WO2012145904A1

WO2012145904A1 - Method for tampered region localization of color images based on jpeg compression noise

Info

Publication number: WO2012145904A1
Application number: PCT/CN2011/073434
Authority: WO
Inventors: 谭铁牛; 董晶; 王伟
Original assignee: 中国科学院自动化研究所
Priority date: 2011-04-28
Filing date: 2011-04-28
Publication date: 2012-11-01

Abstract

A method for tampered region localization of color images based on JPEG compression noise includes: performing JPEG compression on the color image to be tested, then taking the difference between the newly obtained image and the original image, and getting the JPEG compression noise; extracting the high frequency quantization noise from the JPEG compression noise; performing normalization and morphology operation on the high frequency quantization noise, and getting the template for the tampered region localization at last. Since the present invention does not need to embed watermark in the image beforehand, just analyze the image itself, and does not need a lot of training library for machine learning, thus can be used effectively under the environment of content forensics, such as a large scale of data communication, multimedia transmission and so on.

Description

Technical field

The invention relates to the field of image passive blind forensics, in particular to a tamper region localization method for color digital images. Background technique

The 21st century is an era of digitalization. With the rapid development of computers and multimedia technologies, people are increasingly unable to rely on digital media. However, digital media does have a crucial problem to solve - seeing is true. Due to the emergence of a large number of digital image processing software (such as Photoshop) in recent years, digital image tampering and counterfeiting are easy, and digital media is no longer as highly credible as traditional media. Especially in news reports, photography competitions, and even legal evidence, tampering images are often seen. If the problem is not solved well, it will likely disrupt social order and undermine the credibility of the news media and the judiciary. As digital image processing software becomes increasingly powerful and easy to use, images that have been tampered with are increasingly difficult to detect by human eyes. Therefore, the demand for digital image content authentication technology is imminent.

At present, there are two main types of digital image content authentication technologies: active mode and passive mode. The active mode requires embedding watermark (or digital signature) information into the real image generation process (when photographed) or before publishing. When the image authenticity needs to be verified, only the watermark (digital signature) information is extracted from the image. If the watermark (or digital signature) information is destroyed, the image may be tampered with, and the watermark (or digital signature) is destroyed. It is the location where the tampering occurs. However, this method is inconvenient in practical applications, and it is impossible to make all digital imaging devices equipped with watermark (digital signature) embedded modules. Passively collecting data directly from the image itself does not require the hardware module configuration described above, which makes it a great concern. For example, tampering operations destroy the consistency of image imaging [AC Popescu and H. Farid, "Exposing digital forgeries in color filter array interpolated images," IEEE Transactions on Signal Processing, vol. 53, no. 10, pp. 3948- 3959, 2005.] , or to make the camera's inherent noise change or disappear in the image [J. Lukas, J. Fridrich, M. Goljan, "Detecting digital image forgeries using sensor pattern noise, " Society of Photo-Optical Instrumentation Engineers (SPIE ) Conference Series., vol. 6072, pp. 362-372, 2006], or to change the high-order statistics of images [YQ Shi, W. Chen, and C. Chen, "A natural image model approach to splicing detection , " Proceedings of the 9th workshop on Multimedia & security, pp.51-62, 2007] and more. Tampering regional positioning is more practical and more credible than just detecting tampering. It can locate the location where the tampering occurs, not just to determine if the image has tampering. Summary of the invention

It is an object of the present invention to provide a color image tampering region detection and positioning method based on JPEG compression noise, which can quickly and accurately detect and locate a color tamper image generated by a JPEG image operation and finally saved as a lossless compression format. Tampering area. In order to achieve the above object, a color image tampering region localization method based on JPEG compression noise includes:

The JPEG compression is performed on the color image to be measured, and the new image is obtained and the original image is taken to obtain JPEG compression noise;

The high-frequency quantization noise is extracted from the JPEG compression noise to distinguish the tamper region from the non-tamper region; the high-frequency quantization noise is normalized and morphologically operated, and finally the template for locating the tamper region is obtained. Since the present invention does not require embedding watermark information into an image in advance, it is only necessary to analyze the image itself and does not require a large training library for machine learning. Therefore, it can be effectively applied in a content forensics environment such as large-scale data communication and multimedia transmission. DRAWINGS

