CN114554227B - Compressed video source detection method based on multi-scale transform domain self-adaptive wiener filtering - Google Patents

Abstract

The invention discloses a compressed video source detection method based on multi-scale transform domain self-adaptive wiener filtering, which comprises the following steps: respectively intervening decoding the reference and the compressed video to be tested, taking out video frames before a loop filtering module, and obtaining macro block quantization parameter QP values of the corresponding frames to form a matrix; performing multi-scale dual-density dual-tree complex wavelet transformation on each frame, adopting local self-adaptive threshold window wiener filtering to process 32 high-frequency subbands of each layer, and obtaining noise residuals before and after noise reduction after inverse transformation; taking the QP matrix as the weight of the noise residual error, and obtaining the final sensor mode noise by adopting a maximum likelihood estimator; and correlating sensor mode noise between the reference and the compressed video to be tested by using the symbol peak correlation energy, and judging according to the threshold value. The method can extract more sufficient sensor mode noise from the compressed video, effectively improve the recognition effect of the compressed video, and has stronger robustness in the recognition of the compressed video source with shorter duration.

Description

Compressed video source detection method based on multi-scale transform domain adaptive Wiener filtering

Technical Field

The present invention relates to the technical field of video evidence collection, and in particular to a compressed video source detection method based on multi-scale transform domain adaptive Wiener filtering.

Background Art

With the advent of the 5G information age and the massive growth in the number of people communicating on social networks, more and more people are communicating through the Internet. As an important carrier of information exchange, pictures and videos are widely used by people on the Internet. At the same time, some videos that seriously disrupt public security, piracy, malicious tampering and other crimes inevitably appear on the Internet. Videos uploaded on the Internet will be transmitted multiple times, and they will inevitably experience different degrees of compression during the transmission process. Therefore, confirming the source of compressed videos has become a very important research topic in multimedia information security forensics in recent years.

When it is necessary to confirm the source of images and videos, the most direct method is to check the watermark information of the image or video. However, with the continuous development of watermarks in image and video editing software, it has become very simple to change watermarks. Therefore, this method is very unreliable for video source detection. Researchers have begun to focus on the intrinsic features in digital images and videos. For example: Swaminathan A, Wu M, Liu K J R. Nonintrusive component forensics of visual sensors using output images [J]. IEEE Transactions on Information Forensics and Security, 2007, 2 (1): 91-106. (Swaminathan A, Wu M, Liu KJ R. Nonintrusive component forensics of visual sensors using output images [J]. IEEE Transactions on Information Forensics and Security, 2007, 2 (1): 91-106), successively used the noise artifacts caused by the CFA interpolation algorithm to extract and authenticate the source of the digital image. For example: Choi K S, Lam E Y, Wong K Y. Source camera identification using footprints from lens aberration [C] // Proceedings of SPIE. 2006, 6069: 172-179. (Choi KS, Lam EY, Wong KK Y. Source camera identification using footprints from lens aberration [C] // SPIE Proceedings. 2006, 6069: 172-179), the source of the image is determined by taking advantage of the fact that the aberration noise generated by lens mirror distortion affects the statistical characteristics of the image. For example: Dirik AE, Sencar HT, Memon N. Digital single lens reflex camera identification from traces of sensor dust [J]. IEEE Transactions on Information Forensics and Security, 2008, 3 (3): 539-552. (DirikAE, Sencar HT, Memon N. Identification of digital single lens reflex camera from traces of sensor dust [J]. IEEE Transactions on Information Forensics and Security, 2008, 3 (3): 539-552). Digital single lens reflex cameras generate sensor dust due to the interchangeable lenses deployed. The dust particles deposited in front of the imaging sensor will form a persistent pattern in all captured images. This sensor dust pattern is used to extract and detect the source of the image.

A good digital image and video source detection algorithm must have strong robustness and high detection rate. Although the above methods can detect the source of images to a certain extent, they all have the problem of high computational complexity and poor recognition effect. For example, when encountering the same model of camera recognition, the same model of camera will produce the same CFA interpolation operation, and the CFA interpolation recognition will lose its effect. Another example is sensor dust recognition. New cameras have less sensor dust accumulation, so it is very difficult to use sensor dust recognition. With the in-depth understanding of the researchers, a method of multimedia source detection with significant effect is proposed - a digital image and video source detection method based on sensor pattern noise.

