CN103400386A - Interactive image processing method used for video - Google Patents

Abstract

The invention provides an interactive image processing method used for a video. The method comprises the following steps of extracting key frames from a video sequence; overlaying frames between adjacent key frames one by one to form a key frame cluster used for interactive marking; distinguishing the images of the key frames into a foreground area, a background area and an unknown area; carrying out spectral clustering and alpha value estimation on the key frames to obtain the sectional drawing result of the key frames; and finally, transferring the sectional drawing result of the key frames to the whole video sequence to obtain a final sectional drawing result. Because the foreground points are interactively marked on each key frame cluster, spatiotemporal information contained in the video sequence is fully utilized, and the naturalness and the intuition of user interaction are guaranteed on the basis of the user interface of the key frame, thereby conforming to the observation habit of a human visual system. Marking on the key frame clusters replaces independently marking on the key frames, and therefore the method has strong robustness for the situation with the big foreground object local action.

Description

A kind of method of Interactive Image Processing for video

Technical field

The present invention relates to image processing field, particularly a kind of method of Interactive Image Processing based on image blurring connection degree.

Background technology

Digital Matting is a kind of by a small amount of user interactions, foreground object is extracted exactly from image or video sequence technical process.Digital matting is the gordian technique on the image processing field bases such as photo editing or production of film and TV, has just obtained paying close attention to widely and studying at the beginning of computer image processing technology rises.

Stingy diagram technology intactly is divided into foreground area (F) and background area (B) with image in simple terms exactly, and that wherein need process is the color vector C of the pixel p of zone of ignorance (C) _p, color vector C _pBy foreground area pixel F _p, background area pixels B _pWith the transparency parameter alpha _pLinear formation:

C _p=α _pF _p+(1-α _p)B _p，

α wherein _p∈ [0,1], 0 represents background, 1 represents prospect.For most of natural pictures, F and B can not be confined to some specific values, and the α of each pixel, F, B value are unknown undetermined.The color vector C of its three-dimensional for our known information of some pixels _p, unknown information is F _p, B _pAnd α _pSo scratching the figure process is a non-binding problem of asking seven unknown quantitys by three known quantities.

Single image is scratched the diagram technology development and has been produced so far a lot of different algorithms, such as the overall method of sampling, the neighbouring node algorithm of KNN() method, Large Kernel method, Nonlocal method, PSF(Point-spread Function, point spread function) method, Shared are scratched drawing method etc., and being all has in various degree marked improvement on accuracy, MSE performance test or algorithm speed as a result scratching figure.But then, the stingy figure that the stingy figure of video sequence is compared single image has more challenge: the fluency of the big data quantity of video sequence, every two field picture edge treated, the influence factors such as adaptability that object is significantly moved all need not to consider in the stingy diagram technology to single image.present video matting algorithm has Bayes video to scratch drawing system, video matting based on Rotoscoping, video based on Graphcut is cliped and pasted system, Snapcut etc., but in these algorithms, perhaps because existing larger error to cause scratching the figure result, estimation itself can't be satisfied with, perhaps the robustness for the foreground object motion is not strong, the algorithm calculated amount of perhaps using is excessive, can't meet the requirement of video sequence large-scale data stream, the information that perhaps needs the user to input is too complicated and not directly perceived, system needs trained professional's operation, practicality is not strong.

Summary of the invention

The present invention is on the basis of forefathers' research, combining video sequences is scratched the distinctive space-time three-dimensional information of figure, proposed a kind of fast video and scratched drawing method: the frame between the key frame of front and back has been superimposed as key frame bunch, carry out the user interactions mark on key frame bunch, guaranteed naturality and the intuitive of user interactions, and for the larger situation of foreground object activities, stronger robustness has been arranged.Also (Spectral Clustering is a kind of of sub-space learning algorithm to the Application standard spectral clustering, spectral clustering is based upon on spectrogram reason basis, compare with traditional clustering algorithm, it have advantages of can be on the sample space of arbitrary shape cluster and converge on globally optimal solution.) complete cluster process, traditional fuzzy connectedness is cut apart in the three dimensions that is extended to video sequence and completed the estimation of α value.Also further use SURF(Speeded Up Robust Features) detect the match search window, the stingy figure result of key frame is passed to whole video sequence, greatly reduced the time complexity of algorithm.Simultaneously, also according to the different motion situation of foreground object in video sequence, design a kind of video flowing assignment method of the adaptively selected direction of propagation, made algorithm can have and scratch preferably the figure result for different types of foreground object sport video.

