CN103561271B - The moving target of static camera shooting is removed video spatial domain altering detecting method - Google Patents

Abstract

The moving target of a kind of static camera shooting is removed video spatial domain altering detecting method, belongs to electronic evidence-collecting technical field, beneficially the method and can position the tampered position on spatial domain.First, extract and distort frame sequence, then calculate suspicious motion dot image by frame difference method, extract the suspicious motion image block on spatial domain, according to energy suspicious degree exclusive PCR image block, determine object removal block, thus realize the tampering location on spatial domain.This algorithm can detect that whether moving target is removed under fixed background, improves detection efficiency effectively, and the authenticity for video differentiates to provide effective foundation.

Description

The moving target of static camera shooting is removed video spatial domain altering detecting method

Technical field:

The present invention relates to a kind of electronic evidence-collecting technical field, the moving target being specific to static camera shooting is removed Video spatial domain altering detecting method.

Background technology:

Along with the development of multimedia technology, the extensive utilization of Video editing software, people can utilize various video Existing video is distorted by software for editing, and in these are distorted, some brings amusement to the life of people, but there is also very A big part brings harm to society, makes people throw doubt upon the authenticity of digital video, therefore, and video tampering detection skill Art has become an important subject of current information-intensive society.Current existing detection method is both for the specifically side of distorting Formula is authenticated, and scratches the means of distorting removed herein for moving target and proposes a kind of effective video altering detecting method.

Digital multimedia forensic technologies is mainly the analysis that digital medium information carries out authenticity, primitiveness, at present, Research to this technology focuses more on digital picture aspect, and the passive forensic technologies of digital video is because of himself technical system Complexity development lag far behind the passive forensic technologies of digital picture, both at home and abroad digital video is distorted the method detected also Seldom.The research of digital video tampering detection mainly divides two classes, a class to be active certifications, as Shenzhen Research Institute of Sun Yat-Sen University invents One based on compression sensing semi-fragile watermarking video tampering detection patent, active certification technology is to regard in protected numeral Embed watermark or digital signature etc. in Pin and be difficult to the signal specific of perception, distort operation and can destroy these signals, believed by detection Number integrality, signal destroy position and destructiveness judge whether video is tampered, the position distorted and distort type. But active certification exists limitation, many imaging devices do not have the function embedding the signals such as watermark, and the signal embedded is difficult to Guarantee not to be easily removed or again embed.And another kind of video tampering detection technology be not embedded at video specific Detect on the premise of signal whether video passes through the passive authentication techniques distorted, according to distorting the coding characteristic of before and after's video, system The change of meter feature and some other characteristic value judges that whether video is distorted, and has important using value.

It is that video distorts common form that moving target removes from video.At present, video tamper detection method mainly has The most several: sequence of frames of video feature detection based on the operation of MPEG dual compression, special by extracting the residual of fixed type frame Levy and judge whether video is tampered, but the method is just for mpeg format video, detect the insertion of whole frame, deletion or whole frame and replicate viscous The situation of patch, is not suitable for frame internal object and deletes detection；Zhang Jing, Song Yi of University Of Tianjin et al. propose to use Space-time domain Block-matching Method detection remove object, need image block is carried out Secondary Match search location, the most time-consuming, and quantify to make an uproar Sound can affect matching detection accuracy, easily produces flase drop district；Zhang Mingyu proposes to use the method for accumulative difference image to judge fortune Whether moving-target is removed, but the method cannot position the tampered position in time domain, and is easily subject to strong marginal information point and the back of the body The interference at sight spot.

Summary of the invention:

In order to overcome the deficiency of the existing video tamper detection method being removed for moving target, the present invention proposes one Planting towards the video spatial domain altering detecting method that the moving target captured by static camera removes from background, utilization energy can Doubtful degree measures the energy variation degree of each frame of video, and improves traditional moving target and add up difference image tracing, introduces Suspicious motion dot image detects the suspicious motion region on spatial domain, determines that moving target is removed finally according to the suspicious degree of energy Region.The method is applicable to arbitrary format video, can not only position the tampered position on spatial domain, and detection effect is greatly improved Rate and accuracy rate, get rid of the strong marginal information interference problem being frequently encountered by conventional detection method.

