Power defect image hiding method based on deep invertible network and differential coding

Abstract

In order to solve the problems of the quality of the camouflage image, the quality of the recovered image and the security of the transmission in the process of hiding the power inspection defect image, the application discloses a power defect image hiding method based on a deep reversible network and differential coding. Firstly, the trained reversible scaling network is used for down-scaling the power inspection defect image; compared with the compression sensing method and the like, the reversible scaling network can recover the defect image with higher quality. Then, a new embedding algorithm based on differential coding is proposed, and the down-scaled defect image is embedded into the cover image by using the algorithm. The defect image is encoded by using the differential coding, and then the embedding operation is completed by using the least significant bit algorithm. The image value after the differential coding is concentrated in a smaller range, so that the damage of the embedding to the pixel value of the cover image is reduced. In addition, the application has good performance on the anti-steganalysis of SPAM.

Description

Technical Field

[0001] This invention pertains to image hiding technology within the field of information security, specifically relating to a method for hiding power defect images based on deep reversible networks and differential coding. Background Technology

[0002] With the rapid development and widespread adoption of the internet, the privacy and security of information transmitted over networks have attracted considerable attention. Currently, with the increasing prevalence of unmanned aerial vehicle (UAV) inspections of power grids, obtaining high-quality inspection images and utilizing deep learning technology to identify power grid defects has become a crucial means of ensuring stable power grid operation. However, how power companies can ensure the secure transmission and sharing of power grid defect images with power maintenance departments has become an urgent problem. If a defect image is identified and captured by hackers, on the one hand, it may cause power maintenance departments to miss the optimal maintenance window, potentially leading to a larger disaster; on the other hand, a corrupted defect image may lead to misjudgments by power maintenance departments, resulting in even greater disasters. Currently, there are two main types of image protection solutions: image encryption and image hiding. Since image encryption, particularly noise-based methods, is easily noticed and intercepted by hackers during network transmission, it cannot guarantee the security of power grid defect images. In the power industry, image hiding refers to embedding a defect image into a natural image to generate a visually meaningful camouflaged image. Because the camouflaged image and the natural image are visually similar, it is less likely to attract the attention of hackers during network transmission. Image hiding primarily focuses on three aspects: the quality of the camouflaged image, the quality of the restored image, and the resistance to steganalysis. High-quality camouflaged images are undetectable to the naked eye by hackers; high-quality restored images reduce the probability of power maintenance departments misreading defective image information; and high resistance to steganalysis reduces the probability of the camouflaged image being detected by steganalysis tools, thereby increasing the probability of secure transmission. Currently, there are many research results in the field of image hiding.

[0003] First, Lai and Lee proposed an image hiding method based on reversible color transformation. This method first divides the cover image and the secret image into non-overlapping blocks, then establishes a one-to-one match between the cover image blocks and the secret image blocks based on the standard deviation of the blocks, and finally completes the image hiding using color transformation. Since the quality of the camouflaged image obtained by this method is not ideal, Zhang and Hou improved the matching algorithm to reduce the number of block indices that need to be embedded. Zhang divided all blocks into two classes according to their standard deviation for matching. Hou used a clustering algorithm to divide the blocks into K classes for matching. However, the PSNR of the camouflaged image obtained by these methods is still difficult to exceed 30dB.

[0004] To further improve the quality of camouflaged images, several image hiding methods based on compressed sensing have been proposed. These methods first compress the secret image using compressed sensing technology and then embed it into the cover image. Although these methods improve the quality of the camouflaged image by reducing the amount of embedded data, improving the quality of the recovered image has become a major concern for many researchers because compressed sensing is a lossy compression technique. Compressed sensing, parallel compressed sensing, two-dimensional compressed sensing, block compressed sensing, and semi-tensor product compressed sensing have been designed to improve the quality of the recovered image. In addition, with the development of convolutional neural networks (CNNs), Baluja was the first to use CNNs to hide a secret image into a cover image of the same size. Subsequently, Baluja improved the hiding performance by improving the network architecture or the loss function. However, the quality of the recovered image obtained by these methods is still unsatisfactory, i.e., the PSNR of the recovered image is below 40dB, and these methods struggle to obtain both high-quality camouflaged images and high-quality recovered images simultaneously. Therefore, if these methods are used in power line inspection work, low-quality camouflaged images are easily noticed by attackers, and low-quality recovered image size can lead to misjudgments by the power maintenance department. Summary of the Invention

[0005] Objective of the Invention: Addressing the problems of existing technologies, the objective of this invention is to provide a power defect image hiding method based on deep reversible networks and differential coding, so as to simultaneously obtain high-quality camouflaged images and restored images. This invention proposes an image hiding method based on deep reversible networks and differential coding.

