CN106412619A - HSV color histogram and DCT perceptual hash based lens boundary detection method - Google Patents

Abstract

The invention discloses an HSV color histogram and DCT perceptual Hash based lens boundary detection method. The HSV color histogram and DCT perceptual Hash based lens boundary detection method comprises a first step of calculating a histogram difference degree between two adjacent frames of images by using the HSV color space based histogram statistic method; a second step of judging the difference degree between the two adjacent frames by adopting a adaptive threshold value, and judging the place between the two adjacent frames is the lens boundary if the difference degree is greater than the threshold value; and a third step of rechecking an early detection result by adopting a DCT based perceptual Hash algorithm and obtaining a final video lens boundary set. Through the early detection of the HSV color histogram and the DCT perceptual Hash algorithm based rechecking, the lens boundary position in a video frame sequence is detected efficiently and accurately, and a firm foundation for the structural analysis of the multimedia video data is accordingly laid.

Description

A kind of lens boundary detection method perceiving Hash based on hsv color rectangular histogram and DCT

Technical field

The present invention relates to field of video content analysis, specific design is a kind of to perceive Hash based on hsv color rectangular histogram and DCT The lens boundary detection method of algorithm.

Background technology

With the development of computer, communication and multimedia technology, the data volume of multimedia video data is being in blowout Ground increases, and simultaneously traditional video analysis method cannot adapt to the continuous growth of data volume and the further multiple of video data structure Miscellaneous.In order to more efficiently store, manage and analysis video data is it is proposed that video analysis retrieval technique based on content, its In, video shot boundary detection is by the basis of video analysis retrieval.Video data be by series of identical size and when Between upper continuous picture frame composition, the structure of video is descending to be segmented into video sequence, scene, camera lens and picture frame. Camera lens is the elementary cell of video structure, and shot boundary detector is the first step of video data analysis retrieval, shot boundary detector Accuracy directly influence the effect of subsequent treatment.

Up to the present, experts and scholars both domestic and external have done substantial amounts of work in this field and have created a lot of researchs Achievement.The shot detection method of main flow includes：Pixel method, histogram method, mutual information are mensuration, edge feature method etc..But It is, because video data species is various, respectively to have a feature, traditional method all has certain limitation, leads to Detection accuracy Not high.So far, the common issue that Shot Detection technology exists is more sensitive to object of which movement and brightness flop, and this will be The technical barrier that all detection techniques will solve backward.

Content of the invention

The technical problem to be solved in the present invention is：Overcome the deficiencies in the prior art, provide one kind to combine improved hsv color Rectangular histogram, normalized threshold and DCT perceive the Methods for Shot Boundary Detection of Video Sequences with double check mechanism of hash algorithm.

Present invention aim at being directed to the problem of current shot boundary detector common problem it is proposed that one kind is based on Hsv color rectangular histogram and the lens boundary detection method of DCT perception Hash, this method solve existing video shot boundary inspection Method of determining and calculating is sensitive to illumination variation, accuracy rate and the relatively low problem of recall ratio, improves the inspection for different structure video data Survey accuracy rate and recall ratio, the suitability improves.

The present invention solves its technical problem and is adopted the technical scheme that：The present invention first passes through statistics sequence of frames of video Hsv color histogram information, filters out in most of camera lens frame data and provides an initial shot boundary set, Greatly reduce the time complexity of reinspection.Then, hash algorithm is perceived using DCT, extract the DCT perception Hash of frame of video Eigenvalue, different by comparative feature value difference, delete the pseudo- shot boundary in initial survey result, overcome the shadow to detection for the illumination variation Ring, further improve algorithm accuracy rate.

A kind of lens boundary detection method perceiving Hash based on hsv color rectangular histogram and DCT, methods described includes as follows Step：

Step 1：Using the Nogata being calculated based on the statistics with histogram method in hsv color space between adjacent two field pictures Figure diversity factor；

Step 2：Using adaptive threshold, the diversity factor of adjacent two interframe being differentiated, if being more than threshold values, tentatively sentencing It is set to shot boundary；

Step 3：Initial survey result is rechecked and is drawn final video lens using the perception hash algorithm based on DCT Border is gathered.

