CN105913096A

CN105913096A - Extracting method for disordered image key frame

Info

Publication number: CN105913096A
Application number: CN201610510776.7A
Authority: CN
Inventors: 林靖宇; 郑恩; 潘莹; 曹绍昊
Original assignee: Guangxi University
Current assignee: Guangxi University
Priority date: 2016-06-29
Filing date: 2016-06-29
Publication date: 2016-08-31

Abstract

A method for extracting a key frame of an unordered image, characterized in that it comprises the following steps: 1) adopting a clustering algorithm that does not set a K value to cluster the unordered image, and gathering unordered images with similar image information content into 2) Solve the clustering center of each cluster according to the similarity distance, and extract the disordered image closest to the clustering center from each cluster as the representative frame F _k _; With reference to the image quality evaluation, if the representative frame F _k satisfies the stereo vision three-dimensional reconstruction quality requirements, then it is reserved as a key frame; if the representative frame F _k does not meet the stereo vision reconstruction quality requirements, then it is deleted from the disordered image, and the present invention can effectively Extract representative frames of unordered images, filter out unordered images with redundant information, less information, and poor quality, and use the clustering algorithm without K value to automatically cluster into different numbers according to the complexity of the content of unordered images Clusters, high-definition key frames can be obtained by using the no-reference image evaluation method.

Description

A Method for Extracting Key Frames from Unordered Images

技术领域technical field

本发明涉及一种计算机视觉及图像图形领域，尤其是一种无序图像关键帧的提取方法。The invention relates to the fields of computer vision and image graphics, in particular to a method for extracting key frames of disordered images.

背景技术Background technique

近年来人们对室外场景对象立体视觉三维重建的研究方兴未艾。然而用手持相机对场景对象从不同视觉、位置采集图像进行三维重建，由于采集的无序图像数目多、内容冗余大、个别图像信息量少、图像模糊等原因，给三维重建带来困难，甚至采集的图像不能进行重建。In recent years, the research on stereoscopic 3D reconstruction of outdoor scene objects is in the ascendant. However, using a hand-held camera to perform 3D reconstruction of images collected from different visions and positions of scene objects brings difficulties to 3D reconstruction due to the large number of disordered images collected, large content redundancy, small amount of individual image information, and blurred images. Even the acquired images cannot be reconstructed.

对无序图像进行关键帧提取，可以提取出图像的代表帧，滤除信息冗余、信息量少的图像，现在一般用的关键帧提取方法有：基于镜头边界的关键帧提取方法，该方法提取的关键帧不具有代表性；基于运动分析的关键帧提取方法，该方法运算复杂度比较大；基于无序图像K均值聚簇算法提取关键帧，把无序图像聚成K簇，从每簇中提取出离聚簇中心最近的一帧作为关键桢，该算法是三种方法中最好的。但无序图像K均值聚类算法有两个问题，一、运用该算法之前必须提前设定K值，无序图像聚K个簇，从每个簇中提取一个关键帧，往往事先我们并不知道无序图像中有多少个关键帧，直接设K值有一定的盲目性，有些关键帧可能不能有效的提取出来。二、提取的关键帧可能存在模糊不清晰问题，不能满足三维重建、目标检测、识别的质量要求。针对以上两个问题本发明专利提出不设K值聚簇算法和无参考图像质量评价算法可以很好的解决以上两个问题，该算法实行分层提取：第一层采用不舍K值的聚簇算法根据无序图像内容复杂度来决定簇的数目即关键帧的数目。第二层把第一层提取出的关键帧进行无参考图像质量评价，提取出的关键帧清晰度满足三维重建需要则保留，不满足则重新从原来的簇中提取关键帧，再次进行无参考图像质量评价，直到评价后的关键帧满足立体视觉三维重建质量要求为止。Extracting key frames from unordered images can extract the representative frames of the image and filter out images with redundant information and less information. Now the key frame extraction methods generally used are: the key frame extraction method based on the shot boundary, this method The extracted key frames are not representative; the key frame extraction method based on motion analysis has a relatively large computational complexity; the key frames are extracted based on the K-means clustering algorithm for unordered images, and the unordered images are clustered into K clusters, from each The frame closest to the center of the cluster is extracted from the cluster as the key frame. This algorithm is the best among the three methods. However, there are two problems with the K-means clustering algorithm for unordered images. First, the K value must be set in advance before using the algorithm. Unordered images gather K clusters, and a key frame is extracted from each cluster. Often we do not know in advance. Knowing how many key frames there are in the unordered image, directly setting the K value has a certain degree of blindness, and some key frames may not be effectively extracted. Second, the extracted key frames may be blurred and unclear, which cannot meet the quality requirements of 3D reconstruction, target detection, and recognition. In view of the above two problems, the patent of the present invention proposes a clustering algorithm without K value and a no-reference image quality evaluation algorithm, which can solve the above two problems very well. The cluster algorithm determines the number of clusters, that is, the number of key frames, according to the complexity of the unordered image content. In the second layer, the key frames extracted from the first layer are evaluated without reference image quality, and the extracted key frames are retained if they meet the requirements of 3D reconstruction; if not, the key frames are extracted from the original cluster again, and no reference Image quality evaluation until the evaluated key frame meets the quality requirements of stereoscopic 3D reconstruction.