[0006] Technical Solution: To achieve the above objectives, this invention provides a power defect image hiding method based on deep reversible networks and differential coding. First, a reversible scaling network is used to scale the power inspection defect image. Then, a new embedding algorithm capable of obtaining a high-quality camouflage image is proposed. Differential coding is used to encode the secret image, and then the least significant bit algorithm is applied to complete the hiding. The encoded secret image has a smaller numerical range, thus causing less damage to the pixel values ​​of the cover image. The image hiding method includes the following steps:

[0007] A reversible scaling network is used to downscale power inspection defect images according to the sampling rate to obtain downscaled images;

[0008] The cover image is transformed using integer wavelet transform to obtain a low-frequency subband A and three high-frequency subbands H, V, and D;

[0009] Each decimal coefficient in the four sub-bands is converted into a 9-bit binary two's complement, with the highest bit representing the sign bit;

[0010] The downscaled image is encoded using lifting differential coding to obtain the encoded image;

[0011] Each decimal value in the encoded image is converted into a 9-bit two's complement binary code.

[0012] The least significant bit, bits 6 to 8, bits 4 and 5, and the three most significant bits of the 9-bit two's complement of the encoded image are embedded into the least significant bits of the 9-bit two's complement of subbands A, H, V, and D, respectively.

[0013] The 9-bit binary two's complement of the subband after embedding information is converted into decimal coefficients, and then inverse integer wavelet transform is performed to obtain the camouflage image.

[0014] Image restoration includes the following steps:

[0015] Integer wavelet transform is used to transform the camouflaged image to obtain a low-frequency subband A′ and three high-frequency subbands H′, V′, and D′;

[0016] Each decimal coefficient in the four sub-bands is converted into a 9-bit binary two's complement;

[0017] Extract the least significant bit, the three lowest bits, the two lowest bits, and the three lowest bits of the two's complement of subbands A', H', V', and D' respectively, and concatenate them together;

[0018] The concatenated binary two's complement is converted into a decimal value, and then boost differential decoding is performed to obtain the downscaled image;

[0019] A reversible scaling network is used to magnify the downscaled image according to the sampling rate to recover the defect image of the power line inspection.

[0020] Preferably, the 9-bit two's complement of the encoded image is embedded into the lower bits of the two's complement of the four sub-bands, represented as:

[0021]

[0022] in This represents taking the e-th to f-th bits of the (i,j)-th coefficient of matrix x, where x∈{a,h,v,d,ds}, i∈[1,P / 2], j∈[1,Q / 2], a,h,v,d,ds are the two's complement matrices of subbands A, H, V, D and the encoded image DS, respectively, and P and Q are the height and width of the cover image.

[0023] Preferably, the reversible scaling network consists of m superimposed downscaling modules in its forward process; each downscaling module includes a discrete wavelet transform module and n reversible neural network blocks (InvBlocks), using the discrete wavelet transform module to decompose a high-resolution power line inspection defect image X of size H×W into low-frequency subbands. and high-frequency subband and The input is fed into the first lnvBlock and then output. and Repeat this process until all lnvBlock operations are completed; the output of the nth lnvBlock of the last downscaling module is... and The obtained low-resolution image Y and loss information Z are represented by the following: Y has a size of H'×W', where... SR = 1 / 4 m is the sampling rate, and m is the number of downscaling modules.

[0024] Preferably, when the reversible scaling network performs image restoration, the loss information Z' is randomly sampled from a specified distribution. and Input the nth lnvBlock, process it, and output the result. and The final output of the first lnvBlock is used as the recovered low-frequency subband. and high-frequency subband Will and The input is fed into the inverse integer wavelet transform block for transformation to obtain the restored high-resolution image.

[0025] Preferably, the process of improving differential coding is expressed as follows:

[0026]

[0027] Where Y′ represents the differentially encoded image.

