A method, apparatus, device, and storage medium for hidden domain image enhancement.
By processing user images using a pre-trained bi-branch image steganography model and an enhancement model, embedding and extracting images, the problem of cloud-based user privacy images being attacked is solved, achieving privacy protection and visual consistency.
Patent Information
- Application Number
- CN202311162039.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-08
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-09-08
AI Technical Summary
In existing technologies, user privacy images are easily discovered by attackers when processed in the cloud, resulting in ineffective privacy protection. Furthermore, encrypted images have visual differences from the original images, making them easily detectable by attackers.
A pre-trained bi-branch image steganography model is used to embed the original image into the carrier image, and then processed by an image enhancement model to generate wavelet subbands of the enhanced image. The image steganography model is used to extract and reconstruct the target image from the enhanced image, ensuring that the privacy image is not obtained by the service provider during the enhancement process, and that it is indistinguishable from the ordinary image to the naked eye.
It achieves cloud-based protection of users' private images, preventing hackers from detecting them and launching targeted attacks, and ensuring privacy and visual consistency during image enhancement.
Smart Images

Figure CN117354428B_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of image processing technology, and in particular to a method, apparatus, device and storage medium for hidden domain image enhancement. Background Technology
[0002] With the increasing maturity of cloud computing technology, cloud-based multimedia services have developed rapidly. Users can upload images that need enhancement to cloud servers and utilize the computing power of the cloud to complete image processing. However, hackers or malicious service providers can easily access images uploaded to the cloud, compromising user privacy. To address this issue, existing technologies typically use homomorphic encryption on user images and then enhance the encrypted images to protect user privacy. However, encrypted images often differ significantly from the original images, making them easily detectable by attackers and vulnerable to attack. Therefore, how to protect user privacy without being detected by attackers and implement image enhancement processing in the cloud is a technical problem that needs to be solved. Summary of the Invention
[0003] To address the aforementioned technical problems, this disclosure provides a hidden domain image enhancement method, apparatus, device, and storage medium.
[0004] A first aspect of this disclosure provides a hidden domain image enhancement method, the method comprising:
[0005] Get the original image uploaded by the user;
[0006] The original image is embedded into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganographic image;
[0007] The steganographic image is enhanced using a pre-trained image enhancement model to obtain wavelet subbands of the enhanced image.
[0008] Based on the image steganography model, the enhanced target image corresponding to the original image is extracted and reconstructed from the wavelet subband of the enhanced image. The image steganography model is a two-branch network structure, with the forward branch used to perform the embedding operation and the backward branch used to perform the extraction operation. The forward branch and the backward branch use the same parameters. The image steganography model and the image enhancement model are obtained based on supervised learning training.
[0009] A second aspect of this disclosure provides a hidden domain image enhancement apparatus, the apparatus comprising:
[0010] The image acquisition module is used to acquire the original images uploaded by the user;
[0011] The embedding module is used to embed the original image into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganographic image.
[0012] The enhancement module is used to perform image enhancement processing on the stegana based on a pre-trained image enhancement model to obtain wavelet subbands of the enhanced image;
[0013] An extraction module is used to extract and reconstruct the enhanced target image corresponding to the original image from the wavelet subband of the enhanced image based on the image steganography model. The image steganography model is a two-branch network structure, with the forward branch used to perform the embedding operation and the backward branch used to perform the extraction operation. The forward branch and the backward branch use the same parameters. The image steganography model and the image enhancement model are obtained based on supervised learning training.
[0014] A third aspect of this disclosure provides a computer device including a memory and a processor, and a computer program, wherein the memory stores the computer program, and when the computer program is executed by the processor, it implements the hidden domain image enhancement method of the first aspect described above.
[0015] A fourth aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the hidden domain image enhancement method of the first aspect described above.
[0016] The technical solution provided in this disclosure has the following advantages compared with the prior art:
[0017] In the hidden domain image enhancement method, apparatus, device, and storage medium provided in this disclosure, an original image uploaded by a user is acquired, and the original image is embedded into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganographic image. The steganographic image is then enhanced using a pre-trained image enhancement model to obtain wavelet subbands of the enhanced image. Based on the image steganography model, the enhanced target image corresponding to the original image is extracted and reconstructed from the wavelet subbands of the enhanced image. The image steganography model is a two-branch network structure, with a forward branch for embedding operations and a backward branch for extraction operations. The forward and backward branches use the same parameters. The image steganography model and the image enhancement model are trained based on supervised learning, enabling enhancement of the steganographic image. This protects the user's privacy image from being obtained by the service provider during the enhancement process, while ensuring that the image uploaded to the cloud is indistinguishable from a normal image, preventing targeted attacks by hackers. Attached Figure Description
[0018] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.
[0019] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0020] Figure 1 This is a flowchart of a hidden domain image enhancement method provided in an embodiment of this disclosure;
[0021] Figure 2 This is a flowchart of an image steganography method provided in an embodiment of this disclosure;
[0022] Figure 3 This is a flowchart of an image enhancement method provided in an embodiment of this disclosure;
[0023] Figure 4 This is a flowchart illustrating a method for extracting a target image from an enhanced image, as provided in an embodiment of this disclosure.
[0024] Figure 5 This is a schematic diagram of the structure of an image steganography model and an image enhancement model provided in an embodiment of this disclosure;
[0025] Figure 6 This is a flowchart of a method for training an image steganography model provided in an embodiment of this disclosure;
[0026] Figure 7 This is a flowchart of a method for training an image enhancement model provided in an embodiment of this disclosure;
[0027] Figure 8 This is a schematic diagram of the structure of a hidden domain image enhancement device provided in an embodiment of this disclosure;
[0028] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. Detailed Implementation
[0029] To better understand the above-mentioned objectives, features, and advantages of this disclosure, the solutions disclosed herein will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.
[0030] Numerous specific details are set forth in the following description in order to provide a full understanding of this disclosure, but this disclosure may also be implemented in other ways different from those described herein; obviously, the embodiments in the specification are only some, and not all, of the embodiments of this disclosure.
[0031] It should be understood that the steps described in the method embodiments of this disclosure may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of this disclosure is not limited in this respect.
[0032] Figure 1 This is a flowchart of a hidden domain image enhancement method provided in an embodiment of this disclosure. This method can be executed by a hidden domain image enhancement apparatus. Figure 1 As shown, the hidden domain image enhancement method provided in this embodiment includes the following steps:
[0033] S101. Obtain the original image uploaded by the user.
[0034] The original image in this embodiment can be understood as a privacy image of the user to be enhanced.
[0035] In this embodiment of the disclosure, the hidden domain image enhancement device can obtain the original image uploaded by the user through a preset image upload interface.
[0036] S102. The original image is embedded into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganographic image.
[0037] The image steganography model in this embodiment can be understood as a pre-trained image processing model capable of steganographically writing secret images into a carrier image. The image steganography model is a two-branch network structure to achieve reversible steganography of the image. For example, the image steganography model can be based on a reversible steganography algorithm, such as the Difference Extension (DE) algorithm, or it can be based on a reversible neural network structure, such as the Lifting-based Invertible Hiding (LIH) network. No limitation is made here.
[0038] In this embodiment of the disclosure, the original image is a secret image obtained during the image steganography process.
[0039] The carrier image in this embodiment can be understood as the image used as a carrier when performing image steganography on a secret image. The carrier image and the steganographic image generated after steganography are not significantly different when observed with the naked eye.
[0040] In this embodiment of the disclosure, the hidden domain image enhancement device can input the user-uploaded steganalysis image and the pre-acquired carrier image into a pre-trained image steganalysis model to obtain the steganalysis image output by the image steganalysis model.
[0041] In one exemplary embodiment of the present disclosure, the hidden domain image enhancement device may, after receiving the original image uploaded by the user, randomly select an image from a preset carrier image database as the carrier image. The carrier image database may be established based on some publicly available images.
[0042] S103. Based on the pre-trained image enhancement model, perform image enhancement processing on the steganalysis image to obtain the wavelet subband of the enhanced image.
[0043] The image enhancement model in this embodiment can be understood as a pre-trained model for enhancing images. For example, the image enhancement model can be established based on a wavelet-based secure restoration (LSR) network. Image enhancement processing can include denoising, deblurring, and super-resolution processing, etc., without limitation.
[0044] In this embodiment of the present disclosure, the hidden domain image enhancement device can input the steganalyte image output by the image steganalysis model into the image enhancement model to obtain the wavelet subband of the enhanced image output by the image enhancement model.
