Copy image recognition method based on moire pattern analysis

Abstract

The invention provides a duplicated image recognition method based on moire analysis. The method comprises the following steps: S1, preprocessing an image; s2, converting the image from a spatial domain to a frequency domain, and performing frequency spectrum centralization processing on the image to obtain a frequency domain energy distribution condition corresponding to the image; s3, analyzing the frequency domain energy distribution condition to obtain energy corresponding to the moire features, and then obtaining an energy ratio corresponding to the moire features; s4, determining a threshold range of a normal image through threshold training, and judging whether the image belongs to a copied image or not according to preset judgment logic; according to the method, the duplication image is rapidly and accurately detected through the frequency domain analysis technology, the periodic mode of moire patterns is revealed through fast Fourier transform (FFT), a normal image and the duplication image are distinguished by setting a dynamic threshold value, the anti-noise capacity and the real-time performance of an algorithm can be effectively improved, and the false detection rate is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to a method for recognizing a re-photographed image based on moiré analysis. Background Technology

[0002] Reproduction image recognition is an important technology in the field of digital image forensics and copyright protection. In the prior art, the recognition of reproduction images is mainly achieved through the following methods: Method 1: Texture feature-based method: By analyzing the regularity of local texture of the image, the screen texture or repeated pattern that may exist in the re-shot image is detected. However, in this method, the texture feature is easily disturbed by the image content and the false detection rate is high.

[0003] Method 2: Color distortion-based method: By detecting color anomalies in the re-shot image caused by screen reflection or ambient light interference. However, the color distortion method is sensitive to changes in ambient light and lacks robustness.

[0004] Method 3: Deep learning-based method: Use convolutional neural network (CNN) to train the model and learn the features of the re-photographed images from a large amount of data. However, deep learning relies on a large amount of labeled data, has high computational complexity, and is difficult to apply in real time.

[0005] In summary, the existing methods for identifying re-photographed images still have the following defects and deficiencies: 1) Existing methods do not fully utilize moiré features: The moiré patterns generated by the shooting screen in the re-photographed image have unique periodic high-frequency components, but traditional methods do not effectively extract such frequency domain features.

[0006] 2) Insufficient noise resistance: Texture- or color-based methods are easily interfered by image content noise, resulting in missed detection or false detection.

[0007] 3) Poor real-time performance: Deep learning methods require a lot of computing resources and are difficult to execute quickly on low-power devices.

[0008] Therefore, it is urgent to provide a new solution to solve the defects and shortcomings of the above-mentioned prior art. SUMMARY OF THE INVENTION

[0009] In order to solve the defects and shortcomings in the prior art, the present invention provides a method for identifying re-shot images based on moiré analysis.

[0010] The specific solution provided by the present invention is: A method for identifying a reproduced image based on moiré analysis, characterized in that it comprises the following steps: S1: Preprocess the image; S2: Convert the image from the spatial domain to the frequency domain and perform spectrum centering on it to obtain the frequency domain energy distribution corresponding to the image; S3: Analyze the energy distribution in the frequency domain to obtain the energy corresponding to the moiré feature, and then obtain the energy proportion corresponding to the moiré feature; S4: Determine the threshold range of normal images through threshold training, and determine whether the image is a re-shot image based on the preset judgment logic.

[0011] As a further preferred embodiment of the present invention, the step S1 at least includes the following steps: S1.1: Convert color image to grayscale image; S1.2: Filter and reduce noise on the grayscale image; S1.3: Enhance the contrast of the image.

[0012] As a further preferred embodiment of the present invention, in step S1.1, the RGB color image is converted into a grayscale image using a weighted average method, and the conversion is performed according to the following formula:

[0013] Among them, I gray ( x , y ) indicates that the converted grayscale image is at coordinates ( x , y ) has a brightness value ranging from 0 to 255; R ( x , y )、 G ( x , y )、 B ( x , y ) respectively represent the original color image at coordinates ( x , y ) is the brightness value of the red, green, and blue channels, ranging from 0 to 255; As a further preferred embodiment of the present invention, the step S1.2 includes the following steps: S1.21: Use a two-dimensional Gaussian kernel function to convolve the grayscale image and define the convolution function as:

[0014] Among them, g represents Gaussian function; x and y ​Respectively represent the two-dimensional coordinates of each point in the kernel; represents the standard deviation of the Gaussian kernel function; S1.22: Verify the grayscale image after filtering and noise reduction by signal-to-noise ratio: Compare the signal-to-noise ratio of the grayscale image after this processing with the preset signal-to-noise ratio corresponding to the same standard deviation: When the current signal-to-noise ratio is not lower than the preset signal-to-noise ratio, the verification is deemed to have passed; otherwise, the verification is deemed to have failed.

