target image, and the gibbs energy of constructing is minimized to formula, uses figure automatic Segmentation image.
2. a kind of image set dividing method based on sampling according to claim 1, is characterized in that, described gibbs energy minimizes structure and comprises two kinds of situations: the firstth, the label that each pixel is possible is set to penalty term according to its prospect probability; The secondth, consider that the consistance of label between adjacent pixel is set the penalty term of flatness; Described label refers to that pixel belongs to prospect or background, if prospect, label gets 1, otherwise gets 0.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110044790.XA CN102651128B (en) | 2011-02-24 | 2011-02-24 | Image set partitioning method based on sampling |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201110044790.XA CN102651128B (en) | 2011-02-24 | 2011-02-24 | Image set partitioning method based on sampling |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN102651128A CN102651128A (en) | 2012-08-29 |
| CN102651128B true CN102651128B (en) | 2014-10-01 |
Family
ID=46693132
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201110044790.XA Expired - Fee Related CN102651128B (en) | 2011-02-24 | 2011-02-24 | Image set partitioning method based on sampling |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN102651128B (en) |
Families Citing this family (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103177446B (en) * | 2013-03-13 | 2016-03-30 | 北京航空航天大学 | Based on the accurate extracting method of display foreground of neighborhood and non-neighborhood smoothing prior |
| CN103268614B (en) * | 2013-05-31 | 2016-01-20 | 电子科技大学 | A kind of for many prospects be divided into cut prospect spectrum drawing generating method |
| CN104077771B (en) * | 2014-06-19 | 2017-06-20 | 哈尔滨工程大学 | A kind of weighting method realizes the mixed model image partition method of space limitation |
| CN104065932B (en) * | 2014-06-30 | 2019-08-13 | 东南大学 | A kind of non-overlapping visual field target matching method based on amendment weighting bigraph (bipartite graph) |
| CN104899877A (en) * | 2015-05-20 | 2015-09-09 | 中国科学院西安光学精密机械研究所 | Image foreground extraction method based on super-pixels and fast three-division graph |
| CN105701810B (en) * | 2016-01-12 | 2019-05-17 | 湖南中航天目测控技术有限公司 | A kind of unmanned plane image electronic based on click type image segmentation is sketched method |
| CN106056573A (en) * | 2016-04-26 | 2016-10-26 | 武汉科技大学 | Method for optimizing energy function in active contour model and application thereof |
| CN108073871A (en) * | 2016-11-18 | 2018-05-25 | 北京体基科技有限公司 | Method and device based on two kinds of presumptive area identification hand regions |
| CN109255790A (en) * | 2018-07-27 | 2019-01-22 | 北京工业大学 | A kind of automatic image marking method of Weakly supervised semantic segmentation |
| CN109493424A (en) * | 2018-11-07 | 2019-03-19 | 绍兴文理学院 | The all standing sampling method of structural plane 3 d surface topography |
| CN111260667B (en) * | 2020-01-20 | 2023-08-04 | 浙江大学 | Neurofibromatosis segmentation method combined with spatial guidance |
| CN111539993B (en) * | 2020-04-13 | 2021-10-19 | 中国人民解放军军事科学院国防科技创新研究院 | Space target visual tracking method based on segmentation |
| CN111488885B (en) * | 2020-06-28 | 2020-09-25 | 成都四方伟业软件股份有限公司 | Intelligent extraction method and device for theme color system of picture |
| CN111862045B (en) * | 2020-07-21 | 2021-09-07 | 上海杏脉信息科技有限公司 | Method and device for generating blood vessel model |
| CN113192072B (en) * | 2021-04-01 | 2023-11-24 | 北京达佳互联信息技术有限公司 | Image segmentation method, device, equipment and storage medium |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101840577A (en) * | 2010-06-11 | 2010-09-22 | 西安电子科技大学 | Image automatic segmentation method based on graph cut |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8005264B2 (en) * | 2008-06-09 | 2011-08-23 | Arcsoft, Inc. | Method of automatically detecting and tracking successive frames in a region of interesting by an electronic imaging device |