Frame repair process would generally make the gray value of restoring area embody a moving region over time and space, we Frame difference method can be used to come the region that pursuit movement target is removed on spatial domain.Known pursuit movement mesh calibration method is to make With accumulative difference method, the change of pixel same in several two field pictures is added up, but this method is often highly susceptible to edge The interference of the pixel that information is stronger.Therefore the method is improved by the present invention, find by calculate accumulative difference image and The differential chart of accumulative edge image, obtains suspicious motion dot image, it is possible to get rid of the interference of marginal information point.

As follows for realizing the concrete technical scheme of purpose of the present invention employing:

The first step, extracts kth frame and distorts sequence to h frame, be labeled as f₁,f₂,...,f_h-k+1.With f₁Frame is reference frame, meter Calculate accumulative difference image AD_h-k+1With accumulative edge image AE_h-k+1.Calculate suspicious motion dot image AC again_h-k+1。

Second step, from suspicious motion dot image AC_h-k+1Middle extraction suspect image block.

3rd step, calculates from f₁Frame is to f_h-k+1The gross energy suspicious degree average of the whole frame of frame image sequence

$\stackrel{\‾}{NT}=\frac{1}{(h-k+1)}\underset{i=1}{\overset{h-k+1}{\Σ}}\left({\mathrm{NT}}_i\right)$

4th step, calculate spatial filter second step record from f₁Frame is to f_h-k+1The energy of frame each suspect image block can Doubt degree averageWhenIt is less thanTime, it is determined that this image block is that the object removal on spatial domain distorts district, thus gets rid of other Interference image block.

In the described spatial filter first step, accumulative difference image, accumulative edge image, suspicious motion dot image are specifically counted Calculate as follows:

Assume there is n frame image sequence f₁(x,y),f₂(x,y),f₃(x,y),....,f_n(x y), generally makes f₁(x, y) for ginseng Examine frame, the initial value AD of accumulative difference image₁Initial value AE with accumulative edge image₁It is 0, then the accumulative difference diagram of kth frame As AD_kIt is calculated as follows:

Wherein, 1 ＜ k≤n.Each two field picture of image sequence is compared with reference frame image or its previous frame image Relatively, when the difference of same pixel is more than a certain threshold value, the corresponding point on accumulative difference image of this pixel just adds 1.

The accumulative edge image AE of kth frame_kIt is calculated as follows:

Wherein, 1 ＜ k≤n.Each two field picture of image sequence is done rim detection, if a certain pixel is image Marginal information, then the corresponding point on accumulative edge image of this pixel just adds 1.

Suspicious motion dot image AC_nIt is calculated as follows:

To a certain pixel, (x, y), when this gray value has generation large change and be not marginal information point in time domain Time, then it is assumed that this point may be tampered, and is called suspicious motion point.

In described spatial filter second step, from suspicious motion dot image AC_h-k+1The concrete inspection of middle extraction suspect image block Survey step is as follows:

1. statistics suspicious motion dot image AC_h-k+1In laterally suspicious motion count HOR.

2. statistics suspicious motion dot image AC_h-k+1Middle longitudinal suspicious motion is counted VER.

3. according to from suspicious motion dot image AC_h-k+1The horizontal suspicious motion of middle statistics is counted HOR, when t row is to r row In the suspicious motion of the every a line HOR (i) that counts be all higher than a certain threshold value, illustrate that t row is to the interval suspicious fortune of this line of r row Dynamic counting more, be stored in array x1 by t and r, it is interval that t row to r row is referred to as a pair suspicious row.One width suspicious motion point diagram Interval as there may be multipair suspicious row, all start-stop positions, suspicious row interval meeting condition are all stored in array x1.