[0028] Preferably, the decoding process of the lifting differential coding is expressed as follows:

[0029]

[0030] Based on the same inventive concept, this invention provides an image hiding system based on deep reversible networks and differential coding, comprising an image hiding module and an image restoration module; the image hiding module includes:

[0031] The downscaling unit is used to downscale the power inspection defect image according to the sampling rate using a reversible scaling network to obtain a downscaled image.

[0032] The cover image wavelet transform unit is used to transform the cover image using integer wavelet transform to obtain a low-frequency subband A and three high-frequency subbands H, V, and D;

[0033] The encoding unit is used to encode the downscaled image using lifting differential coding to obtain the encoded image;

[0034] The binary two's complement conversion unit is used to convert each decimal coefficient in the four sub-bands into a 9-bit binary two's complement, with the highest bit representing the sign bit; and to convert each decimal value in the encoded image into a 9-bit binary two's complement.

[0035] The information embedding unit is used to embed the least significant bit, bits 6 to 8, bits 4 and 5, and the three most significant bits of the 9-bit binary two's complement of the encoded image into the least significant bit of the 9-bit binary two's complement of subbands A, H, V, and D, respectively.

[0036] The camouflage image generation unit is used to convert the 9-bit binary two's complement of the subband after embedding information into decimal coefficients and perform inverse integer wavelet transform to obtain the camouflage image.

[0037] The image restoration includes:

[0038] The camouflage image wavelet transform unit is used to transform the camouflage image using integer wavelet transform to obtain a low-frequency subband A' and three high-frequency subbands H', V', and D';

[0039] The information extraction unit is used to convert each decimal coefficient in the four sub-bands into a 9-bit binary two's complement; and to extract the least significant bit, the lower three bits, the lower two bits, and the lower three bits of the binary two's complement of sub-bands A', H', V', and D' respectively, and concatenate them together.

[0040] The decoding unit is used to convert the concatenated binary two's complement into a decimal value and perform boost differential decoding to obtain a downscaled image;

[0041] The image restoration unit is used to magnify the downscaled image according to the sampling rate using a reversible scaling network to restore the power inspection defect image.

[0042] Based on the same inventive concept, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the image hiding and / or image restoration steps of the image hiding method based on deep reversible networks and differential coding.

[0043] Based on the same inventive concept, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the image hiding and / or image restoration steps of the image hiding method based on deep reversible networks and differential coding.

[0044] Beneficial Effects: Compared with existing inventions, this invention has the following advantages: (1) This invention uses a reversible scaling network to scale power inspection defect images. Compared with existing compressed sensing technology, the quality of the secret image recovered by the reversible scaling network is higher. (2) This invention proposes a new embedding algorithm that can obtain high-quality camouflage images. Unlike algorithms that directly apply the least significant bit embedding algorithm to the secret image to complete the hiding, we use differential coding to encode the secret image and then apply the least significant bit algorithm to complete the hiding. The numerical range of the encoded secret image is smaller, so it causes less damage to the pixel values ​​of the cover image. (3) Extensive experimental results show that this invention can not only obtain high-quality camouflage images and high-quality recovered images at the same time, but also the anti-steganography analysis capability of the camouflage image is better than other methods. Attached Figure Description

[0045] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0046] Figure 2 This is a visualization of the simulation results.

[0047] Figure 3 Examples of the cover image (a) and the camouflage image (b).

[0048] Figure 4 Histograms of the RGB three channels of the cover image (a)(c)(e) and the camouflage image (b)(d)(f).

[0049] Figure 5 The results of the correlation coefficient analysis are shown in the figure. (a)(c)(e) correspond to the horizontal, vertical and diagonal directions of the cover image, and (b)(d)(f) correspond to the horizontal, vertical and diagonal directions of the camouflage image. Detailed Implementation

[0050] The present invention will now be further described. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0051] like Figure 1 As shown in the figure, an embodiment of the present invention discloses a method for hiding power defect images based on deep reversible networks and differential coding, including image hiding and image restoration processes.

[0052] Image hiding consists of two stages: downscaling and embedding. The specific steps for embedding a power line inspection defect image of size H×W into a cover image of size P×Q are as follows.

[0053] Step 1: First, train the reversible scaling network, and then use the reversible scaling network to downscale the power inspection defect image D according to the sampling rate SR to obtain a downscaled image S of size H′×W'.

[0054] Step 2: Embed the downscaled image into the cover image using an embedding algorithm. The specific process is as follows:

[0055] Step 2.1: Use Haar Integer Wavelet Transform (Haar IWT) to transform the cover image to obtain a low-frequency subband A and three high-frequency subbands H, V and D, each with a size of (P / 2)×(Q / 2).