发明内容Contents of the invention

针对现有技术的不足，本发明提供一种无序图像关键帧的提取方法。Aiming at the deficiencies of the prior art, the present invention provides a method for extracting key frames of an unordered image.

本发明的技术方案为：一种无序图像关键帧的提取方法，其特征在于，包括以下步骤：The technical solution of the present invention is: a method for extracting a key frame of an unordered image, characterized in that it comprises the following steps:

1)采用不设K值的聚簇算法对无序图像进行聚簇处理，把图像信息内容相近的无序图像聚为一簇；1) Use the clustering algorithm without setting the K value to cluster the unordered images, and gather the unordered images with similar image information content into one cluster;

2)根据相似距离求解每簇的聚簇中心，从每簇中把距离聚簇中心最近的无序图像作为代表帧F_k提取出来；2) Solve the cluster center of each cluster according to the similarity distance, and extract the disordered image closest to the cluster center from each cluster as the representative frame _Fk ;

3)对提取出的代表帧F_k进行无参考图像质量评价，如果该代表帧F_k满足立体视觉三维重建质量要求，则作为关键帧保留，如果该代表帧F_k不满足立体视觉三维重建质量要求，则删除该代表帧F_k，重新从原来的簇中提取另一代表帧，并将提取出的该另一代表帧再次进行无参考图像质量评价，直到无参考图像质量评价后的代表帧满足立体视角三维重建质量要求为止。3) Carry out no-reference image quality evaluation on the extracted representative frame _Fk , if the representative frame _Fk meets the quality requirements of stereoscopic 3D reconstruction, it will be reserved as a key frame, if the representative frame _Fk does not meet the quality of stereoscopic 3D reconstruction If required, the representative frame F _k is deleted, another representative frame is extracted from the original cluster, and the extracted representative frame is subjected to no-reference image quality evaluation again until the representative frame after no-reference image quality evaluation Until the quality requirements of stereoscopic 3D reconstruction are met.

上述技术方案中，步骤1)中不设K值的聚簇算法对无序图像聚簇，将每幅无序图像分成M(M＝16)块，每块纹理特征均值m_l、方差每块纹理特征均值m_l为：In the above-mentioned technical scheme, the clustering algorithm without setting the K value in step 1) clusters the disordered images, and divides each disordered image into M (M=16) blocks, and the texture feature mean value m _l and variance of each block The mean value m _l of each texture feature is:

${m m}_{l l} = = E E. (({X x}_{l l})) = = \frac{11}{{D D.}^{22}} {Σ Σ}_{i i = = 00}^{D D. - - 11} {Σ Σ}_{j j = = 00}^{D D. - - 11} {x x}_{l l} ((i i,, j j)),, l l = = 11,, 22,, 33,, ... ...,, M m;;$