According to from suspicious motion dot image AC_h-k+1Longitudinal suspicious motion of middle statistics is counted VER, in t arranges r row The suspicious motion of each row VER (j) that counts is all higher than a certain threshold value, illustrates that t arranges r and arranges the suspicious motion that these row are interval Counting more, be stored in array x2 by t and r, t arranges r row and is referred to as a pair suspicious row interval.One width suspicious motion dot image There may be multipair suspicious row interval, all interval start-stop positions of suspicious row meeting condition are all stored in array x2.

4. extract every pair of suspicious row interval and the image block of every pair of suspicious row interval composition in array x2 in array x1, it is assumed that S block altogether.Add up the suspicious motion in each image block to count C_i=count (AC_h-k+1(i)=0), wherein 1≤i≤s.Work as C_iGreatly When a certain threshold value, then this image block i is likely removed region for moving target, and marking this image block is suspect image block.

In described spatial filter the 3rd step and the 4th step, energy suspicious degree NT is specifically calculated as follows:

$NT=\frac{1}{B}+H$

The lowest high-frequency energy computing formula than B and frequency domain entropy H is as follows:

$B=\frac{1}{r}\sqrt{\underset{i=1}{\overset{r}{\Σ}}{\mathrm{\β}}_i^2}/\left(\frac{1}{m*n-r}\sqrt{\underset{i=r+1}{\overset{m*n}{\Σ}}{\mathrm{\β}}_i^2}\right)$

$H=-\underset{i=1}{\overset{m*n}{\Σ}}\left(\frac{\left|{\mathrm{\β}}_i\right|}{\underset{i=1}{\overset{m*n}{\Σ}}\left|{\mathrm{\β}}_i\right|}{\log }_2\frac{\left|{\mathrm{\β}}_i\right|}{\underset{i=1}{\overset{m*n}{\Σ}}\left|{\mathrm{\β}}_i\right|}\right)$

Block size is that m*n, r represent that image concentrates on the minority low frequency coefficient number in the upper left corner after dct transform.Will Two-dimensional array after dct transform from the upper left corner carry out Z-type scanning dimensionality reduction so as to get one-dimension array coefficient by decreasing energy Mode sort, β_iRepresent i-th DCT coefficient after sequence.

Accompanying drawing explanation

For the clearer explanation embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing In having technology to describe, the required accompanying drawing used is briefly introduced.

Fig. 1 is the spatial filter flow chart of this method

Fig. 2 is experimental series Detection results figure

Detailed description of the invention

Concrete steps and embodiment below in conjunction with the accompanying drawing in the embodiment of the present invention, in detail the narration present invention.

In Fig. 1, it is assumed that tampered image sequence detected in time domain, the kth frame extracting tim e-domain detection distorts sequence to h frame Row, are labeled as f₁,f₂,....,f_h-k+1.With f₁Frame is reference frame, calculates accumulative difference image AD_h-k+1With accumulative edge image AE_h-k+1, then calculate suspicious motion dot image AC_h-k+1.Then from image AC_h-k+1Middle extraction suspect image block.And then calculate From f₁Frame is to f_h-k+1The energy suspicious degree average of frame image sequenceBy comparing the energy of suspect image block suspicious degree averageWithJudge tampered image block, the target delete position being on spatial domain.

In fig. 2, the most tampered under being static background for (a)-(c) video segment；D () is for delete sport foreground After video segment；E suspicious motion dot image that () obtains for using inventive algorithm；F () is for using inventive algorithm detection To spatial domain on moving target delete position.

Embodiment 1

The video tamper detection method pair that this exemplary application judges based on the suspicious degree of energy and suspicious motion dot image calculates The authenticity of suspicious video sequence is collected evidence, and Fig. 1 gives spatial filter flow chart of the present invention, is situated between referring now to Fig. 1 Continue specific operation process:

The first step, the 445th frame extracting tim e-domain detection is distorted sequence to the 538th frame, is labeled as f₁,f₂,....,f₉₄.With f₁ Frame is reference frame, calculates accumulative difference image AD₉₄With accumulative edge image AE₉₄.Calculate suspicious motion dot image AC again₉₄。

Second step, adds up suspicious motion dot image AC₉₄In laterally suspicious motion counts HOR and longitudinal suspicious motion is counted VER。

3rd step, according to from AC₉₄The horizontal suspicious motion of middle statistics is counted HOR, extracts suspicious motion and counts more row Interval is stored in array x1, x1={44, and 191,380,467}；According to from AC₉₄Longitudinal suspicious motion of middle statistics is counted VER, carries Take the row interval that suspicious motion counts more to be stored in array x2, x2={1,640}.