[0056] Step 2.2: Convert each decimal coefficient in subbands A, H, V, and D into 9-bit two's complement binary code, denoted as follows: in i∈[1,P / 2], j∈[1,Q / 2] represent the 1st to 9th bit two's complement of the coefficient in the i-th row and j-th column of subband A. and The meanings are similar. Note that we use 9 bits to represent each coefficient instead of 8 bits because the coefficients of the H, V, and D subbands contain negative, positive, and 0 values, so the highest bit is used to represent the sign bit.

[0057] Step 2.3: Encode the downscaled image S using lifting differential coding to obtain the encoded image DS, which has a size of H'×W'.

[0058] Step 2.4: Convert each decimal value of the encoded image DS into 9-bit two's complement binary code, denoted as i∈[1,H'],j∈[1,W'].

[0059] Step 2.5: Use the least significant bit embedding algorithm to... The lowest digit Embedded The 6th, 7th, and 8th positions Embedded The 4th and 5th positions Embedded The three senior Embedded The embedding process is shown in the following equation:

[0060]

[0061] In all After all the embedding is complete, four sub-bands containing the embedded information are created. It is obtained. In this embodiment, H' = P / 2 and W' = Q / 2 are assumed to be obtained.

[0062] Step 2.6: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Each 9-bit binary two's complement is converted into a decimal value and denoted as A', H', V', D'.

[0063] Step 2.7: Perform inverse Haar integer wavelet transform on the four subbands A′, H′, V′, and D′ to obtain a camouflage image of size P×Q.

[0064] Image restoration process:

[0065] Image restoration consists of two stages: extraction and upscaling. The process of restoring a power line inspection defect image of size H×W from a camouflaged image of size P×Q is as follows.

[0066] Step 1: Use an extraction algorithm to extract and restore the scaled image from the camouflaged image.

[0067] Step 1.1: Use Haar integer wavelet transform to transform the camouflaged image to obtain four subbands A', H', V', and D'.

[0068] Step 1.2: Convert each coefficient of subbands A', H', V', and D' into 9-bit two's complement binary code to obtain...

[0069] Step 1.3: Extraction The lower one extract The lower three extract The two lower ones extract The lower three Then When pieced together, it looks like this:

[0070]

[0071] Where i∈[1,P / 2], j∈[1,Q / 2].

[0072] Step 1.4: Take the results obtained in Step 1.3 Convert to decimal form to obtain the downscaled image S.

[0073] Step 2: At a sampling rate of 1 / SR, use a reversible scaling network to enlarge the downscaled image S to obtain the restored defective image D'.

[0074] Step 1 specifically refers to the forward process of the deep reversible network consisting of m stacked downscaling modules. Each downscaling module includes a discrete wavelet transform module and n reversible neural network blocks (lnvBlocks). The discrete wavelet transform module decomposes a high-resolution image X of size H×W into low-frequency subbands. and high-frequency subband Then, and The input is fed into the l-th lnvBlock and output. and Repeat this process until all lnvBlock operations are completed, i.e., the output of the l-th lnvBlock is... and The input is fed into the (l+1)th lnvBlock. The final output of the nth lnvBlock is... and The operation process of the l-th lnvBlock is as follows:

[0075]

[0076]

[0077] Where ⊙ represents the Hadamard product, exp represents the exponential operation, and φ(·), ρ(·), and η(·) are all implemented by densely connected convolutional blocks (called dense blocks). After scaling, a low-resolution image and loss information It is obtained. The dimension of Y is H'xW', where SR = 1 / 4 m is the sampling rate, and m is the number of downscaling modules. During downscaling, the lost information Z is trained to follow a simple, specified distribution, such as a Gaussian distribution.