[0045] S104. Based on the image steganography model, extract and reconstruct the enhanced target image corresponding to the original image from the wavelet subband of the enhanced image. The image steganography model is a two-branch network structure, with the forward branch used to perform the embedding operation and the backward branch used to perform the extraction operation. The forward branch and the backward branch use the same parameters. The image steganography model and the image enhancement model are obtained based on supervised learning training.
[0046] In this embodiment of the disclosure, the hidden domain image enhancement device can input the wavelet sub-band of the enhanced image into the image steganalysis model after obtaining it. The image steganalysis model extracts and reconstructs the enhanced image corresponding to the original image from the wavelet sub-band of the enhanced image, and outputs it as the target image. Specifically, the image steganalysis model is a two-branch network structure, including a forward branch and a backward branch. The forward branch is used to perform the embedding operation in S102, and the backward branch is used to perform the extraction operation in S104. The forward branch and the backward branch use the same parameters, which are determined during the training process of the image steganalysis model. The image steganalysis model and the image enhancement model are obtained based on supervised learning training.
[0047] In one exemplary embodiment of the present disclosure, the hidden domain image enhancement device can execute S101, S102 and S104 on the user's local electronic device to protect the privacy and security of the user's original image and the enhanced target image, and execute S103 on the cloud server to reduce the computing power requirements of the user's electronic device.
[0048] This embodiment of the disclosure obtains an original image uploaded by a user, embeds the original image into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganography image, performs image enhancement processing on the steganography image based on a pre-trained image enhancement model to obtain wavelet subbands of the enhanced image, and extracts and reconstructs the enhanced target image corresponding to the original image from the wavelet subbands of the enhanced image based on the image steganography model. The image steganography model is a two-branch network structure, with a forward branch for performing the embedding operation and a backward branch for performing the extraction operation. The forward and backward branches use the same parameters. The image steganography model and the image enhancement model are trained based on supervised learning, enabling them to enhance the steganography image, protecting the user's privacy image from being obtained by the service provider during the enhancement process, and ensuring that the image uploaded to the cloud is indistinguishable from a normal image, thus preventing targeted attacks by hackers.
[0049] Figure 2 This is a flowchart of an image steganography method provided in an embodiment of this disclosure, such as... Figure 2 As shown, based on the above embodiments, image steganography can be performed by the following method, wherein the image steganography model includes a first discrete wavelet transform layer, a first wavelet lifting layer, and a wavelet inverse transform layer.
[0050] S201. Input the original image and the carrier image into the first discrete wavelet transform layer to obtain the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image.
[0051] The first discrete wavelet transform layer in this embodiment can be understood as a processing layer that discretizes and samples the image through discrete wavelet transform (DWT), wherein the DWT transform is an invertible transform.
[0052] In this embodiment of the disclosure, the hidden domain image enhancement device can, after obtaining the original image and the carrier image, input them into the first discrete wavelet transform layer to obtain the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image output by the first discrete wavelet transform layer.
[0053] S202. Input the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image into the first wavelet lifting layer, and perform iterative prediction and update processing on the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image to obtain the wavelet sub-bands of the steganalysis image.
[0054] The first wavelet lifting layer in this embodiment can be understood as a processing layer that predicts and updates wavelet subbands through wavelet lifting (WL) transform, wherein the WL transform is a reversible transform.
[0055] In this embodiment of the present disclosure, the hidden domain image enhancement device can, after obtaining the wavelet sub-bands of the original image and the carrier image output by the first discrete wavelet transform layer, input them into the first wavelet lifting layer for iterative prediction and update processing to obtain the wavelet sub-bands of the output steganalysis image.
[0056] In one exemplary embodiment of the present disclosure, the first wavelet lifting layer may include multiple WL blocks. The hidden domain image enhancement device may input the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image into the first WL block and input the output result into the next WL block, and so on, until the last WL block outputs the prediction and update result and determines it as the wavelet sub-band of the steganalysis image. The number of WL blocks may be 24.
[0057] S203. Input the wavelet subband of the steganalysis image into the wavelet inverse transform layer to obtain the steganalysis image.
[0058] The wavelet inverse transform layer in this embodiment can be understood as a processing layer that restores wavelet subbands to an image through the inverse transform of DWT.