[0015] As a further preferred embodiment of the present invention, in step S1.3, the contrast of the image is enhanced by the following steps: S1.31: Divide the image into multiple sub-regions; S1.32: Perform histogram equalization on each sub-region independently; S1.33: Eliminate block boundary effects through bilinear interpolation.

[0016] As a further preferred embodiment of the present invention, in step S2, when converting the image from the spatial domain to the frequency domain, a two-dimensional fast Fourier transform operation is performed on the preprocessed image according to the following formula to obtain a frequency domain complex matrix :

[0017] Among them, N and M Respectively the length and width of the image; and is the frequency domain coordinate; j is an imaginary unit, representing the imaginary part of a complex number, satisfying j 2 = −1.

[0018] As a further preferred embodiment of the present invention, in step S2, when performing spectrum centering processing, the zero frequency component in the obtained frequency domain complex matrix is ​​moved to the center of the spectrum according to the following formula:

[0019] Among them, represents the frequency domain complex matrix after centralization; As a further preferred embodiment of the present invention, in step S3, when obtaining the energy corresponding to the moiré feature, the center of the spectrum is taken as the origin, and the energy integral at different angles θ is calculated according to the following formula: ​​​

[0020] Among them, ; is the energy integral; is the angle of the integral direction, the range is [0°, 180°], and the step length is 1°; r Indicates the mirror distance from the center of the spectrum to the target point, unit: pixel; Then normalize the energy integral: If there is a significant periodic peak, it is determined to be a moiré feature; If there is no significant periodic peak, it is determined not to be a moiré feature.

[0021] As a further preferred embodiment of the present invention, in step S3, the energy proportion corresponding to the moiré feature is obtained according to the following steps: Define the high frequency area as the outer ring of the spectrum, and calculate its energy proportion according to the following formula:

[0022] Among them, is the energy proportion of high frequency area; represents the sum of the energy of all frequency domain points in the high frequency area; Represents the sum of the energy of all frequency domain points in the entire frequency domain.

[0023] As a further preferred embodiment of the present invention, in step S4, When determining the threshold range of normal images through threshold training, use a training set containing multiple normal images and multiple re-photographed images to count the proportion of high-frequency energy respectively and peak intensity E peak The distribution of , and determine the classification boundary of the threshold T by the following formula:

[0024] Among them, is the feature mean of a normal image; is the standard deviation of a normal image; When judging whether an image is a re-photographed image according to the preset judgment logic, the preset judgment logic satisfies: (1) If ≥T and there are at least two significant periodic peaks, then it is determined to be a re-shot image; ​(2) Otherwise, it is judged as a normal image.

[0025] Compared with the prior art, the technical effects that can be achieved by the present invention include: 1) The present invention provides a method for identifying reprinted images based on moiré analysis. The reprinted images are detected quickly and accurately through frequency domain analysis technology, the periodic pattern of moiré is revealed by fast Fourier transform (FFT), and the normal image is distinguished from the reprinted image by setting a dynamic threshold. This can effectively improve the algorithm's anti-noise ability and real-time performance, and reduce the false detection rate.

[0026] 2) The present invention provides a method for identifying re-photographed images based on moiré analysis. Through fast Fourier transform (FFT) and radial integration algorithm, the periodic characteristics of moiré are accurately quantified, and the high-frequency components of moiré are effectively separated, thus avoiding the limitations of traditional methods that rely on artificial feature design and complex texture modeling, and improving the calculation speed.

[0027] 3) The present invention provides a method for identifying re-photographed images based on moiré analysis, which adopts a dynamic threshold mechanism and combines statistical learning to adaptively adjust the judgment threshold. The judgment threshold can be automatically adjusted according to the application scenario, thereby improving the generalization ability and adapting to the image quality differences in different devices and environments.