[0078] In step 2.3 of the image hiding process, differential coding refers to representing each element of a digital data stream, except for the first element, as the difference between that element and its preceding element. This can greatly reduce digital redundancy, i.e., reduce the repetition of identical numbers. If differential coding is applied to a digital image, it means that the pixel value of the next column in each row is subtracted from the pixel value of the previous column. Since adjacent pixel values ​​in a natural image are always the same or similar, differential coding can not only reduce pixel value redundancy but also greatly reduce the range of pixel values. The process of differential coding for a digital image is as follows:

[0079] Y'=Y(:,j)-Y(:,j-1),j∈[2,W']

[0080] Where Y represents the original image, and Y' represents the differentially encoded image. From the above equation, it can be seen that the first column of traditional differential coding remains unchanged, meaning the image encoding is insufficient. Therefore, we perform differential coding separately on the first column of the original image Y, based on the above equation; this is called lifting differential coding. The entire process of lifting differential coding is as follows:

[0081]

[0082] The corresponding decoding process is as follows:

[0083]

[0084] The image restoration step 2 specifically refers to the loss of information. They are randomly sampled from a specified distribution. Then, and The input is processed by the nth InvBlock to output. and The process is repeated as follows. After multiple InvBlock processes, the low-frequency subband is recovered. and high-frequency subband

[0085]

[0086]

[0087] Finally, and The image is input into the inverse Haar DWT conversion block for conversion to obtain a restored high-resolution image.

[0088] The performance analysis of the embodiments is presented below with reference to the accompanying drawings and experimental data:

[0089] 1. Image quality analysis

[0090] This invention conducts simulation experiments on both image quality and security analysis to demonstrate the superior performance of our method. Regarding image quality, we conducted three sets of simulation experiments using randomly selected images and presented the visualization results. Furthermore, two objective quality evaluation metrics, Peak Signal-to-Noise Ratio (PSNR) and Mean Structural Similarity (MSSIM), were selected to evaluate image quality and compared with other state-of-the-art methods. PSNR and MSSIM measure the degree of distortion of an image compared to its corresponding reference image. A higher PSNR and a MSSIM closer to 1 indicate less image distortion, meaning higher image quality. The calculation of PSNR and MSSIM is as follows:

[0091]

[0092]

[0093]

[0094]

[0095] Where X and Y are the pixel matrices of the two images, X is the pixel matrix of the cover image, Y is the pixel matrix of the dense image, L is the difference between the maximum and minimum possible values ​​of the image pixels, M and N are the length and width of the image, respectively, and G is the number of blocks after dividing the image into blocks. k and Y k Let μ be the k-th image patch in X and Y respectively. x and μ y Let x and y be the mean and standard deviation of the image patch, respectively, and σ be the standard deviation of the image patch. xy Let x be the covariance of blocks x and y. k1 and k2 are two parameters, where k1 and k2 are 0.01 and 0.03 respectively.

[0096] 1.1 Visualization Results

[0097] Figure 2 The visualization results of three sets of experiments are presented. A 512×512 power inspection defect image is selected, and a reversible scaling network is used to downscale the defect image at a sampling rate of 0.25 to obtain a downscaled image of size 256×256. Then, our proposed embedding algorithm is used to embed the downscaled image into a 512×512 cover image to obtain a 512×512 camouflaged image. Finally, the defect image is recovered from the camouflaged image, and the recovered defect image is 512×512 in size. Careful observation of the visualization results of the three sets of experiments reveals that the camouflaged image and the cover image, as well as the defect image and the recovered defect image, are visually identical, i.e., there is no visible distortion. This demonstrates that our method can obtain high-quality camouflaged images and recovered defect images. On the one hand, this ensures that the camouflaged image can evade the attention of attackers during transmission, thus ensuring security; on the other hand, the visually distortion-free recovered defect image ensures that power maintenance departments can accurately identify power system problems and carry out timely repairs.

[0098] 1.2 Quality of camouflaged images

[0099] We use two metrics, PSNR and SSIM, to quantify the quality of our camouflaged images and compare them with other state-of-the-art methods. Tables 1 and 2 show the quantitative results. From Table 1, the average PSNR of our camouflaged images is 39.94 dB. Table 2 shows similar results for SSIM. This indicates that our scheme is less likely to attract the attention of attackers during network transmission and has high security. The good camouflaged image quality achieved by this invention is mainly due to the following three points. First, compared with directly embedding defective images in the spatial domain of the cover image, our method embedding defective images in the frequency domain of the cover image causes less damage to the pixel values ​​of the cover image; second, in our method, the defective images encoded by differential coding have smaller values, so the damage to the subband coefficients is less during the embedding process; third, during the embedding process, according to the degree of influence of the modification of different subband coefficients on image quality, every 9 bits of binary data are reasonably distributed in four different subbands. As is well known, modifications to the coefficients of low-frequency subband A have the greatest impact on image quality, modifications to the coefficients of high-frequency subband D have the least impact, and modifications to the coefficients of high-frequency subbands H and V have an impact on image quality that falls between that of subband A and subband D. Therefore, in this invention, each coefficient of subband A is embedded with 1 bit, each coefficient of subband D is embedded with 3 bits, and each coefficient of subbands H and V is embedded with 3 bits and 2 bits, respectively.

