CN1477566A - Method for making video search of scenes based on contents - Google Patents

Abstract

The present invention relates to a method of making video search of scene based on contents. It uses the fuzzy aggregation analysis method in the scene search. As compared with existent method it can obtain higher accuracy and quick searching speed.

Description

A kind of method of camera lens being carried out Content-based Video Retrieval

Technical field

The invention belongs to video search technique area, be specifically related to a kind of method of camera lens being carried out Content-based Video Retrieval.

Background technology

Along with in the great technical progress that obtains aspect multi-medium data manufacturing, storage and the propagation, digital video has become a part indispensable in people's daily life.The problem that people face no longer is to lack content of multimedia, but how to find own needed information in the vast as the open sea multimedia world.At present, traditional video frequency searching of describing based on keyword is because descriptive power is limited, and subjectivity is strong, reasons such as manual mark, demand that can not the satisfying magnanimity video frequency searching.In order to be convenient for people to seek multi-medium data, begin the nineties in last century, content-based video analysis and retrieval technique become the hot issue of research, and Multimedia Content Description Interface MPEG-7 progressively formulates and optimizes, and have promoted the development of Content-based Video Retrieval technology more.

In the prior art, as document " A New Approach to Retrieval Video by ExampleVideo Clip " [X.M.Liu, Y.T.Zhuang, and Y.H.Pan, ACM Multimedia, pp.41-44,1999] described, the conventional method of video frequency searching is at first to carry out shot boundary to detect, with basic structural unit and the retrieval unit of camera lens as video sequence; Represent the content of this camera lens then at each camera lens internal extraction key frame, go out low-level features such as color and texture from key-frame extraction, be used for the index and the retrieval of camera lens.Like this, just content-based searching lens being converted into CBIR solves.The problem that these class methods exist is that camera lens is an image continuous sequence in time, temporal information and the movable information that is present in the video is not fully utilized.(document author is s.H.Kim and R.-H.Park to the document of delivering at IEEE Trans.Circuits and Systems for Video Technology in 2002 " An efficient algorithm forvideo sequence matching using the modified Hausdorff distance and the directeddivergence " in addition, vol.CSVT-12, no.7, page number 592-595) disperses (Cumulative Directed Divergence) method with the orientation of accumulation and extract key frame, obtain two similarity degrees between the camera lens with improved Hausdorff distance (Modified Hausdorff Distance) method, used the YUV color space histogram when extracting key frame and definition camera lens similarity.Owing to set two threshold values when extracting key frame: the threshold value of similar value between the threshold value of front and back frame similar value and present frame and the previous key frame, these two conditions must be satisfied simultaneously and a key frame could be occurred, will influence the accuracy of key-frame extraction like this, finally will certainly influence the correctness of inquiry; In addition, used the YUV color space of using always in the video as visual signature, it is compared with the hsv color space, and is also little consistent with people's visually-perceptible.

Summary of the invention

At the existing defective of existing searching lens method, the objective of the invention is to propose a kind of method of camera lens being carried out Content-based Video Retrieval, this method can improve the accuracy rate of content-based searching lens on the basis of existing technology greatly, keep very fast retrieval rate simultaneously, thereby bring into play the huge effect of searching lens technology in current network information society more fully.

The object of the present invention is achieved like this: a kind of camera lens is carried out the method for Content-based Video Retrieval, may further comprise the steps:

(1) at first video database is carried out camera lens and cut apart, with the basic structural unit and the retrieval unit of camera lens as video;

(2) similarity between two two field pictures of calculating is set up fuzzy similarity matrix R by following method: when i=j, make r _IjBe 1; When i ≠ j, make r _IjBe x _iWith y _jBetween similarity;

(3) utilize the transitive closure method to calculate the equivalent matrice of fuzzy similarity matrix R

(4) threshold value λ is set and determines cut set, the transitive closure matrix of R matrix

Carry out fuzzy clustering, calculate $\left[x\right]=\left\{y\right|\hat{R}(x,y)\≥\mathrm{\λ}\},$ Set [x] is the equivalence class of fuzzy clustering, and each frame is similar in each equivalence class set, so we can get in each set arbitrary frame as key frame;

(5) with key frame { r _I1, r _I2..., r _IkExpression camera lens s _i, gather with key frame and to measure two similaritys between the camera lens.

