Image watermark embedding and extraction method, device, equipment and readable storage medium

Abstract

The application discloses an image watermark embedding and extracting method, device and equipment and a readable storage medium, and belongs to the technical field of multimedia content security. The application obtains first watermark information and an original picture, converts the first watermark information and the original picture into two-dimensional watermark features and original picture frequency domain features, inputs the two-dimensional watermark features and the original picture frequency domain features into an invertible neural network in a forward direction, and outputs a watermark picture and loss information; converts the watermark picture and the loss information into watermark picture frequency domain features and a matrix, inputs the watermark picture frequency domain features and the matrix into the invertible neural network in a reverse direction, and outputs second watermark information; and compares the first watermark information with the second watermark information to obtain a similarity verification result. The similarity verification result is used to determine the copyright ownership of the original picture. In combination with a discrete wavelet transform, watermark message processing and the invertible neural network, a high-robustness deep learning image watermark method is realized on the basis of ensuring image quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multimedia content security, in particular to an image watermark embedding and extraction method and device, equipment and a readable storage medium. BACKGROUND

[0002] With the rapid development of the Internet, a large number of application requirements for image digital watermarking have emerged. Digital watermarking is an important technology in the field of information hiding. It hides watermark information with specific meaning in digital products such as images, videos, audios, and documents through digital embedding. Through extraction and restoration of watermark information, it can prove the ownership of the work to the owner of the work and serve as evidence to identify illegal infringement, thereby becoming an effective means of copyright protection and digital multimedia product anti-counterfeiting.

[0003] According to whether the embedding method of digital watermarking uses a deep neural network, digital watermarking can be roughly divided into two categories: one is a traditional method that embeds watermarking by artificially designing image features, and the other is a deep learning method that learns image features by using a deep neural network to embed watermarking. However, the traditional method relies heavily on artificially designed image features and cannot fully utilize the redundancy information of the picture, resulting in poor robustness of the traditional watermark embedding method. The deep learning method applies a deep neural network to digital image watermarking and improves the robustness of the watermarking method by adding various common image noise in training. However, most deep learning methods perform poorly in terms of robustness against JPEG compression, which is a non-differentiable and commonly used lossy image compression method on the Internet.

[0004] Therefore, there is a technical problem that existing digital watermarking technology cannot resist non-differentiable noise.

[0005] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0006] The main purpose of the present application is to provide an image watermark embedding and extraction method, device, equipment and readable storage medium, which aims to design and implement a deep learning digital watermarking method that is more robust than existing digital watermarking methods under various attacks.

[0007] To achieve the above purpose, the present application provides an image watermark embedding and extraction method, which comprises the following steps:

[0008] Obtain first watermark information and an original picture, the original picture being a picture to be embedded with watermarking;

[0009] transforming the original picture and the first watermark information into original picture frequency domain features and two-dimensional watermark features with redundant information, wherein the redundant information is evenly spread on the two-dimensional watermark features;

[0010] forward inputting the two-dimensional watermark features and the original picture frequency domain features into a reversible neural network, and outputting a watermark picture and loss information;

[0011] transforming the watermark picture and the loss information into watermark picture frequency domain features and a matrix, the matrix being a matrix with the same dimension as the loss information generated by Gaussian sampling;

[0012] backward inputting the watermark picture frequency domain features and the matrix into the reversible neural network, and outputting second watermark information;

[0013] comparing the first watermark information with the second watermark information to obtain a similarity verification result, wherein the similarity verification result is used to determine the copyright ownership of the original picture.

[0014] Optionally, the step of transforming the original picture and the first watermark information into original picture frequency domain features and two-dimensional watermark features with redundant information, wherein the redundant information is evenly spread on the two-dimensional watermark features, comprises:

[0015] based on a compression and excitation network, forward transforming the first watermark information into two-dimensional watermark features with redundant information, the first watermark information being one-dimensional watermark information;

[0016] forward transforming the original picture into the original picture frequency domain features through discrete wavelet transform, wherein embedding the watermark through the frequency domain can improve the image quality of the watermark picture.

