CN110121109A - Towards the real-time source tracing method of monitoring system digital video, city video monitoring system - Google Patents
Towards the real-time source tracing method of monitoring system digital video, city video monitoring system Download PDFInfo
- Publication number
- CN110121109A CN110121109A CN201910224267.1A CN201910224267A CN110121109A CN 110121109 A CN110121109 A CN 110121109A CN 201910224267 A CN201910224267 A CN 201910224267A CN 110121109 A CN110121109 A CN 110121109A
- Authority
- CN
- China
- Prior art keywords
- video
- real
- monitoring system
- noise
- class
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 62
- 238000012544 monitoring process Methods 0.000 title claims abstract description 35
- 238000000605 extraction Methods 0.000 claims abstract description 10
- 238000000354 decomposition reaction Methods 0.000 claims description 9
- 239000000284 extract Substances 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 238000002790 cross-validation Methods 0.000 claims description 5
- 238000012360 testing method Methods 0.000 claims description 5
- 238000012549 training Methods 0.000 claims description 5
- 239000012141 concentrate Substances 0.000 claims description 4
- 239000000654 additive Substances 0.000 claims description 3
- 230000000996 additive effect Effects 0.000 claims description 3
- 238000013178 mathematical model Methods 0.000 claims description 3
- 238000006243 chemical reaction Methods 0.000 claims 1
- 238000012545 processing Methods 0.000 abstract description 11
- 238000005516 engineering process Methods 0.000 abstract description 10
- 230000009466 transformation Effects 0.000 abstract description 8
- 238000012706 support-vector machine Methods 0.000 description 10
- 238000002474 experimental method Methods 0.000 description 9
- 238000010586 diagram Methods 0.000 description 7
- 238000004458 analytical method Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000004069 differentiation Effects 0.000 description 1
- 238000013095 identification testing Methods 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012806 monitoring device Methods 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/30—Noise filtering
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N17/00—Diagnosis, testing or measuring for television systems or their details
- H04N17/004—Diagnosis, testing or measuring for television systems or their details for digital television systems
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
- H04N21/23—Processing of content or additional data; Elementary server operations; Server middleware
- H04N21/234—Processing of video elementary streams, e.g. splicing of video streams or manipulating encoded video stream scene graphs
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
- H04N21/23—Processing of content or additional data; Elementary server operations; Server middleware
- H04N21/234—Processing of video elementary streams, e.g. splicing of video streams or manipulating encoded video stream scene graphs
- H04N21/23418—Processing of video elementary streams, e.g. splicing of video streams or manipulating encoded video stream scene graphs involving operations for analysing video streams, e.g. detecting features or characteristics
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/43—Processing of content or additional data, e.g. demultiplexing additional data from a digital video stream; Elementary client operations, e.g. monitoring of home network or synchronising decoder's clock; Client middleware
- H04N21/44—Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
- H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
- H04N21/43—Processing of content or additional data, e.g. demultiplexing additional data from a digital video stream; Elementary client operations, e.g. monitoring of home network or synchronising decoder's clock; Client middleware
- H04N21/44—Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs
- H04N21/44008—Processing of video elementary streams, e.g. splicing a video clip retrieved from local storage with an incoming video stream or rendering scenes according to encoded video stream scene graphs involving operations for analysing video streams, e.g. detecting features or characteristics in the video stream
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N7/00—Television systems
- H04N7/18—Closed-circuit television [CCTV] systems, i.e. systems in which the video signal is not broadcast
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to field of information security technology, disclose one kind and H.264 analyze towards the real-time source tracing method of monitoring system digital video, city video monitoring system, including to the common coding mode of monitor video, obtain the key frame of monitor video；The extraction of PRNU noise is carried out to the key frame of monitor video；Real-time grading is carried out using statistical nature of the real-time classification method to PRNU noise.The present invention improves the accuracy rate of video identifing source, while characteristic dimension can be made to reduce using SFFS algorithm and not lose precision, this, which helps to reduce, calculates time and system complexity；Computation complexity is low and can be applied to video of different sizes.In addition, being also robust for the video processing and geometric transformation seemed；It is very steady that image by primary image processing or geometric transformation is all showed.It is very important in actual application scenarios.Therefore, which can be used as the effective ways of video source video camera identification.
Description
Technical field
The invention belongs to field of information security technology, more particularly to it is a kind of towards the monitoring system digital video side of tracing to the source in real time
Method, city video monitoring system.
Background technique
Currently, the immediate prior art: along with the increasingly increase of monitoring system scale, safe city video monitoring skill
Requirement of the large-scale application and people of art to public safety is increasingly increased, and skillfully sharply increasing for video monitoring equipment is caused
Long, from urban construction to public transport, from community security protection to production safety, none does not need a large amount of video monitoring equipment.?