-
2011
- 2011-02-24 CN CN201110044790.XA patent/CN102651128B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101840577A (en) * | 2010-06-11 | 2010-09-22 | 西安电子科技大学 | Image automatic segmentation method based on graph cut |
Non-Patent Citations (6)
| Title |
|---|
| Carsten Rother等."GrabCut":interactive foreground extraction using iterated graph cuts.《ACM Transactions on Graphics (TOG) - Proceedings of ACM SIGGRAPH 2004 》.2004,第23卷(第3期),第309-314页. |
| Carsten Rother等."GrabCut":interactive foreground extraction using iterated graph cuts.《ACM Transactions on Graphics (TOG)- Proceedings of ACM SIGGRAPH 2004 》.2004,第23卷(第3期),第309-314页. * |
| Dhruv Batra等.iCoseg:Interactive co-segmentation with intelligent scribble guidance.《2010 IEEE Conference on Computer Vision and Pattern Recognition》.2010,第3169-3176页. |
| iCoseg:Interactive co-segmentation with intelligent scribble guidance;Dhruv Batra等;《2010 IEEE Conference on Computer Vision and Pattern Recognition》;20100618;第3169-3176页 * |
| 一种基于视频聚类的关键帧提取方法;朱映映等;《计算机工程》;20040229;第30卷(第4期);第12-13页及第121页 * |
| 朱映映等.一种基于视频聚类的关键帧提取方法.《计算机工程》.2004,第30卷(第4期),第12-13页及第121页. |
Also Published As
| Publication number | Publication date |
|---|---|
| CN102651128A (en) | 2012-08-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN102651128B (en) | Image set partitioning method based on sampling | |
| Li et al. | A novel green apple segmentation algorithm based on ensemble U-Net under complex orchard environment | |
| CN102592268B (en) | Method for segmenting foreground image | |
| Liu et al. | Adaptive partial differential equation learning for visual saliency detection | |
| Arteta et al. | Interactive object counting | |
| CN104217225B (en) | A kind of sensation target detection and mask method | |
| CN100573523C (en) | A kind of image inquiry method based on marking area | |
| CN103942794B (en) | A kind of image based on confidence level is collaborative scratches drawing method | |
| CN112750106B (en) | Nuclear staining cell counting method based on incomplete marker deep learning, computer equipment and storage medium | |
| CN106372648A (en) | Multi-feature-fusion-convolutional-neural-network-based plankton image classification method | |
| CN110059697A (en) | A kind of Lung neoplasm automatic division method based on deep learning | |
| CN110363233B (en) | Fine-grained image recognition method and system of convolutional neural network based on block detector and feature fusion | |
| CN106023145A (en) | Remote sensing image segmentation and identification method based on superpixel marking | |
| CN109117703B (en) | Hybrid cell type identification method based on fine-grained identification | |
| Bhagat et al. | Eff-UNet++: A novel architecture for plant leaf segmentation and counting | |
| CN102663757A (en) | Semi-automatic image cutting method based on nuclear transfer | |
| WO2021073279A1 (en) | Staining normalization method and system for digital pathological image, electronic device and storage medium | |
| CN110348435A (en) | A kind of object detection method and system based on clipping region candidate network | |
| Zhu et al. | A novel approach for apple leaf disease image segmentation in complex scenes based on two-stage DeepLabv3+ with adaptive loss | |
| CN108846404A (en) | A kind of image significance detection method and device based on the sequence of related constraint figure | |
| CN107527054A (en) | Prospect extraction method based on various visual angles fusion | |
| Zhou et al. | An efficient two-stage region merging method for interactive image segmentation | |
| CN116580203A (en) | Unsupervised cervical cell instance segmentation method based on visual attention | |
| CN103578107A (en) | Method for interactive image segmentation | |
| CN108090913B (en) | Image semantic segmentation method based on object-level Gauss-Markov random field |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C14 | Grant of patent or utility model | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20141001 Termination date: 20180224 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |