CN103034993A - Digital video transcode detection method - Google Patents

Abstract

The invention belongs to the technical field of video detection, and relates to a digital video transcode detection method. The method comprises the following steps: building a sample bank; carrying out complete decoding to video data in a video file and obtaining relevant digital video content data in the video file, and building a video file dataset; carrying out respective data modeling to a discrete cosine transformation (DCT) coefficient of all frames and a motion vector and a motion error of a P frame and a B frame, and carrying out feature extraction to the dataset, wherein five features are extracted, namely a periodic feature and a singular point of a DCT curve, an image resampling feature, motion vector similarity and motion error curve periodicity; forming a feature vector, and building a classifier and carrying out training; reading a to-be-detected video file and obtaining a feature vector, and carrying out classification detection to the detected video data using the classifier. By means of the digital video transcode detection method, any pre-process, like adding digital watermarking information and the like, is not needed to a digital video, and the digital video transcode detection method has the advantages of being strong in practicability and small in calculated quantity.

Description

A kind of digital video transcoding detection method

Affiliated technical field

The invention belongs to the video detection technology field, be specifically related to a kind of digital video resource transcoding detection method, whether it is come by other video resource transcoding by analyzing video resource, comes the authenticity of discriminating digit video resource.

Background technology

The digital video coding technology is maked rapid progress in the past few decades, develop rapidly, and be widely used, Network Video Transmission, digital video broadcasting (DVB), digital video disk (DVD), Digital Video, video conference, remote teaching etc. are exactly some typical examples.Different application scenarioss has different requirements to Video Codec, video data Store form, web-transporting device etc., and the coding standard of working out for specific area also is not quite similar.H.26x standard mainly applies to the occasions such as visual telephone and video conference; The MEPG-2 index plane is to occasions such as digital video broadcasting, high definition digital television (HDTV) and DVD; The ultralow bit rate coding mode of MPEG-4 standard more is applicable to mobile multimedia and uses.Along with popularizing of digital video application, require video data format to adopt the requirement of different coding standard to become more and more urgent with applied environment is different.For many real-time application, usually require video data can be between heterogeneous network, different access device and between the different multimedia data layout can seamless link, it is particularly important that Video Transcoding Technology just seems, for example: in video on-demand system, video server is transcoded into the MPEG-4 form that is applicable to low code check transmission with high-quality video, and then people just can and obtain respective resources by the wireless network access video server; In video monitoring system, contextual data reaches the purpose of remote monitoring through just can Internet Transmission behind the transcoding; In the digital television broadcasting field, if want any multimedia terminal can receiving digital broadcast, the equipment that then must add a kind of similar set-top box or video gateway be finished the transcoding function of code check and resolution.

In general, the digital video transcoding process can think to be transformed into from a kind of video compression format the end to end processing procedure of another kind of video compression format, the input, the output that are transcoder all are compressed videos, and the compressed video bit stream behind the transcoding can be more suitable for the demand of transmission bandwidth and receiving end.Video code translator of a great variety if divide according to standard under the video before and after the transcoding, can be divided into two classes: similar video code conversion and variety classes video code conversion.Similar video code conversion refers to the video data stream of certain standard format is converted to the same standard compression form that is suitable for another transmission environment from the compressed format that is fit to a certain transmission environment, mainly comprises picture size (spatial resolution) conversion, frame per second (temporal resolution) conversion, code check conversion; The variety classes video code conversion be will compression video code flow be converted into the code stream of another kind of video compression coding standard from a certain compression and coding standard, the interoperability between the different compression and coding standard system is provided.

Video Transcoding Technology can reasonably be utilized existing video resource, and the MPEG-2 format video resource of former wide-scale distribution can be converted to video format such as the MPEG-4 form of adaptation Internet communication and low consumption terminal device, but also is with simultaneously the problem of serving.Some have the video resource of copyright protection after compressing through transcoding, and copyright protection information will be lost or be damaged, and the primitiveness of video resource also just is difficult to checking; In transcoding process, people also can carry out a series of operation of distorting simultaneously, and such as region duplication and deletion, the authenticity of video resource also just is difficult to judge.The video code conversion process may be destroyed the copyright protection of former video and carry out other the operation of distorting, and has destroyed primitiveness and the authenticity of video, and therefore, whether operation is the important evidence of checking video primitiveness through transcoding to detect digital video.Therefore, the video code conversion detection technique is an important step of video authentication, for judiciary and relevant departments provide necessary evidence.

Summary of the invention

The object of the present invention is to provide a kind of video code conversion detection method, after the method utilizes video resource through the transcoding operation, the regularity that its frame is interior, the statistical distribution of conversion coefficient, motion vector and the pixel domain of interframe occurs changes, extract many stack features amounts from five aspects such as distribution character of resampling, motion vector and the kinematic error of the periodic characteristic of transform coefficient data distribution curve and singular point, pixel domain, and make up corresponding detection model.Statistical nature by these digital videos itself, in conjunction with ripe classification and Detection algorithm, whether the process video code conversion operates to detect digital video resource to be measured, reaches to filter out the purpose that may pass through the video file of distorting operation, realizes the effectively supervision and management to digital video resource.Technical scheme of the present invention is as follows:

A kind of digital video transcoding detection method may further comprise the steps:

(1) sets up a Sample Storehouse, comprise the video file that some have known original compression and the video file that operates through various transcodings;

(2) video data in the video file is carried out complete decoding, and obtain wherein correlated digital video content data, set up the data set of video file, comprise five kinds of data in the data set: reconstruction DCT coefficient, the quantization scale factor, motion vector, kinematic error and pixel value information in the video;

(3) according to the digital video frame type video data being divided into the intraframe coding frame data is that I frame data, forward-predictive-coded frames data are that P frame data and bi-directional predictive coding frame data are B frame data three major types data set;

(4) DCT coefficient and P, the motion vector of B frame, the kinematic error of all frames are carried out respectively data modeling, method is as follows:

I.DCT coefficient modeling: will rebuild the DCT coefficient block according to the quantization scale factor and be divided into N subset Y1, be that reconstruction DCT coefficient block in every subset has the identical quantization scale factor, then to the coefficient block in every subset, positional information according to transform domain, the non-zero transform coefficient of low frequency region is made up respectively DCT coefficient histogram curve, and normalization is designated as H (q; P, n), every curve is corresponding to a quantization scale factor and a low frequency position, and wherein q represents the quantization scale factor of piece, and p represents the low frequency position, and n represents to rebuild the value of DCT coefficient;

Ii. motion vector modeling: take frame as unit, add up respectively the motion vector of its each macro block in the size of horizontal and vertical direction, and deposit by macro block position information, be designated as V (x, y), wherein (x, y) is the positional information of macro block;

Iii. kinematic error modeling: take frame as unit, add up respectively the mean motion error size of its macro block, and make up the corresponding sports graph of errors;

(5) data set is carried out feature extraction, extract altogether five category features: the periodic characteristic of DCT curve and singular point, image resampling feature (claiming again deformation characteristics), the similarity of motion vector and the periodicity of kinematic error curve, wherein the first two feature is the ubiquity feature, and rear three kinds of features are specific aim features.

(6) periodic characteristic of DCT curve extracts:

I. with DCT coefficient histogram H (q; P, n) extract H (q according to n order from small to large; P, n) be not equal to zero point and be rearranged for a histogram H (q; P, m), then to histogram H (q; P, m) the FFT conversion of doing one dimension obtains frequency curve HT (q; P, ω);

Ii. define the curvature representation operator $k(q;p,\mathrm{\ω})=\left|\frac{{H_T}^{\′\′}(q;p,\mathrm{\ω})}{{(1+{\left({H_T}^{\′}(q;p,\mathrm{\ω})\right)}^2)}^{\frac{3}{2}}}\right|$ ω ∈ (0, π), wherein, H ' _T(q; P, ω) and H " _T(q; P, ω) be respectively H _T(q; P, ω) first order derivative and second derivative;

Iii. select k (q; P, ω) maximal value is as histogram H (q; P, n) characteristic quantity;

(7) the singular point feature extraction of DCT coefficient:

I. define protruding Characteristics Detection function $t(q;p,m)=\left\{\begin{array}{cc}0& H^{\&Prime;}(q;p,m)\&GreaterEqual;0\\ \frac{-H^{\&Prime;}(q;p,m)}{H(q;p,m)}& H^{\&Prime;}(q;p,m)<0\end{array}\right.$ m＝2,3

Ii. to obtaining a series of DCT histogram among every subset Y1, select wherein maximum protruding characterisitic parameter as main distinguishing feature;

(8) deformation characteristics is extracted: after the video information complete decoding, by greatest hope algorithm (EM algorithm) these two kinds of signals are carried out cluster analysis, whether detected image contains the resampling vestige;