Furthermore, in the step (1) video database is carried out the method that camera lens cuts apart and be preferably space-time section algorithm.Calculate x in the step (2) _iWith y _jBetween similarity can calculate with the friendship of two image histograms: $\mathrm{Inter}\sec t(x_i,y_j)=\frac{1}{A(x_i,y_j)}\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}\min \{H_i(h,s,v),H_j(h,s,v)\}$ $A(x_i,x_j)=\min \{\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_i(h,s,v),\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_j(h,s,v)\}$

H _i(h, s v) are the histograms in hsv color space, and we use H, S, the V component is statistic histogram in 18 * 3 * 3 three dimensions, with 162 numerical value after the normalization as color feature value, Intersect (x _i, y _j) two histogrammic friendships of expression, judge the similarity of two key frames with it, use A (x _i, y _j) normalize between 0,1.

Further again, in the step (3), calculate the equivalent matrice of fuzzy similarity matrix R The transitive closure method can adopt quadratic method: $R\→R^2\→{\left(R^2\right)}^2\→\·\·\·\→R^{2^k}=\hat{R},$ Its time complexity is O (n ³Log ₂N), if the n value is big especially, will certainly influence total computing time, so adopt the compose operation of the fuzzy clustering optimal algorithm compute matrix that calculates based on figure connected component, recursion is as follows: $r_{\mathrm{ij}}^{\left(0\right)}=r_{\mathrm{ij}}---0\≤i,j\≤n$ $r_{\mathrm{ij}}^{\left(k\right)}=\max \{r_{\mathrm{ij}}^{(k-1)},\min [r_{\mathrm{ik}}^{(k-1)},r_{\mathrm{kj}}^{(k-1)}\left]\right\}----0\≤i,j\≤n;0\≤k\≤n$

The time complexity T (n) of this algorithm satisfies O (n)≤T (n)≤O (n ²).

In order to realize purpose of the present invention better, when carrying out searching lens, right

The method of carrying out fuzzy clustering is as follows:

(1) determines n sample X=(X ₁..., X _n) on fuzzy resembling relation R and a cut set threshold alpha;

(2) transform R as an equivalent matrice by following calculating;

RoR＝R ²

R ²oR ²＝R ⁴

... $R^{2^k}o{R}^{2^k}=R^{2^{(K+1)}}$

Up to existing a k to satisfy $R^{2^k}=R^{2^{(k+1)}}$

In the above-mentioned formula, RoR is the compose operation of fuzzy relation, is under the hypothesis of similar matrix at R, and having proved to have such k to exist, and satisfies k≤log n;

(3) set of computations $\left[x\right]=\left\{y\right|\hat{R}(x,y)\≥\mathrm{\α}\},$ [x] is fuzzy clustering, and algorithm finishes;

After n sample space carried out fuzzy cluster analysis, obtain several equivalence classes, in each equivalence class, choose a sample as key frame.Measuring similarity between such two camera lenses just becomes the similarity measurement between the key frame set.

In the step (5) of this method, can be camera lens s _iAnd s _jSimilarity be defined as $\mathrm{Sim}(s_i,s_j)=\frac{1}{2}\{M(s_i,s_j)+\hat{M}(s_i,s_j)\},$ M represents the maximal value that key frame is similar, The similar second largest value of expression key frame, wherein, $M(s_i,s_j)=\underset{p=\{\mathrm{1,2},...\}}{\max }\underset{q=\{\mathrm{1,2},...\}}{\max }\left\{\mathrm{Inter}\sec t(r_{\mathrm{ip}},r_{\mathrm{jq}})\right\}$ $\hat{M}(s_i,s_j)=\underset{p=\{\mathrm{1,2},...\}}{\overset{^}{\max }}\underset{q=\{\mathrm{1,2},...\}}{\max }\left\{\mathrm{Inter}\sec t(r_{\mathrm{ip}},r_{\mathrm{jq}})\right\}$ $\mathrm{Inter}\sec t(r_i,r_j)=\frac{1}{A(r_i,r_j)}\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}\min \{H_i(h,s,v),H_j(h,s,v)\}$ $A(r_i,r_j)=\min \{\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_i(h,s,v),\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_j(h,s,v)\}.$

Effect of the present invention is: adopt and of the present invention camera lens is carried out the method for Content-based Video Retrieval, can obtain higher accuracy rate, keep very fast retrieval rate simultaneously.