Technical matters to be solved by this invention is: the defect that overcomes prior art, a kind of Interactive Image Processing method that is used for video that user interactions simple, intuitive, robustness are strong and stingy figure is effective is provided, the method is extracted key frame in video sequence, and the frame between adjacent key frame is superposeed one by one and is formed for the key frame bunch of mutual mark, be divided into foreground area, background area and zone of ignorance with this image area with described key frame; Then described key frame is carried out spectral clustering and the estimation of α value, obtain the stingy figure result of described key frame; Finally, the stingy figure result of described key frame is passed to whole video sequence, obtains final stingy figure result.

According to embodiment, the present invention also can adopt following preferred technical scheme:

Described stingy figure result with described key frame is passed to whole video sequence and comprises: a. detects the edge line of foreground object at key frame, b. travel through described edge line several search windows are set, c. find the described search window of frame coupling before and after mark with the SURF feature point detecting method, the stingy figure result in d. search window that described key frame is corresponding assignment is in turn given search window corresponding to other intermediate frames.

In described steps d, if prospect is single or separate moving object, the video flowing of assignment both direction propagation forwards, backwards; If in video sequence, prospect is a plurality of objects that relative motion is arranged, the video flowing of assignment is only propagated from front to back.

Described SURF feature point detecting method comprises the steps:

1) calculated product component;

2) build the Hessian matrix;

3) build metric space;

4) location feature point;

5) determine rotation principal direction;

6) calculate SURF unique point descriptor.

Described extraction key frame is to extract a frame as key frame every 10～20 frames.

Described spectral clustering adopts the standardization Spectral Clustering.

The step of described standardization spectral clustering is as follows:

1) build similarity matrix W by raw data set, $W_{i,j}=\left\{\begin{array}{c}{w}_{i,j},i\≠j\\ 0,i=j\end{array}\right.,$

w _i,jBe used for the similarity between the expression data;

2) each column element addition of W is obtained the N number, build one and form diagonal line by this N number, other elements are all 0 N * N matrix D, $D_{i,j}=\left\{\begin{array}{c}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{i,j},i=j\\ 0,i\≠j\end{array}\right.;$

3) build similarity matrix W by raw data set and build Laplacian Matrix L=D-W, $L_{i,j}=\left\{\begin{array}{c}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{i,j},i=j\\ -w_{i,j},i\≠j\end{array}\right.,$ To obtain after the L standardization

Solve front k the eigenwert of L '

And characteristic of correspondence vector

Will

Multiply by one by one

Obtain front k proper vector of matrix L;

4) k proper vector lined up and formed the matrix of a N * k, regard every delegation wherein as in the k dimension space a vector, carry out cluster;

Wherein, w _i,jBe used for the similarity between two pixel numbers certificates of expression, i, j represent two different pixels, and N represents the data amount check on diagonal line, k representation feature vector number.

The choice for use heuristic of described k, namely all less from the eigenwert of the 1st to m, since m+1, eigenwert has the change on the order of magnitude large, and k gets m.

Cluster in described step 4) adopts K-means algorithm or Mean-shift algorithm or GMM algorithm.

Described α value estimates to adopt the α value method of estimation based on three-dimensional fuzzy connectedness, be used for calculating the fuzzy connectedness between zone of ignorance and known region pixel, and according to basic figure formula C=α F+ (1-α) B that scratches, obtain the stingy figure result of key frame foreground object.

The beneficial effect that the present invention is compared with the prior art is: because be background dot before mutual mark on each key frame bunch, take full advantage of the space time information that video sequence comprises, naturality and the intuitive of user interactions have been guaranteed based on the user interface of key frame, the observation habit that meets the human visual system, combine simultaneously the temporal information of video flowing, stack forms key frame bunch, replace the mark on key frame separately by the mark on key frame bunch, for the larger situation of foreground object activities, stronger robustness is arranged.

In a preferred technical scheme, also comprise the step that described stingy figure result is passed to whole video sequence, comprise that a. detects the foreground object edge at key frame, b. arranges the search window along described edge, and c. finds the described search window of mark front and back frame coupling with the SURF feature point detecting method.Because the SURF detection method has been used loaded down with trivial details summation process repeatedly in the general rectangular area of integrogram replacement, greatly reduced calculated amount on the basis that guarantees matching performance, accelerated processing speed.