[0100] 1.3 Restoring Image Quality

[0101] When image hiding is applied to power line inspection, high-quality restored defect images are crucial for accurate location and timely repair by power maintenance personnel. Low-quality restored defect images can mislead maintenance personnel, leading to greater losses. Tables 3 and 4 show the PSNR and MSSIM of the images restored by our method and compare them with other advanced methods. From Table 3, we find that the quality of the restored images by our method is above 40 dB. Therefore, this invention is suitable for application in the field of power line inspection because the high-quality restored defect images can reduce misjudgments of power system problems by maintenance departments. Since the embedding method is completely reversible, the superior quality of our restored defect images is only related to the high performance of the deep reversible network. The reversible scaling network can retain more information during the downscaling process of the image. This is because the reversible scaling network uses latent variables of a specific distribution to capture the distribution of lost information during downscaling, and then simulates this specific distribution to plot the latent variables during upscaling.

[0102] Table 1 PSNR of camouflage images and cover images

[0103]

[0104] Table 2 MSSIM of camouflage images and cover images

[0105]

[0106] Table 3 PSNR of restored and secret images

[0107]

[0108] Table 4 MSSIM for Recovered and Secret Images

[0109]

[0110] 2. Security Analysis

[0111] The ability to transmit camouflaged images over a network is crucial for power maintenance departments to receive them promptly and accurately reconstruct images of power inspection defects. Therefore, we conducted steganalysis, histogram analysis, and correlation coefficient analysis on the camouflaged images to verify the security of the method presented in this invention.

[0112] 2.1 Resistance to Steganalysis

[0113] While high-quality camouflaged images can evade attackers during network transmission, the ability of camouflaged images to resist detection by common steganalysis tools has become a crucial criterion for security, given the rapid development of steganalysis technology. Steganalysis tools are used to identify whether information is hidden in an image. In the field of image hiding, researchers evaluate the steganalysis resistance of a method in terms of detection precision. Since the probability of correctly guessing an image is 50%, a detection precision of approximately 50% indicates a stronger steganalysis resistance. SPAM was trained on two classic steganalysis methods (WOW and S-UNIWARD), and then the steganalysis resistance performance of our invention was tested. The first 6000 images from BOSSbase_1.01 were used for training, and the last 2000 were used for testing. Table 5 shows the detection precision of our method. In Table 5, the average detection precision of our method is 34.20%. Therefore, our camouflaged images can evade detection by steganalysis tools during network transmission.

[0114] Table 5. Steg resistance capabilities of SPAM

[0115]

[0116] 2.2 Histogram Analysis

[0117] Histogram analysis is a method for detailing the distribution of pixel values ​​in an image. Different images typically exhibit different distribution trends. Once the pixel values ​​of an image are changed, its histogram distribution may alter. Therefore, observing whether the histogram distribution of an image is similar to or identical to that of the original image has become a classic method for attackers to identify whether an image hides secret information. Based on this, ensuring the similarity of the histograms of the spoofed image and the cover image is crucial for guaranteeing the security of defective image transmission and for ensuring that the spoofed image can evade attackers' attention. Figure 4 They were shown respectively Figure 2 The three-channel histograms of the cover image and the camouflage image in the first set of experiments. Our method ensures that the histograms of the cover image and the camouflage image are visually consistent, enabling the camouflage image to evade attackers and be securely transmitted to the power maintenance department.

[0118] 2.3 Correlation Coefficient Analysis

[0119] The correlation coefficient is a classic method for evaluating the correlation between adjacent pixels in an image. Typically, a natural image with unmodified pixel values ​​has strong pixel correlation. However, if pixel values ​​are altered due to information embedding, the pixel values ​​between pixels may be disrupted, and the correlation coefficient will decrease. Therefore, the correlation coefficient is also an important aspect of measuring whether a spoofed image can evade attacker detection and thus be securely transmitted to power maintenance departments. Figure 5 And Table 6 shows Figure 2 The correlation results between the cover image and the camouflage image in the first group of experiments. First, from... Figure 5 The study found that the correlation distributions of the cover image and the camouflage image were basically consistent in all three directions. Furthermore, Table 6 shows that the correlation coefficients between the cover image and the camouflage image are very similar. This indicates that our method causes minimal disruption to the correlation of image pixel values, making it less likely to be detected by attackers and thus ensuring transmission security.