(9) the motion vector correlative character extracts:

I. define motion vector angle similarity θ and length similarity L:

$\mathrm{\&theta;}=\mathrm{ar}\cos \left(\frac{V_1(x,y)*V_2(x,y)}{\left|V_1(x,y)\right|*\left|V_2(x,y)\right|}\right)$

$L=\frac{\left|V_1(x,y)\right|}{\left|V_2(x,y)\right|}$

Wherein (x, y) is the positional information of macro block, V ₁(x, y) and V ₂(x, y) is the motion vectors of adjacent two frames on same position (x, y);

Ii. define similarity and detect operator

Iii. in an image sets, each predictive frame can obtain a series of motion vector similarity value, when the similarity detection operator that macro blocks more than half are arranged in the frame was 0, the similarity that then defines this frame was 0, represents that this video has probably passed through the video code conversion operation;

(10) periodicity of kinematic error curve: the FFT that the mean motion graph of errors is done one dimension changes, if new secondary lobe occurred on its frequency domain map of magnitudes, has illustrated that then kinematic error exists periodically, also points out that video has passed through simultaneously to delete frame or interleave operation;

(11) the above-mentioned characteristic quantity of comprehensive three class frames forms proper vector, sets up sorter and trains;

(12) read video file to be detected, repeating step (2) obtains the proper vector of video file to be detected to (11), utilize sorter that the video data that detects is carried out classification and Detection, it is divided into two classes: through the video resource file of transcoding compressed encoding and the video resource file of original compression coding.

The present invention differentiates the primitiveness of video resource and authenticity, differentiates video resource and whether passes through the transcoding squeeze operation, distorts to detect for omnibearing video and lays a good foundation.Its evident characteristic of the present invention comprises:

(1) practicality: the video code conversion detection scheme is to utilize the internal data statistical property of video code flow to differentiate video, does not need in advance digital video to be done any pre-service as add digital watermark information etc. in video, and is practical.

(2) gradability: two category features are arranged in the characteristic extracting module of the present invention: ubiquity feature and specific aim feature, when the needs fast detecting, can utilize the ubiquity feature, its utilization be the statistical distribution Changing Pattern of conversion coefficient, need not carry out the complete decoding operation, reduced greatly operand, execution speed is fast; When needs detect targetedly, can utilize the specific aim feature, as the resampling feature can detection space resolution conversion, the motion vector similarity can detection time the conversion of resolution, the periodicity of kinematic error can detect deletes frame or interleave operation; When needs obtain high precision detection effect, can unite the primitiveness that ubiquity feature and specific aim feature detect video.

(3) applicability: two category features that the present invention proposes almost can detect all video code conversion operations and change, delete the operation of (inserting) frame and standard handovers such as code check conversion, temporal resolution conversion, spatial resolution, and practicality is wide.

(4) novelty: the feature that the present invention proposes series of novel is described operator, proposed ubiquity feature and specific aim feature, has fully reflected the impact that the digital video transcoding squeeze operation causes the statistical distribution of video data.

Description of drawings

Fig. 1 is the overall flow figure of video code conversion detection system of the present invention;

Fig. 2 is the process flow diagram of date statistical modeling module of the present invention;

Fig. 3 is the low frequency region synoptic diagram of 8 * 8 dct transform domains;

Fig. 4 is the process flow diagram of low frequency region DCT coefficient of the present invention modeling;

Fig. 5 is the process flow diagram of characteristic extracting module of the present invention;

Fig. 6 is the figure after a representative DCT coefficient histogram is reset;

Fig. 7 is a histogrammic frequency curve of representative DCT.

Embodiment

Fig. 1 illustrates the overall flow figure of video code conversion detection system of the present invention.In steps A, realize the complete decoding of video resource by data-analyzing machine, and obtain picture material data wherein, comprise information such as rebuilding DCT coefficient, the quantization scale factor, motion vector, kinematic error and pixel value.At step B, DCT coefficient, motion vector, kinematic error are carried out respectively modeling, and be depicted as corresponding curve.At step C, the statistic of video data is carried out signature analysis, and extract various characteristic quantities, the last comprehensive proper vector that forms.At step D, the sorter by a technology maturation is divided into two classes with video resource: the video resource file of encoding through video resource file and the original compression of transcoding compression.