4th step, the interval image block that each row interval forms with array x2 of each row in extraction array x1, totally 2 pieces. The position of image block 1 is: row 44-191, arranges 1-640；The position of image block 2 is: row 380-467, arranges 1-640.Add up each figure As the suspicious motion in block is counted C_i=count (AC_h-k+1(i)=0), wherein 1≤i≤2.Due to C₁And C₂It is all higher than arranging threshold Value, so image block 1 and image block 2 are likely removed region for moving target, mark image block 1 and image block 2 are suspicious Image block.

5th step, calculates from f₁Frame is to f₉₄The gross energy suspicious degree average of the whole frame of frame image sequence

$\stackrel{\‾}{NT}=\frac{1}{(h-k+1)}\underset{i=1}{\overset{h-k+1}{\Σ}}\left({\mathrm{NT}}_i\right)$

6th step, calculates each suspect image block from f₁Frame is to f₉₄The energy suspicious degree average of frameDue to suspect image The energy suspicious degree average of block 2Less than gross energy suspicious degree averageIt is determined that suspect image block 2 is the mesh on spatial domain Mark removes distorts district, thus gets rid of the interference of suspect image block 1.

In fig. 2, (e) is the suspicious motion dot image using inventive algorithm to obtain, and (f) obtains for using inventive algorithm To tested video in moving target delete position testing result figure.Experiment uses roberts operator extraction to detection video Marginal information.According to the suspicious motion dot image obtained, count by adding up the suspicious motion on horizontal and vertical, extract suspicious The suspect image block that motion is counted more, relatively the energy suspicious degree average of each image block, determines object removal image block, gets rid of Interference image block, result such as table 1

Table 1 spatial filter table

Shown in.

Claims (4)

1. the moving target of static camera shooting is removed a video spatial domain altering detecting method, the concrete technical side of employing Case is as follows:

The first step, the kth frame extracting tim e-domain detection is distorted sequence to h frame, is labeled as f₁,f₂,...,f_h-k+1, with f₁Frame is ginseng Examine frame, calculate accumulative difference image AD_h-k+1With accumulative edge image AE_h-k+1, then calculate suspicious motion dot image AC_h-k+1；

Second step, from suspicious motion dot image AC_h-k+1Middle extraction suspect image block；

3rd step, calculates from f₁Frame is to f_h-k+1The gross energy suspicious degree average of the whole frame of frame image sequence,

$\stackrel{\‾}{NT}=\frac{1}{(h-k+1)}\underset{i=1}{\overset{h-k+1}{\Σ}}\left({\mathrm{NT}}_i\right);$

4th step, calculate spatial filter second step record from f₁Frame is to f_h-k+1The suspicious degree of energy of frame each suspect image block AverageWhenIt is less thanTime, it is determined that this image block is that the object removal on spatial domain distorts district, thus gets rid of other interference Image block.

The moving target of a kind of static camera the most according to claim 1 shooting is removed video spatial domain tampering detection side Method, it is characterised in that in the described spatial filter first step, accumulative difference image, accumulative edge image, suspicious motion dot image Specifically it is calculated as follows:

Assume there is n frame image sequence f₁(x,y),f₂(x,y),f₃(x,y),...,f_n(x y), generally makes f₁(x, y) is reference frame, The initial value AD of accumulative difference image₁Initial value AE with accumulative edge image₁It is 0, then the accumulative difference image AD of kth frame_k It is calculated as follows:

Wherein, 1 ＜ k≤n；

Each two field picture of image sequence is compared with reference frame image or its previous frame image, when same pixel When difference is more than a certain threshold value, the corresponding point on accumulative difference image of this pixel just adds 1；

The accumulative edge image AE of kth frame_kIt is calculated as follows:

Wherein, 1 ＜ k≤n；

Each two field picture of image sequence is done rim detection, and edge detection operator uses roberts operator, if a certain pixel Point is the marginal information of image, then the corresponding point on accumulative edge image of this pixel just adds 1, suspicious motion dot image AC_n It is calculated as follows:

To a certain pixel, (x, y), when this point adds up difference gray value AD_n(x is not y) 0 and when not being marginal information point, then recognizes May be tampered for this point, be called suspicious motion point.