In a further preferred technical scheme, described stingy figure result is passed in the step of whole video sequence, if prospect is single or separate moving object, the video flowing of assignment both direction propagation forwards, backwards, can guarantee that by two-way propagation the movable information of foreground object is by the more complete whole video sequence that is passed to; If in video sequence, prospect is a plurality of objects that relative motion is arranged, the video flowing of assignment is only propagated from front to back, can avoid when propagating stingy figure information due to the wrong backpropagation information of overlapping generation in the relative motion of object process.

Description of drawings

Fig. 1 is the use procedure block diagram of an embodiment of disposal route of the present invention.

Fig. 2 is that the graph key frame extracts schematic diagram.

Fig. 3 is the schematic diagram that 6 adjacent pixels of connection pixel in video sequence are set up the three-dimensional space-time model.

Fig. 4 is the stingy figure result schematic diagram of a pixel of the zone of ignorance of an embodiment.

Fig. 5 is that the video flowing of assignment in an embodiment only allows along the schematic diagram of time orientation forward-propagating.

Fig. 6 is the schematic diagram that the video flowing of assignment in another embodiment can be propagated from key frame along positive and negative both direction.

Embodiment

A kind of Interactive Image Processing method that is used for video that user interactions simple, intuitive, robustness are strong and stingy figure is effective, the method is mainly by extracting key frame in video sequence, and the frame between adjacent key frame is superposeed one by one and is formed for the key frame bunch of mutual mark, with this, prospect and background separation are obtained to scratch the figure result.The processing that it is basic and use flow process can be expressed as FB(flow block) as shown in Figure 1, carry out key frame stack step after being included in input video sequence, spectral clustering step after user interactions (input), step based on the estimation of fuzzy connectedness α value, SURF finds the step of corresponding search window, and the step of adaptive video stream assignment.

Below contrast accompanying drawing and in conjunction with preferred embodiment, preferred embodiment of the present invention be explained in detail.

1. key frame stack

the input original video sequence, key frame in abstraction sequence, frame between adjacent key frame is superposeed one by one, form some groups of key frames bunch, the user is on each key frame bunch before mutual mark, background dot, take full advantage of the space time information that video sequence comprises, naturality and the intuitive of user interactions have been guaranteed based on the user interface of key frame, the observation habit that meets the human visual system, combine simultaneously the temporal information of video flowing, stack forms key frame bunch, replace the mark on key frame separately by the mark on key frame bunch, for the larger situation of foreground object activities, stronger robustness is arranged.Be described as follows:

Input video sequence, extract key frame, and the key frame that uses in this algorithm is to extract a frame every 10 frames, as shown in Figure 2.Those skilled in the art can extract the interval of key frame according to the motion conditions adjustment of foreground object, generally can select between 10～20 frames, all can realize goal of the invention of the present invention.Such as,, in the situation that the foreground object motion is little, can extract a frame every 20 frames.Frame between the key frame of front and back is superposeed one by one, form a series of key frames bunch, as Fig. 2, the user marks on each key frame bunch before, background dot, through after user interactions, can be divided into three zones on each key frame images: foreground area, background area and zone of ignorance, the main target of follow-up stingy figure are namely the color distribution of determining the zone of ignorance pixel.

2. standardization spectral clustering

The Application standard spectral clustering carries out cluster to the key frame picture after aforementioned key frame stack, the standardization spectral clustering utilizes the similar matrix of sample data to carry out feature decomposition, then carry out cluster with the proper vector that obtains, only need the similarity matrix of image data just can complete cluster, usually represent original data with the proper vector unit of data, played important dimensionality reduction effect, can identify the sample space of arbitrary shape and converge on globally optimal solution, and computation complexity is less than general clustering algorithm, and performance is particularly evident on high dimensional data.