[0059] In this embodiment of the disclosure, the hidden domain image enhancement device can input the wavelet subband of the steganalyte image output by the first wavelet lifting layer into the wavelet inverse transform layer to obtain the steganalyte image output by the wavelet inverse transform layer.
[0060] This embodiment of the disclosure inputs the original image and the carrier image into the first discrete wavelet transform layer to obtain wavelet sub-bands of the original image and the carrier image. The wavelet sub-bands of the original image and the carrier image are then input into the first wavelet lifting layer. Iterative prediction and update processing is performed on the wavelet sub-bands of the original image and the carrier image to obtain the wavelet sub-bands of the steganalysis image. The wavelet sub-bands of the steganalysis image are then input into the inverse wavelet transform layer to obtain the steganalysis image. This method can naturally separate low-frequency and high-frequency details through wavelet lifting, facilitating subsequent image enhancement of high-frequency details and achieving better image enhancement results. Furthermore, by completing image steganography through reversible wavelet transform, the stability of the secret image during steganography and restoration can be maintained. Information can also be hidden in the high-frequency sub-bands of the wavelet domain, thus minimizing the difference between the steganalysis image and the carrier image when observed by the naked eye, achieving a better steganography effect.
[0061] In some embodiments of this disclosure, the first wavelet lifting layer includes a predictor and an updater. S202 specifically includes: inputting the wavelet sub-band of the carrier image into the predictor to obtain a first prediction result; iteratively generating a new wavelet sub-band of the original image based on the wavelet sub-band of the original image and the first prediction result; inputting the wavelet sub-band of the original image into the updater to obtain a first update result; iteratively generating a new wavelet sub-band of the carrier image based on the wavelet sub-band of the carrier image and the first update result; and determining the iterated wavelet sub-band of the carrier image as the wavelet sub-band of the steganographic image.
[0062] The predictor and updater can be based on residual networks, VGG networks, densely connected convolutional networks (DenseNet), or other neural networks; no restrictions are imposed here.
[0063] In some embodiments of this disclosure, the hidden domain image enhancement apparatus can, after obtaining the wavelet sub-bands of the original image and the carrier image output by the first discrete wavelet transform layer, input the wavelet sub-bands of the carrier image into a predictor to obtain a first prediction result output by the predictor. Based on the wavelet sub-bands of the original image and the first prediction result output by the predictor, iteratively generate new wavelet sub-bands of the original image. Input the wavelet sub-bands of the original image into an updater to obtain a first update result output by the updater. Based on the wavelet sub-bands of the carrier image and the first update result output by the updater, iteratively generate new wavelet sub-bands of the carrier image. The wavelet sub-bands of the carrier image generated after multiple prediction updates and multiple iterations are determined as the wavelet sub-bands of the steganalysis image. The specific process is as follows:
[0064]
[0065]
[0066] in, The wavelet subband represents the original image after i updates. The wavelet subband representing the carrier image after i updates, The wavelet subband representing the original image after i+1 iterations. The wavelet subband represents the carrier image after i+1 iterations, P represents the predictor, and U represents the generator.
[0067] Figure 3 This is a flowchart of an image enhancement method provided in an embodiment of this disclosure, such as... Figure 3 As shown, based on the above embodiments, image enhancement can be performed by the following method, wherein the image enhancement model includes a second discrete wavelet transform layer, a second wavelet lifting layer, an enhancement processing layer, and a first inverse wavelet lifting layer.
[0068] S301. Input the steganalytic image into the second discrete wavelet transform layer to obtain the first high-frequency sub-band and the first low-frequency sub-band.
[0069] The second discrete wavelet transform layer in this embodiment can be understood as a processing layer that discretizes and samples the image through DWT transform.
[0070] In this embodiment of the disclosure, the hidden domain image enhancement device can input the stegana image into the second discrete wavelet transform layer after obtaining the stegana image to obtain the first high-frequency sub-band and the first low-frequency sub-band output by the second discrete wavelet transform layer.
[0071] S302. Input the first high-frequency sub-band and the first low-frequency sub-band into the second wavelet lifting layer, and perform iterative prediction and update processing on the first high-frequency sub-band and the first low-frequency sub-band to obtain the second high-frequency sub-band and the second low-frequency sub-band.