[0120] Table 6 Correlation coefficients

[0121]

[0122] Based on the same inventive concept, this invention discloses an image hiding system based on deep reversible networks and differential coding, comprising an image hiding module and an image restoration module; the image hiding module includes:

[0123] The downscaling unit is used to downscale the power inspection defect image according to the sampling rate using a reversible scaling network to obtain a downscaled image.

[0124] The cover image wavelet transform unit is used to transform the cover image using integer wavelet transform to obtain a low-frequency subband A and three high-frequency subbands H, V, and D;

[0125] The encoding unit is used to encode the downscaled image using lifting differential coding to obtain the encoded image;

[0126] The binary two's complement conversion unit is used to convert each decimal coefficient in the four sub-bands into a 9-bit binary two's complement, with the highest bit representing the sign bit; and to convert each decimal value in the encoded image into a 9-bit binary two's complement.

[0127] The information embedding unit is used to embed the least significant bit, bits 6 to 8, bits 4 and 5, and the three most significant bits of the 9-bit binary two's complement of the encoded image into the least significant bit of the 9-bit binary two's complement of subbands A, H, V, and D, respectively.

[0128] The camouflage image generation unit is used to convert the 9-bit binary two's complement of the subband after embedding information into decimal coefficients and perform inverse integer wavelet transform to obtain the camouflage image.

[0129] The image restoration includes:

[0130] The camouflage image wavelet transform unit is used to transform the camouflage image using integer wavelet transform to obtain a low-frequency subband A' and three high-frequency subbands H', V', and D';

[0131] The information extraction unit is used to convert each decimal coefficient in the four sub-bands into a 9-bit binary two's complement; and to extract the least significant bit, the lower three bits, the lower two bits, and the lower three bits of the binary two's complement of sub-bands A', H', V', and D' respectively, and concatenate them together.

[0132] The decoding unit is used to convert the concatenated binary two's complement into a decimal value and perform boost differential decoding to obtain a downscaled image;

[0133] The image restoration unit is used to magnify the downscaled image according to the sampling rate using a reversible scaling network to restore the power inspection defect image.

[0134] Based on the same inventive concept, an electronic device disclosed in this invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the image hiding and / or image restoration steps of the image hiding method based on deep reversible networks and differential coding.

[0135] Based on the same inventive concept, embodiments of the present invention disclose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the image hiding and / or image restoration steps of the image hiding method based on deep reversible networks and differential coding.

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any transformations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for hiding power defect images based on deep reversible networks and differential coding, comprising image hiding and image restoration steps; characterized in that, Image hiding includes the following steps: A reversible scaling network is used to downscale power inspection defect images according to the sampling rate to obtain downscaled images; The cover image is transformed using integer wavelet transform to obtain a low-frequency subband A and three high-frequency subbands H, V, and D; Each decimal coefficient in the four sub-bands is converted into a 9-bit binary two's complement, with the highest bit representing the sign bit; The downscaled image is encoded using lifting differential coding to obtain the encoded image; Each decimal value in the encoded image is converted into a 9-bit two's complement binary code. The least significant bit, bits 6 to 8, bits 4 and 5, and the three most significant bits of the 9-bit two's complement of the encoded image are embedded into the least significant bits of the 9-bit two's complement of subbands A, H, V, and D, respectively. The 9-bit binary two's complement of the subband after embedding information is converted into decimal coefficients, and then inverse integer wavelet transform is performed to obtain the camouflage image. Image restoration includes the following steps: Integer wavelet transform is used to transform the camouflaged image to obtain a low-frequency subband. and three high-frequency subbands , , ; Each decimal coefficient in the four sub-bands is converted into a 9-bit binary two's complement; Extract sub-bands separately , , , The least significant bit, the three least significant bits, the two least significant bits, and the three least significant bits of the two's complement binary code are concatenated together. The concatenated binary two's complement is converted into a decimal value, and then boost differential decoding is performed to obtain the downscaled image; A reversible scaling network is used to magnify the downscaled image according to the sampling rate to recover the defect image of the power line inspection.