[0017] Optionally, the step of based on a compression and excitation network, forward transforming the first watermark information into two-dimensional watermark features with redundant information, the first watermark information being one-dimensional watermark information, comprises:

[0018] increasing the length of the first watermark information through a fully connected layer;

[0019] transforming the first watermark information into two-dimensional watermark information, the two-dimensional watermark information being single-channel;

[0020] based on the compression and excitation network, transforming the two-dimensional watermark information into two-dimensional watermark features with redundant information, the two-dimensional watermark features being multi-channel.

[0021] Optionally, the step of based on the compression and excitation network, transforming the two-dimensional watermark information into two-dimensional watermark features with redundant information, comprises:

[0022] Based on the compression and excitation network, the two-dimensional watermark information is expanded to generate the two-dimensional watermark feature.

[0023] Optionally, the step of forward inputting the two-dimensional watermark feature and the original picture frequency domain feature to the reversible neural network and outputting a watermark picture and loss information comprises:

[0024] Forward inputting the two-dimensional watermark feature and the original picture frequency domain feature to the reversible neural network and outputting a watermark picture frequency domain feature and loss information;

[0025] Converting the watermark picture frequency domain feature into a watermark picture through inverse discrete wavelet transform.

[0026] Optionally, the step of backward inputting the watermark picture frequency domain feature and the matrix to the reversible neural network and outputting second watermark information comprises:

[0027] Backward inputting the watermark picture frequency domain feature and the matrix to the reversible neural network and outputting second watermark feature and recovered picture frequency domain information;

[0028] Based on the compression and excitation network, the second watermark feature is converted into second watermark information.

[0029] Optionally, the step of forward inputting the two-dimensional watermark feature and the original picture frequency domain feature to the reversible neural network and outputting a watermark picture and loss information comprises:

[0030] Adding noise to the watermark picture through a noise layer, the noise comprising non-differentiable noise and or differentiable noise, wherein the noise layer adds the differentiable noise to the watermark picture using a Standard ASL (standard attack simulation layer) and or adds the non-differentiable noise to the watermark picture using a Forward ASL (forward attack simulation layer).

[0031] In addition, to achieve the above object, the present application also provides an image watermark embedding and extracting device, the device comprising:

[0032] An acquisition module is configured to acquire first watermark information and an original picture, the original picture being a picture to be embedded with a watermark.

[0033] A conversion module is configured to convert the original picture and the first watermark information into an original picture frequency domain feature and a two-dimensional watermark feature with redundant information, wherein the redundant information is diffused on the two-dimensional watermark feature averagely.

[0034] The watermark picture and the loss information are converted into watermark picture frequency domain features and a matrix, and the matrix is a matrix with the same dimension as the loss information generated by Gaussian sampling.

[0035] An embedding module is configured to forwardly input the two-dimensional watermark features and the original picture frequency domain features into an invertible neural network, and output a watermark picture and loss information.

[0036] An extraction module is configured to reversely input the watermark picture frequency domain features and the matrix into the invertible neural network, and output second watermark information.

[0037] A comparison module is configured to compare the first watermark information and the second watermark information to obtain a similarity verification result, wherein the similarity verification result is used to determine the copyright ownership of the original picture.

[0038] In addition, to achieve the above-mentioned purpose, the present application also provides an image watermark embedding and extraction device, which comprises a memory, a processor, and an image watermark embedding and extraction program stored in the memory and executable on the processor, and the image watermark embedding and extraction program is configured to implement the steps of the image watermark embedding and extraction method according to any one of claims 1 to 7.

[0039] In addition, to achieve the above-mentioned purpose, the present application also provides a readable storage medium, which stores an image watermark embedding and extraction program, and the image watermark embedding and extraction program is executed by a processor to implement the steps of the image watermark embedding and extraction method according to any one of claims 1 to 7.

