CN102693522A - Method for detecting region duplication and forgery of color image - Google Patents

Abstract

The present invention proposes a color image region copy tampering detection method, first removes noise from the detected color image; divides the noise-processed image into blocks, and only one row or one column of pixels is different between adjacent image blocks; Quantize the pixels of the red, green and blue channels of each image block, divide each channel into M segments according to the gray level, count the number of pixels in each segment as the feature value of the image block, and divide the obtained three All the eigenvalues of each channel are combined as the eigenvector D of the image block; using the eigenvectors of each image block, one image block is similarly matched with the rest of the image blocks to obtain the correct matching pair of image blocks; Detect the binary image with the same image size and the gray level of each pixel is all zero, mark the position of the obtained correct matching image block into the binary image, and in the binary image generated by the obtained detection result Detects color image region copy tampering.

Description

A color image area copy tampering detection method

technical field

The invention belongs to the technical field of information security, and particularly relates to a method for detecting whether a color image area has been copied and tampered and locating the copied area and the tampered area. It can effectively detect the copying and tampering of regions under geometric transformations such as zooming and flipping, and can be used for authenticity identification and forensics of digital images in network environments.

Background technique

The rapid development of communication technology, computers and the Internet has enabled people to share resources and easily obtain images, audio and video on the Internet. However, due to the popularity and application of various image processing software, ordinary users can tamper with digital images at will without leaving traces of processing that can be seen by human eyes. This makes seeing not necessarily believing. If these false pictures are widely disseminated on the Internet or some forged images are used in judicial evidence collection, news reports, photography competitions and other occasions, it will lead to misjudgments, wrong reports, etc. Incalculable losses will be caused. Therefore, it is of great significance to identify the authenticity and integrity of digital images.

At present, digital image authentication methods are mainly divided into two categories: active authentication and passive authentication. Active authentication is divided into two categories: one is image authentication based on fragile digital watermarks. A fragile digital watermark is pre-embedded in the image. If the image is tampered, the digital watermark will be damaged and the damaged part will expose the tampering behavior. The second is image authentication based on digital signature, which uses image content to generate authentication code or digital signature. The method based on digital watermark needs to embed the watermark in the image, which will have a certain impact on the perceptual performance of the image, and the watermark is vulnerable to attack, and cannot protect the image without the embedded watermark; although the method based on digital signature will not The perceptual performance of the image is affected, but auxiliary information needs to be generated in advance and the signature is easily lost. Most images on the Internet do not embed digital watermarks or generate auxiliary information in advance. Due to the limitations of active authentication, the method of passive image authentication has attracted more and more attention. This method does not need to embed digital watermarks in the image in advance, nor does it need to generate auxiliary information in advance. It can be judged only based on the information of the image itself. Whether it has been falsified by tampering or synthesis. Passive authentication of digital images has become a cutting-edge research topic in the field of image processing, and it is of great significance for multimedia information security, screening of fake news, and judicial evidence collection.

Image area copying and tampering is to copy an area in an image and paste it to another area in the image to achieve the purpose of covering up a specific target or highlighting a certain scene. Since the copy and paste operations are performed in the same image, the tampered area has the same texture, color, noise and other characteristics as the entire image, so it is difficult for the naked eye to distinguish whether the image has been tampered with. Literature: Fridrich J, Soukal D, Lukas J. Detection of Copy-Move Forgery in Digital Images. Proceedings of Digital Forensic Research Workshop, Cleveland, 2003. A fuzzy matching method for discrete cosine transform (DCT) quantization coefficients is proposed, and the image is divided into blocks Finally, the DCT quantization coefficients of each image block are calculated, and the quantized DCT coefficients are lexicographically sorted to detect tampered areas of the image. Literature: Popescu A, Farid H. Exposing Digital Forgeries by Detecting Duplicated Image Regions. Dartmouth College, USA, TR2004-515, 2004. A detection method based on principal component analysis (PCA) was proposed, which uses PCA to analyze the The dimensionality reduction of the image feature vector can effectively improve the detection rate. In the invention patent: Luo Weiqi, Huang Jiwu. A Robust Detection Method for Copying and Tampering of Image Areas (Patent No.: ZL200610036600.9), Luo Weiqi et al. proposed a robust area copying and tampering detection method based on image block similarity comparison .