One line city, the quantity of video monitoring equipment are even more in terms of hundreds of thousands of, and so large-scale video monitoring system can all have daily
Video information of the number in terms of PB generates.Therefore, some safety problems in video monitoring system are also gradually exposed.Because of number
Being worth video and image has real-time and understandable form, ready-made for one of current main media method.Digital technology
Progress expedite the emergence of low-cost Digital Video collection equipment, such as mobile phone, digital camera, video monitoring equipment etc., this
So that the generation of digital audiovisual data becomes to be more easier.And since image, Video editing software become more easily to use,
It is very easy for modifying the content of video.Due to the source of video diversity and be easy the characteristic being tampered, number view
It the certification of frequency content and traces to the source and becomes more difficult.Therefore, it is necessary to new techniques or methods to identify the source of image, video simultaneously
Assess their authenticity and validity.Digital video forensic technologies are to detect to the source of video and originality, and know
The other operation that it is carried out.Assessment and detection to digital video or image can provide in court for various criminal offences can
By ground scientific basis.Digital video tracing technology is the branch of digital video evidence obtaining, it is intended to which identification is for capturing digital video
Source camera.
Although video monitoring system is widely used in safe city at present builds, but the peace of video monitoring system itself
Full problem but needs further to be studied.At present in terms of digital multimedia forensic technologies, the research of multimedia tracing technology is main
Concentrate in terms of being image, compare and say for the research in terms of video will be backward very much, reason mainly has following:
Digital video has more complexity, either digital collection or compressed encoding compared to digital picture；Digital video is before transmission
Lossy compression is carried out because data volume is big, so that part system Character losing；Video processing technique is more and more various, traces back
Source difficulty is also increasing.
In conclusion problem of the existing technology is: existing video is traced to the source flat applied to safe city intelligent security guard
The correlative study of the video processing and analysis of platform is simultaneously few, and the video tracing technology towards safe city video monitoring system is ground
Study carefully and is still in infancy.
The difficulty of solution above-mentioned technical problem: one, video has the complexity of time shaft relative to image；Two, number view
Camera feature extraction difficulty belonging to frequency is big；Three, it is increasing that increasingly diversified video processing technique makes video trace to the source.
It solves the meaning of above-mentioned technical problem: safe city can be made safer, it, can be in case of special event
The source camera for quickly finding digital video, determines its source, to especially passively evidence obtaining has very great help with digital evidence obtaining.
Summary of the invention
In view of the problems of the existing technology, the present invention provides one kind towards the monitoring system digital video side of tracing to the source in real time
Method, city video monitoring system.
The invention is realized in this way one kind is towards the real-time source tracing method of monitoring system digital video, which is characterized in that institute
It states and includes: towards the real-time source tracing method of monitoring system digital video
H.264, the first step analyzes the common coding mode of monitor video, obtains the key frame of monitor video；
Second step carries out the extraction of PRNU noise to the key frame of monitor video；
Third step carries out real-time grading using statistical nature of the real-time classification method to PRNU noise.
Further, H.264 the first step analyzes the common coding mode of monitor video, obtains monitor video
Key frame selects the extraction of I frame progress PRNU noise.
Further, the extraction that the second step carries out PRNU noise to the key frame of monitor video specifically includes:
(1) PRNU noise extracts, and the mathematical model of Image Acquisition indicates are as follows:
yij=fij(xij+ηij)+cij+εij；
Wherein, yijIt is sensor output, xijThe quantity of the photon of sensor, η are hit in expressionijIt is shot noise, cijIt is dark
Current noise, εijIt is Complex-valued additive random noise, fijIndicate PRNU noise contribution；
Image is removed dryness using the filter that removes dryness based on small echo；It is the original of dominant noise component from PRNU is had
The noise residual error that the image after removing dryness obtains given image is subtracted in input picture；With IkIndicate original input picture, size
For M × N,It is to IkIt is being obtained after being removed dryness as a result,It is from original image IkIn the noise residual error extracted；
(2) feature selecting and classification shown with two group profiles, first group include estimation PRNU noise high-order small echo system
Meter, second group include original video frame statistical nature；Feature is combined, can be used for the identification of video source camera；Respectively
Three-level wavelet decomposition is carried out using Haar small echo to each Color Channel in video frame, obtains 9 subbands；For all subbands
The statistical nature for calculating separately single order and higher order finally obtains 108 features；Feature set is calculated using before sequence to feature selecting
Method feature selecting.
Further, specific step is as follows for the process that removes dryness based on small echo:
(1) it executes multilevel wavelet to the key frame of video extracted using 8-tap Daubechies QMF to decompose, by this
Frame is divided into level Four subband, wherein vertically, horizontal, diagonal subband respectively indicates are as follows: h (i, j), v (i, j), d (i, j)；
(2) original nothing is calculated with the MAP of the every level-one sub-band coefficients obtained in square W × W neighborhood estimation to make an uproar video frame
Local variance, wherein (3,5,7,9) W ∈；
(3) take the minimum value in four variances being calculated as the final estimated value of local variance.
σ2(i, j)=min (σ3 2(i,j),σ5 2(i,j),σ7 2(i,j),σ9 2(i,j)),(i,j)∈J；
(4) the vertical component h (i, j) of wavelet coefficient is removed dryness with Weiner filter, formula are as follows:
(5) above procedure is repeated to each Color Channel of all subbands of each decomposition rank and video frame, then carried out
You obtain denoising image after wavelet transformation.
Further, the third step using real-time classification method carries out real-time grading to the statistical nature of PRNU noise
It specifically includes:
(1) the more mode classifications of SVM have:
1) a pair of all；
M class is given, needs to train m two classification device；Classifier i therein is class that the setting of i class data is positive, separately
M-1 outer other classes other than i class are all set to negative class；One two classifier is trained to each class, finally
Available m classifier altogether；There is the data x of classification demand for one, class belonging to x is determined by the way of ballot
Not；
2) all to all；
M class is given, a classifier is trained two-by-two to class all in m class, the number of such two classifier is in total
There is m (m-1)/2；There is the data x of classification demand for one, the prediction of each classifier will be passed through, and uses ballot
Mode determine class belonging to x.
(2) use the mode of ten folding cross validations to assess the performance of classifier, data set is divided into ten parts, so that each
Part includes 1/10th that initial data concentrates sample size；From the quantity of the class samples of two different classes of middle selections each
It is almost equal in part；It is developed in turn using nine training datasets for classifier, and uses remaining a test data
Assess the classifier of building.
Another object of the present invention is to provide described in a kind of application towards the real-time source tracing method of monitoring system digital video
City video monitoring system.
In conclusion advantages of the present invention and good effect are as follows: the invention proposes the PRNU estimations based on Wavelet Denoising Method
Method, and indicated with high-order small echo statistics (HOWS).Experiments have shown that this method improves the accuracy rate of video identifing source, simultaneously
Characteristic dimension can be made to reduce using SFFS algorithm and not lose precision, this, which helps to reduce, calculates time and system complexity.
The algorithm computation complexity is low and can be applied to video of different sizes.Furthermore, it is not necessary that any preprocess method.And it is right
It is handled in the video seemed and geometric transformation is also robust.PRNU estimation method based on Wavelet Denoising Method has video identifing source
Ability, even video is captured from similar video camera and similar scene.The present invention is based on PRNU features
It is very steady that new method all shows the image by primary image processing or geometric transformation.This is in actual applied field
It is very important in scape.Therefore, the effective ways of video source video camera identification be can be used as.
Detailed description of the invention
Fig. 1 is provided in an embodiment of the present invention towards the real-time source tracing method flow chart of monitoring system digital video.
Fig. 2 is provided in an embodiment of the present invention respectively using the accuracy rate comparison schematic diagram of tracing to the source of I, GOP, P frame.
Fig. 3 is the estimation schematic diagram of PRNU noise provided in an embodiment of the present invention.
Fig. 4 is calculating set of image characteristics schematic diagram provided in an embodiment of the present invention.
Fig. 5 is the HOWS feature schematic diagram provided in an embodiment of the present invention from two video cameras.
Fig. 6 is the HOWS feature schematic diagram provided in an embodiment of the present invention from three video cameras.
Fig. 7 is feature selecting algorithm flow chart provided in an embodiment of the present invention.
Fig. 8 is the video frame sample schematic diagram of use (a) HK1 and (b) TD1 shooting provided in an embodiment of the present invention.
Fig. 9 is that use (a) HK3 provided in an embodiment of the present invention illustrates with the similar scene video frame sample that (b) TD1 is shot
Figure.
Figure 10 is inventive algorithm provided in an embodiment of the present invention and traditional algorithm performance comparison schematic diagram.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
It traces to the source for existing video and is ground applied to the correlation of the video processing and analysis of safe city intelligent security platform
Study carefully and few problem.The present invention is based on the new methods of PRNU feature for the figure by primary image processing or geometric transformation
It is very steady as what is all showed.This is very important in actual application scenarios.It can be used as the identification of video source video camera
Effective ways.
Application principle of the invention is explained in detail with reference to the accompanying drawing.
It towards the real-time source tracing method of monitoring system digital video include following as shown in Figure 1, provided in an embodiment of the present invention
Step:
S101: H.264 the common coding mode of monitor video is analyzed, the key frame of monitor video is obtained；
S102: the extraction of PRNU noise is carried out to the key frame of monitor video；
S103: real-time grading is carried out using statistical nature of the real-time classification method to PRNU noise.
Application principle of the invention is further described with reference to the accompanying drawing.
1, key frame of video choosing method
Major video monitoring device is all made of H.264 coding mode at present, and three kinds of frames are defined in H.264 agreement, complete
The frame for reorganizing code is I frame, and the frame only encoded comprising difference section generated with reference to I frame before is P frame, also before a kind of reference
The coding of frame is B frame afterwards.I frame is intracoded frame, it can be understood as the complete reservation of this frame picture, therefore only needed when decoding
The data of itself can be completed, without necessarily referring to other frames.Therefore, I frame can be used as the key frame for extracting PRNU noise.Such as Fig. 2
It is shown, PRNU noise is extracted in GOP combined frames, needs accumulative about 200 frames that can just extract complete PRNU noise, and
PRNU noise is extracted in P frame then needs more P frames.And PRNU noise is extracted in I frame only needs about 60 frames that can extract
Complete PRNU noise.It can be seen that extracting PRNU noise in I frame is the highest extracting mode of efficiency.Therefore, the present invention selects
Select the extraction that I frame carries out PRNU noise.
2, sensor PRNU noise extracting method
1. PRNU noise extracts
The mathematical model of Image Acquisition can indicate are as follows:
yij=fij(xij+ηij)+cij+εij；
Wherein, yijIt is sensor output, xijThe quantity of the photon of sensor, η are hit in expressionijIt is shot noise, cijIt is dark
Current noise, εijIt is Complex-valued additive random noise, fijIndicate PRNU noise contribution.Determine that the noise component(s) f in image is determining image
The necessary condition of source camera.Compared with the statistical nature of original image, the statistical nature for extracting PRNU noise is more advantageous to knowledge
The source camera of other image.Currently, most effective PRNU noise extracting mode is to remove dryness to estimate that PRNU makes an uproar from image using small echo
Sound.Using being a little for noise residual error, low frequency component present in image can be inhibited automatically.Calculating process is as shown in Figure 3.
Image is removed dryness using the filter that removes dryness based on small echo.It is the original of dominant noise component from PRNU is had
The image after removing dryness is subtracted in input picture to obtain the noise residual error of given image.With IkIndicate original input picture, it is big
Small is M × N,It is to IkIt is being obtained after being removed dryness as a result,It is from original image IkIn the noise residual error extracted.
Process resume in two steps is removed dryness based on small echo.The estimation of image variance is executed in the first step, second step then uses
Weiner filter obtains denoising image.Steps are as follows for specific algorithm:
(1) it executes multilevel wavelet to the key frame of video extracted using 8-tap Daubechies QMF to decompose, by this
Frame is divided into level Four subband, wherein vertically, horizontal, diagonal subband respectively indicates are as follows: h (i, j), v (i, j), d (i, j).
(2) original nothing is calculated with the MAP of the every level-one sub-band coefficients obtained in square W × W neighborhood estimation to make an uproar video frame
Local variance, wherein (3,5,7,9) W ∈.
(3) take the minimum value in four variances being calculated as the final estimated value of local variance.
σ2(i, j)=min (σ3 2(i,j),σ5 2(i,j),σ7 2(i,j),σ9 2(i,j)),(i,j)∈J；
(4) the vertical component h (i, j) of wavelet coefficient is removed dryness with Weiner filter, formula are as follows:
(5) and so on, each Color Channel of all subbands and video frame to each decomposition rank repeats the above mistake
Journey, then denoising image is obtained after carrying out your wavelet transformation.σ is chosen under normal conditions0It=5 the case where, in the case can be
PRNU value is extracted in complicated noise contribution.
2. feature selecting and classification
After estimating PRNU modal noise according to video frame, the feature by can be used for classifying is needed to indicate PRNU
Noise.The present invention is quasi- to indicate them with two groups of features, and first group includes that the high-order small echo of PRNU noise of estimation counts, second
Group includes the statistical nature of original video frame.These features are combined, can be used for the identification of video source camera.
Source video camera recognition methods based on feature proposed by the invention uses HOWS feature, can be caught by this feature
Obtain the coefficient distribution of the PRNU noise of estimation.HOWS is built upon on the basis of multi-scale wavelet decomposition, for analyzing nature figure
The statistical model of picture.The decomposition (QMF) of PRNU image of this method based on separable quadrature mirror filter, by frequency space
It is divided into multiple scales and direction, level (LH) is generated in each decomposition rank, vertical (HL) and diagonal line (HH) three carries.
The statistical natures such as mean value, variance, skewness and kurtosis are calculated for each included coefficient, as shown in Figure 4:
This method carries out three-level wavelet decomposition using Haar small echo to each Color Channel in video frame respectively, obtains 9
A subband.The statistical nature that single order and higher order are calculated separately for all subbands finally obtains 108 features.These features
Including space domain characteristic, such as the mean value and variance of original video frame and image PRNU, a total of 120 features.These features
Showing between the video of different cameras shooting has apparent difference, therefore can be used for the source camera identification of video.Figure
5, Fig. 6 respectively illustrate the feature distribution for belonging to the video of two and the shooting of three different cameras.