[0072] The second wavelet lifting layer in this embodiment can be understood as a processing layer that predicts and updates wavelet subbands through WL transform.
[0073] In this embodiment of the present disclosure, the hidden domain image enhancement device can input the first high-frequency sub-band and the first low-frequency sub-band output by the second discrete wavelet transform layer into the second wavelet lifting layer for iterative prediction and update processing to obtain the output second high-frequency sub-band and the second low-frequency sub-band. The specific processing method is similar to S202 and will not be described in detail here.
[0074] In one exemplary embodiment of this disclosure, the number of WL blocks may be four.
[0075] S303. Input the second high-frequency sub-band into the enhancement processing layer to obtain the enhanced third high-frequency sub-band.
[0076] The enhancement processing layer in this embodiment can be understood as a processing layer for enhancing wavelet subbands of an image. For example, the enhancement processing layer may include an encoder and a decoder.
[0077] In one exemplary embodiment of this disclosure, the hidden domain image enhancement device can input the updated second high-frequency subband output by the second wavelet lifting layer into the enhancement processing layer, and the enhancement processing layer can enhance the second high-frequency subband during the encoding and decoding process to obtain the enhanced third high-frequency subband.
[0078] S304. Input the second low-frequency sub-band into the second discrete wavelet transform layer and perform the above processing procedure until the number of executions reaches the preset number, thereby obtaining the first target low-frequency sub-band and the enhanced first target high-frequency sub-band.
[0079] The preset number of times in this embodiment can be understood as the number of times the loop is executed in advance. During the loop execution, new high-frequency sub-bands and low-frequency sub-bands can be further separated from the low-frequency sub-bands, thereby enhancing the high-frequency sub-bands. The specific value of the preset number of times can be set as needed. The more preset numbers, the better the enhancement effect.
[0080] In this embodiment of the present disclosure, the hidden domain image enhancement device can input the updated second low-frequency sub-band output by the second wavelet lifting layer back into the second discrete wavelet transform layer and execute the prediction and update process in S302. The high-frequency sub-band output by the second discrete wavelet transform layer is input into the enhancement processing layer, and the enhancement process in S303 is executed. This process is repeated until the number of executions reaches a preset number, and the first target low-frequency sub-band output by the second discrete wavelet transform layer and the enhanced first target high-frequency sub-band output by the enhancement processing layer are obtained at the last execution.
[0081] S305. Input the first target low-frequency sub-band and the first target high-frequency sub-band into the first inverse wavelet lifting layer to obtain the second target low-frequency sub-band.
[0082] The inverse wavelet lifting layer in this embodiment can be understood as a processing layer that combines two wavelet subbands to obtain a new wavelet subband through WL inverse transform.
[0083] In this embodiment of the present disclosure, the hidden domain image enhancement device can, after obtaining the first target low-frequency sub-band and the first target high-frequency sub-band, input them into the first inverse wavelet lifting layer to obtain the newly generated wavelet sub-band, and determine it as the second target low-frequency sub-band.
[0084] S306. Input the second target low-frequency sub-band and the enhanced second target high-frequency sub-band generated in the previous execution into the first inverse wavelet lifting layer to obtain the third target low-frequency sub-band, until the enhanced target high-frequency sub-band generated in the first execution is input into the first inverse wavelet lifting layer to obtain the wavelet sub-band of the enhanced image.
[0085] In this embodiment of the present disclosure, after generating the second target low-frequency sub-band, the hidden domain image enhancement device can input it together with the enhanced second target high-frequency sub-band generated in the previous execution during the execution of S302-S304 into the first inverse wavelet lifting layer to obtain the third target low-frequency sub-band. The loop continues to be executed until the target low-frequency sub-band generated by the first inverse wavelet lifting layer is input together with the enhanced target high-frequency sub-band generated during the first execution of S303 into the first inverse wavelet lifting layer to obtain the output result, and the output result is determined as the wavelet sub-band of the enhanced image.