[0040] Compared with other deep learning digital watermarking methods that directly add watermarks in the spatial domain, the present application embeds and extracts watermarks in the original picture frequency domain, achieving better image quality. On this basis, the present application further designs an invertible neural network responsible for embedding and extracting two-dimensional watermark features in the original picture frequency domain, which improves the robustness of the watermark image in terms of anti-compression transcoding, picture cropping, Gaussian noise, and other classic image noise while ensuring image quality. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 is a structural schematic diagram of an image watermark embedding and extraction device of a hardware running environment related to the embodiment scheme of the present application;

[0042] Figure 2 is a flowchart of the first embodiment of the image watermark embedding and extraction method of the present application;

[0043] Figure 3The reversible neural network internal processing diagram of the image watermark embedding and extracting method first embodiment of the present application;

[0044] Figure 4 The reversible neural network internal processing diagram of the image watermark embedding and extracting method first embodiment of the present application;

[0045] Figure 5 The watermark feature generation schematic diagram of the image watermark embedding and extracting method second embodiment of the present application;

[0046] Figure 6 The watermark feature analysis schematic diagram of the image watermark embedding and extracting method third embodiment of the present application;

[0047] Figure 7 The function module schematic diagram of the image watermark embedding and extracting device first embodiment of the present application.

[0048] The implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0049] It should be understood that the specific embodiments described herein are merely intended to explain the present application and are not intended to limit the present application.

[0050] Reference Figure 1 , Figure 1 The image watermark embedding and extracting device structure schematic diagram of the hardware running environment involved in the embodiment scheme of the present application.

[0051] As Figure 1 shown, the image watermark embedding and extracting device can include a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 can include a display screen, an input unit such as a keyboard, and can also include a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a wireless fidelity (WIreless-FIdelity, WI-FI) interface). The memory 1005 can be a high-speed random access memory (RAM) memory, or a stable non-volatile memory (Non-Volatile Memory, NVM), such as a magnetic disk memory. The memory 1005 can also be a storage device independent of the aforementioned processor 1001.

[0052] Those skilled in the art can understand that Figure 1 The structure shown in the figure does not constitute a limitation on the image watermark embedding and extracting device, and can include more or fewer components than the figure, or combine certain components, or different component arrangements.

[0053] As Figure 1 As shown, the memory 1005 as a storage medium can include an operating system, a data storage module, a network communication module, a user interface module, and an image watermark embedding and extracting program.

[0054] In Figure 1 In the image watermark embedding and extracting device shown in the figure, the network interface 1004 is mainly used for data communication with other devices; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the image watermark embedding and extracting device of the present application can be arranged in the image watermark embedding and extracting device, and the image watermark embedding and extracting device calls the image watermark embedding and extracting program stored in the memory 1005 through the processor 1001, and executes the image watermark embedding and extracting method provided by the embodiments of the present application.

[0055] The embodiments of the present application provide an image watermark embedding and extracting method, which refers to Figure 2 , Figure 2 The flowchart of a first embodiment of an image watermark embedding and extracting method of the present application.

[0056] In this embodiment, the image watermark embedding and extracting method comprises:

[0057] Step S10: obtaining first watermark information and an original picture, the original picture being a picture to be embedded with a watermark;

[0058] Step S20: converting the original picture and the first watermark information into original picture frequency domain features and two-dimensional watermark features with redundant information, wherein the redundant information is evenly spread on the two-dimensional watermark features;

[0059] Step S30: forward inputting the two-dimensional watermark features and the original picture frequency domain features into an invertible neural network, and outputting a watermark picture and loss information;

[0060] Further, the step S30 comprises:

[0061] Step S31: forward inputting the two-dimensional watermark features and the original picture frequency domain features into an invertible neural network, and outputting watermark picture frequency domain features and loss information;

[0062] Step S32: converting the watermark picture frequency domain features into a watermark picture through inverse discrete wavelet transform.

[0063] Reversible neural network shows strong embedding and extracting information ability in the field of image steganography, but image steganography does not consider the case that the image is attacked, and because the design idea of reversible neural network is to provide a bidirectional lossless information transmission, it does not have the ability to resist external noise. Therefore, the digital image watermark embedding and extraction is decomposed into two sub-problems of watermark feature generation and analysis through SENet (compression and excitation network) and watermark feature embedding and extraction through reversible neural network. After simplifying the problem, by generating watermark features with redundancy and error correction ability, it is beneficial to better design and adjust the neural network structure, and finally make the application realize high robustness.