The date statistical modeling module at first is divided into intracoded frame (I frame) data, forward-predictive-coded frames (P frame) data and bi-directional predictive coding frame (B frame) data three major types data set to data according to the digital video frame type, five kinds of data in each data set, have been comprised: rebuild DCT coefficient, the quantization scale factor, motion vector, kinematic error and pixel value, wherein the I frame does not adopt estimation, so its motion vector and kinematic error data are zero entirely.And DCT coefficient, motion vector, kinematic error carried out respectively modeling:

What Fig. 2 described is the date statistical modeling process.At first according to the digital video frame type data are divided into intracoded frame (I frame) data, forward-predictive-coded frames (P frame) data and bi-directional predictive coding frame (B frame) data three major types data set.Then in every frame, respectively low frequency DCT coefficient, motion vector and kinematic error are carried out modeling:

(1) low frequency region DCT coefficient modeling: as shown in Figure 3, it is the low frequency region synoptic diagram of 8 * 8 dct transform domains, white colour partly represents low frequency region among the figure, and dark colour partly represents the medium-high frequency zone, and the medium-high frequency zone is owing to data quantize seriously will not add up.In decode procedure, each piece has their a quantization scale factor, therefore we can be divided into some subset Y1 to all pieces of every frame according to it, the piece of every subset all has the identical quantization scale factor, and this is conducive to eliminate the aliasing effect that causes because of quantization scale factor difference.Then to the piece in every subset, according to the positional information of transform domain, the non-zero transform coefficient of low frequency region is made up respectively histogram curve, and normalization, whole process can be depicted as Fig. 4.Therefore, to a video to be measured, we can obtain a series of normalization histogram curve and be designated as H (q; P, n), every curve is corresponding to a quantization scale factor and a low frequency position, and wherein q represents the quantization scale factor of piece, and p represents the low frequency position, and n represents to rebuild the value of DCT coefficient.

(2) motion vector modeling: take frame as unit, add up respectively the motion vector of its each macro block in the size of horizontal and vertical direction, and deposit by the positional information of macro block, in order to carry out next step feature extraction.

(3) kinematic error modeling: take frame as unit, add up respectively the mean motion error of its macro block, and make up the corresponding sports graph of errors.

Fig. 5 has described the process flow diagram of characteristic extracting module of the present invention.The present invention will extract respectively feature from three class data be I frame feature, P frame feature, B frame feature, after obtaining DCT coefficient histogram curve, motion vector distribution, kinematic error curve and picture material, can extract respectively feature from following five aspects, then comprehensive all Characteristics creation proper vectors are input to proper vector at last and whether can finish video in the sorter through transcoding compression discriminating.

(1) periodic characteristic of DCT coefficient:

In order to embody better the periodic feature of curve, we are with DCT coefficient histogram H (q; P, n) extract H (q according to n order from small to large; P, n) be not equal to zero point and be rearranged for a histogram H (q; P, m), as shown in Figure 6.Then frequency curve H is obtained in the FFT conversion of this curve being done one dimension _T(q; P, ω), as shown in Figure 7, the DCT coefficient had obvious periodic characteristic after we can find video code conversion, and this is distinct with the original compression DCT coefficient spectrum curve of dullness decline.

For the existence in new cycle of surveying every spectrum curve, we have defined a curvature representation operator:

$k(q;p,\mathrm{\&omega;})=\left|\frac{{H_T}^{\&prime;\&prime;}(q;p,\mathrm{\&omega;})}{{(1+{\left({H_T}^{\&prime;}(q;p,\mathrm{\&omega;})\right)}^2)}^{\frac{3}{2}}}\right|$ ω∈(0,π)（1）

H ' wherein _T(q; P, ω) and H " _T(q; P, ω) be respectively H _T(q; P, ω) first order derivative and second derivative, k (q; P, ω) to have described under certain quantization scale factor q, the smooth degree of the DCT coefficient spectrum curve on the p of low frequency position selects maximal value as the characteristic quantity of this quantization scale factor.

$K\left(q\right)=\underset{p,\mathrm{\&omega;}}{\max }k(q;p,\mathrm{\&omega;})---\left(2\right)$

(2) singular point of DCT coefficient:

In the video code conversion process, can relate to twice quantizing process, quantization scale factor q in its twice quantizing process of same pixel macroblock generally is different, frame is interior, the interframe quantization matrix is also different, see once whole video code conversion process as special quantizing process, then its quantization step presents between periodic size and replaces, and shows as that concavo-convex transformation has occured in certain some position on the DCT histogram.Because DCT coefficient data suppression ratio is very fast, locational DCT data statistics amount may be not after the 2nd cycle, be difficult to produce a desired effect, it is whether 2,3 locational DCT numbers concavo-convex transformation has occured that the present invention only detects first period position, as shown in Figure 6,2 obvious concavo-convex conversion has occured, and by contrast, the video DCT histogram of original compression is rendered as a laplacian curve.