每块纹理特征方差为：The variance of each texture feature is:

${e e}_{l l}^{22} = = E E. [[{(({X x}_{l l} - - E E. (({X x}_{l l}))))}^{22}]] = = \frac{11}{{D D.}^{22}} {Σ Σ}_{i i = = 00}^{D D. - - 11} {Σ Σ}_{j j = = 00}^{D D. - - 11} [[{x x}_{l l} ((i i,, j j)) - - E E. {(({X x}_{l l}))}^{22}]],, l l = = 11,, 22,, 33,, ... ...,, M m;;$

把每块的纹理特征均值m_l、方差合起来作为该无序图像的特征向量F，并对特征向量F进行归一化处理，其中：The texture feature mean m _l and variance of each block Take them together as the feature vector F of the unordered image, and normalize the feature vector F, where:

$F f = = [[{m m}_{11,,} {e e}_{11,,}^{22} {m m}_{22,,} {e e}_{22,,}^{22} ... ...,, {m m}_{M m,,} {e e}_{M m}^{22}]];;$

设原始向量[f₁,f₂,f₃,…f_M]，归一化公式：Suppose the original vector [f ₁ ,f ₂ ,f ₃ ,…f _M ], the normalization formula:

${F f}_{i i} = = \frac{{f f}_{i i} - - m m}{e e},, ((i i = = 11,, 22,, ... ...,, M m)),,$

其中e、m为原始特征向量标准差和均值，归一化后特征向量为[F₁,F₂,F₃,…F_M]，任意两帧图像F_a和F_b归一化后的特征向量为：Where e and m are the standard deviation and mean of the original feature vector, the feature vector after normalization is [F ₁ , F ₂ , F ₃ ,...F _M ], the normalized features of any two frames of images F _a and F _b The vector is:

F_a＝[F_a1,F_a2,…F_aM]和F_b＝[F_b1,F_b2,…F_bM]，F _a = [F _a1 , F _a2 , ... F _aM ] and F _b = [F _b1 , F _b2 , ... F _bM ],

任意两帧图像F_a和F_b之间的相似距离为：The similarity distance between any two frames of images F _a and F _b is:

$d d i i s the s t t (({F f}_{a a},, {F f}_{b b})) = = {[[{Σ Σ}_{i i = = 11}^{M m} {(({F f}_{a a i i} - - {F f}_{b b i i}))}^{22}]]}^{\frac{11}{22}},,$

阈值T是任意两张无序图像相似距离之和的平均值，其计算式为：The threshold T is the average value of the sum of similar distances between any two unordered images, and its calculation formula is:

$T T = = [[\frac{11}{N N * * ((N N - - 11))} {Σ Σ}_{i i &NotEqual; &NotEqual; j j}^{N N} d d i i s the s t t (({F f}_{i i},, {F f}_{j j}))]],,$

其中，N为无序图像数目；Among them, N is the number of unordered images;

上述技术方案中，步骤1)中不设K值得聚簇步骤为：In above-mentioned technical scheme, in step 1), do not set K to be worth clustering step to be:

a)、获取第一帧图像F₁并把其划分到簇K₁中，并将第一帧图像F₁作为簇K₁的聚簇中心；a), obtain the _first frame image F1 and divide it into cluster K1, and use the _first frame image F1 as the cluster center _of cluster K1 _;

b)、获取下一帧图像F_I,(I＝2,3,…,N)，其中，N为无序图像数；b), acquiring the next frame of image F _I , (I=2,3,...,N), where N is the number of unordered images;