The moving target of a kind of static camera the most according to claim 1 shooting is removed video spatial domain tampering detection side Method, it is characterised in that in described spatial filter second step, from suspicious motion dot image AC_h-k+1The tool of middle extraction suspect image block Body detecting step is as follows:

(1). statistics suspicious motion dot image AC_h-k+1In laterally count HOR, HOR (i) of suspicious motion represent the suspicious fortune of the i-th row Move and count；

(2). statistics suspicious motion dot image AC_h-k+1Middle longitudinal suspicious motion count VER, VER (j) represent jth row suspicious fortune Move and count；

(3). according to from suspicious motion dot image AC_h-k+1The horizontal suspicious motion of middle statistics is counted HOR, when in t row to r row The suspicious motion of the every a line HOR (i) that counts is all higher than a certain threshold value, illustrates that t row is to the interval suspicious motion of this line of r row Counting more, be stored in array x1 by t and r, it is interval that t row to r row is referred to as a pair suspicious row, a width suspicious motion dot image There may be multipair suspicious row interval, all start-stop positions, suspicious row interval meeting condition are all stored in array x1；

According to from suspicious motion dot image AC_h-k+1Longitudinal suspicious motion of middle statistics is counted VER, each in t arranges r row The suspicious motion VER (j) that counts of row is all higher than a certain threshold value, illustrates that t arranges the suspicious motion that r arranges these row interval and counts More, t and r is stored in array x2, t arranges r row and is referred to as a pair suspicious row interval；One width suspicious motion dot image may There is multipair suspicious row interval, all interval start-stop positions of suspicious row meeting condition are all stored in array x2；

(4). extract every pair of suspicious row interval and the image block of every pair of suspicious row interval composition in array x2 in array x1, it is assumed that be total to S block, adds up the suspicious motion in each image block and counts C_i=count (AC_h-k+1(i)=0), wherein 1≤i≤s, works as C_iIt is more than During a certain threshold value, then this image block i is likely removed region for moving target, and marking this image block is suspect image block.

The moving target of a kind of static camera the most according to claim 1 shooting is removed video spatial domain tampering detection side Method, it is characterised in that in described spatial filter the 3rd step and the 4th step, energy suspicious degree NT is specifically calculated as follows:

$NT=\frac{1}{B}+H$

The lowest high-frequency energy computing formula than B and frequency domain entropy H is as follows:

$B=\frac{1}{r}\sqrt{\underset{i=1}{\overset{r}{\Σ}}{\mathrm{\β}}_i^2}/\left(\frac{1}{m*n-r}\sqrt{\underset{i=r+1}{\overset{m*n}{\Σ}}{\mathrm{\β}}_i^2}\right)$

$H=-\underset{i=1}{\overset{m*n}{\Σ}}\left(\frac{\left|{\mathrm{\β}}_i\right|}{\underset{i=1}{\overset{m*n}{\Σ}}\left|{\mathrm{\β}}_i\right|}{\log }_2\frac{\left|{\mathrm{\β}}_i\right|}{\underset{i=1}{\overset{m*n}{\Σ}}\left|{\mathrm{\β}}_i\right|}\right)$

Block size is that m*n, r represent that image concentrates on the minority low frequency coefficient number in the upper left corner after dct transform, is become by DCT Two-dimensional array after changing from the upper left corner carry out Z-type scanning dimensionality reduction so as to get one-dimension array coefficient by the side of decreasing energy Formula sorts, β_iRepresent i-th DCT coefficient after sequence.