As one of the important components of camera imaging, the sensor will produce some unique noise artifacts during imaging due to defects in its manufacturing materials and production processes. Even cameras of the same model will produce different noise artifacts. This unique noise artifact is called sensor pattern noise by researchers. The main component of sensor pattern noise is photo response non-uniformity (PRNU) noise, so PRNU is also considered sensor pattern noise, which can be used as a fingerprint of the camera to detect the source of unknown images and videos.

As domestic and foreign researchers have conducted in-depth exploration of sensor pattern noise extraction, they have found that sensor pattern noise is a very weak signal relative to the original content of the image. Researchers have proposed using filtering algorithms to extract sensor pattern noise, such as: Lukas J, Fridrich J, Goljan M. Digital camera identification from sensor pattern noise [J]. IEEE Transactions on Information Forensics and Security, 2006, 1 (2): 205-214. (Lukas J, Fridrich J, Goljan M. Identifying digital cameras from sensor pattern noise [J]. IEEE Transactions on Information Forensics and Security, 2006, 1 (2): 205-214.), and proposed using wavelet transform combined with Wiener filtering to extract noise residuals from images. Conotter V, Boato G. Analysis of sensor fingerprint for source camera identification [J]. Electronics letters, 2011, 47 (25): 1366-1367. (Conotter V, Boato G. Analysis of sensor fingerprint for source camera identification [J]. Electronics letters, 2011, 47 (25): 1366-1367.), introduced a complex BM3D (block matching and 3D filtering) filter to extract noise residuals from the image. BM3D works by identifying similar blocks in the image and combining them together. This method has been shown to be effective in extracting noise from images. Kang X, Chen J, Lin K, et al. A context-adaptive SPN predictor for trustworthy source camera identification [J]. EURASIP Journal on Image and Video Processing, 2014, 2014 (1): 1-11. (Kang X, Chen J, Lin K, et al. A context-adaptive SPN predictor for trustworthy source camera identification [J]. EURASIP Journal on Image and Video Processing, 2014, 2014 (1): 1-11.) and Zeng H, Kang X. Fast source camera identification using content adaptive guided image filter [J]. Journal of forensic sciences, 2016, 61 (2): 520-526. (Zeng H, Kang X. Fast source camera identification based on content adaptive guided image filter [J]. Journal of Forensic Sciences, 2016, 61 (2): 520-526.) proposed context-adaptive interpolation algorithm and adaptive guided image filtering algorithm respectively, which can further extract noise residuals. Subsequently, Lawgaly A, Khelifi F. Sensor pattern noise estimation based on improved locally adaptive DCT filtering and weighted averaging for source camera identification and verification[J]. IEEE Transactions on Information Forensics and Security, 2016, 12(2): 392-404. (Lawgaly A, Khelifi F. Sensor pattern noise estimation based on improved locally adaptive DCT filtering and weighted averaging for source camera identification and verification[J]. IEEE Transactions on Information Forensics and Security, 2016, 12(2): 392-404.) and Zeng H, Wan Y, Deng K, et al. Source camera identification with Dual-Tree complex wavelet transform[J]. IEEE Access, 2020, 8: 18874-18883. (Zeng H, Wan Y, Deng K et al. Source camera identification based on dual-tree complex wavelet transform [J]. IEEE Access, 2020, 8: 18874-18883.) proposed improved local adaptive discrete cosine transform filtering and dual complex wavelet transform combined with local adaptive window Wiener filtering methods, which achieved better filtering effects than the former. After using the filtering method to obtain the noise residual, some researchers found that the noise residual not only contains sensor pattern noise, but also contains other non-unique components to be filtered out, such as CFA interpolation noise, speckle noise, image texture details, etc. These non-unique components will be more numerous and more complex in compressed video. In Chen M, Fridrich J, Goljan M, et al. Determining image origin and integrity using sensor noise[J]. IEEE Transactions on Information Forensics and Security, 2008, 3(1): 74-90. (Chen M, Fridrich J, Goljan M, et al. Determining image origin and integrity using sensor noise[J]. IEEE Transactions on Information Forensics and Security, 2008, 3(1): 74-90.), it is proposed to use the maximum likelihood estimation algorithm to estimate a sensor pattern noise from 50 noisy residual images, and use the zero mean combined with frequency domain Wiener filtering method to further filter out other noises from the estimated sensor pattern noise. Kang X, Li Y, Qu Z, et al. Enhancing source camera identification performance with a camera reference phase sensor pattern noise [J]. IEEE Transactions on Information Forensics and Security, 2011, 7 (2): 393-402. (Kang X, Li Y, Qu Z, et al. Using camera reference phase sensor pattern noise to improve source camera identification performance [J]. IEEE Information Forensics and Security Transactions, 2011, 7 (2): 393-402.) proposed to use only the phase information of the noise residual in the Fourier domain for identification. Finally, this purified sensor pattern noise is used for matching and identification. There are many different methods for extracting sensor pattern noise. Lin X, Li C T. Preprocessing reference sensor pattern noise via spectrum equalization[J]. IEEE Transactions on Information Forensics and Security, 2015, 11(1): 126-140. (Lin X, Li C T. Preprocessing reference sensor pattern noise via spectrum equalization[J]. IEEE Transactions on Information Forensics and Security, 2015, 11(1): 126-140.) A spectrum equalization method is proposed to suppress local peaks to achieve a smoother effect. Most of the above algorithms can improve the quality of the sensor pattern noise of the image, but because the video undergoes multiple compressions during Internet transmission, its noise is more complex and the sensor pattern noise is also suppressed a lot, making it difficult for traditional image extraction algorithms to extract sensor pattern noise from compressed videos. Therefore, the recognition effect of these algorithms applied to compressed videos is extremely limited.