c)、根据公式计算F_I与已得到簇K_j(j＝1,2,…,N_c)聚簇中心的相似距离dist(F_I,K_jc),其中K_jc为簇K_j的聚簇中心，计算F_I与聚簇中心K_jc的最小距离min dist(F_I,K_jc)，如果min dist(F_I,K_jc)≤T，则把F_I划分到具有min dist(F_I,K_jc)值的簇K_j中，并且计算该簇新的聚簇中心，该簇新的聚簇中心为该簇中所有无序图像归一化后特征向量之和除以该簇中无序图像的数目；c), according to the formula Calculate the similarity distance dist(F _I , K _jc ) between F _I and the cluster center of the obtained cluster K _j (j=1,2,…,N _c ), where K _jc is the cluster center of cluster K _j , and calculate F The minimum distance min dist(F _I ,K _jc ) between _I and the cluster center K _jc , if min dist(F _I ,K _jc )≤T, then divide F _I into one with min dist(F _I ,K _jc ) value In the cluster K _j of the cluster, and calculate the new cluster center of the cluster, the new cluster center of the cluster is the sum of the normalized feature vectors of all unordered images in the cluster divided by the number of unordered images in the cluster;

d)、如果min dist(F_I,K_jc)>T,则说明F_I与存在的任何一个簇都不相似，把F_I划分到另一个新簇K_j中；d) If min dist(F _I , K _jc )>T, it means that F _I is not similar to any existing cluster, and divide F _I into another new cluster K _j ;

e)、然后根据步骤b、c、d将所有的无序图像划分到不同的簇K_j中。e), and then divide all unordered images into different clusters K _j according to steps b, c, and d.

上述技术方案中，步骤3)中对提取的代表帧F_k进行无参考图像质量评价，其步骤为：In the above-mentioned technical scheme, step 3) carries out no reference image quality evaluation to the extracted representative frame F _k , and its steps are:

A)对提取的代表帧F_k为进行二次模糊处理得到模糊图像b；A) performing secondary fuzzy processing on the extracted representative frame _Fk to obtain fuzzy image b;

b_V＝h_v*F_k，b_H＝h_h*F_k，h_h＝(h_v)^T＝h_v，b _V =h _v *F _k , b _H =h _h *F _k , h _h = (h _v ) ^T = h _v ,

其中，b_V、b_H为图像F_k经过垂直、水平低通滤波后得到的模糊图像，h_h、h_v为滤波器垂直和水平模型；Among them, b _V , b _H are blurred images obtained by image F _k after vertical and horizontal low-pass filtering, h _h , h _v are filter vertical and horizontal models;

B)分别计算代表帧F_k滤波前相邻像素垂直绝对误差Df_V(i,j)、水平绝对误差Df_H(i,j)和滤波后模糊图像b相邻像素的垂直绝对误差Db_V(i,j)、水平绝对误差Db_H(i,j)，得到相邻像素值的变化；B) Calculate the vertical absolute error Df _V (i, j) and the horizontal absolute error Df _H (i, j) of the adjacent pixels representing the frame F _k before filtering, and the vertical absolute error Db _V ( i, j), horizontal absolute error Db _H (i, j), to obtain the change of adjacent pixel values;

DF_V(i,j)＝abs(F_k(i,j)-F_k(i-1,j))，DF_H(i,j)＝abs(F_k(i,j)-F_k(i,j-1))，DF _V (i,j)=abs(F _k (i,j)-F _k (i-1,j)), DF _H (i,j)=abs(F _k (i,j)-F _k ( i,j-1)),

Db_V(i,j)＝abs(b_V(i,j)-b_V(i-1,j))，Db_H(i,j)＝abs(b_H(i,j)-b_H(i,j-1))；Db _V (i,j)=abs(b _V (i,j)-b _V (i-1,j)), Db _H (i,j)=abs(b _H (i,j)-b _H ( i,j-1));

C)对代表帧F_k中相邻像素差进行求和处理:C) Summing the difference between adjacent pixels in the representative frame F _k :

${sf sf}_{V V} = = {Σ Σ}_{i i,, j j = = 11}^{m m - - 11,, n no - - 11} {DF DF}_{V V} ((i i,, j j)),,$

${sf sf}_{H h} = = {Σ Σ}_{i i,, j j = = 11}^{m m - - 11,, n no - - 11} {DF DF}_{H h} ((i i,, j j)),,$