1 is a flow chart of a method for locating a color image tampering region based on JPEG compression noise; FIG. 2 is a diagram showing an example of a result of positioning an image to be detected and a tamper region thereof in the present invention, wherein (a) is an image to be tested, and the image is a tamper image. (b) JPEG compression noise when the image to be measured is Q=95, (c) high frequency quantization noise, (d) normalized high frequency quantization noise, (e) tampering area positioning template;

Figure 3 is a diagram showing an example of high frequency quantization noise extraction of the method of the present invention. detailed description

For any tampering image generated by a JPEG image operation and finally saved as a lossless compression format, the tampering region can be located by using the JPEG compression noise of the non-tamper region and the tamper region. The specific expression is: Image JPEG compression noise can be calculated by the difference between the image itself and its JPEG compressed version. In JPEG compression, DCT transform is its key operation, and JPEG compression noise mainly comes from self-DCT. The quantized loss of the coefficients after the transformation. For a color tampering image that has been tampered with by a JPEG image and saved as a lossless compression format, the non-tampering area, which must have experienced at least one JPEG compression. Another JPEG compression is required to calculate the JPEG compression noise in this region, and most of the high frequency coefficients in the DCT coefficients of this region have been quantized to zero in the previous compression. Therefore, its JPEG compression noise is mainly derived from the quantization loss of the mid-low frequency DCT coefficients. For its tampering region, if it comes from a lossless compressed image (which has not experienced the quantization loss in JPEG compression before), its JPEG compression noise comes from the quantization loss of all (high, medium and low frequency) DCT coefficients; if from JPEG compressed image, due to The probability that the tamper region and the non-tamper region are aligned according to the 8×8 DCT block is almost 0. When calculating the JPEG compression noise of the tamper region, the high frequency DCT coefficient is no longer zero. In this case, the JPEG compression noise of the tamper region is mainly derived from the quantization loss of the high, medium, and low frequency DCT coefficients. Therefore, the biggest difference between the tamper region and the non-tamper region is that the value of the high frequency quantization noise in the JPEG compression noise is different, the high frequency quantization noise of the non-tamper region is less than zero, and the high frequency quantization noise value of the tamper region is Larger, that is, the high-frequency quantization noise of the tampering region is stronger than the high-frequency quantization noise of the non-tamper region. The present invention utilizes this difference to achieve the purpose of locating the tampering area. In a given image, as long as a region whose value is greater than a certain threshold and relatively concentrated is found in its high-frequency quantization noise, it can be determined as a tamper region.

The present invention utilizes the difference in JPEG compression noise between the non-tamper region and the tamper region to locate the tamper region.

As shown in Figure 1, the following steps are included:

Calculate the JPEG compression noise step S 1 : Use the standard JPEG compression algorithm with quality factor Q (such as Q=95) to perform JPEG compression on the image to be measured, and obtain a new image and subtract it from the original image to obtain JPEG compression noise. The present invention can also calculate noise using non-standard JPEG compression, such as JPEG quantization in photoshop.

Extracting high-frequency quantization noise Step S2: JPEG compression noise of the image to be measured, using the principal component analysis (PCA) method to extract its high-frequency quantization noise. The present invention can also extract high frequency quantization noise using an Independent Component Analysis (ICA) method.

Post-processing step S3: normalization processing and morphological operation of the high-frequency quantization noise of the image to be measured, and finally obtaining a template for locating the tamper region (binarized image, white showing the location of the tampering region).

The process of extracting the high frequency quantization noise step S2 is as follows - step S21: treating each pixel point in the image to be tested as a sample point, and the color value (R, G, B) of the pixel point is regarded as a variable of the sample , construct the original dataset X. X is a matrix of N X 3, where N is the number of pixels in the image. Each row of X represents a pixel in the image, listed as the color value corresponding to the pixel. .

Step S22: Using the method of principal component analysis (PCA), the data set X is re-expressed to obtain a new matrix Y such that the column vectors in Y are orthogonal to each other. Step S23: Extracting the least important component in Y (the column with the smallest variance) as the high frequency quantization noise of the image to be tested. If the image to be tested is a tamper image, the high-frequency quantization noise of the non-tamper region is less than zero, and the high-frequency quantization noise of the tamper region is large, and there is a significant difference between the two.

The post-processing step S3 is as follows:

Step S31: The sigmoid function is used to normalize the high-frequency quantization noise of the image to be measured, and the normalized high-frequency quantization noise is obtained.

Step S32: Perform morphological operation on the normalized high-frequency quantization noise, open and close the image, and obtain a template for locating the tamper region.