Why the present invention has so significant technique effect, its reason is: the method for utilization fuzzy cluster analysis, the camera lens content is divided into a plurality of equivalence classes, these equivalence classes have well been described the camera lens content change, and the similarity between the camera lens then shows as the similarity between the key frame combination.Similarity measurement has been considered to use the hsv color histogram to represent the shortcoming of key frame between the camera lens: if two key frames have similar color distribution, even their content is different, can think that also these two key frames are similar.So the mean value that uses maximum similar value and second largest similar value adds the robustness of strong algorithms.Contrast and experiment has confirmed that the present invention proposes the validity of method.

Description of drawings

Fig. 1 is the schematic flow sheet that camera lens is carried out the method for Content-based Video Retrieval;

Fig. 2 is 7 semantic category example synoptic diagram of searching lens in the experiment contrast;

Fig. 3 is the result for retrieval synoptic diagram of method of the present invention to the swimming camera lens.

Embodiment

Fig. 1 is an overall framework of the present invention, is the schematic flow sheet of each one step process among the present invention.As shown in Figure 1, a kind of searching lens method based on fuzzy cluster analysis may further comprise the steps:

1, camera lens is cut apart

At first use space-time section algorithm (spatio-temporal slice), video database is carried out camera lens to be cut apart, with basic structural unit and the retrieval unit of camera lens as video, can list of references " Video Partitioning by Temporal Slice Coherency " [C.W.Ngo about the detailed description of space-time section algorithm, T.C.Pong, and R.T.Chin, IEEE Transactions on Circuits andSystems for Video Technology, Vol.11, No.8, pp.941-953, August, 2001].

2, set up fuzzy similarity matrix R

Set up between the camera lens internal image to set up fuzzy similarity matrix R method as follows: when i=j, make r _IjBe 1, when i ≠ j, make r _IjBe x _iWith y _jBetween similarity, similarity then adopts following method to calculate: $\mathrm{Inter}\sec t(x_i,y_j)=\frac{1}{A(x_i,y_j)}\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}\min \{H_i(h,s,v),H_j(h,s,v)\}$ $A(x_i,x_j)=\min \{\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_i(h,s,v),\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_j(h,s,v)\}$

H _i(h, s v) are the histograms in hsv color space, and we use H, S, the V component is statistic histogram in 18 * 3 * 3 three dimensions, with 162 numerical value after the normalization as color feature value.Intersect (x _i, y _j) two histogrammic friendships of expression, judge the similarity of two key frames with it, use A (x _i, y _j) normalize between 0,1.

3, ask the transitive closure of similar matrix R, obtain equivalent matrice

In the present embodiment, ask the transitive closure of similar matrix to adopt quadratic method: $R\→R^2\→{\left(R^2\right)}^2\→\·\·\·\→{R^2}^k=\hat{R}$

Its time complexity is O (n ³Log ₂N), if the n value is big especially, will certainly influence total computing time.So adopt the compose operation of the fuzzy clustering optimal algorithm compute matrix that calculates based on figure connected component, recursion is as follows: $r_{\mathrm{ij}}^{\left(0\right)}=r_{\mathrm{ij}}---0\≤i,j\≤n$ $r_{\mathrm{ij}}^{\left(k\right)}=\max \{r_{\mathrm{ij}}^{(k-1)},\min [r_{\mathrm{ik}}^{(k-1)},r_{\mathrm{kj}}^{(k-1)}\left]\right\}----0\≤i,j\≤n;0\≤k\≤n.$

The time complexity T (n) of this algorithm satisfies O (n)≤T (n)≤O (n ²).

4, threshold value λ is set and determines cut set, the transitive closure matrix of R matrix Carry out fuzzy clustering.