A kind of described lens boundary detection method being based on hsv color rectangular histogram and DCT perception Hash, step 1 Concrete grammar as follows：

A) read video sequence M, calculate the normalization tone of each two field picture, saturation and bright according to formula (1)-(3) Angle value, i.e. normalization HSV information；

Wherein, R, three color components of red, green, blue of G, B difference representative image pixel, it is empty that H, S, V represent hsv color Between tone, saturation and luma component values, span [0,1]；

B) differentiation quantization is carried out to HSV three-component, be quantized into 16 ranks, 8 ranks and 2 ranks respectively；And the three-dimensional by image Color space is compressed in the one-dimensional space, and computing formula is as follows：

L (i, j)=16H (i, j)+2S (i, j)+V (i, j) (4)

Wherein, i, j are the location of pixels transverse and longitudinal coordinates of image, H (i, j), S (i, j), and V (i, j) is current pixel location Tri- component values of HSV；

C) combine variate-value L (i, j) of each location of pixels, calculate the statistics with histogram value of every two field picture, and calculate phase Card side's distance between adjacent two frames, computing formula is as follows：

Wherein, j represents jth two field picture, and d (j, j+1) represents the Histogram distance between adjacent two frames, H_jI () represents The statistical value of the i component in jth two field picture histogram information, the span [0,255] of i.

A kind of described lens boundary detection method being based on hsv color rectangular histogram and DCT perception Hash, step 2 Concrete grammar as follows：

A) a kind of normalized camera lens initial survey thresholding T1 is proposed, formula is as follows：

T1=m+ λ σ (6)

Wherein, m, σ are average and the standard deviations of Histogram distance sequence, and λ is constant, and span can be { 1,2 }；

B) using the threshold value calculating, carry out shot boundary initial survey：

If Histogram distance is more than thresholding, judges there is shot boundary at this two field picture, and boundary value j is stored first See in border set P, take j=j+1, judged again；Otherwise, it is determined that image is camera lens frame interior herein, j=j+1, in due order Sequence judges successively until j=N, and N is sequence of frames of video totalframes.

A kind of described lens boundary detection method being based on hsv color rectangular histogram and DCT perception Hash, step 3 Concrete grammar as follows：

A) the every two field picture in initial survey border set P is converted into gray level image from coloured image；

B) call resize () function pair image to reset size, reduce algorithm amount of calculation；

C) discrete cosine transform is utilized to calculate the DCT coefficient matrix of gray level image frame sequence, in conjunction with image DCT transform Characteristic, take coefficient matrix upper left side sub-fraction nonzero element submatrix C；

D) the average of design factor Matrix C, if the element value of correspondence position is more than average, puts 1, otherwise sets to 0, Suo Youyuan Element combines DCT perception Hash sequence G just constituting image in order；

E the DCT perception Hash diversity factor that the every two field picture in P is adjacent between picture frame, diversity factor definition) are calculated It is as follows,

Wherein, j represents the shot boundary position in initial survey set, and D represents the DCT perception Hash difference between adjacent two frames Degree, Num is image perception Hash sequence length, and G is image perception Hash sequence；

F) rechecked by diversity factor size, if diversity factor be more than 30%; judge be herein genuine video mirror in front Boundary, and j is stored in border set R, otherwise judge at this as flase drop；All elements in traversal set P, are sentenced respectively Fixed, finally return to video boundaries set R.

The present invention is applied to multimedia video data analysis and retrieval.

Beneficial effects of the present invention：

1st, the lens boundary detection method of the present invention employs the histogramming algorithm of main flow, according to human eye to HSV three-component Sensitivity, carried out the quantization operation of different proportion respectively, inhibited the impact to algorithm for the brightness to a certain extent, improve The accuracy rate of algorithm.

2nd, it is directed to the video that existing major part algorithm may be only available for some ad hoc structures, present invention employs one kind and return The threshold judgement method of one change, algorithm motility is higher, goes for various video structure.

3 present invention utilizes the low algorithm complex of histogramming algorithm and the advantage of relatively low time complexity and sense Know the high advantage of hash algorithm data compression rate.Although present invention uses initial survey rechecks the mechanism of double check, detection The spent time does not significantly increase, and simultaneously effective improves the accuracy of detection method.

Brief description

Fig. 1 is the inventive method flow chart.

Fig. 2 is video frame image hsv color statistics with histogram information schematic diagram in the present invention.

Fig. 3 is that the DCT perception Hash of the present invention rechecks algorithm flow chart.

Specific embodiment

Below in conjunction with specific embodiment, the present invention is described in detail.

As shown in figure 1, the Methods for Shot Boundary Detection of Video Sequences of the present invention, comprise the following steps：

1) read video sequence M, obtain the RGB information matrix of every two field picture.Because hsv color model is more directly perceived, more Meet the visual experience of human eye, so obtain the hsv color information matrix of picture frame by following conversion formula：

Wherein, R, three color components of red, green, blue of G, B difference representative image pixel, it is empty that H, S, V represent hsv color Between tone, saturation and luma component values, span [0,1].

2) data-measuring compression.Multimedia video data is huge famous with data volume, during analysis video data, Reducing data volume under conditions of not impact analysis precision is a critically important link.