The JPEG compression noise refers to subtracting a new image obtained by compressing the image to be tested with the image to be tested, as shown in the formula (1).

S = Jpeg(I,Q) -I ( 1 ) where S is the JPEG compression noise of the image I to be tested, Jpeg() is the JPEG compression operation, and Q is the quality factor used in the JPEG compression operation.

The use of the principal component analysis method to extract high-frequency quantization noise refers to finding a most linear transformation P for the original data set X (JPEG compression noise), so that X obtains a new data expression Y by the transformation. As shown in equation (2). Make the column vectors in Y orthogonal to each other.

Y = PX ( 2) Principal Component Analysis (PCA) is a classic dimensionality reduction method in pattern recognition that discriminates the least important components (variables) because they have little effect on the data distribution. However, the present invention uses the least important component obtained by PCA as the high frequency quantization noise to perform tamper region localization.

The normalization processing of the high-frequency quantization noise of the image to be tested is performed using the sigmoid function in the equation (3).

1

P ⁼ ] + _e -a(tb) ( 3 ) where a and b are constants and are artificially set according to the specific situation (eg a=3, b is the sum of the mean value of the high-frequency quantization noise and its 3 standard deviation). t is the high frequency quantization noise, and P(t) is the normalized high frequency quantization noise.

Referring again to FIG. 1, the image file to be processed is first determined to be a color image format and then enters the flow. Secondly, the image is JPEG-compressed, and its JPEG compression noise is calculated. Then, based on the present invention, the high-frequency quantization noise is extracted by using PCA, as shown in FIG. Finally, the high-frequency quantization noise is normalized, and the final tamper-area localization template is obtained through morphological operations.

In the embodiment, the image in FIG. 2 will be described as an example. First, JPEG compression (compression factor Q=95) is performed on the image to obtain a new image, and the original image is subtracted to obtain JPEG compression noise (Fig. 2b).

Then, the PCA is used to obtain high-frequency quantization noise (Fig. 2c). It can be seen from the figure that the high-frequency quantization noise in the non-tamper region is small, almost zero; and the high-frequency quantization noise in the tamper region is large. There is a clear difference between the two.

Secondly, the sigmoid function is used to normalize the high frequency quantization noise. Where a=3, b is the sum of the mean of the high-frequency quantization noise and its standard deviation of 3 times.

Finally, the obtained normalized high-frequency quantization noise is subjected to morphological operation, and the final tamper region positioning template is obtained after the image is opened and closed.

The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art who are familiar with the technology can make changes or substitutions within the technical scope of the present invention. The transformations and substitutions made are intended to be included within the scope of the invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

Rights request

A color image tampering region localization method based on JPEG compression noise, comprising:

The JPEG compression is performed on the color image to be measured, and a new image is obtained, and the difference between the original image and the original image is obtained to obtain JPEG compression noise;

The high-frequency quantization noise is extracted from the JPEG compression noise to distinguish the tamper region from the non-tamper region; the high-frequency quantization noise is normalized and morphologically operated, and finally the template for locating the tamper region is obtained.

2. The method according to claim 1, wherein the high frequency quantization noise of the non-tamper region is almost zero, and the high frequency quantization noise of the tamper region is large.

The method according to claim 1, characterized in that the image to be measured is subjected to JPEG compression using a standard JPEG compression algorithm with a quality factor of Q.

The method according to claim 1, characterized in that the image to be measured is JPEG-compressed using a JPEG compression algorithm in photoshop or other non-standard JPEG compression algorithm.

5. The method of claim 1, the extracting high frequency quantization noise comprising - constructing an original data set X;

Solve a new matrix Y such that Y is obtained by linear transformation of X, and the column vectors of Y are positive to each other.

, ■

father.

The column with the smallest variance in Y is extracted as the high-frequency quantization noise of the image to be tested.

6. Method according to claim 5, characterized in that the new matrix Y is solved by means of principal component analysis.

7. Method according to claim 5, characterized in that the new matrix 求解 is solved by means of independent component analysis.

8. The method of claim 1 wherein said normalizing said high frequency quantization noise comprises:

The sigmoid function is used to normalize the high-frequency quantization noise of the image to be measured.

9. The method according to claim 8, wherein the normalized high-frequency quantization noise is subjected to a morphological operation, and an image is opened and closed to obtain a final tamper region positioning template.