[0064] Specifically, the deep neural network of the present application is trained on 10000 COCO training dataset pictures, and the generalization of the model is ensured by evaluating the neural network through 5000 COCO test dataset pictures. The embodiment is realized by PyTorch and runs on NVIDIA A100. Refer to Figure 3 , Figure 3 is the reversible neural network architecture diagram of the first embodiment of the image watermark embedding and extraction method of the present application. The original picture is converted into original picture frequency domain features after DWT, and the first watermark information is discretely wavelet transformed to generate two-dimensional watermark features with the same height and width as the original picture frequency domain features. The two-dimensional watermark features and the original picture frequency domain features are used as the input of the forward process of the reversible neural network. The reversible neural network includes N reversible neural network blocks, and N is an adjustable hyperparameter. In the forward process, that is, the watermark feature embedding process, refer to Figure 4 , Figure 4 is the internal processing diagram of the reversible neural network of the first embodiment of the image watermark embedding and extraction method of the present application. represents the input of the original picture features in the i-th reversible neural network block, represents the input of the two-dimensional watermark features in the i-th reversible neural network block, and After passing through the reversible neural network block, the output is and While and will be used as the input of the next reversible neural network block, and the final output After IWT, it is our watermark picture, and The calculation formula is as follows:

[0065]

[0066]

[0067] In the reversible neural network of the embodiment, φ, ρ, η are calculated using a dense block. Finally, the reversible neural network outputs watermark picture frequency domain features and loss information, and the watermark picture frequency domain features are converted into a watermark picture through inverse wavelet transform (IWT).

[0068] Step S40: converting the watermark picture and the loss information into watermark picture frequency domain features and a matrix, the matrix being a matrix consistent with the dimension of the loss information generated through Gaussian sampling;

[0069] Step S50: inputting the watermark picture frequency domain features and the matrix into the reversible neural network in reverse, and outputting second watermark information;

[0070] Specifically, the watermark picture is converted into watermark picture frequency domain features through discrete wavelet transform, and then enters the reversible neural network as input in the forward process, and outputs the second watermark information extracted from the watermark picture. In the reverse process, and The calculation formula is as follows:

[0071]

[0072]

[0073] The image watermark embedding and extraction method of the present application has two goals, one is that the second watermark information recovered from the watermark picture and the first watermark information are as same as possible, and the other is that the watermark picture and the original picture are as same as possible. Therefore, in order to keep the dimensions of input and output consistent, in the reverse process, in addition to the watermark picture frequency domain features, there is also a matrix consistent with the dimension of the loss information, which is generated through Gaussian sampling. Let x cover represent the original picture, x watermark represent the first watermark information, x steg represent the watermark picture, x watemark_rev represent the second watermark information recovered from the watermark picture, then the loss function can be represented as

[0074] L = λ g MSE(x cover , x watermark ) + λ r MSE(x watermark , X watermark_rev )

[0075] Wherein, MSE is the mean square error function, λ g and λ r are adjustable hyperparameters.

[0076] Step S60: comparing the first watermark information with the second watermark information to obtain a similarity verification result. The similarity verification result is used to determine the copyright ownership of the original picture.

[0077] Specifically, there are two indexes for evaluating the image watermark embedding and extracting method. One is the robustness of the watermark, which can be obtained by comparing the difference between the first watermark information and the second watermark information, i.e. the bit error rate (BER). The other is the quality of the generated watermark image, i.e. the noise image, which can be obtained by calculating the peak signal-to-noise ratio (PSNR) of the original image and the noise image.

[0078] In order to test the robustness of the image watermark embedding and extracting method, the bit error rate of the image watermark embedding and extracting method of the present application under six different image noises (Identity, Cropout(0.3), Dropout(0.3), Crop(0.035), Gassian Filter(2), JPEG(50)) is tested, and the method is compared with Hidden[1] and MBRS[6]. The details of the six image noises are as follows:

[0079] Identity: no processing is performed on the watermark image.

[0080] Cropout(0.3): a rectangular region in the watermark image is randomly selected, and the pixels outside the rectangular region are replaced with the pixels at the corresponding positions in the original image, and the size of the rectangular region is 30% of the size of the watermark image.

[0081] Dropout(0.3): 70% of the probability of each pixel point in the watermark image is replaced with the pixel point at the corresponding position in the original watermark image.

[0082] Crop(0.035): a rectangular region in the watermark image is randomly selected, and the pixel values outside the rectangular region are set to 0, and the size of the rectangular region is 3.5% of the size of the watermark image.

[0083] Gassian Filter(2): Gaussian filtering is performed on the watermark image, and the value of σ in the Gaussian filtering function is 2.

[0084] JPEG(50): JPEG compression is performed on the watermark image, and the quality factor in the JPEG compression algorithm is 50.

[0085] The following table is the comparison result of the three methods under the condition that the length of the watermark message is 30 bits:

[0086] PSNR ldentity Cropout Dropout Crop GF JPEG Hidden 33.5 0％ 6％ 7％ 12％ 4％ 37％ MBRS 33.5 0％ 0.0027％ 0.0087％ 4.15％ 0.011％ 4.48％ 0urs 34.8 0％ 0.002％ 0.0013％ 3.74％ 0％ 0％

[0087] From the experimental results, it can be seen that under the premise of higher image quality than the first two methods, the bit error rate of the digital image watermark embedding and extraction of the application is lower on the six image noises, and it shows stronger robustness.

[0088] In this embodiment, compared with other deep learning digital watermark methods directly watermarking in the spatial domain, the application embeds and extracts the watermark in the original picture frequency domain, which realizes better image quality. On this basis, the scheme further designs a reversible neural network model responsible for embedding and extracting two-dimensional watermark features in the original picture frequency domain, realizing a high-robustness image watermark embedding and extraction method combined with discrete wavelet transform, watermark message processing, and reversible neural network, which overcomes the non-differentiable noise to ensure image quality. The robustness of the watermark generated by the neural network to classic image noise such as compression transcoding, picture cropping, and Gaussian noise is the best.

[0089] Further, based on the above embodiment, the second embodiment of the application is provided, and the step S20 comprises:

[0090] Step S21: based on the compression and excitation network, the first watermark information is converted into a two-dimensional watermark feature with redundant information, and the first watermark information is one-dimensional watermark information;

[0091] Further, step S21 comprises:

[0092] Step S22: increase the length of the first watermark information through a fully connected layer;

[0093] Step S23: convert the first watermark information into two-dimensional watermark information, and the two-dimensional watermark information is a single channel;

[0094] Step S24: based on the compression and excitation network, the two-dimensional watermark information is converted into a two-dimensional watermark feature with redundant information, and the two-dimensional watermark feature is a multi-channel.

[0095] Further, step S24 comprises:

[0096] Step S25: based on the compression and excitation network, the two-dimensional watermark information is expanded to generate the two-dimensional watermark feature.

[0097] Specifically, the generation of the watermark feature refers to Figure 5 , Figure 5The schematic diagram of the watermark feature generation of the second embodiment of the image watermark embedding and extraction method of the present application is shown. Based on the SENet (compression and excitation network), after receiving the one-dimensional first watermark information, first, the length of the first watermark information is expanded through a fully connected layer, then the one-dimensional first watermark information is changed into a two-dimensional single-channel watermark feature, and finally the SENet generates a two-dimensional watermark feature with the same size and scale as the original picture. The generated watermark feature has redundant information (more channel quantity), so that the watermark feature has robustness against various noises after being embedded into the watermark picture.

[0098] Step S26: converting the original picture into the original picture frequency domain feature through forward discrete wavelet transform, wherein the embedding of the watermark through the frequency domain can improve the image quality of the watermark picture.