Concrete steps are as follows:

1) become a figure G=(V, E) according to data configuration, wherein vertex set and the Bian Ji of V and E difference presentation graphs G, scheme corresponding data point of each pixel of G.Similar point is coupled together, suppose that two different pixels points of limit e are i and j, weight is w _i,j, w _i,jBe used for the similarity between two pixel numbers certificates of expression.According to the similarity definition, build similarity matrix W by raw data set,

$W_{i,j}=\left\{\begin{array}{c}{w}_{i,j},i\≠j\\ 0,i=j\end{array}\right.,$

2) each column element addition of similarity matrix W is obtained the N number, building one, to form diagonal line, other elements by this N number be all 0 N * N matrix D,

$D_{i,j}=\left\{\begin{array}{c}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{i,j},i=j\\ 0,i\≠j\end{array}\right.,$

3) by similarity matrix build Laplacian Matrix L=D ?W,

$L_{i,j}=\left\{\begin{array}{c}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{i,j},i=j\\ -w_{i,j},i\≠j\end{array}\right.,$

To obtain after the L standardization

$L^{\′}=D^{-\frac{1}{2}}{\mathrm{LD}}^{-\frac{1}{2}},$

Solve front k (arranging from small to large) eigenwert of L '

And characteristic of correspondence vector

Will

Multiply by one by one

Obtain front k proper vector of matrix L.Wherein, N and k are fixed values to every two field picture, and N is the data amount check on diagonal line, and k is the proper vector number.

A preferred way is: select the process of k to use heuristic,, if namely all less from the eigenwert of the 1st to m, since m+1 eigenwert, have the change on the order of magnitude large, k gets m.

4) k proper vector lined up and formed the matrix of a N * k, regard every delegation wherein as in the k dimension space a vector, and use general clustering algorithm to carry out cluster, such as K ?means algorithm etc., the classification in cluster result under every delegation is the affiliated classification of pixel in former figure G.

Need to prove: adopt general spectral clustering can realize goal of the invention of the present invention, in this step 2, the accepted standard spectral clustering is in order conveniently to calculate and/or obtain better technique effect, also namely 3) standardization in is selectable operation.

3. estimate based on the α value of three-dimensional fuzzy connectedness

In video sequence, each pixel of each key frame has 6 neighborhood pixels: comprise the neighborhood pixels on space in 4 same frames, temporal neighborhood pixels between 2 front and back frames.So just built the three-dimensional space-time model of video sequence, calculate on this basis the fuzzy connectedness between zone of ignorance and known region pixel, it is the maximal value of a section of similarity minimum in the point-to-point access path, the connection degree design of this being similar to " wooden barrel short slab " makes when having calculated zone of ignorance a bit after the fuzzy connectedness of known region, result of calculation before other points of the same area can be continued to use to the connection degree of known region, thereby can greatly reduce calculated amount and algorithm required time.

Specifically: connect 6 neighborhood pixels of each pixel in video sequence, set up the three-dimensional space-time model as shown in Figure 3, calculate the fuzzy connectedness FC between zone of ignorance and known region pixel in this model, suppose that the pixel that zone of ignorance will calculate is p ₁, a certain pixel of known region is q ₁, p ₁With q ₁Between fuzzy connectedness FC be:

FC(p ₁,q ₁)=maxmin{μ _κ(p ₁,r),μ _κ(r,q ₁)}

Wherein, r is p ₁To q ₁Any point on path, and μ _κBe two similarities between pixel:

${\mathrm{\μ}}_{\mathrm{\κ}}(x,y)=\exp \{-\frac{1}{2}[I\left(x\right)-I\left(y\right){]}^T{\mathrm{\Σ}}^{-1}[I\left(x\right)-I\left(y\right)]\},$

I(x), I(y) the three-dimensional color vector of expression pixel x, y, what T represented is matrix transpose.

Can obtain fast pixel p in image thus ₁The α value:

$\mathrm{\α}\left(p_1\right)=\frac{F{C}^f\left(p_1\right)}{F{C}^f\left(p_1\right)+F{C}^b\left(p_1\right)},$

Wherein, FC ^fAnd FC ^bRespectively p ₁To prospect and p ₁Fuzzy connectedness to the background known region.

After trying to achieve α (opacity) value of zone of ignorance pixel, just being easy to can be according to basic stingy figure formula: C=α F+ (1-α) B obtains the color distribution of zone of ignorance pixel, has namely obtained the stingy figure result of key frame foreground object, as Fig. 4.Wherein, C, F, B represent respectively the vector that a RGB three-dimensional value forms.