Based on the above reasons, in order to extract more sensor pattern noise from compressed videos and improve the recognition effect of compressed videos, it is necessary to study a compressed video source detection algorithm for compressed videos.

Summary of the invention

The purpose of the present invention is to overcome the shortcomings of the prior art and provide a compressed video source detection method based on multi-scale transform domain adaptive Wiener filtering, which can extract more sufficient sensor pattern noise from the compressed video, effectively improve the recognition effect of the compressed video, and also has strong robustness in the recognition of compressed videos of shorter duration.

To achieve the above purpose, the technical solution provided by the present invention is:

The method for detecting the source of compressed video based on multi-scale transform domain adaptive Wiener filtering comprises the following steps:

S1, obtaining sensor pattern noises of multiple reference compressed videos, and storing the sensor pattern noises in a sensor pattern noise database;

S2, obtaining the sensor pattern noise of the test compressed video;

S3. Use the symbol peak correlation energy to calculate the correlation between the sensor pattern noise of the test compressed video and the sensor pattern noise of the reference compressed video in the sensor pattern noise database. If the symbol peak correlation energy is greater than or equal to the set threshold, it is determined that the test compressed video comes from the camera that recorded the reference video. Otherwise, the test compressed video is not from the camera that recorded the reference compressed video.

Furthermore, in the steps S1 and S2, the specific steps of obtaining the sensor pattern noise are as follows:

A1. Convert the video back into bitstream data, intervene in the decoding process, output all video frames before the loop filter of the codec, and obtain the macroblock quantization parameter QP value of the corresponding frame to form a QP matrix;

A2. Obtain G video frames, and number the video frames as I ₁ , I ₂ , ..., I _G ; perform multi-layer double-density double-tree complex wavelet decomposition transform on each video frame;

A3. Apply the local adaptive threshold window Wiener filtering method to filter and estimate all high-frequency subbands to obtain filtered wavelet subbands; perform inverse transform of double-density double-tree complex wavelet decomposition to obtain denoised video frames; subtract the input video frame from the denoised video frame to obtain the noise residual of each video frame;

A4. Use the maximum likelihood estimation algorithm weighted by the QP matrix to estimate the set of noise residuals of I ₁ , I ₂ , ..., I _G to obtain the multiplication factor K of the preliminary sensor pattern noise of the video;

A5. Use zero averaging to remove the CFA interpolation artifacts of the multiplication factor K to obtain the multiplication factor K without CFA interpolation artifacts. Then use the frequency domain Wiener filtering algorithm to filter the multiplication factor K without CFA interpolation artifacts to further remove other non-unique noise components. The multiplication factor is multiplied with each input video frame and then averaged to obtain the sensor pattern noise.