The above method can only detect the image that has not undergone geometric transformation in the tampered area. However, in order to cover up the traces of tampering, the tamperer may perform certain geometric deformations on the copied target, such as rotating, scaling, flipping and other transformations on the area. Make the above method impotent to such zone copy tampering.

Contents of the invention

In order to overcome the above-mentioned defects, the present invention provides an image tampering detection method that can detect the copied area after being processed by rotation, scaling, flipping, noise or blurring, etc., and can effectively detect images that have been tampered with by copying areas in the Internet environment authenticity. The method can detect the suspected tampered image and judge whether it has been tampered with, and if it has been tampered with, mark the tampered area and the duplicated area, so as to determine the authenticity of the image.

In order to achieve the above-mentioned purpose of the invention, a method for detecting color image region duplication and tampering is proposed, which is characterized in that it comprises the following steps:

Step 1): removing noise from the detected color image;

Step 2): The image processed in step 1) is divided into blocks, the image blocks overlap each other, and only one row or one column of pixels is different between adjacent image blocks;

Step 3): the red (R), green (G) and blue (B) three channels of the red (R), green (G) and blue (B) of each image block obtained in the step 2) are respectively divided into M segment, the pixel value of each segment is counted as the feature value of the image block, and all the feature values of the obtained three channels are combined as the feature vector D of the image block; where M is a positive integer, and each image block has a total of Obtain 3*M eigenvalues;

Step 4): using the eigenvectors of each image block obtained in the step 3), each image block is similarly matched with the remaining image blocks to obtain a correct pair of image blocks;

Step 5): Set up a binary image with the same size as the detected image and the gray value of each pixel is all zero, and mark the position of the correct matching image block obtained in step 4) into the binary image , in the obtained binary image generated from the detection results, the copy tampering of the color image area is marked.

More preferably, step 6) is also included; said step 6): the binary image generated from the detection result obtained in step 5) is filtered to remove false matches, and combined with morphological processing to obtain the final detection result.

More preferably, the step 1) adopts a Gaussian low-pass filtering method to filter the image to remove noise.

More preferably, in the step 2), the processed image is decomposed into circular image blocks with a radius b.

More preferably, the specific working steps of said step 3) are:

Step 31): Divide pixels between the red (R), green (G) and blue (B) channels of each image block with gray levels between [0T] into M segments, these M gray levels The intervals are $\left[\begin{array}{cc}0& \frac{1}{m}\×T-1\end{array}\right],$ $\left[\begin{array}{cc}\frac{1}{m}\×T& \frac{2}{m}\×T-1\end{array}\right],...\left[\begin{array}{cc}\frac{m-1}{m}\×T& T-1\end{array}\right],$ Among them, 180≤T≤255;

Step 32): Counting the pixel value of each channel grayscale interval in the step 31), the pixel value is the feature value of the corresponding channel of the image block;

Step 33): Use all the eigenvalues of the three channels obtained in the step 32) to combine as the eigenvector D of the corresponding image block, and form the eigenvectors obtained in all image blocks of the entire image into a vector group S.

More preferably, when marking the binary image in step 5), only the center position of the image block and the points around it are marked up, down, left, and right.

More preferably, in the step 6), a filtering method is used to remove the wrongly matched image blocks, and the filtering process includes:

First, generate a square window of h×h;

Then, the binary image generated by the detection result obtained in the step 5) is filtered, and the image is detected;

When it is detected that the number of white points in the window is greater than or equal to the set threshold, no processing is done; when the number of white points in the detected window is less than the threshold, the gray level of the pixels in the window is All are set to zero; among them, the step size of window movement is h.

The advantage of the present invention is that it can detect the suspected tampered image, and judge whether it has passed through the area copy tampering operation, and if it has been tampered with, locate the tampered position and the copied position, thereby judging the authenticity of the digital image . Compared with some existing methods, this method proposes a region copy tampering detection method based on the quantized histogram of the color image, which can not only detect the region tampering problem of direct translation, but also detect the linearity of the region after rotation, scaling, flipping, etc. The area copy tampering problem of geometric transformation can also deal with area copy tampering in the case of non-linear affine deformation such as oblique cutting, local distortion of the area, and projective transformation. The method is simple in calculation, and only needs to match the image blocks in the air domain, which improves the calculation efficiency.