While calculating fuzzy connectedness FC, the connection degree design owing to being similar to " wooden barrel short slab ", make the process that simplification can be arranged in calculation procedure: as Fig. 4, when calculating p ₁After point arrives all fuzzy connectedness in other zones, FC (p ₁, q ₁) and FC (p ₁, p ₂) be all known.Based on easily knowing for the mathematical design of fuzzy connectedness FC before, at any three pixels (three pixels in the three-dimensional space-time model of above-mentioned foundation, be also three spatial neighbor pixels in above-mentioned figure G), for summit, fuzzy connectedness FC between any two are in the spatial triangle that forms of limit, certainly exist that both sides equate and less than the structural relation on the 3rd limit.In Fig. 4, because the definition of fuzzy connectedness, FC(p1, q1) refer on a certain path from p1 to q1, find out the most weak one section of weight, namely be equivalent to the short slab of a wooden barrel, the short slab of more every paths, find out a paths of short slab maximum, the short slab of that on this paths is as the FC that will calculate, and namely finds out the short slab length of the strongest wooden barrel of short slab.Get back in Fig. 4, in the situation of known FC (q1, p1) and FC (p1, p2), according to the transitivity on mathematical definition, it is less in both that FC (q1, p2) necessarily equals.Also namely, if both are unequal, FC (q1, p1) equals less in FC (p1, p2) and FC (q1, p2), if FC is (p ₁, q ₁)＜FC (p ₁, p ₂), FC (p ₂, q ₁)=FC (p ₁, q ₁); If FC is (p ₁, q ₁) FC (p ₁, p ₂), FC (p ₂, q ₁)=FC (p ₁, p ₂); Only have when both equate, i.e. FC (p ₁, q ₁)=FC (p ₁, p ₂) time, FC (p ₂, q ₁) need to recalculate, only under the use of this skill, just calculated amount can be reduced to and originally travel through one by one 1/3 of pixel calculating.

4.SURF find corresponding search window

Obtain the stingy figure result of video sequence key frame via above step, need this result is passed to whole video sequence.At first, detect the edge of foreground object on key frame, and along described edge, the search window is set; Then, use SURF point-of-interest detection method to find the search window of mark front and back frame coupling.Because the SURF detection method has been used loaded down with trivial details summation process repeatedly in the general rectangular area of integrogram replacement, greatly reduced calculated amount on the basis that guarantees matching performance, accelerated algorithm speed.Specifically describe as follows:

Obtain the stingy figure result of key frame via above step after, use the Sobel operator (in rim detection, a kind of template commonly used is the Sobel operator, the Sobel operator has two, one is the detection level edge, another is the detection of vertical edge) detect the edge line of foreground object, clockwise every the central point of n pixel selected point as the search window on described edge line, travel through several square search windows of whole edge line extraction, the length of side of search window is generally got 1/10 of the minimum boundary rectangle length of side of foreground object, and n gets half of the search window length of side.

After setting up the search window, select SURF to detect search window corresponding to front and back frame, follow-up video flowing assignment procedure is all carried out in window in each search, and the stingy figure result in subsequent step in search window that key frame is corresponding assignment is in turn given search window corresponding to other intermediate frames (the corresponding window that the search window that described intermediate frame is corresponding is namely obtained by the coupling of the search window in key frame by above step).Wherein the concrete steps of SURF detection are as follows:

1) calculated product component;

2) build the Hessian matrix;

3) build metric space;

4) accurate location feature point;

5) determine rotation principal direction;

6) calculate SURF unique point descriptor.

5. adaptive video flows assignment

The stingy figure result that will search for window by key frame in the process of other intermediate frame assignment, if scratching prospect in the figure result is single or separate moving object, the video flowing of assignment should be propagated from former and later two directions, as shown in Figure 5, guarantee that by two-way propagation the movable information of foreground object is by the more complete whole video sequence that is passed to; If in video sequence, foreground object is a plurality of objects that relative motion is arranged, the video flowing of assignment should only be propagated from front to back, as shown in Figure 6, can avoid when propagating stingy figure information due to the overlapping backpropagation information that produces mistake in the relative motion of object process.Described intermediate frame refers to each frame between each adjacent key frame.

At first, judge whether foreground object is a plurality of objects that relative motion is arranged, if so, as shown in Figure 5, the video flowing that assignment is set only allows along the time orientation forward-propagating; If not, as shown in Figure 6, the video flowing that assignment is set can be propagated along positive and negative both direction from key frame.