A) combine the visual signature of human eye, respectively tri- kinds of component even amounts of H, S, V are melted into 16,8 and 2 ranks, greatly reduce Computation complexity.

B) data matrix of the video frame image after quantifying is mapped to one-dimensional color space from three-dimensional color space：

L (i, j)=16H (i, j)+2S (i, j)+V (i, j) (4)

Wherein, i, j are the location of pixels transverse and longitudinal coordinates of image, H (i, j), S (i, j), and V (i, j) is current pixel location Tri- component values of HSV.

3) obtain the histogram difference degree between adjacent image frame successively.

A) count the histogram information of each image.

Wherein, H_j(i) represent jth two field picture histogram information in the appearance of i component the frequency, i span [0, 255], m and n represents the transverse and longitudinal dimension of image respectively.The hsv color statistics with histogram information that Fig. 2 gives a picture frame is shown It is intended to.

B) combine statistics with histogram amount, calculate the card side's distance between adjacent two frames, formula is as follows：

Wherein, j represents jth two field picture, and d (j, j+1) represents the Histogram distance between adjacent two frames.Since then, pass through Above step can obtain the hsv color Histogram distance of sequence of frames of video.

4) define a kind of normalized camera lens initial survey thresholding T1, formula is as follows：

T1=m+ λ σ (7)

Wherein, m, σ are average and the standard deviations of Histogram distance sequence, and λ is constant, and span can be [1,2].

By formula (7), draw the threshold value of just detecting method, and provide decision method：

D (j, j+1) >=T1 (8)

If image frame pitch is from more than threshold T 1, preliminary judgement is shot boundary value at this, and this frame position is deposited Store up in set P.Otherwise it is assumed that being the picture frame within camera lens at this.Judge the frame pitch of each adjacent position image successively From, and draw initial survey border set P.

5) although improving to traditional histogram method in invention, improved method yet suffers from quick to illumination The frame difference that illumination variation produces can be judged into by accident the border of camera lens by the problem of sense.Therefore, the present invention utilizes DCT perception to breathe out Uncommon algorithm is rechecked to initial survey result, and improves method accuracy rate further, and specific algorithm flow chart is as shown in Figure 3.Specifically Step is as follows：

A) call rgb2gray () function, the every frame original image in initial survey border set P is converted into from coloured image Gray level image, and preserve result；

B) call resize () function, picture size is adjusted to 32*32；

C) discrete cosine transform is utilized to calculate the DCT coefficient Matrix C of gray level image frame sequence, in conjunction with image DCT transform Characteristic, take coefficient matrix upper left side sub-fraction nonzero element submatrix C'；

D) design factor Matrix C ' average,

Wherein, i, j are the location of pixels transverse and longitudinal coordinates of image.If the element value of correspondence position is more than average, correspondence position Put 1, otherwise set to 0, all elements combine DCT perception Hash sequence G just constituting image, i.e. frame DPHA system in order Number；

E) calculate the DCT perception Hash diversity factor that the every two field picture in initial survey border set P is adjacent between picture frame, Diversity factor definition is as follows,

Wherein, j represents the shot boundary position in initial survey set, and D represents the DCT perception Hash difference between adjacent two frames Degree, Num is image perception Hash sequence length, and G is image perception Hash sequence.

F) rechecked by comparing difference degree size, if diversity factor is more than sequence length 30% (thresholding T2), judged Video shot boundary exists herein, and j is stored in border set R, otherwise judges at this as flase drop.In traversal set P All elements, carry out detection successively and judge, finally return to video boundaries set R.

The present invention proposes a kind of lens boundary detection method perceiving Hash based on hsv color rectangular histogram and DCT, utilizes The statistical property of histogram method, rotational invariance and using perception hash algorithm overcome shot boundary detection algorithms to illumination become The impact changed.By shot boundary detector is carried out to several different classes of video segments, and carry out with two kinds of existing algorithms Contrast, is obtained for effective raising (table 1) on algorithm Detection accuracy, recall ratio.

Table 1

It should be appreciated that for those of ordinary skills, can be improved according to the above description or be converted, And all these modifications and variations all should belong to the protection domain of claims of the present invention.

Claims (4)

1. a kind of the lens boundary detection method of Hash is perceived it is characterised in that described side based on hsv color rectangular histogram and DCT Method comprises the steps：

Step 1：Poor using the rectangular histogram being calculated between adjacent two field pictures based on the statistics with histogram method in hsv color space Different degree；

Step 2：Using adaptive threshold, the diversity factor of adjacent two interframe is differentiated, if being more than threshold values, preliminary judgement is Shot boundary；

Step 3：Initial survey result is rechecked and is drawn final video shot boundary using the perception hash algorithm based on DCT Set.