${sb sb}_{V V} = = {Σ Σ}_{i i,, j j = = 11}^{m m - - 11,, n no - - 11} {Db DB}_{V V} ((i i,, j j)),,$

${sb sb}_{H h} = = {Σ Σ}_{i i,, j j = = 11}^{m m - - 11,, n no - - 11} {Db DB}_{H h} ((i i,, j j)),,$

归一化得：Normalized to get:

${bF f}_{V V} = = \frac{{sF f}_{V V} - - {sb sb}_{V V}}{{sF f}_{V V}},,$

${bF f}_{H h} = = \frac{{sF f}_{H h} - - {sb sb}_{H h}}{{sF f}_{H h}},,$

clear＝max(bF_V,bF_H)；clear=max(bF _V ,bF _H );

D)、评价值范围在(0，1)之间,如果0.3＜clear＝max(bF_V,bF_H)<1，将代表帧F_k作为关键帧提取出来，如果clear＝max(bF_V,bF_H)≤0.3，则删除该代表帧，并从原簇中重新提取代表帧，然后根据步骤A、B、C计算该重新提取的代表帧。D), the evaluation value range is between (0, 1), if 0.3<clear=max(bF _V ,bF _H )<1, the representative frame F _k is extracted as a key frame, if clear=max(bF _V , bF _H )≤0.3, delete the representative frame, and re-extract the representative frame from the original cluster, and then calculate the re-extracted representative frame according to steps A, B, and C.

本发明的有益效果为：可以有效的提取无序图像的代表帧，滤除信息冗余、信息量少的无序图像，并且采用不设K值的聚簇算法可以根据无序图像内容复杂度自动聚成不同数目的簇，通过采用无参考图像评价法可以获得高清晰的关键帧。The beneficial effects of the present invention are: the representative frame of the disordered image can be effectively extracted, the disordered image with redundant information and less information can be filtered out, and the clustering algorithm without K value can be used according to the content complexity of the disordered image Automatically gather into different numbers of clusters, and high-definition key frames can be obtained by using the no-reference image evaluation method.

附图说明Description of drawings

图1为本发明的流程图；Fig. 1 is a flow chart of the present invention;

具体实施方式detailed description

下面结合附图对本发明的具体实施方式作进一步说明：The specific embodiment of the present invention will be further described below in conjunction with accompanying drawing:

如图1所示，一种无序图像关键帧的提取方法，其特征在于，包括以下步骤：As shown in Figure 1, a method for extracting key frames of an unordered image is characterized in that it comprises the following steps:

每块纹理特征方差为：Variance of texture features per block for:

A)对提取的关键代表帧F_k为进行二次模糊处理得到模糊图像b；A) performing secondary fuzzy processing on the extracted key representative frame _Fk to obtain fuzzy image b;

C)对代表帧F_k中相邻像素差求和处理:C) Summing the difference between adjacent pixels in the representative frame F _k :

归一化得：Normalized to get:

${bF f}_{V V} = = \frac{{sF f}_{V V} - - {sb sb}_{V V}}{{sF f}_{V V}},,$

${bF f}_{H h} = = \frac{{sF f}_{H h} - - {sb sb}_{H h}}{{sF f}_{H h}},,$

clear＝max(bF_V,bF_H)；clear=max(bF _V ,bF _H );

D)、评价值范围在(0，1)之间,如果0.3＜clear＝max(bF_V,bF_H)＜1，将代表帧F_k作为关键帧提取出来，如果clear＝max(bF_V,bF_H)≤0.3，则删除该代表帧，并从原簇中重新提取代表帧，然后根据步骤A、B、C计算该重新提取的代表帧。D), the evaluation value range is between (0, 1), if 0.3<clear=max(bF _V , bF _H )<1, the representative frame F _k will be extracted as a key frame, if clear=max(bF _V , bF _H )≤0.3, delete the representative frame, and re-extract the representative frame from the original cluster, and then calculate the re-extracted representative frame according to steps A, B, and C.