[0028] 4) The present invention provides a method for re-photographed image recognition based on moiré analysis, which achieves lightweight and real-time performance. The whole process does not require complex model training, and the processing time of a single image is ≤50ms. It can run in real time in embedded devices (such as mobile phones and security cameras).

[0029] 5) The present invention provides a method for identifying re-photographed images based on moiré analysis, which pre-processes the image, converts the image into a grayscale image, and uses the synergistic effect of Gaussian filtering and histogram equalization to enhance feature robustness.

[0030] 6) The present invention provides a method for identifying re-photographed images based on moiré analysis, which has the advantages of high detection accuracy (tested on the public data set RAISE, the accuracy rate is 98.2%, and the false detection rate is less than 2%), strong real-time performance (single image processing time ≤50ms (CPU i5-8250U), 10 times faster than the deep learning model), excellent noise resistance (when adding Gaussian noise (SNR=20dB), the detection accuracy is still maintained above 95%) and a wide range of applications. It can be widely used in copyright protection, identity authentication, judicial evidence collection, field photography and clocking in and other fields. Brief Description of the Figures

[0031] Figure 1 Shown is a flowchart of the method steps provided by the present invention.

[0032] Figure 2The figure shows the structure of four adjacent sub-regions in the bilinear interpolation method in step S1.3. Specific implementation method

[0033] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] In the description of the present invention, it should be noted that the terms "upper", "lower", "inner", "outer", "front end", "rear end", "two ends", "one end", "the other end" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance.

[0035] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "provided with", "connected", etc. should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal connection of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0036] Since the periodic interference (moiré) generated by the screen when the copied image is shot will appear as abnormal high-frequency components or strip-shaped energy distribution in the frequency domain, the present invention captures these features through frequency domain analysis and combines dynamic thresholds to achieve efficient recognition.

[0037] [First embodiment] The first embodiment of the present invention provides a method for identifying a re-photographed image based on moiré analysis, such as Figure 1 As shown, the following steps are included: S1: Preprocess the image; the purpose of preprocessing the image is to eliminate environmental noise and enhance moiré features, providing high-quality input for subsequent frequency domain analysis.

[0038] In this embodiment, preprocessing the image includes at least the following steps: ​​​S1.1: Convert the color image to a grayscale image; in this step, the RGB color image is converted to a grayscale image using the weighted average method, and the conversion is performed according to the following formula:

[0039] Among them, I gray ( x , y ) indicates that the converted grayscale image is at coordinates ( x , y ) has a brightness value ranging from 0 to 255; R ( x , y )、 G ( x , y )、 B ( x , y ) respectively represent the original color image at coordinates ( x , y ) is the brightness value of the red, green, and blue channels, ranging from 0 to 255; The coefficients 0.299, 0.587 and 0.114 are weights designed based on the sensitivity of the human eye to different colors of light.

[0040] Because the cone cells in the human eye's retina are most sensitive to green light (wavelength about 550 nm), followed by red light (about 630 nm), and least sensitive to blue light (about 440 nm). This difference in perception leads to: Green (G) contributes the most: weight 0.587, accounting for about 60%; Red (R) is second: weight 0.299, accounting for about 30%; Blue (B) is the smallest: weight 0.114, accounting for about 10%.

[0041] The coefficient weights in this formula are more in line with the sensitivity of the human eye to different colors and are often used in actual image processing. It has been verified through psychophysical experiments (such as brightness matching experiments) that this weighting method can make grayscale images closer to the subjective perception of brightness by the human eye and can effectively avoid visual distortion caused by the direct averaging method (Y=R+G+B3).

[0042] The selection of the above weight coefficients is based on the biological characteristics of the human eye's sensitivity to different colors, giving priority to retaining the brightness information while reducing the complexity of the calculation.

[0043] S1.2: Filter and reduce noise on the grayscale image; specifically, the following steps are included: ​S1.21: Use a two-dimensional Gaussian kernel function to convolve the grayscale image and define the convolution function as:

[0044] Among them, g represents Gaussian function; x and y represent the two-dimensional coordinates of each point in the kernel; represents the standard deviation of the Gaussian kernel function; x in the formula 2 、y 2 They are the squares of the coordinate values, which are used to calculate the square of the Euclidean distance between the current point and the center point. This operation can convert the coordinates into non-negative values, providing a basis for the subsequent calculation of the exponential function.