2. a kind of shot boundary detector being based on hsv color rectangular histogram and DCT perception Hash according to claim 1 Method is it is characterised in that the concrete grammar of step 1 is as follows：

A) read video sequence M, calculate normalization tone, saturation and the brightness value of each two field picture according to formula (1)-(3), I.e. normalization HSV information；

$V=\frac{\max (R,G,B)}{255}---\left(1\right)$

$S=\left\{\begin{array}{cc}\frac{V-\min (R,G,B)}{V},& V\≠0\\ 0,& V=0\end{array}\right.---\left(2\right)$

$H=\left\{\begin{array}{cc}\frac{(G-B)}{6(V-\min (R,G,B\left)\right)},& R=V\\ \frac{1}{3}+\frac{(B-R)}{6(V-\min (R,G,B\left)\right)},& G=V\\ \frac{2}{3}+\frac{(R-G)}{6(V-\min (R,G,B\left)\right)},& B=V\end{array}\right.---\left(3\right)$

Wherein, R, three color components of red, green, blue of G, B difference representative image pixel, H, S, V represent hsv color space Tone, saturation and luma component values, span [0,1]；

B) differentiation quantization is carried out to HSV three-component, be quantized into 16 ranks, 8 ranks and 2 ranks respectively；And the three-dimensional color by image , in the one-dimensional space, computing formula is as follows for space compression：

L (i, j)=16H (i, j)+2S (i, j)+V (i, j) (4)

Wherein, i, j are the location of pixels transverse and longitudinal coordinates of image, and H (i, j), S (i, j), V (i, j) are the HSV of current pixel location Three component values；

C) combine variate-value L (i, j) of each location of pixels, calculate the statistics with histogram value of every two field picture, and calculate adjacent two Card side's distance between frame, computing formula is as follows：

$d(j,j+1)=\underset{i=0}{\overset{255}{\mathrm{\Σ}}}\frac{{\left(H_j\right(i)-H_{j+1}(i\left)\right)}^2}{max\left(H_j\right(i),H_{j+1}(i\left)\right)}---\left(5\right)$

Wherein, j represents jth two field picture, and d (j, j+1) represents the Histogram distance between adjacent two frames, H_jI () represents jth frame The statistical value of the i component in image histogram information, the span [0,255] of i.

3. a kind of shot boundary detector being based on hsv color rectangular histogram and DCT perception Hash according to claim 1 Method is it is characterised in that the concrete grammar of step 2 is as follows：

A) a kind of normalized camera lens initial survey thresholding T1 is proposed, formula is as follows：

T1=m+ λ σ (6)

Wherein, m, σ are average and the standard deviations of Histogram distance sequence, and λ is constant, and span can be { 1,2 }；

B) using the threshold value calculating, carry out shot boundary initial survey：

If Histogram distance is more than thresholding, judges there is shot boundary at this two field picture, and boundary value j is stored first meeting side In boundary's set P, take j=j+1, judged again；Otherwise, it is determined that image is camera lens frame interior herein, j=j+1, in order according to Until j=N, N is sequence of frames of video totalframes for secondary judgement.

4. a kind of shot boundary detector being based on hsv color rectangular histogram and DCT perception Hash according to claim 1 Method is it is characterised in that the concrete grammar of step 3 is as follows：

A) the every two field picture in initial survey border set P is converted into gray level image from coloured image；

B) call resize () function pair image to reset size, reduce algorithm amount of calculation；

C) discrete cosine transform is utilized to calculate the DCT coefficient matrix of gray level image frame sequence, in conjunction with the spy of image DCT transform Property, take coefficient matrix upper left side sub-fraction nonzero element submatrix C；

D) the average of design factor Matrix C, if the element value of correspondence position is more than average, puts 1, otherwise sets to 0, all elements are pressed Sequential combination just constitutes DCT perception Hash sequence G of image together；

E) calculate the DCT perception Hash diversity factor that the every two field picture in P is adjacent between picture frame, diversity factor definition is such as Under,

$D\left(j\right)=\underset{i=1}{\overset{Num}{\mathrm{\Σ}}}|G\left(i\right)-G(i+1)|---\left(7\right)$

Wherein, j represents the shot boundary position in initial survey set, and D represents the DCT perception Hash diversity factor between adjacent two frames, Num is image perception Hash sequence length, and G is image perception Hash sequence；

F) rechecked by diversity factor size, if diversity factor is more than 30%, judged it is genuine video shot boundary herein, and J is stored in border set R, otherwise judges at this as flase drop；All elements in traversal set P, are judged, the most respectively Return video boundaries set R afterwards.