Furthermore, in step A2, the specific process of the dual-density dual-tree complex wavelet decomposition transformation is as follows:

First, each video frame is decomposed into two dual trees by dual tree decomposition. The one-dimensional data of each row of each tree is decomposed twice by one-dimensional wavelet decomposition to obtain high-frequency, sub-high-frequency and low-frequency parts.

Then, the one-dimensional data composed of each column of the high-frequency, sub-high-frequency and low-frequency information formed by the decomposition is decomposed twice by one-dimensional wavelet decomposition, and finally nine sub-band images are obtained, namely eight high-frequency sub-bands L ₀ H ₁ , L ₀ H ₂ , H ₁ L ₀ , H ₁ H ₁ , H ₁ H ₂ , H ₂ L ₀ , H ₂ H ₁ , H ₂ H ₂ and one low-frequency sub-band L ₀ L ₀ ; each high-frequency sub-band is divided into two parts, the real part and the imaginary part, and 32 high-frequency components are obtained in each layer.

Furthermore, in step A3, the local adaptive threshold window Wiener filtering method is as shown in formula (1):

和子带方差

由式子(2)和(3)得到：In formula (1), _Win represents the wavelet coefficient before filtering, _Wout represents the wavelet coefficient after filtering, and the noise variance estimation

and subband variance

From equations (2) and (3), we can get:

In formula (2), median() represents the median estimator, W _temp represents the first high-frequency subband of the first layer decomposition;

In formula (3), _Nh is a local window with a size of hxh and centered at (u, v); max() means taking the maximum value between 0 and the variance estimate, and min() function means taking the minimum value of all window estimation results.

Furthermore, in step A4, the weighted maximum likelihood estimation algorithm of the quantization parameter value is expressed as formula (4):

In formula (4), G is the number of video frames in a single video used to estimate the sensor pattern noise factor K, _Nz represents the noise residual of the z-th video frame, _Iz represents the z-th video frame of the video, _WQP represents the correlation calculated according to different quantization parameters QP, and the QP-SPCE curve is drawn. The weight matrix is formulated using the curve relationship for weighting; δ is a set value used to prevent the denominator from being 0.

Furthermore, in step A5,

The zero-meaning process is to subtract the column mean from each pixel in the column, and then subtract the row mean from each pixel in the row;

The frequency domain Wiener filtering process transforms the sensor pattern noise without CFA interpolation artifacts into the frequency domain and estimates it using the Wiener filtering operation.

Furthermore, in step S3, the symbol peak correlation energy is expressed as formula (5):

Where sign() is the sign function, _CRQ (a,b) is the two-dimensional cyclic cross-correlation between the sensor pattern noise R of the reference compressed video and the sensor pattern noise Q of the test compressed video, β is a small area around (0, 0), |β| is the dimensional product of the area, and MN is the dimensional product of the matching sensor pattern noise.

Compared with the existing technology, the principles and advantages of this solution are as follows:

1) The loop filter module of the codec will filter out some sensor pattern noise in the video frame. This solution changes the codec process and extracts the video frame before reaching the loop filter module, which can preserve more sensor pattern noise in the video frame.

2) Sensor pattern noise belongs to medium and high frequency noise, and the wavelet decomposition method can better separate the high and low frequency information of the image signal. The double-density dual-tree complex wavelet transform has the advantages of double-density wavelet transform and double-tree complex wavelet transform, and can provide signals in 16 main directions, and each main direction has two wavelets (real wavelet and imaginary wavelet) to represent. Compared with wavelet transform, dual-tree complex wavelet transform and double-density wavelet transform, double-density dual-tree complex wavelet transform can further improve the decomposition and reconstruction accuracy of video frames.