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 the general technical staff of the technical field of the invention, without departing from the inventive concept of the premise, make some being equal to substitute or obvious modification, and performance or purposes identical, all should be considered as belonging to protection scope of the present invention.

Claims (10)

1. Interactive Image Processing method that is used for video, it is characterized in that: extract key frame in video sequence, and the frame between adjacent key frame is superposeed one by one and is formed for the key frame bunch of mutual mark, be divided into foreground area, background area and zone of ignorance with this image area with described key frame; Then described key frame is carried out spectral clustering and the estimation of α value, obtain the stingy figure result of described key frame; Finally, the stingy figure result of described key frame is passed to whole video sequence, obtains final stingy figure result.

2. the method for the Interactive Image Processing for video as claimed in claim 1, it is characterized in that: described stingy figure result with described key frame is passed to whole video sequence and comprises: a. detects the edge line of foreground object at key frame, b. travel through whole described edge line several search windows are set, c. find the described search window of frame coupling before and after mark with the SURF feature point detecting method, the stingy figure result in d. search window that described key frame is corresponding assignment is in turn given search window corresponding to other intermediate frame.

3. the method for the Interactive Image Processing for video as claimed in claim 2 is characterized in that: in described steps d, if prospect is single or separate moving object, the video flowing of assignment forwards, backwards both direction propagate; If in video sequence, prospect is a plurality of objects that relative motion is arranged, the video flowing of assignment is only propagated from front to back.

4. the method for the Interactive Image Processing for video as claimed in claim 2, is characterized in that described SURF feature point detecting method comprises the steps:

1) calculated product component;

2) build the Hessian matrix;

3) build metric space;

4) location feature point;

5) determine rotation principal direction;

6) calculate SURF unique point descriptor.

5. the method for the Interactive Image Processing for video as claimed in claim 1 is characterized in that: described extraction key frame is to extract frames as key frame every 10～20 frames.

6. the method for the Interactive Image Processing for video as claimed in claim 1, is characterized in that: described spectral clustering employing standardization Spectral Clustering.

7. the method for the Interactive Image Processing for video as claimed in claim 6, it is characterized in that: the step of described standardization spectral clustering is as follows:

1) build similarity matrix W by raw data set,

$W_{i,j}=\left\{\begin{array}{c}{w}_{i,j},i\≠j\\ 0,i=j\end{array}\right.,$

2) each column element addition of W is obtained the N number, build one and form diagonal line by this N number, other elements are all 0 N * N matrix D,

$D_{i,j}=\left\{\begin{array}{c}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{i,j},i=j\\ 0,i\≠j\end{array}\right.;$

3) build similarity matrix W by raw data set and build Laplacian Matrix L=D-W, $L_{i,j}=\left\{\begin{array}{c}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{w}_{i,j},i=j\\ -w_{i,j},i\≠j\end{array}\right.,$

To obtain after the L standardization

$L^{\′}=D^{-\frac{1}{2}}{\mathrm{LD}}^{-\frac{1}{2}},$

Solve front k the eigenwert of L ' And characteristic of correspondence vector

Will

Multiply by one by one

Obtain front k proper vector of matrix L;

4) k proper vector lined up and formed the matrix of a N * k, regard every delegation wherein as in the k dimension space a vector, carry out cluster;

Wherein, w _i,jBe used for the similarity between two pixel numbers certificates of expression, i, j represent two different pixels, and N represents the data amount check on diagonal line, k representation feature vector number.

8. the method for the Interactive Image Processing for video as claimed in claim 7, it is characterized in that: the choice for use heuristic of described k, even all less from the eigenwert of the 1st to m, since m+1, eigenwert has the change on the order of magnitude large, and k gets m.

9. the method for the Interactive Image Processing for video as claimed in claim 7, is characterized in that: cluster employing K-means algorithm or Mean-shift algorithm or GMM algorithm in described step 4).

10. the method for the Interactive Image Processing for video as claimed in claim 1, it is characterized in that: described α value estimates to adopt the α value method of estimation based on three-dimensional fuzzy connectedness, be used for calculating the fuzzy connectedness between zone of ignorance and known region pixel, and according to basic figure formula C=α F+ (1-α) B that scratches, obtain the stingy figure result of key frame foreground object.

Images