Scatter plot depicts the feature of different cameras, it can be seen that the feature of different cameras can be distinguished obviously in figure
Out.But with the increase of source number of cameras, some overlappings can be generated between the feature of different cameras.Each extract
Feature be not all have to distinguish various types of another characteristic, and some of them be characterized in it is extra.In order to optimize this
Big feature set, the present invention are used for feature selecting to feature selecting (SFFS) algorithm using before sequence.
To the big measure feature of feature selecting program analysis to eliminate redundancy feature before sequence, and only select effective to classifying
Important feature.The algorithm is applied to the entire feature set extracted from all videos simultaneously.Algorithm flow such as Fig. 7.
In each step, support vector machines (SVM) assesses the performance of selected feature set according to recognition accuracy.
3.SVM and cross validation mode
The present invention utilizes the feature training from known source category using the source video camera of SVM classifier identification test video
Classifier identifies to complete source video camera.And the performance of classifier is assessed using a variety of different kernel functions and cost parameter
To select optimal parameter.SVM is a kind of typical two classification device.However it to be solved the problems, such as in reality, often more points
The problem of class.In the present invention, as soon as when need to identify in multiple equipment when, polytypic SVM must be used.Use SVM
More classification problems are solved, main method is exactly multiple two classifiers of training.SVM is more, and mode classification has following two:
(1) a pair of all (One-Versus-All, OVA)
M class is given, needs to train m two classification device.Classifier i therein is class that the setting of i class data is positive, separately
M-1 outer other classes other than i class are all set to negative class.In this way, two classifiers are trained to each class,
In this way, finally having available m classifier altogether.There is the data x of classification demand for one, is determined by the way of ballot
Classification belonging to x.Such as data x is predicted using classifier i, if obtain the result is that positive class, illustrates classifier i
Classification results to data x are that x belongs to i class, then i class obtains a ticket.If obtaining the result is that negative class, illustrates that x is not belonging to i
Class, then other classes other than i class will obtain a ticket.It is final to count the most class of gained vote quantity, class belonging to as x.
(2) all to all (All-Versus-All, AVA)
M class is given, a classifier is trained two-by-two to class all in m class, the number of such two classifier is in total
There is m (m-1)/2.There is the data x of classification demand for one, it will pass through the prediction of each classifier, and use throwing
The mode of ticket determines class belonging to x.But the shortcomings that the method, is, the classifier quantity needed is big, and is being divided
When class is predicted, it is possible to can have the identical situation of several class polls, such data x may belong to multiple classes, thus seriously
Influence the precision of classification.
In addition, the analysis for classifier performance, the present invention uses the mode of ten folding cross validations to assess classifier
Performance.In this process, entire data set is divided into ten parts, so that every portion includes that initial data concentrates sample size very
One of.It is almost equal in every portion from the quantity of the class sample of two different classes of middle selections.Nine trained numbers are used in turn
It is developed according to collection for classifier, and uses the classifier of remaining a test data assessment building.Assessment is tested all every time
It can obtain corresponding accuracy, the accuracy obtained to ten experimental results does average value, obtains the fine estimation of algorithm, leads to
Multiple ten foldings cross validation is also carried out in normal situation, is then averaging again, the estimation to algorithm accuracy can be obtained.
Application effect of the invention is described in detail below with reference to experiment.
The laboratory of video camera source tracing method proposed by the present invention based on PRNU is by using from 11 different camera shootings
The video that machine is captured is performed.Table 1 summarizes video camera and its attribute used in experiment, including 6 differences
The video camera of brand, such as HKVISION, DAHUA, Tiandy, uniview, HANBANG and JOVISION.Include in same brand
The video camera of different model.Such as the model of three HKVISION, DAHUA and each two models of Tiandy.Select same equipment
Two size sensors are identical and the identical equipment of resolution ratio is the performance for parser in the worst cases.
Equipment and attribute used in the experiment of table 1
As shown in figure 8, the sample image of the data set, from two different video cameras, two video cameras are shot respectively
Different scenes.In addition, experiment is also acquired from two different cameras (i.e. HKVISION DS-2CD6233F and Tiandy
TC-NC9201S3E-2MP-E-I3S) image of the similar scene shot is as another data set.The sample of the data set
Video frame is as shown in Figure 9.
Firstly, the performance of the algorithm proposed using the video evaluations for imaging unit from two, images unit such as 2 institute of table
Show.