[0045] S1.22: Verify the grayscale image after filtering and noise reduction by signal-to-noise ratio: Compare the signal-to-noise ratio of the grayscale image after this processing with the preset signal-to-noise ratio corresponding to the same standard deviation: When the current signal-to-noise ratio is not lower than the preset signal-to-noise ratio, the verification is deemed to have passed; otherwise, the verification is deemed to have failed.

[0046] In this embodiment, after experimental determination, the preset signal-to-noise ratio corresponding to each standard deviation satisfies: Standard Deviation = 0.5, the preset signal-to-noise ratio SNR (db) = 15.2; Standard Deviation = 1.0, the preset signal-to-noise ratio SNR (db) = 18.7; Standard Deviation = 1.5, the preset signal-to-noise ratio SNR (db) = 20.3; Standard Deviation = 2.0, the preset signal-to-noise ratio SNR (db) = 19.1; Standard Deviation = 2.5, the preset signal-to-noise ratio SNR (db) = 17.5; When the SNR improvement value is higher, the noise suppression effect is better. When σ<1, it indicates that the noise suppression is insufficient; and when σ>2, it indicates that the image details are blurred. Therefore, in this embodiment, the standard deviation of the Gaussian kernel function σ=1.5 is selected and experimental verification is carried out. Through the SNR test, when σ=1.5, the SNR can be effectively improved by about 22dB, so it is determined that the verification is passed.

[0047] S1.3: Enhance the contrast of the image. In this embodiment, the following steps are used to enhance the contrast of the image: S1.31: Divide the image into multiple sub-regions;​ S1.32: Perform histogram equalization on each sub-region independently; S1.33: Eliminate block boundary effects through bilinear interpolation.

[0048] In contrast-limited adaptive histogram equalization (CLAHE), the image is divided into multiple sub-regions (such as 8×8 blocks), and each sub-region is independently histogram equalized. However, this block processing may lead to sudden changes in brightness or contrast between adjacent blocks (i.e., "inter-block boundary effect"). Bilinear interpolation achieves a natural transition by smoothing the mapping results of adjacent blocks.

[0049] The following are the specific steps of step S1.3: 1) Divide the image into M*N sub-regions (e.g. 8*8); 2) Perform histogram equalization on each sub-region independently to generate the corresponding grayscale function T k ( x , y ) where k is the sub-area number; 3) For any point in the image ( x , y ), find the center points of the four adjacent sub-regions, such as Figure 2 Shown: Center of upper left sub-area: ( x TL , y TL ); Center coordinates: ( x 1 , y 1 ); Center of the upper right sub-area: ( x TR , y TR ); Center coordinates: ( x 2 , y 1 ); Center of lower left sub-area: ( x BL , y BL ); Center coordinates: ( x 1 , y 2 ); Center of lower right sub-area: ( x BR , y BR); Center coordinates: ( x 2 , y 2 ); If Figure 2 As shown, the target point ( x , y ) is located at the junction of four adjacent sub-areas; Calculate the horizontal weight and vertical weight of the bilinear difference respectively: Horizontal weight: the horizontal offset ratio of the target point relative to the center of the upper left sub-region w x Satisfy:

[0050] Vertical weight: the vertical offset ratio of the target point relative to the center of the upper left sub-region w y Satisfy:

[0051] The gray value of the target point can be obtained by weighting the mapping functions of the four adjacent sub-regions by distance, and the closer the sub-region is, the greater the contribution is. Therefore, the mapping functions of the four adjacent sub-regions can be interpolated according to the weights to obtain the final gray value of the target point I out ( x , y ), the calculation formula is as follows:

[0052] In the formula, Corresponding to the upper left sub-region, it means that when the target point is close to the upper left sub-region, the contribution of the upper left sub-region is the largest, and at this time both the horizontal and vertical weights are small; corresponds to the upper right sub-region, indicating that when the target point is close to the upper right sub-region, the contribution of the upper right sub-region is the largest, at this time the horizontal weight is large and the vertical weight is small; Corresponding to the lower left sub-region, it means that when the target point is close to the lower left sub-region, the contribution of the lower left sub-region is the largest, at this time the horizontal weight is small and the vertical weight is large; Corresponding to the lower right sub-region, it means that when the target point is close to the lower right sub-region, the contribution of the lower right sub-region is the largest, and at this time both the horizontal and vertical weights are large; For image boundary areas: If the target point is close to the edge of the image and lacks adjacent sub-regions (for example, there is only one sub-region in the upper left corner), the interpolation corresponds to: One-sided interpolation: that is, linear interpolation is performed using only the existing sub-regions; or Direct copy: Directly use the mapping result of the nearest sub-area.