3) The noise variance of Wiener filtering is a fixed value, and it is difficult to accurately and effectively estimate video frames with different compression levels. The median estimator is used to estimate the noise variance of the wavelet subband, and the wavelet in the field of image denoising generally selects the highest frequency wavelet subband for noise variance estimation. Since this scheme is to extract medium and high frequency sensor pattern noise, selecting the medium and high frequency subband (LH ₁ ) for noise variance estimation will obtain better extraction effect.

4) Compared with the original maximum likelihood estimation, the maximum likelihood estimation based on QP value weighting can give different weights to each pixel value according to different compression levels, which can better suppress artifacts caused by more complex compression levels and effectively improve the quality of sensor pattern noise.

5) Compared with the existing sensor noise extraction algorithm, this scheme can retain more sensor pattern noise for the input video frame. In particular, more sensor pattern noise can be extracted in the filtering process, and other noise components contained in the sensor pattern noise can be effectively suppressed in the subsequent processing. Therefore, this scheme has higher recognition effect and robustness.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the services required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is a principle flow chart of a method for detecting a source of a compressed video based on multi-scale transform domain adaptive Wiener filtering according to the present invention (step S1 is omitted);

FIG. 2 is a flow chart showing the principle of obtaining sensor pattern noise in the compressed video source detection method based on multi-scale transform domain adaptive Wiener filtering according to the present invention.

DETAILED DESCRIPTION

The present invention will be further described below in conjunction with specific embodiments:

As shown in FIG1 , the compressed video source detection method based on multi-scale transform domain adaptive Wiener filtering described in this embodiment includes the following steps:

S1, obtaining sensor pattern noises of multiple reference compressed videos, and storing the sensor pattern noises in a sensor pattern noise database;

S2, obtaining the sensor pattern noise of the test compressed video;

As shown in FIG. 2 , in steps S1 and S2 , the specific steps for obtaining the sensor pattern noise are as follows:

A1. Convert the video back into bitstream data, intervene in the decoding process, and output all video frames before the loop filter of the codec;

A2. Obtain G video frames, and number the video frames as I ₁ , I ₂ , ..., I _G ; perform a 4-layer double-density double-tree complex wavelet decomposition transform on each video frame;

The specific process of double-density double-tree complex wavelet decomposition transform is:

First, each video frame is decomposed into two dual trees by dual tree decomposition. The one-dimensional data of each row of each tree is decomposed twice by one-dimensional wavelet decomposition to obtain high-frequency, sub-high-frequency and low-frequency parts.

Then, the one-dimensional data composed of each column of the high-frequency, sub-high-frequency and low-frequency information formed by the decomposition is decomposed twice by one-dimensional wavelet decomposition, and finally nine sub-band images are obtained, namely eight high-frequency sub-bands L ₀ H ₁ , L ₀ H ₂ , H ₁ L ₀ , H ₁ H ₁ , H ₁ H ₂ , H ₂ L ₀ , H ₂ H ₁ , H ₂ H ₂ and one low-frequency sub-band L ₀ L ₀ ; each high-frequency sub-band is divided into two parts, the real part and the imaginary part, and 32 high-frequency components are obtained in each layer.

A3. Apply the local adaptive threshold window Wiener filtering method to filter and estimate all high-frequency subbands to obtain filtered wavelet subbands; perform inverse transform of double-density double-tree complex wavelet decomposition to obtain denoised video frames; subtract the input video frame from the denoised video frame to obtain the noise residual of each video frame;

The local adaptive threshold window Wiener filtering method is shown in formula (1):

和子带方差

由式子(2)和(3)得到：In formula (1), _Win represents the wavelet coefficient before filtering, _Wout represents the wavelet coefficient after filtering, and the noise variance estimation

and subband variance

From equations (2) and (3), we can get:

In formula (2), median() represents the median estimator, W _temp represents the first high-frequency subband of the first layer decomposition;

In formula (3), _Nh is a local window with a size of hxh and centered at (u, v); max() means taking the maximum value between 0 and the variance estimate, and min() function means taking the minimum value of all window estimation results.