The grouping of 2 experimental facilities of table
Classification results are obtained by experiment, as shown in table 3.1st group, the 2nd group, the 3rd group, the 4th group obtains 100% respectively,
99.25%, 95.25%, 96.50% Average Accuracy.These results indicate that the algorithm proposed can identify that source images
Machine, and do not influenced by video camera brand, model and photographed scene content.When source, video camera belongs to different manufacturers and model
When, the performance of classification is more ideal.And when being tested with brand with the video camera of model using two, classifying quality is slightly worse
It is some.
The classification accuracy of 3 two video cameras of table
Analyze only that there are two the sorting algorithm performances under the conditions of video camera above.Below by the quantity for increasing video camera
Carry out the performance of assessment algorithm, and is classified using more classification SVM.Table 4 shows the camera shooting for using different number in an experiment
The performance of algorithm when machine.From the results, it was seen that the classification performance of algorithm is declined slightly when the quantity of source video camera increases.
This is because caused by the overlapping for using the video features from multiple source video cameras.Therefore, even if the results show
With greater number of video camera classification, it is still separable for belonging to various video camera class another characteristics.
Table 4 carries out performance evaluation using the video camera of different number
Algorithm performance for further evaluation, experiment choose the video from 8 different camerals to form a large amount of camera
Group, the SFFS algorithm picks 56 differentiation features for classification.Table 5 shows the classification results of 8 camera shooting units.
The algorithm performance assessment of 58, table camera shooting units
The average recognition accuracy tested above is 96.375%.As can be seen from the results, common mistake, which is classified, is
Between identical video camera model or between the different model of same brand.From the video camera of same model or brand
The video of capture has some common traits, these features lead to a small number of mistake classification.
It is proposed by the present invention based on the denoising method of small echo compared with traditional denoising method based on Gauss, for 8
The identifing source problem of a camera shooting unit, the method based on Wavelet Denoising Method obtain 96.375% average recognition accuracy, compare
There is promotion by a relatively large margin in the average recognition accuracy of the denoising method obtained 85% based on Gauss.Therefore, using base
It is better than traditional Gauss denoising method for video performance of tracing to the source in the method for Wavelet Denoising Method estimation PRNU.
In a practical situation, most of videos may be by some basic video processing, such as are sized, and rotate,
H.264 compressed encoding etc..Therefore, when using processed video in an experiment, it is necessary to the robustness of parser.To view
Frequency data set carries out various processing using different parameter, and be allowed to these processed videos together with original video with
Machine is blended in training dataset and test data is concentrated.For each video camera, 700 key frames are extracted, wherein 100 are former
Beginning image, 300 images are adjusted size operation, and in addition 300 images carry out rotation process.As shown in table 6.
6 key frame of table handles type and its value
It is tested, is analyzed respectively in 2,3,4,5 number of cameras using video after treatment
Recognition accuracy.In the case where number of cameras is 2, the recognition accuracy of inventive algorithm is 99.75%, and traditional algorithm is
93.5%；In the case that number of cameras is 3, the recognition accuracy of inventive algorithm is 98.47%, and traditional algorithm is
91.32%；In the case that number of cameras is 4, the recognition accuracy of inventive algorithm is 95.28%, and traditional algorithm is
88.3%；In the case that number of cameras is 5, the recognition accuracy of inventive algorithm is 92.7%, traditional algorithm 85.6%.
Experimental result is as shown in Figure 10.
It can be seen that this hair when testing the video data concentration video that includes that treated from the above experimental result
The algorithm of bright proposition is higher compared to the discrimination of traditional algorithm.This shows that proposed video traces to the source algorithm for primary image
Processing and geometric transformation are robusts.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (6)
1. one kind is towards the real-time source tracing method of monitoring system digital video, which is characterized in that described to be regarded towards monitoring system number
Source tracing method includes: frequency in real time
H.264, the first step analyzes the common coding mode of monitor video, obtains the key frame of monitor video；
Second step carries out the extraction of PRNU noise to the key frame of monitor video；
Third step carries out real-time grading using statistical nature of the real-time classification method to PRNU noise.
2. as described in claim 1 towards the real-time source tracing method of monitoring system digital video, which is characterized in that the first step
H.264, the common coding mode of monitor video is analyzed, the key frame selection I frame for obtaining monitor video carries out PRNU noise
Extraction.