[0086] In this embodiment, the steganalyte is input into the second discrete wavelet transform layer to obtain a first high-frequency subband and a first low-frequency subband. The first high-frequency subband and the first low-frequency subband are then input into the second wavelet lifting layer. Iterative prediction and update processing is performed on the first high-frequency subband and the first low-frequency subband to obtain a second high-frequency subband and a second low-frequency subband. The second high-frequency subband is then input into the enhancement processing layer to obtain an enhanced third high-frequency subband. The second low-frequency subband is then input into the second discrete wavelet transform layer, and the above processing is repeated until a preset number of iterations is reached to obtain the first... The target low-frequency subband and the enhanced first target high-frequency subband are input into the first inverse wavelet lifting layer to obtain the second target low-frequency subband. The second target low-frequency subband and the enhanced second target high-frequency subband generated in the previous execution are then input into the first inverse wavelet lifting layer to obtain the third target low-frequency subband. This process continues until the enhanced target high-frequency subband generated in the first execution is input into the first inverse wavelet lifting layer to obtain the wavelet subband of the enhanced image. This method can restore high-frequency details lost during image degradation and improve the image enhancement effect.
[0087] In other embodiments of this disclosure, the image enhancement processing includes denoising, deblurring, and super-resolution processing. The enhancement processing layer includes a first encoder, a second encoder, a third encoder, a first decoder, a second decoder, and a third decoder constructed based on a nonlinear inactivation-free network. Before S102, the method further includes: obtaining the enhancement type of the image enhancement processing input by the user; S303 specifically includes: inputting the second high-frequency subband into the first encoder to obtain a first encoding result; if the enhancement type is denoising, then inputting the first encoding result into the first decoder to obtain an enhanced third high-frequency subband; if the enhancement type is not denoising, then inputting the first encoding result into the second encoder to obtain a second encoding result; if the enhancement type is deblurring, then inputting the second encoding result into the second decoder to obtain an enhanced third high-frequency subband; if the enhancement type is super-resolution processing, then inputting the second encoding result into the third encoder to obtain a third encoding result; and inputting the third encoding result into the third decoder to obtain an enhanced third high-frequency subband.
[0088] Among them, the nonlinear non-active network is a NAF network. The first encoder, the first decoder, the second decoder and the third decoder can form a U-Net structure with shortcuts. The first encoder includes a 3*3 convolutional layer, a NAF layer and a downsampling layer. The first decoder, the second decoder and the third decoder all include a 3*3 convolutional layer, a NAF layer and an upsampling layer. The second encoder and the third encoder include a NAF layer.
[0089] In other embodiments of this disclosure, the hidden domain image enhancement apparatus can first obtain the type of image enhancement processing the user wants to perform before steganography on the original image uploaded by the user. When the second high-frequency subband is input into the enhancement processing layer for enhancement processing, the second high-frequency subband is input into the first encoder to obtain the first encoding result. The subsequent operation is determined according to the type of image enhancement processing input by the user. If the type is denoising processing, the first encoding result is input into the first decoder for decoding, and the decoding result is determined as the enhanced third high-frequency subband. If the type is deblurring processing and super-resolution processing, the first encoding result is input into the second encoder to obtain the second encoding result. For deblurring processing, the second encoding result is input into the second decoder for decoding, and the decoding result is determined as the enhanced third high-frequency subband. For super-resolution processing, the second encoding result is input into the third encoder to obtain the third encoding result, and the third encoding result is input into the third decoder for decoding, and the decoding result is determined as the enhanced third high-frequency subband.
[0090] Figure 4 This is a flowchart illustrating a method for extracting a target image from an enhanced image, as provided in an embodiment of this disclosure. Figure 4 As shown, based on the above embodiments, the target image can be extracted from the enhanced image using the following method.
[0091] S401. Input the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image into the first wavelet lifting layer, and perform iterative prediction and update processing on the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image to obtain the wavelet sub-bands and additional information of the steganographic image.
[0092] The additional information in this disclosure can be understood as information used to store content lost during the steganography process.
[0093] In this embodiment of the present disclosure, the hidden domain image enhancement device can obtain the wavelet sub-band of the steganalysis image output by the first wavelet lifting layer and additional information when the wavelet sub-band of the original image and the wavelet sub-band of the carrier image are input into the first wavelet lifting layer for iterative prediction and update processing. Specifically, after the iteration is completed, the wavelet sub-band of the newly generated carrier image can be determined as the wavelet sub-band of the steganalysis image, and the wavelet sub-band of the newly generated original image can be determined as additional information.
[0094] S402. Determine auxiliary variables that have the same normal distribution as the additional information.