[0099] Specifically, compared with other deep learning digital watermark methods that directly embed watermarks in the spatial domain, the image watermark embedding and extraction method of the present application uses discrete wavelet transform (DWT) to change the picture. The discrete wavelet transform can convert the image from the spatial domain to the frequency domain, and the embedding of the watermark through the frequency domain can improve the image quality of the watermark picture. In this embodiment, the discrete wavelet transform is realized by Haar wavelet, and the dimension of the image of (B, C, H, W) is changed to (B, 4C, H / 2, W / 2) after DWT, wherein B is the batch size, C is the channel number, H is the picture height, and W is the picture width. IWT is the inverse process of DWT.

[0100] In this embodiment, compared with other deep learning digital watermark methods that directly embed watermarks in the spatial domain, the present application uses discrete wavelet transform to change the original picture, and embeds and extracts the watermark in the frequency domain of the original picture, thereby achieving better image quality. And the one-dimensional first watermark information is processed into a two-dimensional watermark feature with redundant information by SENet, which gets rid of the artificial design of the coding method.

[0101] Further, based on the above embodiments, referring to Figure 6 , the third embodiment of the present application is provided, and the step S40 comprises:

[0102] Step S41: inputting the watermark picture frequency domain feature and the matrix to the invertible neural network in reverse, and outputting the second watermark feature and the restored picture frequency domain information;

[0103] Step S42: converting the second watermark feature into the second watermark information based on the compression and excitation network.

[0104] Specifically, the invertible neural network outputs the second watermark feature and the restored picture frequency domain information. Referring to Figure 6 , Figure 6The figure is a schematic diagram for analyzing the watermark features of the third embodiment of the image watermark embedding and extraction method of the present application. After receiving the two-dimensional second watermark features, the second watermark features are first converted into single-channel features by the SENet, and then converted into one-dimensional watermark information. Finally, a fully connected layer is used to restore the second watermark information to be consistent with the length of the first watermark information.

[0105] In this embodiment, the SENet is used to analyze the watermark features, so as to extract the second watermark information. The technical effect of making the second watermark information as consistent as possible with the first watermark information is achieved.

[0106] Further, based on the above embodiments, the fourth embodiment of the present application is provided, which includes the following steps after step S30:

[0107] Step S33: adding noise to the watermark picture by a noise layer, the noise including non-differentiable noise and or differentiable noise. The noise layer adds the differentiable noise to the watermark picture by using a Standard ASL (standard attack simulation layer), or adds the non-differentiable noise to the watermark picture by using a Forward ASL (forward attack simulation layer).

[0108] It can be understood that the watermark picture may be attacked to become a noise picture during public use, and JPEG is the most commonly used lossy image compression method on the Internet at present. However, the learning of the deep neural network is achieved by gradient direction propagation, and the JPEG compression is non-differentiable, which causes the gradient transmitted to the encoder part of the deep neural network to be zero. At this time, the encoder cannot obtain the learning direction with enhanced robustness according to the feedback of the loss function.

[0109] Therefore, in this embodiment, a noise layer is set to simulate the noise attack in reality, and a kind of noise is randomly added to the watermark picture. For the differentiable noise, the Standard ASL (standard attack simulation layer) is used to add noise, that is, the noise is directly added to the watermark picture. For the non-differentiable noise, the training method proposed by the Forward ASL (forward attack simulation layer) is used. The implementation process includes: generating an initial noise picture by the Standard ASL (standard attack simulation layer) from the watermark picture; obtaining the difference between the watermark picture and the initial noise picture, so that the difference does not participate in the back propagation; adding the difference to the watermark picture to obtain a noise picture.

[0110] Specifically, the embodiment first makes the watermark picture pass through the attack simulation layer to generate a noise picture by Standard ASL, then calculates the residual of the noise picture and the watermark picture, residual = noise picture - watermark picture, then makes the residual not participate in back propagation, and then obtains the final noise picture = watermark picture + residual not participating in back propagation.