A4. Use the maximum likelihood estimation algorithm weighted by the QP matrix to estimate the set of noise residuals of I ₁ , I ₂ , ..., I _G to obtain the multiplication factor K of the preliminary sensor pattern noise of the video;

The maximum likelihood estimation algorithm for weighted quantization parameter values is expressed as formula (4):

In formula (4), G is the number of video frames in a single video used to estimate the sensor pattern noise factor K, _Nz represents the noise residual of the z-th video frame, _Iz represents the z-th video frame of the video, _WQP represents the correlation calculated according to different quantization parameters QP, and the QP-SPCE curve is drawn. The weight matrix is formulated using the curve relationship for weighting; δ is a set value used to prevent the denominator from being 0.

A5. Use zero averaging to remove CFA interpolation artifacts of the multiplication factor K to obtain a multiplication factor K without CFA interpolation artifacts. Then use the frequency domain Wiener filter algorithm to filter the multiplication factor K without CFA interpolation artifacts to further remove other non-unique noise components. The multiplication factor is multiplied with each input video frame and then averaged to obtain the sensor pattern noise.

The zero-meaning process is to subtract the column mean from each pixel in the column, and then subtract the row mean from each pixel in the row;

The frequency domain Wiener filtering process transforms the sensor pattern noise without CFA interpolation artifacts into the frequency domain and estimates it using the Wiener filtering operation.

S3. Use the symbol peak correlation energy to calculate the correlation between the sensor pattern noise of the test compressed video and the sensor pattern noise of the reference compressed video in the sensor pattern noise database. If the symbol peak correlation energy is greater than or equal to the set threshold, it is determined that the test compressed video comes from the camera that recorded the reference video. Otherwise, the test compressed video is not from the camera that recorded the reference compressed video.

In this step, the symbol peak correlation energy is expressed as formula (5):

In formula (5), sign() is the sign function, _CRQ (a,b) is the two-dimensional cyclic cross-correlation between the sensor pattern noise R of the reference compressed video and the sensor pattern noise Q of the test compressed video, β is a small area around (0, 0), |β| is the dimensional product of the area, and MN is the dimensional product of the matching sensor pattern noise.

The embodiments described above are only preferred embodiments of the present invention and are not intended to limit the scope of implementation of the present invention. Therefore, all changes made according to the shape and principle of the present invention should be included in the protection scope of the present invention.

Claims (6)

1. The compressed video source detection method based on the multi-scale transform domain self-adaptive wiener filtering is characterized by comprising the following steps:

s1, obtaining sensor mode noises of a plurality of reference compressed videos, and storing the sensor mode noises into a sensor mode noise database;

s2, solving sensor mode noise of the test compressed video;

s3, using the symbol peak correlation energy to calculate the correlation between the sensor pattern noise of the test compressed video and the sensor pattern noise of the reference compressed video in the sensor pattern noise database, if the symbol peak correlation energy is greater than or equal to a set threshold, judging that the test compressed video is from a camera recording the reference compressed video, otherwise, the test compressed video is not from the camera recording the reference compressed video;

in the steps S1 and S2, the specific steps for obtaining the sensor pattern noise are as follows:

a1, changing the video into bit stream data again, intervening in the decoding process, outputting all video frames before a loop filter of a coder-decoder, and simultaneously obtaining macro block quantization parameter QP values of corresponding frames to form a QP matrix;

a2, obtaining G Zhang Shipin frames, and enabling the video frames to be numbered as I ₁ ，I ₂ ，...，I _G The method comprises the steps of carrying out a first treatment on the surface of the Respectively carrying out multi-layer double-density double-tree complex wavelet decomposition transformation on each video frame;

a3, performing filtering estimation on all high-frequency sub-bands by using a local self-adaptive threshold window wiener filtering method to obtain filtered wavelet sub-bands; performing double-density double-tree complex wavelet decomposition inverse transformation to obtain a denoised video frame; subtracting the input video frame from the denoised video frame to obtain a noise residual error of each video frame;