[0053] For situations where the sub-regions are not evenly divided, for example, if the image size cannot be evenly divided, the interpolation weights need to be adjusted to match the actual coordinates.

[0054] For example: Assume the target point ( x , y ) is located at the junction of the four sub-areas and satisfies: Top left sub-area mapping value T TL = 100; Upper right sub-area mapping value T TR =120; Lower left sub-area mapping value T BL =110; Lower right sub-area mapping value T BR =130; Weight w x = 0.5, w y =0.5; The final grayscale value is: I out =0.25×100+0.25×120+0.25×110+0.25×130=115 Thus, a smoothly transitioned grayscale value is generated at the junction through interpolation, rather than directly using the result of a single sub-area (such as 100 or 130); Bilinear interpolation achieves a balance between smoothing effect and computational efficiency, which is suitable for the real-time processing requirements of CLAHE. Bilinear interpolation effectively eliminates the block boundary effect caused by CLAHE block processing by fusing the mapping results of adjacent sub-regions. Its core is to achieve a smooth transition of grayscale values ​​through distance weights, taking into account the needs of noise suppression and detail retention.

[0055] By using histogram equalization to enhance the contrast of the image, the local overexposure problem caused by global equalization can be avoided, ensuring that the periodic structure of the moiré pattern is clearly visible.

[0056] S2: Convert the image from the spatial domain to the frequency domain and perform spectrum centering processing on it to obtain the frequency domain energy distribution corresponding to the image; to reveal the periodic interference pattern of the moiré pattern.

[0057] ​​In this step, when converting the image from the spatial domain to the frequency domain, the preprocessed image is subjected to a two-dimensional fast Fourier transform (FFT) operation according to the following formula to obtain a frequency domain complex matrix :

[0058] Among them, N and M Respectively the length and width of the image; and is the frequency domain coordinate; j is an imaginary unit, representing the imaginary part of a complex number, satisfying j 2 = −1.

[0059] In this step, when performing spectrum centering processing, the zero-frequency component in the obtained frequency domain complex matrix is ​​moved to the center of the spectrum according to the following formula:

[0060] Among them, represents the frequency domain complex matrix after centralization; After the spectrum centering process, the low-frequency energy can be concentrated in the center of the spectrum, and the high-frequency energy can be distributed around the periphery, so as to facilitate the subsequent detection of strip-shaped structures.

[0061] S3: Analyze the energy distribution in the frequency domain to obtain the energy corresponding to the moiré feature, and then obtain the energy proportion corresponding to the moiré feature; In this step, when obtaining the energy corresponding to the moiré feature, the center of the spectrum is taken as the origin, and the energy integral at different angles θ is calculated according to the following formula:

[0062] Among them, ; is the energy integral; is the angle of the integral direction, the range is [0°, 180°], and the step length is 1°; r Indicates the mirror distance from the center of the spectrum to the target point, unit: pixel; Then normalize the energy integral: If there are significant periodic peaks (e.g. =45° or 135° direction, etc.), it is determined to be a moiré feature; ​​If there is no significant periodic peak, it is judged not to be a moiré feature; In this step, the energy proportion corresponding to the moiré feature is obtained by following the steps below: Define the high frequency area as the outer ring of the spectrum, and calculate its energy proportion according to the following formula:

[0063] Among them, is the energy proportion of high frequency area; represents the sum of the energy of all frequency domain points in the high frequency area; Represents the sum of the energy of all frequency domain points in the entire frequency domain.

[0064] The energy proportion of high-frequency area in the re-photographed image Usually higher than normal images due to: For normal images: Natural images usually have rich low-frequency information (such as smooth background) and a small amount of high-frequency information (such as detailed texture). Therefore, the proportion of high-frequency energy in normal images is relatively low (its average value is generally 0.12±0.04).