Description of drawings

Fig. 1 is a flow chart of a method for detecting color image region duplication and tampering proposed by the present invention;

Figure 2 is the tampered image and its red, green, and blue histograms; among them, (a) is the tampered image, (b) is the histogram of the red component, (c) is the histogram of the green component, (d) is the blue component histogram;

Figure 3 is the original image, tampered image and detection result diagram; among them, (a) is the original image, (b) is the tampered image, (c) is the filter diagram, (d) is the initial detection result diagram, (e) is the filter After the result figure, (f) is the final result figure;

Fig. 4 is a tamper diagram and a detection result diagram of a tampered area rotated 20 degrees; wherein, (a) is a tampered diagram rotated 20 degrees of a tampered area, and (b) is a detection result diagram;

Fig. 5 is a tamper diagram and a detection result diagram of a tampered area rotated 90 degrees; wherein, (a) is a tampered diagram rotated 90 degrees of a tampered area, and (b) is a detection result diagram;

Fig. 6 is a tampering diagram and a detection result diagram of a horizontal flip of a tampered area; wherein, (a) is a tampered diagram of a horizontal flip of a tampered area, and (b) is a diagram of a detection result;

Fig. 7 is a tampering diagram and a detection result diagram of a vertical flip of a tampered region; wherein, (a) is a tampered diagram of a vertical flip of a tampered region, and (b) is a diagram of a detection result;

Fig. 8 is a tampering diagram and a detection result diagram in which the tampering area is reduced to 0.8 times the original size; wherein (a) is a tampering diagram in which the tampering area is reduced to 0.8 times the original size, and (b) is a detection result diagram;

Fig. 9 is a tampering diagram and a detection result diagram in which the tampering area is enlarged to an original size of 1.2; where (a) is a tampering diagram in which the tampering area is enlarged to 1.2 times the original size, and (b) is a detection result diagram;

Figure 10 is a tampered image with JPEG compression, Gaussian white noise and Gaussian blur added; where (a) is a tampered image, and pictures (b), (c), and (d) are the quality factors of JPEG compression respectively 70, 80, and 90 detection results. Figures (e) and (f) are the detection results of adding Gaussian white noise with SNR of 15 and 25 respectively. Figures (g) and (h) are the window size 3×3, standard Detection maps with Gaussian blur with variances of 3 and 5, respectively.

Detailed ways

The technical solution of the present invention will be described in further detail below in conjunction with the accompanying drawings.

The technical solution of the present invention comprises the following steps: (1) performing Gaussian low-pass filtering on the color image to be detected; (2) dividing the filtered image into circular image blocks of a fixed size, and there is only one line between adjacent image blocks Or a column is different; (3) Calculate the quantized histogram of the three components of the red, green and blue (RGB) components of the circular image block, and form a combined histogram, which is used as the feature of the image block; (4) the image block Match the features, find the most similar image blocks and mark them on the image; (5) remove the wrong match through filtering and morphological processing, the whole detection process is shown in flow chart 1.

(1) Filter the color image

Let f be the color image that is suspected to have been tampered with by region copying, and perform Gaussian low-pass filtering on the suspected tampered image f to extract the low-frequency part of the image. When filtering the image, use a window with a template of 3×3, and the standard deviation is taken as The value is 3, the tampered image is shown in Figure 3(b), and the filtered image is shown in Figure 3(c); the interference can be effectively reduced by filtering, and it is prepared for the subsequent operation of the image.

(2) Block the color image

Assuming that the size f of the suspected tampered color image is W×H, first decompose the image f into circular image blocks with a radius of b=8, and ensure that only one row or column of pixels between adjacent image blocks is different, thus It can be seen that the total number of image blocks available for image f is B=(W-2b+1)×(H-2b+1); the circular image has some advantages compared with image blocks of other shapes: when the tampered area occurs The effect on matching is relatively small when the rotation operation is performed.