本发明的有益效果为：可以有效的提取无序图像的代表帧，滤除信息冗余、信息量少、质量差的无序图像，并且采用不设K值的聚簇算法可以根据无序图像内容复杂度自动聚成不同数目的簇，通过采用无参考图像评价法可以获得高清晰的关键帧。The beneficial effects of the present invention are: the representative frame of the disordered image can be effectively extracted, the disordered image with redundant information, less information, and poor quality can be filtered out, and the clustering algorithm without K value can be used to filter out the disordered image according to the The content complexity is automatically clustered into different numbers of clusters, and high-definition key frames can be obtained by using the no-reference image evaluation method.

上述实施例和说明书中描述的只是说明本发明的原理和最佳实施例，在不脱离本发明精神和范围的前提下，本发明还会有各种变化和改进，这些变化和改进都落入要求保护的本发明范围内。What described in above-mentioned embodiment and description just illustrate the principle of the present invention and preferred embodiment, under the premise of not departing from the spirit and scope of the present invention, the present invention also can have various changes and improvements, and these changes and improvements all fall into within the scope of the claimed invention.

Claims

1. a unordered graph is as the extracting method of key frame, it is characterised in that comprise the following steps:

1) use and do not set the clustering algorithms of K value and unordered graph picture is clustered process, unordered graph close for image information content As gathering for cluster；

2) solve the center that clusters of every bunch according to similarity distance, the nearest unordered graph picture in the center that from every bunch, distance clustered as Represent frame F_kExtract；

3) to the representative frame F extracted_kCarry out non-reference picture quality appraisement, if this represents frame F_kMeet stereo vision three-dimensional Reconstruction quality requirement, then retain as key frame, if this represents frame F_kIt is unsatisfactory for stereo vision three-dimensional rebuilding prescription, then Delete this and represent frame F_k, from original bunch, again extract another represent frame, and will be extracted another represent frame and again carry out nothing Reference image quality appraisement, until the representative frame after non-reference picture quality appraisement meets perspective view three-dimensional reconstruction prescription Till.

A kind of unordered graph the most according to claim 1 is as the extracting method of key frame, it is characterised in that: step 1) in do not set Unordered graph picture is clustered by the clustering algorithms of K value, and every width unordered graph picture is divided into M (M=16) block, every piece of textural characteristics average m_l、 VarianceEvery piece of textural characteristics average m_lFor:

m_{l} = E (X_{l}) = \frac{1}{D^{2}} Σ_{i = 0}^{D - 1} Σ_{j = 0}^{D - 1} x_{l} (i, j), l = 1, 2, 3, ..., M;

Every piece of textural characteristics varianceFor:

e_{l}^{2} = E [{(X_{l} - E (X_{l}))}^{2}] = \frac{1}{D^{2}} Σ_{i = 0}^{D - 1} Σ_{j = 0}^{D - 1} [x_{l} (i, j) - E {(X_{l})}^{2}], l = 1, 2, 3, ..., M;

Textural characteristics average m of every piece_l, varianceIt is together as characteristic vector F of this unordered graph picture, and to characteristic vector F It is normalized, wherein:

F = [m_{1,} e_{1,}^{2} m_{2,} e_{2,}^{2} ..., m_{M,} e_{M}^{2}];

If original vector [f₁, f₂, f₃... f_M], normalization formula:

F_{i} = \frac{f_{i} - m}{e} (i = 1, 2, ..., M),

Wherein e, m are original feature vector standard deviation and average, and after normalization, characteristic vector is [F₁, F₂, F₃... F_M], any two Two field picture F_aAnd F_bCharacteristic vector after normalization is:

F_a=[F_a1, F_a2... F_aM] and F_b=[F_b1, F_b2... F_bM],

Any two two field picture F_aAnd F_bBetween similarity distance be:

d i s t (F_{a}, F_{b}) = {[Σ_{i = 1}^{M} {(F_{a i} - F_{b i})}^{2}]}^{\frac{1}{2}},