[0095] In this embodiment of the disclosure, since the additional information output by the first wavelet lifting layer cannot be directly applied to the second inverse wavelet lifting layer, the hidden domain image enhancement device can, after obtaining the additional information, randomly select a variable as an auxiliary variable from the variables that have the same normal distribution as the additional information.
[0096] S403. Input the wavelet sub-band of the enhanced image and the auxiliary variable into the second inverse wavelet lifting layer to obtain the wavelet sub-band of the target image.
[0097] In this embodiment of the disclosure, the hidden domain image enhancement device can input the wavelet subband and auxiliary variables of the enhanced image into the second inverse wavelet lifting layer after obtaining them to obtain the newly generated wavelet subband and determine it as the wavelet subband of the target image. The specific method is similar to S305 and will not be described in detail here.
[0098] S404. Input the wavelet subband of the target image into the wavelet inverse transform layer to obtain the target image.
[0099] In this embodiment of the disclosure, the hidden domain image enhancement device can input the wavelet subband of the target image output by the second inverse wavelet lifting layer into the wavelet inverse transform layer to obtain the target image output by the wavelet inverse transform layer. The specific method is similar to S203 and will not be described in detail here.
[0100] This embodiment of the disclosure embeds the original image into a pre-acquired carrier image using a pre-trained image steganography model to obtain a steganographic image and additional information. An auxiliary variable with the same normal distribution as the additional information is determined. The wavelet sub-band of the enhanced image and the auxiliary variable are input into a second inverse wavelet lifting layer to obtain the wavelet sub-band of the target image. The wavelet sub-band of the target image is then input into the inverse wavelet transform layer to obtain the target image. This method can extract the enhanced target image corresponding to the user's original image from the enhanced image, minimizing the loss of image content during the enhancement process while enhancing the image's intensity.
[0101] Figure 5 This is a schematic diagram of the structure of an image steganography model and an image enhancement model provided in the embodiments of this disclosure. The following is a combination of... Figure 5 The image enhancement methods will be further elaborated.
[0102] The hidden domain image enhancement device inputs the original image X into the image steganalysis model LIH. secret and carrier image X cover The first discrete wavelet transform layer performs DWT transform to obtain the wavelet subband x of the original image. secret and the wavelet subband x of the carrier image cover The data is then input into the first wavelet lifting layer, where multiple WL Blocks in the first wavelet lifting layer perform iterative prediction and update processing (only one WL Block is shown in the figure) to obtain the wavelet subband x of the steganalyte image. stego And additional information R, the wavelet subband x of the steganographic image stego The input wavelet inverse transform layer is used for DWT transform to obtain the steganalytic image X. stego Then the steganographic image X stego The second discrete wavelet transform layer in the input image enhancement model LSR performs DWT transform to obtain the first high-frequency subband H1 and the first low-frequency subband L1. These are then input into multiple WL blocks in the second wavelet lifting layer for iterative prediction and update processing (only one WL block is shown in the figure) to obtain the second high-frequency subband H1. 11 Second low frequency subband L 11 The second high-frequency subband H 11 The input enhancement processing layer is encoded by the first encoder ENC1. Depending on the enhancement type selected by the user, it is determined whether to pass through the second encoder ENC2 and the third encoder ENC3. After decoding by the first decoder DEC1, the second decoder DEC2, or the third decoder DEC3, the enhanced third high-frequency subband H1' is obtained. The second low-frequency subband L... 11The second discrete wavelet transform layer is then input for DWT transformation to obtain the high-frequency subband H2 and the low-frequency subband L2. Multiple WL blocks are then input for iterative prediction and update processing to obtain the high-frequency subband H. 21 and low-frequency subband L 21 , will the high-frequency subband H 21 The input enhancement processing layer, after passing through ENC1, ENC2, ENC3, and DEC, yields the enhanced high-frequency subband H2'. This process is repeated until the preset number of executions is reached (twice in the diagram). The first target low-frequency subband L generated during the last execution is then processed. 21 The first target high-frequency subband H2' is input into multiple WL Blocks in the first inverse wavelet lifting layer to obtain the output second target low-frequency subband L1'. L1' is then input together with the enhanced second target high-frequency subband H1' generated in the previous execution into multiple WL Blocks in the first inverse wavelet lifting layer. This process continues until the enhanced target high-frequency subband H1' generated in the first execution (since the execution count in the figure is twice, both the second target high-frequency subband and the target high-frequency subband are the third high-frequency subband H1') is input into multiple WL Blocks in the first inverse wavelet lifting layer to obtain the wavelet subband x of the enhanced image. stego Based on the additional information R, determine the auxiliary variable Z that has the same normal distribution, and further compare Z with x. stego The second inverse wavelet lifting layer in the LIH steganography model of the input image, after processing by multiple WL blocks, yields the wavelet subband x of the target image. secret 'and the wavelet subband x of the restored carrier image cover The target image X is obtained by inputting it into the inverse wavelet transform layer and performing DWT transform. secret The 'and restored carrier image Xcover' enables the entire process of image steganography, steganography-based image enhancement, and steganographic image restoration. It is worth noting that the DWT processing and WL Block processing are reversible, while the encoding and decoding processes are unidirectional.