In the present embodiment, concrete grammar is as follows:

(1) determines n sample X=(x ₁..., x _n) on fuzzy resembling relation R and a cut set threshold alpha;

(2) transform R as an equivalent matrice by following calculating;

RoR＝R ²

R ²oR ²＝R ⁴

... $R^{2^k}o{R}^{2^k}=R^{2^{(K+1)}}$

Up to existing a k to satisfy $R^{2^k}=R^{2^{(k+1)}}$

In the above-mentioned formula, RoR is the compose operation of fuzzy relation, is under the hypothesis of similar matrix at R, and having proved to have such k to exist, and satisfies k≤log n;

(3) set of computations $\left[x\right]=\left\{y\right|\hat{R}(x,y)\≥\mathrm{\α}\},$ [x] is fuzzy clustering, and algorithm finishes

5, obtain the camera lens key frame with Fuzzy Cluster Analysis method after, carry out searching lens based on these key frames then.On this basis, with key frame { r _I1, r _I2..., r _Ik) the expression camera lens, s _iCamera lens s _iAnd s _jSimilarity be defined as $\mathrm{Sim}(s_i,s_j)=\frac{1}{2}\{M(s_i,s_j)+\hat{M}(s_i,s_j)\},$ Wherein, $M(s_i,s_j)=\underset{p=\{\mathrm{1,2},...\}}{\max }\underset{q=\{\mathrm{1,2},...\}}{\max }\left\{\mathrm{Inter}\sec t(r_{\mathrm{ip}},r_{\mathrm{jq}})\right\}$ $\hat{M}(s_i,s_j)=\underset{p=\{\mathrm{1,2},...\}}{\overset{^}{\max }}\underset{q=\{\mathrm{1,2},...\}}{\max }\left\{\mathrm{Inter}\sec t(r_{\mathrm{ip}},r_{\mathrm{jq}})\right\}$ $\mathrm{Inter}\sec t(r_i,r_j)=\frac{1}{A(r_i,r_j)}\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}\min \{H_i(h,s,v),H_j(h,s,v)\}$ $A(r_i,r_j)=\min \{\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_i(h,s,v),\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_j(h,s,v)\}.$

Represent second largest value, use

Be that its shortcoming is if two key frames have similar color distribution, even their content is different, can think that also these two key frames are similar because this paper uses the hsv color histogram to represent key frame, in order to overcome this defective, use M and

Mean value add the robustness of strong algorithms.H _i(h, s v) are the histograms in hsv color space, this paper H, S, the V component is statistic histogram in 18 * 3 * 3 three dimensions, with 162 numerical value after the normalization as color feature value.Intersect (r _i, r _j) two histogrammic friendships of expression, this paper judges the similarity of two key frames with it.

Following experimental result shows that the present invention has obtained than the better effect of existing method, and retrieval rate is very fast simultaneously, has confirmed the validity of fuzzy cluster analysis algorithm in searching lens.

The experimental data of searching lens is the Asian Games programs in 2002 from television recording, always has 41 minutes, 777 camera lenses, 62132 two field pictures.It comprises multiple sports items, as various ball game, weight lifting, swim and the advertising programme that intercuts etc.We have selected 7 semantic categories as the inquiry camera lens, and they are weight lifting, vollyball, swimming, judo, row the boat, gymnastics, football, as shown in Figure 2.

In order to verify validity of the present invention, we have tested following 3 kinds of methods and have done the experiment contrast:

(1) the first frame of Chang Yong each camera lens of use is done the searching lens algorithm of key frame;

(document author is s.H.Kim and R.-H.Park to the document of delivering at IEEE Trans.Circuits and Systems for Video Technology in (2) 2002 years " An efficient algorithm for video sequence matching using the modifiedHausdorffdistance and the directed divergence ", vo1.CSVT-12, no.7, page number 592-595) the middle algorithm of describing;

(3) use the fuzzy cluster analysis algorithm to obtain key frame and carry out searching lens (only using color characteristic);

Above-mentioned preceding 3 kinds of methods have all only been used color characteristic, and therefore last experimental result can prove the superiority of the disclosed method of the present invention from the measure of shot similarity.Fig. 3 has provided the user interface of experimental arrangement, delegation is the browsing area of inquiry video above the right, and the 1st of each camera lens the key frame is used for representing each camera lens in the display video, the camera lens that the user can therefrom select to want to inquire about is retrieved, and is the Query Result zone below the right.Fig. 3 is the 1st camera lens of delegation above selecting, it is a swimming camera lens, and the first two field picture 022430.bmp represents by this camera lens, the authority of the similarity that calculates according to method of the present invention, arrange Query Result (from left to right, arranging from top to bottom) from big to small.Below, the left side is a simple and easy broadcast phase, double-clicks that section video that the result for retrieval image can be play corresponding camera lens correspondence.