[0065] For the re-photographed images: when the screen is photographed, periodic interference (moiré) will be generated. This structure is manifested as strip-shaped energy concentration in the high-frequency area in the frequency domain. Therefore, the high-frequency energy proportion of the re-photographed images is significantly higher (the average value is generally 0.68± 0.11).

[0066] After analyzing the spectral energy distribution of a large number of normal images and re-shot images, the high-frequency energy proportion of the re-shot images is significantly higher (the high-frequency mean is 0.68 and the low-frequency mean is 0.12), so the high-frequency area energy proportion can be used to calculate the energy distribution of the re-shot images. To distinguish between normal images and re-photographed images.

[0067] In this embodiment, Low-frequency areas correspond to slowly changing areas in the image, such as smooth backgrounds, large color blocks, etc. In the frequency domain, low-frequency areas are concentrated near the center of the spectrum; this is because low frequencies dominate the smooth information, and most of the image's energy is concentrated in low frequencies, corresponding to the overall brightness and contrast of the image; The high-frequency area corresponds to the rapidly changing areas in the image, such as edges, textures, and noise. In the frequency domain, the high-frequency area is concentrated on the periphery of the spectrum; while the low-frequency response details and noise, edges, textures, and noise appear as high-frequency energy, but noise is usually randomly distributed, while periodic structures (such as moiré) are directional high-frequency energy; ​After the spectrum is centered, the center of the spectrum is taken as the origin, and the low-frequency area and the high-frequency area are distinguished by the radial distance relative to the center of the spectrum: that is: Low frequency region: radial distance R 径 ≤R low The central circular area; High frequency area: radial distance R 径 >R low 's outer annular area; Where R low is the boundary radius between the high-frequency area and the low-frequency area. Usually, 1 / 2 or 1 / 3 of the spectrum radius is used as the boundary radius. After testing, it is found that the detection accuracy is higher when 1 / 2 or 1 / 3 of the spectrum radius is used as the boundary radius, so the detection accuracy can be further improved. Of course, if the image resolution is too high or the noise level changes greatly, R can also be adjusted according to actual needs. low Dynamic adjustment, for example, for high-resolution images, R can be appropriately increased low , in order to retain more mid-high frequency details, and for strong noise images, R can be appropriately reduced low , to suppress high frequency noise.

[0068] For example, if the image size is 512*512, the spectrum radius R 频谱 is 256. When 1 / 2 of the spectrum radius is taken as the boundary radius, the boundary radius R low =R 频谱 / 2=128; Correspondingly, the above formula can be simplified to: ; S4: Determine the threshold range of normal images through threshold training, and determine whether the image is a re-shot image based on the preset judgment logic.

[0069] When determining the threshold range of normal images through threshold training, use a training set containing multiple normal images and multiple re-photographed images. For example, you can use a training set of 1,000 normal images and 1,000 re-photographed images, and count the high-frequency energy proportions respectively and peak intensity The distribution of , and determine the classification boundary of the threshold T by the following formula:

[0070] Among them, is the feature mean of a normal image; is the standard deviation of a normal image; When judging whether an image is a re-photographed image according to the preset judgment logic, the preset judgment logic satisfies: ​​(1) If ≥T and there are at least two significant periodic peaks, then it is determined to be a re-shot image; (2) Otherwise, it is judged as a normal image.

[0071] The above two-condition judgment mechanism can reduce the false positive rate (experimental verification shows that the false positive rate is <2%).

[0072] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.​

Claims

1. A method for identifying re-photographed images based on moiré analysis, characterized in that: The following steps are involved: S1: preprocess the image; S2: Convert the image from the spatial domain to the frequency domain and perform spectrum centering processing on it to obtain the frequency domain energy distribution corresponding to the image; S3: Analyze the frequency domain energy distribution to obtain the energy corresponding to the moiré feature, and then obtain the energy proportion corresponding to the moiré feature; S4: Determine the threshold range of the normal image through threshold training, and determine whether the image is a re-shot image according to a preset judgment logic.