[0111] In the embodiment, compared with the existing deep learning training method that cannot truly replace natural noise with simulated noise, the application solves the training problem of simulated non-differentiable noise by combining the noise adding method provided by Forward ASL, thereby realizing the robustness of the watermark image in terms of classic image noise such as anti-compression transcoding, picture cropping, and Gaussian noise on the basis of ensuring image quality and adding a reversible neural network.

[0112] In addition, the embodiment of the application also provides an image watermark embedding and extracting device, which refers to Figure 7 , Figure 7 The figure is a functional module schematic diagram of the first embodiment of the image watermark embedding and extracting device of the application. The image watermark embedding and extracting device comprises:

[0113] The acquisition module 10 is configured to acquire first watermark information and an original picture, wherein the original picture is a picture to be embedded with a watermark.

[0114] The conversion module 20 is configured to convert the original picture and the first watermark information into original picture frequency domain features and two-dimensional watermark features with redundant information, wherein the redundant information is evenly spread on the two-dimensional watermark features.

[0115] The conversion module 20 is also configured to convert the watermark picture and the loss information into watermark picture frequency domain features and a matrix, wherein the matrix is a matrix with the same dimension as the loss information generated by Gaussian sampling.

[0116] The embedding module 30 is configured to input the two-dimensional watermark features and the original picture frequency domain features to the reversible neural network in a forward direction, and output a watermark picture and loss information.

[0117] The extraction module 40 is configured to input the watermark picture frequency domain features and the matrix to the reversible neural network in a reverse direction, and output second watermark information.

[0118] The comparison module 50 is configured to compare the first watermark information with the second watermark information to obtain a similarity verification result. The similarity verification result is used to determine the copyright ownership of the original picture.

[0119] In the image watermark embedding and extracting device of the application, the specific embodiments performed by each module are basically the same as those of the above-mentioned image watermark embedding and extracting method, and thus are not described here.

[0120] In addition, the embodiment of the present application further provides a readable storage medium.

[0121] The readable storage medium of the present application stores an image watermark embedding and extracting program, and the image watermark embedding and extracting program is executed by a processor to realize the steps of the image watermark embedding and extracting method.

[0122] In the embodiment, the image watermark embedding and extracting program stored in the readable storage medium of the present application is executed by the processor, and the specific implementation is basically the same as the above-mentioned image watermark embedding and extracting method, and is not described here.

[0123] It should be noted that in this paper, the term "including", "containing" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or system including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or system. Without more limitations, the element defined by the statement "including an image watermark embedding and extraction" does not exclude the presence of another same element in the process, method, article or system including the element.

[0124] The above-mentioned serial numbers of the embodiments of the present application are only for description, not representing the advantages and disadvantages of the embodiments.

[0125] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment method can be realized by software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a plurality of instructions to make a terminal device (which can be a mobile phone, computer, server, or network device, etc.) execute the method described in each embodiment of the present application.

[0126] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

Claims

1. An image watermark embedding and extracting method characterized by, The image watermark embedding and extracting method comprises the following steps: Obtain first watermark information and an original picture, the original picture being a picture to be embedded with a watermark; Increase the length of the first watermark information through a fully connected layer, the first watermark information being one-dimensional watermark information; Convert the first watermark information into two-dimensional watermark information, the two-dimensional watermark information being single-channel; Convert the two-dimensional watermark information into two-dimensional watermark features with redundant information based on a compression and excitation network, the two-dimensional watermark features being multi-channel, wherein the redundant information is evenly diffused on the two-dimensional watermark features; Convert the original picture into original picture frequency domain features through discrete wavelet transform, wherein embedding a watermark in the frequency domain can improve the image quality of the watermark picture; Forward input the two-dimensional watermark features and the original picture frequency domain features into a reversible neural network, and output a watermark picture and loss information, the reversible neural network model being used for embedding and extracting two-dimensional watermark features in the original picture frequency domain; Add noise to the watermark picture through a noise layer, the noise including non-differentiable noise and differentiable noise, wherein the noise layer adds the differentiable noise to the watermark picture using a standard attack simulation layer (Standard ASL) or adds the non-differentiable noise to the watermark picture using a forward attack simulation layer (Forward ASL); Convert the watermark picture and the loss information into watermark picture frequency domain features and a matrix, the matrix being a matrix with the same dimension as the loss information generated through Gaussian sampling; Reverse input the watermark picture frequency domain features and the matrix into the reversible neural network, and output second watermark information; Compare the first watermark information with the second watermark information to obtain a similarity verification result, wherein the similarity verification result is used to determine the copyright ownership of the original picture.