(3) Extract the features of the image block

Firstly, the pixels whose gray levels of the red R, green G, and blue B channels of the obtained image block are between [0 and 199] are respectively quantized, so that each channel is divided into 5 gray levels. Figure 2 (b), (c), and (d) are the histograms of the red, green, and blue components of Figure 2 (a), respectively. The color range of the histogram color image of the image block is too large, and in the RGB color space , the region color histogram contains 256×256×256=2 ²⁴ colors. The main distribution interval of the gray levels of the pixels of the three colors of R, G, and B in the image block is [0 199], and the pixels with gray levels greater than 199 are relatively few. Considering the algorithm efficiency and statistical characteristics, quantize each color channel to 5 gray levels [0 39], [40 79], [80 119], [120 159], [159 199], and count each interval The pixel value of each interval is counted as the feature value of the image block, and all the feature values of the obtained three channels are combined as the feature vector of the image block. A total of 15 eigenvalues are obtained for each image block, and assuming that the entire eigenvector group is S, a total of (W-2b+1)×(H-2b+1)×15 eigenvalues can be obtained for the whole image.

(4) Feature matching and positioning of matching blocks

First, dimensionality reduction is performed on the feature vector group S. Since S is a three-dimensional vector, it is not easy to perform dictionary sorting. Adjust the dimension of the feature vector group to make it a two-dimensional vector and generate a number of rows (W-2b+1) ×(H-2b+1) array A with 15 columns. Then sort the (W-2b+1)×(H-2b+1) eigenvectors of array A lexicographically, so that the eigenvectors in the array are arranged from small to large in order from top to bottom and record the image corresponding to each row of eigenvectors The position of the block in the original image.

When performing similarity matching on the feature vectors of image blocks, three thresholds need to be preset: similarity threshold T _s , distance threshold T _d and area area threshold T _a .

Similarity threshold T _s : the similarity threshold is a parameter to measure the similarity of the feature vectors of two image blocks. In order to measure the similarity of feature vectors A _{i, s} , A _{j, s,} the present invention adopts Euclidean distance, which is defined as

$\mathrm{<span\; class="notranslate">\; TS</span>}<span\; class="notranslate">\; =</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; 15</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

Firstly, the feature comparison range is set. Since all feature vectors are sorted lexicographically, the position of feature vectors of similar image blocks in array A is relatively close, and the feature comparison range is set to 20 rows. From the first row of data, compare it with the 20 rows of feature vectors below and calculate its Euclidean distance; the more similar the image blocks, the smaller the calculated Euclidean distance. During the experiment, we randomly selected 100 images for area copy and tampering, the size of the tampered area was 80×80, and then performed different post-processing operations on the tampered images, including adding Gaussian white noise, Gaussian blur, rotation, and scaling and JPEG compression. For each tampered image, when the similarity threshold T _s is 5.5, the detection effect is better and the tampered area and the duplicated area can be detected effectively.

Distance threshold T _d : The image block of the tampered image has only one row or one column of pixels different from its adjacent image block. , even if it is not the correct matching block, it will get a high similarity, resulting in a false match. To this end, a distance threshold T _d is proposed. The radius of the matching block selected during the experiment is 8. In order to reduce the influence caused by adjacent matching blocks, when the distance between the center coordinates of the matching block pair is greater than or equal to the distance threshold, the comparison is made. In this method, we take The distance threshold T _d is 16.

Area area threshold T _a : In natural images, except for large flat areas (such as blue sky, grassland, road, white clouds, etc.), the possibility of similar large areas (including color, shape, texture, etc.) is not very high. If large similar regions are detected in a color image, it is likely that the image has been tampered with by region replication. In the literature: Luo Weiqi, Huang Jiwu, Qiu Guoping. Robust region copy image tampering detection technology. Journal of Computer Science, 2007, 30(11): 1998-2007. In the author pointed out that the large area area is not less than 0.85% of the original image size, In the detection process, the detected image block and the original tampered block often have certain differences in shape, so the area threshold T _a can be defined as:

T _a ≥ W×H×0.85%×η (3)

Among them, η is the damage coefficient.