Two kinds of evaluation indexes in the MPEG-7 standardization activity have been adopted in experiment: the average adjusted retrieval order of normalization ANMRR (average normalized modified retrieval rank) and recall level average AR (average recall).AR is similar to traditional recall ratio (recall), and ANMRR compares with traditional precision ratio (precision), not only can reflect correct result for retrieval ratio, and can reflect correct result's arrangement sequence number.The ANMRR value is more little, means that the rank of the correct camera lens that retrieval obtains is forward more; The AR value is big more, and it is big more to mean that in the individual Query Result of preceding K (K is the cutoff value of result for retrieval) similar camera lens accounts for the ratio of all similar camera lenses.Table 1 is AR and the ANMRR comparisons of above-mentioned 3 kinds of methods to 7 semantic camera lens classes.

Table 1 the present invention and the contrast and experiment that has two kinds of methods now

Classification	Method 1		Method 2		Method 3
							????AR	????ANMRR	????AR	????ANMRR	????AR	????ANMRR
	Weight lifting	????0.8824	????0.3098	????0.8824	????0.1539	????0.9412							????0.2186
Vollyball							????0.6333	????0.4974	????0.7895	????0.3264	????0.8556	????0.3279
	Swimming	????0.8400	????0.2676	????0.8250	????0.3164	????0.9200							????0.2175
Judo							????0.7000	????0.4310	????0.8214	????0.2393	????0.8000	????0.3093
	Row the boat	????0.8750	????0.3407	????0.6875	????0.3570	????0.8125							????0.2223
Gymnastics							????0.7857	????0.3445	????0.9600	????0.1759	????0.7857	????0.2056
	Football	????0.5789	????0.4883	????0.6889	????0.2815	????0.8421							????0.2614
Average							????0.7565	????0.3827	????0.8078	????0.2642	????0.8510	????0.2518

No matter as can be seen from Table 1, adopt method of the present invention, be AR, or ANMRR, all obtained than existing two kinds of better effects of algorithm, confirmed that the present invention is used for the Fuzzy Cluster Analysis method method validity of searching lens.The method of method utilization fuzzy cluster analysis of the present invention is divided into a plurality of equivalence classes to the camera lens content, and these equivalence classes have well been described the camera lens content change, and the similarity between the camera lens then shows as the similarity between the key frame combination.Similarity measurement has been considered to use the hsv color histogram to represent the shortcoming of key frame between the camera lens: if two key frames have similar color distribution, even their content is different, can think that also these two key frames are similar.So the mean value that uses maximum similar value and second largest similar value adds the robustness of strong algorithms.Contrast and experiment has confirmed that the present invention proposes the validity of method.In addition, at CPU 500M PIII, on the PC of 256 MB of memory, algorithm ART of the present invention is 22.557 seconds, and for the video library of 777 camera lenses, the retrieval rate of two kinds of algorithms of the present invention all is very fast.

Claims (6)

1, a kind of camera lens is carried out the method for Content-based Video Retrieval, it is characterized in that this method may further comprise the steps:

(1) at first video database is carried out camera lens and cut apart, with the basic structural unit and the retrieval unit of camera lens as video;

(2) similarity between two two field pictures of calculating is set up fuzzy similarity matrix R by following method: when i=j, make r _IjBe 1; When i ≠ j, make r _IjBe x _iWith y _j, between similarity;

(3) utilize the transitive closure method to calculate the equivalent matrice of fuzzy similarity matrix R

(4) threshold value λ is set and determines cut set, the transitive closure matrix of R matrix

Carry out fuzzy clustering, calculate $\left[x\right]=\left\{y\right|\hat{R}(x,y)\≥\mathrm{\λ}\},$ Set [x] is the equivalence class of fuzzy clustering, and each frame is similar in each equivalence class set, so we can get in each set arbitrary two field picture as key frame;

(5) with key frame { r _I1, r _I2..., r _IkExpression camera lens s _i, key frame is gathered and is measured two similaritys between the camera lens.

2, as claimed in claim 1ly a kind of camera lens is carried out the method for Content-based Video Retrieval, it is characterized in that: in the step (1), it is space-time section algorithm that video database is carried out the method that camera lens cuts apart.