2. The power defect image hiding method based on deep reversible networks and differential coding according to claim 1, characterized in that, The 9-bit two's complement of the encoded image is embedded into the lower bits of the two's complement of the four sub-bands, represented as: ； in Indicates taking the matrix The The coefficient of the first The position reached the first Bit binary, , , , These are the two's complement matrices of subbands A, H, V, D and the encoded image DS, respectively, and P and Q are the height and width of the cover image.

3. The power defect image hiding method based on deep reversible networks and differential coding according to claim 1, characterized in that, The reversible scaling network's forward process consists of m stacked downscaling modules; each downscaling module includes a discrete wavelet transform module and n reversible neural network blocks (InvBlocks), utilizing the discrete wavelet transform module to scale the network to a size of... High-resolution power line inspection defect images X are decomposed into low-frequency subbands. and high-frequency subband ; and The input is fed into the first InvBlock and then output. and Repeat this process until all InvBlock operations are completed; the output of the nth InvBlock of the last downscaling module is... and As the obtained low-resolution image Y and the loss information Z, the size of Y is ,in , , is the sampling rate, and m is the number of downscaling modules.

4. The power defect image hiding method based on deep reversible networks and differential coding according to claim 3, characterized in that, When the reversible scaling network performs image restoration, it loses information. Random samples will be taken from a specified distribution. and Input the nth InvBlock, process it, and output the result. and The first InvBlock output is used as the recovered low-frequency subband. and high-frequency subband ;Will and The input is fed into the inverse integer wavelet transform block for transformation to obtain the restored high-resolution image.

5. The power defect image hiding method based on deep reversible networks and differential coding according to claim 1, characterized in that, The process of improving differential coding is represented as follows: ； Where Y represents the original image, This represents the image after differential coding. It is the column number of the original image. It is the row number of the original image.

6. The power defect image hiding method based on deep reversible networks and differential coding according to claim 5, characterized in that, The decoding process of the lifting differential coding is represented as follows: 。 7. A power defect image hiding system based on deep reversible networks and differential coding, comprising an image hiding module and an image restoration module; characterized in that, The image hiding module includes: The downscaling unit is used to downscale the power inspection defect image according to the sampling rate using a reversible scaling network to obtain a downscaled image. The cover image wavelet transform unit is used to transform the cover image using integer wavelet transform to obtain a low-frequency subband A and three high-frequency subbands H, V, and D; The encoding unit is used to encode the downscaled image using lifting differential coding to obtain the encoded image; The binary two's complement conversion unit is used to convert each decimal coefficient in the four sub-bands into a 9-bit binary two's complement, with the highest bit representing the sign bit; and to convert each decimal value in the encoded image into a 9-bit binary two's complement. The information embedding unit is used to embed the least significant bit, bits 6 to 8, bits 4 and 5, and the three most significant bits of the 9-bit binary two's complement of the encoded image into the least significant bit of the 9-bit binary two's complement of subbands A, H, V, and D, respectively. The camouflage image generation unit is used to convert the 9-bit binary two's complement of the subband after embedding information into decimal coefficients and perform inverse integer wavelet transform to obtain the camouflage image. The image restoration module includes: The camouflage image wavelet transform unit is used to transform the camouflage image using integer wavelet transform to obtain a low-frequency subband. and three high-frequency subbands , , ; The information extraction unit is used to convert each decimal coefficient in the four sub-bands into 9-bit binary two's complement; and to extract the sub-bands respectively. , , , The least significant bit, the three least significant bits, the two least significant bits, and the three least significant bits of the two's complement binary code are concatenated together. The decoding unit is used to convert the concatenated binary two's complement into a decimal value and perform boost differential decoding to obtain a downscaled image; The image restoration unit is used to magnify the downscaled image according to the sampling rate using a reversible scaling network to restore the power inspection defect image.

8. The power defect image hiding system based on deep reversible networks and differential coding according to claim 7, characterized in that, The process of improving differential coding is represented as follows: ； Where Y represents the original image, This represents the image after differential coding. It is the column number of the original image. It is the row number of the original image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the power defect image hiding and / or image restoration steps of the image hiding method based on deep reversible networks and differential coding according to any one of claims 1-6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the image hiding and / or image restoration steps of the power defect image hiding method based on deep reversible networks and differential coding according to any one of claims 1-6.

Images