Threshold value T be any two unordered graphs as the meansigma methods of similarity distance sum, its calculating formula is:

T = [\frac{1}{N * (N - 1)} Σ_{i &NotEqual; j}^{N} d i s t (F_{i}, F_{j})],

Wherein, N is unordered picture number；

In technique scheme, step 1) in do not set K value cluster step as:

A) the first two field picture F, is obtained₁And it is divided into a bunch K₁In, and by the first two field picture F₁As a bunch K₁The center that clusters；

B), next frame image F is obtained_I, (I=2,3 ..., N), wherein, N is unordered picture number；

C), according to formulaCalculate F_IWith obtain a bunch K_j(j=1,2 ..., N_c) cluster Similarity distance dist (the F at center_I,K_jc), wherein K_jcFor a bunch K_jThe center that clusters, calculate F_IWith the center K that clusters_jc(j=1, 2,…,N_c) minimum range min dist (F_I,K_jc), if min is dist (F_I,K_jc)≤T, then F_IIt is divided into and there is min dist(F_I,K_jc) bunch K of value_jIn, and calculate this brand new center that clusters, what this was brand new cluster center for institute in this bunch with or without After sequence image normalization, characteristic vector sum is divided by the number of unordered graph picture in this bunch；

If d) min dist (F_I,K_jc) > T, then F is described_IAll dissimilar, F with any one bunch existed_IIt is divided into another Individual new bunch of K_jIn；

E), then according to step b, c, d, all of unordered graph picture is divided into different bunch K_jIn.

A kind of unordered graph the most according to claim 1 is as the extracting method of key frame, it is characterised in that: step 3) in carrying The representative frame F taken_kCarry out non-reference picture quality appraisement, the steps include:

A) to the representative frame F extracted_kCarry out secondary Fuzzy Processing and obtain broad image b:

b_V=h_v*F_k, b_H=h_h*F_k,h_h=(h_v)^T=h_v,

Wherein, b_V、b_HFor image F_kThe broad image obtained after vertical, horizontal low pass ripple, h_h、h_vVertical for wave filter and Horizontal model；

B) calculating represents frame F respectively_kVertical absolute error Df of neighbor before filtering_V(i, j), horizontal absolute error Df_H(i,j) With vertical absolute error Db of broad image b neighbor after filtering_V(i, j), horizontal absolute error Db_H(i j), obtains adjacent The change of pixel value:

DF_V(i, j)=abs (F_k(i,j)-F_k(i-1, j)), DF_H(i, j)=abs (F_k(i,j)-F_k(i, j-1)),

Db_V(i, j)=abs (b_V(i,j)-b_V(i-1, j)), Db_H(i, j)=abs (b_H(i,j)-b_H(i,j-1))；

C) to representing frame F_kMiddle neighbor difference carries out summation process:

{sf}_{V} = Σ_{i, j = 1}^{m - 1, n - 1} {DF}_{V} (i, j),

{sf}_{H} = Σ_{i, j = 1}^{m - l, n - 1} {DF}_{H} (i, j),

{sb}_{V} = Σ_{i, j = 1}^{m - 1, n - 1} {Db}_{V} (i, j),

{sb}_{H} = Σ_{i, j = 1}^{m - l, n - 1} {Db}_{H} (i, j),

Normalization obtains:

{bF}_{V} = \frac{{sF}_{V} - {sb}_{V}}{{sF}_{V}},

{bF}_{H} = \frac{{sF}_{H} - {sb}_{H}}{{sF}_{H}},

Clear=max (bF_V,bF_H)；

D), evaluation of estimate scope between (0,1), if 0.3 ＜ clear=max (bF_V,bF_H) < 1, frame F will be represented_kAs key Frame extracts, if clear=max is (bF_V,bF_H)≤0.3, then delete this and represent frame, and again extract representative from former bunch Frame, then calculates this representative frame again extracted according to step A, B, C.