3, as claimed in claim 1ly a kind of camera lens is carried out the method for Content-based Video Retrieval, it is characterized in that: in the step (2), calculate x _iWith y _jBetween similarity can calculate with the friendship of two image histograms: $\mathrm{Inter}\sec t(x_i,y_j)=\frac{1}{A(x_i,y_j)}\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}\min \{H_i(h,s,v),H_j(h,s,v)\}$ $A(x_i,x_j)=\min \{\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_i(h,s,v),\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_j(h,s,v)\}$

H _i(h, s v) are the histograms in hsv color space, and we use H, S, the V component is statistic histogram in 18 * 3 * 3 three dimensions, with 162 numerical value after the normalization as face color characteristic value, Intersect (x _i, y _j) two histogrammic friendships of expression, judge the similarity of two key frames with it, use A (x _i, y _j) normalize between 0,1.

4, as claimed in claim 1ly a kind of camera lens is carried out the method for Content-based Video Retrieval, it is characterized in that: in the step (3), calculate the equivalent matrice of fuzzy similarity matrix R The transitive closure method adopt quadratic method: $R\→R^2\→{\left(R^2\right)}^2\→\·\·\·\→R^{2^k}=\hat{R},$ Its time complexity is O (n ³Log ₂N), if the n value is big especially, will certainly influence total computing time, so adopt the compose operation of the fuzzy clustering optimal algorithm compute matrix that calculates based on figure connected component, recursion is as follows: $r_{\mathrm{ij}}^{\left(0\right)}=r_{\mathrm{ij}}---0\≤i,j\≤n$ $r_{\mathrm{ij}}^{\left(k\right)}=\max \{r_{\mathrm{ij}}^{(k-1)},\min [r_{\mathrm{ik}}^{(k-1)},r_{\mathrm{kj}}^{(k-1)}\left]\right\}----0\≤i,j\≤n;0\≤k\≤n$

The time complexity T (n) of this algorithm satisfies O (n)≤T (n)≤O (n ²).

5, as claimed in claim 1ly a kind of camera lens is carried out the method for Content-based Video Retrieval, it is characterized in that: right The method of carrying out fuzzy clustering is as follows:

(1) determines n sample x=(x ₁..., x _n) on fuzzy resembling relation R and a cut set threshold alpha;

(2) transform R as an equivalent matrice by following calculating

RoR＝R ²

R ²oR ²＝R ⁴

...? $R^{2^k}o{R}^{2^k}=R^{2^{(k+1)}}$

Up to existing a k to satisfy $R^{2^k}=R^{2^{(k+1)}}$

In the above-mentioned formula, RoR is the compose operation of fuzzy relation, is under the hypothesis of similar matrix at R, and having proved to have such k to exist, and satisfies k≤log n;

(3) set of computations $\left[x\right]=\left\{y\right|\hat{R}(x,y)\≥\mathrm{\α}\},$ [x] is fuzzy clustering, and algorithm finishes;

After n sample space carried out fuzzy cluster analysis, obtain several equivalence classes, choose a sample as key frame in each equivalence class, the measuring similarity between such two camera lenses just becomes the similarity measurement between the key frame set.

6, describedly a kind of camera lens is carried out the method for Content-based Video Retrieval as claim 1 or 5, it is characterized in that: can be camera lens s _iAnd s _jSimilarity be defined as $\mathrm{Sim}(s_i,s_j)=\frac{1}{2}\{M(s_i,s_j)+\hat{M}(s_i,s_j)\},$ M represents the maximal value that key frame is similar, The similar second largest value of expression key frame, wherein, $M(s_i,s_j)=\underset{p=\{\mathrm{1,2},...\}}{\max }\underset{q=\{\mathrm{1,2},...\}}{\max }\left\{\mathrm{Inter}\sec t(r_{\mathrm{ip}},r_{\mathrm{jq}})\right\}$ $\hat{M}(s_i,s_j)=\underset{p=\{\mathrm{1,2},...\}}{\overset{^}{\max }}\underset{q=\{\mathrm{1,2},...\}}{\max }\left\{\mathrm{Inter}\sec t(r_{\mathrm{ip}},r_{\mathrm{jq}})\right\}$ $\mathrm{Inter}\sec t(r_i,r_j)=\frac{1}{A(r_i,r_j)}\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}\min \{H_i(h,s,v),H_j(h,s,v)\}$ $A(r_i,r_j)=\min \{\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_i(h,s,v),\underset{h}{\mathrm{\Σ}}\underset{s}{\mathrm{\Σ}}\underset{v}{\mathrm{\Σ}}{H}_j(h,s,v)\}.$