3. as described in claim 1 towards the real-time source tracing method of monitoring system digital video, which is characterized in that the second step
The extraction for carrying out PRNU noise to the key frame of monitor video specifically includes:
(1) PRNU noise extracts, and the mathematical model of Image Acquisition indicates are as follows:
yij=fij(xij+ηij)+cij+εij；
Wherein, yijIt is sensor output, xijThe quantity of the photon of sensor, η are hit in expressionijIt is shot noise, cijIt is dark current
Noise, εijIt is Complex-valued additive random noise, fijIndicate PRNU noise contribution；
Image is removed dryness using the filter that removes dryness based on small echo；From being originally inputted for dominant noise component with PRNU
The noise residual error that the image after removing dryness obtains given image is subtracted in image；With IkIndicate original input picture, size be M ×
N,It is to IkIt is being obtained after being removed dryness as a result,It is from original image IkIn the noise residual error extracted；
(2) feature selecting and classification shown with two group profiles, first group include estimation PRNU noise high-order small echo statistics,
Second group include original video frame statistical nature；Feature is combined, can be used for the identification of video source camera；Respectively to view
Each Color Channel in frequency frame carries out three-level wavelet decomposition using Haar small echo, obtains 9 subbands；Distinguish for all subbands
The statistical nature for calculating single order and higher order, finally obtains 108 features；Feature set is using special to feature selecting algorithm before sequence
Sign selection.
4. as claimed in claim 3 towards the real-time source tracing method of monitoring system digital video, which is characterized in that described based on small
Specific step is as follows for the process that removes dryness of wave:
(1) multilevel wavelet is executed to the key frame of video extracted using 8-tap Daubechies QMF to decompose, which is drawn
Be divided into level Four subband, wherein vertically, horizontal, diagonal subband respectively indicates are as follows: h (i, j), v (i, j), d (i, j)；
(2) original nothing is calculated with the MAP of the every level-one sub-band coefficients obtained in square W × W neighborhood estimation to make an uproar the office of video frame
Portion's variance, wherein (3,5,7,9) W ∈；
(3) take the minimum value in four variances being calculated as the final estimated value of local variance；
σ2(i, j)=min (σ3 2(i,j),σ5 2(i,j),σ7 2(i,j),σ9 2(i,j)),(i,j)∈J；
(4) the vertical component h (i, j) of wavelet coefficient is removed dryness with Weiner filter, formula are as follows:
(5) above procedure is repeated to each Color Channels of all subbands of each decomposition rank and video frame, then to carry out you small
Denoising image is obtained after wave conversion.
5. as described in claim 1 towards the real-time source tracing method of monitoring system digital video, which is characterized in that the third step
Real-time grading carried out to the statistical nature of PRNU noise using real-time classification method specifically include:
(1) the more mode classifications of SVM have:
1) a pair of all；
M class is given, needs to train m two classification device；Classifier i therein is class that the setting of i class data is positive, in addition
M-1 other classes other than i class are all set to negative class；One two classifier is trained to each class, it is final to have altogether
Available m classifier；There is the data x of classification demand for one, classification belonging to x is determined by the way of ballot；
2) all to all；
M class is given, a classifier, a total of m of number of such two classifier are trained two-by-two to class all in m class
(m-1)/2；There is the data x of classification demand for one, the prediction of each classifier will be passed through, and uses the side of ballot
Formula determines class belonging to x；
(2) use the mode of ten folding cross validations to assess the performance of classifier, data set is divided into ten parts, so that every a packet
Include initial data concentrates sample size 1/10th；From the quantity of the class samples of two different classes of middle selections in every portion
It is almost equal；It is developed using nine training datasets for classifier, and is assessed using remaining a test data in turn
The classifier of building.
6. a kind of city using towards the real-time source tracing method of monitoring system digital video described in Claims 1 to 5 any one
Video monitoring system.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910224267.1A CN110121109A (en) | 2019-03-22 | 2019-03-22 | Towards the real-time source tracing method of monitoring system digital video, city video monitoring system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910224267.1A CN110121109A (en) | 2019-03-22 | 2019-03-22 | Towards the real-time source tracing method of monitoring system digital video, city video monitoring system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110121109A true CN110121109A (en) | 2019-08-13 |
Family
ID=67520656
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910224267.1A Pending CN110121109A (en) | 2019-03-22 | 2019-03-22 | Towards the real-time source tracing method of monitoring system digital video, city video monitoring system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110121109A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160122A (en) * | 2019-12-11 | 2020-05-15 | 桂林长海发展有限责任公司 | Airplane type identification method and device and storage medium |
| CN111178204A (en) * | 2019-12-20 | 2020-05-19 | 深圳大学 | Video data editing and identifying method and device, intelligent terminal and storage medium |