When the detected matching block satisfies the set thresholds T _s and T _d , it is considered to be a pair of correctly matched image blocks. In the 20 pairs of matching blocks that are compared, there may be multiple matching blocks that meet the conditions. If there are multiple pairs of matching blocks that meet the conditions within the comparison range, only the image block with the smallest similarity threshold is selected as the correct matching block pair. And locate it, if there is no one that meets the conditions, no processing will be done.

(5) Filter and morphologically process the image

Create a binary image with the same size as the original image and the gray level of each pixel is all zero, mark the position of the correct matching pair of image blocks obtained in step (4) into the binary image, and only mark The center position of the image block and the pixels at five positions including top, bottom, left and right, the result of direct marking is shown in Figure 3(d). There are many isolated white pixels in the binary image generated by the detection results, these points are actually wrongly matched image blocks, so these points should be removed as much as possible. In the present invention, we establish an 8×8 filtering window and use this window to filter the entire image. First, start detection from the upper left corner of the image. When the number of white points detected in the window is greater than or equal to the set threshold of 20, no processing will be done. If it is less than the threshold of 20, the gray value of the pixels in the window will be all Zero. This method can not only effectively filter out those isolated wrong matching blocks, but also protect those correctly matched image blocks from interference, and improve the effect of image region copy tampering detection.

If the binary image 3(e) still has some larger mismatched regions after filtering, the mathematical morphology opening operation can be performed on the image to remove these matched regions; if there are no such mismatched regions, this step can be omitted. Because when the image is marked, only the center position of the image block and the pixels on the five positions in total, up, down, left, and right are marked, so that the detected tampered area and copied area are smaller than the original tampered area and copied area, so for Figure 3 ( e) Perform mathematical morphology closing operation processing.

To demonstrate the effectiveness of our invention, we present some tampered images and corresponding detection results.

As shown in Figure 3, (a) is the original image, (b) is the tampered image, (c) is the filtered image of the tampered image, (d) is the initial detection result, and figure (e) is the The result of filtering by the proposed filter, (f) is the final detection result obtained after closing the graph (e). As can be seen from Figure (f), the detection works well without any geometric transformation and attack. Figure 4, Figure 5, Figure 6, and Figure 7 are the tampered images and detection results after the tampered area has been rotated by 20 degrees, 90 degrees, horizontally flipped, and vertically flipped. It can be seen from the detection results that the detection effect of the tampered area is still very good even after these geometric transformations. Figure 8(a) and Figure 9(a) are images after the tampered area has been reduced to 0.8 times the original size and enlarged to 1.2 times the original size, respectively. This zoom processing detection effect is still very good. Figure 10 shows the corresponding detection results of tampered images after lossy JPEG compression, adding Gaussian white noise and Gaussian blur. Figures (b), (c), and (d) are the detection results after lossy JPEG compression with quality factors of 70, 80, and 90 respectively. Figures (e) and (f) are tampered images with SNRs of 15db and 25db The results after Gaussian white noise processing, Figures (g) and (h) are the detection results of tampered images after Gaussian blur processing with a template of 3×3 and standard deviations of 3 and 5 respectively. From the above test results, it can be seen that the present invention has good robustness against various conventional signal processing attacks and geometric deformation of the region.

Performance testing and experimental analysis

In order to verify the robustness of the present invention more effectively, we test the method of the present invention in two databases, where the correct matching rate F _r and the wrong matching rate F _w are defined, and the formula is as follows,

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&cap;</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&cap;</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&cup;</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; \&cup;</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; |</span>}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; C</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; m</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; |</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; f</span>}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

Among them, C ₁ and C ₂ are the size of the copied area of the tampered image and the size of the detected copy area, respectively, M ₁ and M ₂ are the size of the tampered area of the tampered image and the size of the detected tampered area, respectively.

In order to evaluate the detection effect, we tested 100 images in two databases: G.Schaefer and M.Stich, UCID-An Uncompressed Color Image Database, Technical Report, School of Computing and Mathematics , Nottingham Trent University, U.K., 2003. The database and a database we built ourselves, because the database UCID is relatively large, we only randomly selected 100 images. In the test experiment, we selected an image block with a size of 80×80 to paste from one area to another area of the same image, and performed Gaussian noise, lossy JPEG compression, and Gaussian blur attacks on the tampered image. From the experimental data obtained in Table 1 and Table 2, it can be seen that in the case of these attacks tested, the correct detection rate is above 0.9, and the error rate is relatively low.