For the convex function characteristic of quantitative test quantization DCT coefficient histogram curve, we have defined a detection function t (q; P, m):

$t(q;p,m)=\left\{\begin{array}{cc}0& H^{\&prime;\&prime;}(q;p,m)\&GreaterEqual;0\\ \frac{-H^{\&prime;\&prime;}(q;p,m)}{H(q;p,m)}& H^{\&prime;\&prime;}(q;p,m)<0\end{array}\right.$ m＝2,3(3)

To obtaining a series of DCT histogram among every subset Y1, select large protruding characterisitic parameter as main distinguishing feature.

$T\left(q\right)=\underset{p,m}{\max }t(q;p,m);---\left(4\right)$

(3) image resampling feature:

When the video code conversion process comprised the conversion of temporal resolution, up-sampling or down-sampling will cause existing between the image pixel correlativity in certain cycle, that is to say, the image after the resampling is in fact the stack of original image signal and certain periodic signal.After the video information complete decoding, by EM(Expectation maximization) algorithm carries out cluster analysis to these two kinds of signals, and just can detect image and whether contain the resampling vestige.

(4) motion vector correlativity:

After the temporal resolution conversion has occured in transcoding process, the similar degree of adjacent two frames of video will descend greatly, and the motion vector that shows as adjacent two frames has larger gap in direction and size.For quantitative these gaps, we have defined similarity and have detected operator s (x, y):

Wherein (x, y) is the positional information of macro block, and θ and L are respectively angle and the length similarity of motion vector, are defined as follows V ₁(x, y) and V ₂(x, y) is the motion vectors of adjacent two frames on same position (x, y).

$\mathrm{\&theta;}=\mathrm{ar}\cos \left(\frac{V_1(x,y)*V_2(x,y)}{\left|V_1(x,y)\right|*\left|V_2(x,y)\right|}\right)---\left(6\right)$

$L=\frac{\left|V_1(x,y)\right|}{\left|V_2(x,y)\right|}---\left(7\right)$

In an image sets, each predictive frame can obtain a series of motion vector similarity value, have its motion vector of macro block of half dissimilar in a frame, the similarity that then defines this frame is 0, is also representing this video simultaneously and is probably passing through the video code conversion operation.

Num is the macro block number that a frame comprises.

(5) periodicity of kinematic error:

The operation of frame or interleave has occured to delete in transcoding process, and then the frame type in subsequently the video sequence can be changed on generating period ground, and this can cause the periodicity pattern of kinematic error.Exist for accurately detecting is periodic, the FFT that the mean motion graph of errors is done one dimension changes, if new secondary lobe occurred on its frequency domain map of magnitudes, has illustrated that then kinematic error exists periodically, has also pointed out that video has passed through simultaneously to delete frame or interleave operation.When carrying out the digital video transcoding compressed detected, at first to set up the Sample Storehouse for training, Sample Storehouse comprises the video file that some have known original compression and the video file that operates through various transcodings.Then each video file in the Sample Storehouse is carried out complete decoding and feature extraction, obtain proper vector, and set up sorter and train.For video file to be detected, adopting uses the same method obtain its proper vector after, utilize sorter that the video data that detects is carried out classification and Detection, it is divided into two classes: through the video resource file of transcoding compression and the video resource file of original compression coding, thereby finish the video discriminating whether the process transcoding compresses.

Claims (1)

1. digital video transcoding detection method may further comprise the steps:

(1) sets up a Sample Storehouse, comprise the video file that some have known original compression and the video file that operates through various transcodings;

(2) video data in the video file is carried out complete decoding, and obtain wherein correlated digital video content data, set up the data set of video file, comprise five kinds of data in the data set: reconstruction DCT coefficient, the quantization scale factor, motion vector, kinematic error and pixel value information in the video;

(3) according to the digital video frame type video data being divided into the intraframe coding frame data is that I frame data, forward-predictive-coded frames data are that P frame data and bi-directional predictive coding frame data are B frame data three major types data set;

(4) DCT coefficient and P, the motion vector of B frame, the kinematic error of all frames are carried out respectively data modeling, method is as follows:

I.DCT coefficient modeling: will rebuild the DCT coefficient block according to the quantization scale factor and be divided into N subset Y1, be that reconstruction DCT coefficient block in every subset has the identical quantization scale factor, then to the coefficient block in every subset, positional information according to transform domain, the non-zero transform coefficient of low frequency region is made up respectively DCT coefficient histogram curve, and normalization is designated as H (q; P, n), every curve is corresponding to a quantization scale factor and a low frequency position, and wherein q represents the quantization scale factor of piece, and p represents the low frequency position, and n represents to rebuild the value of DCT coefficient;

Ii. motion vector modeling: take frame as unit, add up respectively the motion vector of its each macro block in the size of horizontal and vertical direction, and deposit by macro block position information, be designated as V (x, y), wherein (x, y) is the positional information of macro block;

Iii. kinematic error modeling: take frame as unit, add up respectively the mean motion error size of its macro block, and make up the corresponding sports graph of errors;

(5) data set is carried out feature extraction, extract altogether five category features: the periodic characteristic of DCT curve and singular point, image resampling feature (claiming again deformation characteristics), the similarity of motion vector and the periodicity of kinematic error curve, wherein the first two feature is the ubiquity feature, and rear three kinds of features are specific aim features.

(6) periodic characteristic of DCT curve extracts:

I. with DCT coefficient histogram H (q; P, n) extract H (q according to n order from small to large; P, n) be not equal to zero point and be rearranged for a histogram H (q; P, m), then to histogram H (q; P, m) the FFT conversion of doing one dimension obtains frequency curve H _T(q; P, ω);

Ii. define the curvature representation operator $k(q;p,\mathrm{\&omega;})=\left|\frac{{H_T}^{\&prime;\&prime;}(q;p,\mathrm{\&omega;})}{{(1+{\left({H_T}^{\&prime;}(q;p,\mathrm{\&omega;})\right)}^2)}^{\frac{3}{2}}}\right|$ ω ∈ (0, π), wherein, H ' _T(q; P, ω) and H " _T(q; P, ω) be respectively H _T(q; P, ω) first order derivative and second derivative;

Iii. select k (q; P, ω) maximal value is as histogram H (q; P, n) characteristic quantity;

(7) the singular point feature extraction of DCT coefficient:

I. define protruding Characteristics Detection function $t(q;p,m)=\left\{\begin{array}{cc}0& H^{\&prime;\&prime;}(q;p,m)\&GreaterEqual;0\\ \frac{-H^{\&prime;\&prime;}(q;p,m)}{H(q;p,m)}& H^{\&prime;\&prime;}(q;p,m)<0\end{array}\right.$ m＝2,3

Ii. to obtaining a series of DCT histogram among every subset Y1, select wherein maximum protruding characterisitic parameter as main distinguishing feature;

(8) deformation characteristics is extracted: after the video information complete decoding, by greatest hope algorithm (EM algorithm) these two kinds of signals are carried out cluster analysis, whether detected image contains the resampling vestige;

(9) the motion vector correlative character extracts:

I. define motion vector angle similarity θ and length similarity L:

$\mathrm{\&theta;}=\mathrm{ar}\cos \left(\frac{V_1(x,y)*V_2(x,y)}{\left|V_1(x,y)\right|*\left|V_2(x,y)\right|}\right)$

$L=\frac{\left|V_1(x,y)\right|}{\left|V_2(x,y)\right|}$

Wherein (x, y) is the positional information of macro block, V ₁(x, y) and V ₂(x, y) is the motion vectors of adjacent two frames on same position (x, y);

Ii. define similarity and detect operator

Iii. in an image sets, each predictive frame can obtain a series of motion vector similarity value, when the similarity detection operator that macro blocks more than half are arranged in the frame was 0, the similarity that then defines this frame was 0, represents that this video has probably passed through the video code conversion operation;

(10) periodicity of kinematic error curve: the FFT that the mean motion graph of errors is done one dimension changes, if new secondary lobe occurred on its frequency domain map of magnitudes, has illustrated that then kinematic error exists periodically, also points out that video has passed through simultaneously to delete frame or interleave operation;

(11) the above-mentioned characteristic quantity of comprehensive three class frames forms proper vector, sets up sorter and trains;

(12) read video file to be detected, repeating step (2) obtains the proper vector of video file to be detected to (11), utilize sorter that the video data that detects is carried out classification and Detection, it is divided into two classes: through the video resource file of transcoding compressed encoding and the video resource file of original compression coding.

Images