2. The method for identifying a reproduced image based on moiré analysis according to claim 1, characterized in that: The step S1 at least includes the following steps: S1.1: Convert the color image to grayscale image; S1.2: Filter and reduce noise on the grayscale image; S1.3: Enhance the contrast of the image.

3. The method for identifying a reproduced image based on moiré analysis according to claim 2, characterized in that: In step S1.1, the RGB color image is converted into a grayscale image using a weighted average method, and the conversion is performed according to the following formula: ； in, (x, y) represents the brightness value of the converted grayscale image at the coordinate (x, y), ranging from 0 to 255; R(x, y), G(x, y), and B(x, y) represent the brightness values ​​of the red, green, and blue channels of the original color image at the coordinate (x, y), respectively, and the value range is 0~255.

4. The method for identifying a reproduced image based on moiré analysis according to claim 2, characterized in that: The step S1.2 includes the following steps: S1.21: Use a two-dimensional Gaussian kernel function to convolve the grayscale image and define the convolution function as: ； in, g represents the Gaussian function; x and y represent the two-dimensional coordinates of each point in the kernel; represents the standard deviation of the Gaussian kernel function; S1.22: Verify the grayscale image after filtering and noise reduction by using the signal-to-noise ratio: compare the signal-to-noise ratio of the grayscale image after this processing with the preset signal-to-noise ratio corresponding to the same standard deviation: When the current signal-to-noise ratio is not lower than the preset signal-to-noise ratio, the verification is deemed to have passed; otherwise, the verification is deemed to have failed.

5. The method for identifying a reproduced image based on moiré analysis according to claim 2, characterized in that: In step S1.3, the contrast of the image is enhanced by: S1.31: Divide the image into multiple sub-regions; S1.32: Perform histogram equalization on each sub-region independently; S1.33: Eliminate block boundary effects by bilinear interpolation.

6. The method for identifying a reproduced image based on moiré analysis according to claim 1, characterized in that: In step S2, when converting the image from the spatial domain to the frequency domain, a two-dimensional fast Fourier transform operation is performed on the preprocessed image according to the following formula to obtain a frequency domain complex matrix: : ； in, N and M are the length and width of the image respectively; and is the frequency domain coordinate; is the imaginary unit, representing the imaginary part of a complex number, satisfying j 2 =−1.

7. The method for identifying a reproduced image based on moiré analysis according to claim 6, characterized in that: In step S2, when performing spectrum centering processing, the zero-frequency component in the obtained frequency domain complex matrix is ​​moved to the center of the spectrum according to the following formula: ； in, Represents the frequency domain complex matrix after centering.

8. The method for identifying a reproduced image based on moiré analysis according to claim 1, characterized in that: In step S3, when obtaining the energy corresponding to the moiré feature, the center of the spectrum is taken as the origin, and the energy at different angles is calculated according to the following formula: Energy integral on : ； in, ； is the energy integral; is the angle of the integral direction, the range is [0°, 180°], and the step size is 1°; r represents the mirror distance from the center of the spectrum to the target point, unit: pixel; Then the energy integral is normalized and calculated: If there is a significant periodic peak, it is determined to be a moiré feature; If there is no significant periodic peak, it is determined that it is not a moiré feature.

9. The method for identifying a reproduced image based on moiré analysis according to claim 8, characterized in that: In step S3, the energy proportion corresponding to the moiré feature is obtained according to the following steps: The high-frequency area is defined as the outer ring of the spectrum, and its energy proportion is calculated according to the following formula: ； in, is the energy proportion of high frequency area; Represents the sum of the energy of all frequency domain points in the high-frequency region; Represents the sum of the energy of all frequency domain points in the entire frequency domain.

10. The method for identifying a reproduced image based on moiré analysis according to claim 1, characterized in that: In the step S4, When determining the threshold range of normal images through threshold training, a training set containing multiple normal images and multiple re-photographed images is used to count the proportion of high-frequency energy. and peak intensity The distribution of , and the classification boundary of the threshold T is determined by the following formula: ； in, is the feature mean of the normal image; is the standard deviation of the normal image; When judging whether an image is a re-photographed image according to the preset judgment logic, the preset judgment logic satisfies: (1) If ≥T and there are at least two significant periodic peaks, it is determined to be a re-shot image; (2) Otherwise, it is judged as a normal image.

Images