Table 1 UCID database detection results

Blur=3 Blur=5 SNR＝15 SNR＝20 SNR＝30 SNR＝40 JPEG＝70 JPEG＝80 JPEG＝90 F _r 0.995 0.995 0.975 0.998 0.998 0.998 0.917 0.982 0.988 _f 0.039 0.041 0.093 0.104 0.105 0.095 0.140 0.083 0.051

Table 2 Test results of self-built database

Blur=3 Blur=5 SNR＝15 SNR＝20 SNR＝30 SNR＝40 JPEG＝70 JPEG＝80 JPEG＝90 F _r 0.995 0.995 0.9965 0.998 0.998 0.998 0.954 0.990 0.992 _f 0.039 0.039 0.081 0.096 0.096 0.095 0.138 0.084 0.072

Claims (7)

1. a color image region is duplicated altering detecting method, it is characterized in that, may further comprise the steps:

Step 1): coloured image to be detected is removed noise;

Step 2): the image after the said step 1) processing is carried out piecemeal, overlapped between the image block, and only have the pixel of delegation or row different between the adjacent image block;

The pixel that gray scale is between [0 T] in the redness (R) of each image block that step 3): respectively to said step 2) obtains, green (G) and blue (B) three passages is divided into the M section; The pixel number that counts every section combines all eigenwerts of three passages that obtain as the proper vector D of this image block as the eigenwert of this image block; Wherein, M is a positive integer, and each image block obtains 3*M eigenwert altogether;

Step 4): utilize the proper vector of each image block of said step 3) acquisition, each image block and remaining image piece are carried out the similarity coupling respectively, obtain correct match to image block;

Step 5): set up and each gray values of pixel points identical and be entirely zero binary map with image to be detected size; The correct match that obtains in the said step 4) to the position mark of image block in binary map, the mark color image region duplication of behaving excellently is distorted in the binary map that is generated by testing result that obtains.

2. according to claims 1 described color image region copy detection method, it is characterized in that, also comprise step 6); Said step 6): the binary map that is generated by testing result to said step 5) obtains is removed erroneous matching through filtering, and combining form processing obtains final detection result.

3. color image region copy detection method according to claim 1 and 2 is characterized in that, said step 1) adopts Gauss's low pass filtering method that image is carried out filtering, removes noise.

4. color image region copy detection method according to claim 1 and 2 is characterized in that, said step 2) in picture breakdown after will handling be that radius is the circular image piece of b.

5. color image region copy detection method according to claim 1 and 2 is characterized in that, the concrete job step of said step 3) is:

Step 31): the pixel that gray level in the redness (R) of each image block, green (G) and blue (B) three passages is between [0 T] is divided into the M section, is respectively between this M gray area $\left[\begin{array}{cc}0& \frac{1}{M}\&times;T-1\end{array}\right],$ $\left[\begin{array}{cc}\frac{1}{M}\&times;T& \frac{2}{M}\&times;T-1\end{array}\right],...\left[\begin{array}{cc}\frac{M-1}{M}\&times;T& T-1\end{array}\right],$ 180≤T≤255 wherein;

Step 32): add up said step 31 pixel number) between each passage gray area, this pixel number are the eigenwert of image block respective channel;

Step 33): utilize said step 32 all eigenwerts of three passages that) obtain combine as the proper vector D of correspondence image piece, the proper vector composition of vector group S that obtains in the entire image all images piece.

6. color image region copy detection method according to claim 1 and 2 is characterized in that, when in the said step 5) binary map being carried out mark, and a marking image piece center and up and down point on every side thereof.

7. color image region copy detection method according to claim 2 is characterized in that, adopts filtering method that the image block of erroneous matching is removed in the said step 6), and said filtering comprises:

At first, generate the square window of a h * h;

Then, the binary map that is generated by testing result that said step 5) obtains is carried out filtering, image is begun to detect;

When the number of white point is more than or equal to preset threshold in detecting window, then be left intact; When the number of white point is less than threshold value in detecting window, then the whole zero setting of the gray level of the pixel in the window; Wherein, the step-length that moves of window is h.

Images