| CN111553848A (en) * | 2020-03-20 | 2020-08-18 | 西安电子科技大学 | Monitoring video tracing processing method, system, storage medium and video monitoring terminal |
| CN112991345A (en) * | 2021-05-11 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Image authenticity detection method and device, computer equipment and storage medium |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104408728A (en) * | 2014-12-03 | 2015-03-11 | 天津工业大学 | Method for detecting forged images based on noise estimation |
| CN108154080A (en) * | 2017-11-27 | 2018-06-12 | 北京交通大学 | A kind of method that video equipment is quickly traced to the source |
| CN108259791A (en) * | 2018-01-05 | 2018-07-06 | 南京信息工程大学 | A kind of method that PRNU noises are extracted from video file |
| CN108257123A (en) * | 2018-01-11 | 2018-07-06 | 西京学院 | Multiband radar image change detection method based on higher order statistical theory |
-
2019
- 2019-03-22 CN CN201910224267.1A patent/CN110121109A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104408728A (en) * | 2014-12-03 | 2015-03-11 | 天津工业大学 | Method for detecting forged images based on noise estimation |
| CN108154080A (en) * | 2017-11-27 | 2018-06-12 | 北京交通大学 | A kind of method that video equipment is quickly traced to the source |
| CN108259791A (en) * | 2018-01-05 | 2018-07-06 | 南京信息工程大学 | A kind of method that PRNU noises are extracted from video file |
| CN108257123A (en) * | 2018-01-11 | 2018-07-06 | 西京学院 | Multiband radar image change detection method based on higher order statistical theory |
Non-Patent Citations (5)
| Title |
|---|
| J.LUKAS等: "Digital camera identification from sensor pattern noise", 《IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY》 * |
| 周牒岚: "基于PRNU与多重分形谱特征的数字图像来源取证", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
| 邓磊，关佶红: "《基于机器学习的蛋白质相互作用与功能预测》", 31 May 2018 * |
| 陈竟博: "基于颜色特征的计算机辅助舌诊分类模型", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
| 雎悦: "面向监控系统的视频源识别及伪造检测技术研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111160122A (en) * | 2019-12-11 | 2020-05-15 | 桂林长海发展有限责任公司 | Airplane type identification method and device and storage medium |
| CN111178204A (en) * | 2019-12-20 | 2020-05-19 | 深圳大学 | Video data editing and identifying method and device, intelligent terminal and storage medium |
| CN111178204B (en) * | 2019-12-20 | 2023-05-09 | 深圳大学 | Video data editing and identifying method and device, intelligent terminal and storage medium |
| CN111553848A (en) * | 2020-03-20 | 2020-08-18 | 西安电子科技大学 | Monitoring video tracing processing method, system, storage medium and video monitoring terminal |
| CN111553848B (en) * | 2020-03-20 | 2023-04-07 | 西安电子科技大学 | Monitoring video tracing processing method, system, storage medium and video monitoring terminal |
| CN112991345A (en) * | 2021-05-11 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Image authenticity detection method and device, computer equipment and storage medium |
| WO2022237397A1 (en) * | 2021-05-11 | 2022-11-17 | 腾讯科技（深圳）有限公司 | Method and apparatus for evaluating image authenticity, computer device, and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Park et al. | Double JPEG detection in mixed JPEG quality factors using deep convolutional neural network | |
| CN110121109A (en) | Towards the real-time source tracing method of monitoring system digital video, city video monitoring system | |
| Ferrara et al. | Image forgery localization via fine-grained analysis of CFA artifacts | |
| Pan et al. | Exposing image forgery with blind noise estimation | |
| Saha et al. | Utilizing image scales towards totally training free blind image quality assessment | |
| Muhammad et al. | Blind copy move image forgery detection using dyadic undecimated wavelet transform | |
| CN111445454B (en) | Image authenticity identification method and application thereof in license identification | |
| CN111553848B (en) | Monitoring video tracing processing method, system, storage medium and video monitoring terminal | |
| Gu et al. | Structural similarity weighting for image quality assessment | |
| Su et al. | A novel forgery detection algorithm for video foreground removal | |
| Singh et al. | Detection and localization of copy-paste forgeries in digital videos | |
| CN105120294B (en) | A kind of jpeg format image sources discrimination method | |
| CN111709930A (en) | Pattern noise based picture provenance and tampering identification method | |
| Luo | A training-based no-reference image quality assessment algorithm | |
| Zhu et al. | Towards automatic wild animal detection in low quality camera-trap images using two-channeled perceiving residual pyramid networks | |
| CN112801037A (en) | Face tampering detection method based on continuous inter-frame difference | |
| Babu et al. | An efficient image dahazing using Googlenet based convolution neural networks | |
| CN111027573B (en) | Image authenticity identification method based on blind evidence obtaining technology | |
| Cong et al. | Image quality assessment with gradient siamese network | |
| CN115546598A (en) | Depth forged image detection method and system based on frequency domain transformation | |
| CN106611417A (en) | A method and device for classifying visual elements as a foreground or a background | |
| Qiao et al. | Classifying between computer generated and natural images: An empirical study from RAW to JPEG format | |
| Mou et al. | Reduced reference image quality assessment via sub-image similarity based redundancy measurement | |
| Wan et al. | A video forensic technique for detecting frame integrity using human visual system-inspired measure | |
| CN109508701B (en) | Face recognition and tracking method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190813 |
|
| RJ01 | Rejection of invention patent application after publication |