2. The image watermark embedding and extracting method of claim 1, wherein, The step of converting the two-dimensional watermark information into two-dimensional watermark features with redundant information based on the compression and excitation network comprises: Expanding the two-dimensional watermark information based on the compression and excitation network to generate the two-dimensional watermark features.

3. The image watermark embedding and extracting method of claim 1, wherein, The step of forward inputting the two-dimensional watermark features and the original picture frequency domain features into the reversible neural network and outputting a watermark picture and loss information comprises: Forward input the two-dimensional watermark features and the original picture frequency domain features into the reversible neural network, and output watermark picture frequency domain features and loss information; Convert the watermark picture frequency domain features into a watermark picture through inverse discrete wavelet transform.

4. The image watermark embedding and extracting method of claim 1, wherein, The step of reverse inputting the watermark picture frequency domain features and the matrix into the reversible neural network and outputting second watermark information comprises: Reverse input the watermark picture frequency domain features and the matrix into the reversible neural network, and output second watermark features and recovered picture frequency domain information; Convert the second watermark features into second watermark information based on the compression and excitation network.

5. An image watermark embedding and extracting apparatus characterized by comprising: The device comprises: An acquisition module configured to obtain first watermark information and an original picture, the original picture being a picture to be embedded with a watermark; The conversion module is configured to increase the length of the first watermark information by a full connection layer, the first watermark information being one-dimensional watermark information; The conversion module is further configured to convert the first watermark information into two-dimensional watermark information, the two-dimensional watermark information being single-channel; The conversion module is further configured to convert the two-dimensional watermark information into two-dimensional watermark features with redundant information based on a compression and excitation network, the two-dimensional watermark features being multi-channel, wherein the redundant information is evenly diffused on the two-dimensional watermark features; The conversion module is further configured to convert the original picture into original picture frequency domain features through discrete wavelet transform, wherein embedding the watermark in the frequency domain can improve the image quality of the watermark picture; The conversion module is further configured to convert the watermark picture and the loss information into watermark picture frequency domain features and a matrix, the matrix being a matrix with the same dimension as the loss information generated by Gaussian sampling; The embedding module is configured to input the two-dimensional watermark features and the original picture frequency domain features in a forward direction to a reversible neural network, and output a watermark picture and loss information, the reversible neural network model being used for embedding and extracting the two-dimensional watermark features in the original picture frequency domain; The embedding module is further configured to add noise to the watermark picture through a noise layer, the noise including non-differentiable noise and differentiable noise, wherein the noise layer adds the differentiable noise to the watermark picture by using a standard attack simulation layer (Standard ASL) or adds the non-differentiable noise to the watermark picture by using a forward attack simulation layer (Forward ASL); The extraction module is configured to input the watermark picture frequency domain features and the matrix in a reverse direction to a reversible neural network, and output second watermark information; The comparison module is configured to compare the first watermark information with the second watermark information to obtain a similarity verification result, wherein the similarity verification result is used to determine the copyright ownership of the original picture.

6. An image watermark embedding and extracting apparatus characterized by comprising: The device includes a memory, a processor, and an image watermark embedding and extracting program stored on the memory and executable on the processor, the image watermark embedding and extracting program being configured to implement the steps of the image watermark embedding and extracting method according to any one of claims 1 to 4.

7. A readable storage medium, characterized by, The readable storage medium stores an image watermark embedding and extracting program, the image watermark embedding and extracting program being executable on a processor to implement the steps of the image watermark embedding and extracting method according to any one of claims 1 to 4.

Images