a4, weighting I by using QP matrix maximum likelihood estimation algorithm ₁ ,I ₂ ,...,I _G Estimating the set of noise residuals to obtain a multiplication factor K of the preliminary sensor mode noise of the video;

and A5, removing the CFA interpolation artifact of the multiplication factor K by using zero-averaging operation to obtain the multiplication factor K without the CFA interpolation artifact, filtering the multiplication factor K without the CFA interpolation artifact by using a frequency domain wiener filtering algorithm to further remove other non-unique noise components, multiplying the filtered multiplication factor K without the CFA interpolation artifact with each input video frame, and then averaging to obtain the sensor mode noise.

2. The compressed video source detection method based on multi-scale transform domain adaptive wiener filtering according to claim 1, wherein in the step A2, the specific process of the dual-density dual-tree complex wavelet decomposition transform is as follows:

performing dual tree decomposition on each video frame to obtain two dual trees, and performing two-dimensional wavelet decomposition on one-dimensional data formed by each row of each tree to obtain high-frequency, sub-high-frequency and low-frequency parts;

then carrying out two-time one-dimensional wavelet decomposition on the one-dimensional data formed by each row of the high-frequency information, the sub-high-frequency information and the low-frequency information formed by decomposition to finally obtain nine sub-band images, namely eight high-frequency sub-bands L ₀ H ₁ 、L ₀ H ₂ 、H ₁ L ₀ 、H ₁ H ₁ 、H ₁ H ₂ 、H ₂ L ₀ 、H ₂ H ₁ 、H ₂ H ₂ And a low frequency subband L ₀ L ₀ The method comprises the steps of carrying out a first treatment on the surface of the Each high frequency subband is in turn divided into two parts, real and imaginary, each layer yielding 32 high frequency components.

3. The compressed video source detection method based on multi-scale transform domain adaptive wiener filtering according to claim 2, wherein in the step A3, the local adaptive threshold window wiener filtering method is as shown in formula (1):

in the formula (1), W _in Expressed as wavelet coefficients before filtering, W _out Expressed as a filtered wavelet coefficient, noise squareDifference estimation

And subband variance->

Obtained from the formulas (2) and (3): />

In formula (2), mean () represents the median estimator, W _temp A first high frequency subband representing a first layer decomposition;

in the formula (3), N _h A local window of size hxh with (u, v) as the center point; max () represents the maximum value of 0 and variance estimation, and min () represents the minimum value of all window estimation results.

4. The method for detecting a compressed video source based on multi-scale transform domain adaptive wiener filtering according to claim 1, wherein in the step A4, the maximum likelihood estimation algorithm weighted by quantization parameter values is expressed as equation (4):

in the formula (4), G is the number of video frames in a single video for estimating the sensor mode noise factor K, nz represents the noise residual of the z-th video frame, I _z Representing a z-th video frame of the video, wherein WQP represents calculating correlation according to different quantization parameters QP, drawing to obtain QP-SPCE curves, and weighting by using a weighting matrix formulated by the curve relation; delta is a set value for preventing denominator from being 0.

5. The method for compressed video source detection based on multi-scale transform domain adaptive wiener filtering of claim 1, wherein in step A5,

the zero-averaging process is to subtract a column average value from each pixel in a column and then a row average value from each pixel in a row;

the frequency domain wiener filtering process is to transform the sensor mode noise without CFA interpolation artifact into the frequency domain and then estimate the sensor mode noise by using wiener filtering operation.

6. The method for detecting a compressed video source based on multi-scale transform domain adaptive wiener filtering according to claim 1, wherein in the step S3, the symbol peak correlation energy is expressed as equation (5):

wherein sign () is a sign function, C _RQ (a, b) is a two-dimensional cyclic cross-correlation between sensor pattern noise R of the reference compressed video and sensor pattern noise Q of the test compressed video, β is a small region around (0, 0), β is the dimensional product of the region, and MN is the dimensional product of the matching sensor pattern noise.

Images