[0103] Figure 6 This is a flowchart of a method for training an image steganography model provided in an embodiment of this disclosure, such as... Figure 6 As shown, based on the above embodiments, the image steganography model can be trained using the following method.
[0104] S601. Obtain a set of user training images and carrier training images for training, wherein the user training images include a first training image and a second training image with the same image content, and the image intensity of the first training image is less than the image intensity of the second training image.
[0105] In this embodiment of the disclosure, the first training image may be a low-quality image with lower intensity generated based on the second training image. For example, the first training image may be an image generated by adding noise, blurring, and reducing the resolution of the second training image.
[0106] In this embodiment of the disclosure, the hidden domain image enhancement device can acquire a set of user training images and carrier training images for training. The user training images and carrier training images may have a corresponding relationship or may not be related. This is not limited here. One second training image in the user training images may correspond to at least one first training image.
[0107] S602. Input the first training image and the carrier training image into the forward branch of a preset image steganalysis network to obtain steganalysis training images and additional training information.
[0108] The additional training information in this embodiment can be understood as information used to store content lost during the steganography process.
[0109] In this embodiment of the present disclosure, the hidden domain image enhancement device can input the first training image and the carrier training image from the user training image into the forward branch of a preset image steganalysis network, and the image steganalysis network will steg the first training image into the carrier training image to obtain the steganalysis training image and additional training information.
[0110] S603. Determine auxiliary training variables that have the same normal distribution as the additional training information.
[0111] In this embodiment of the present disclosure, the hidden domain image enhancement device may, after obtaining additional training information, randomly select a variable from variables that have the same normal distribution as the additional training information as an auxiliary training variable.
[0112] S604. Input the steganalysis training image and the auxiliary training variables into the reverse branch of the image steganalysis network to obtain the recovered training image.
[0113] In this embodiment of the disclosure, the hidden domain image enhancement device can, after obtaining the steganalytic training image and auxiliary training variables, input them into the reverse branch of the image steganalytic network to obtain the output restored training image.
[0114] S605. Based on the steganalysis training image and the carrier training image, calculate the first steganalysis loss.
[0115] The first steganalysis loss in this embodiment can be understood as a loss introduced to ensure the stealth performance of the image steganalysis model. Specifically, it can be understood as a loss introduced to ensure that the carrier image before and after embedding the secret image is small during the steganalysis process.
[0116] In this embodiment of the disclosure, the hidden domain image enhancement device can, after obtaining the steganalysis training image, compare the carrier training image with the steganalysis training image, and obtain a first steganalysis loss based on the comparison result. Specifically, the calculation method of the first steganalysis loss can be expressed as follows:
[0117]
[0118] Among them, L con Let l represent the first steganalysis loss, N represent the number of training samples, and l2 represent the mean squared error function. This represents the nth steganalysis training image. This represents the nth training image.
[0119] S606. Based on the first training image and the recovered training image, calculate the second steganalysis loss.
[0120] The second steganalysis loss in this embodiment can be understood as a loss introduced to ensure the revealing performance of the image steganalysis model. Specifically, it can be understood as a loss introduced to ensure that the difference between the secret image and the recovered image after steganalysis and extraction is small during the steganalysis process.
[0121] In this embodiment of the disclosure, the hidden domain image enhancement device can, after obtaining the recovered training image, compare the recovered training image with the first training image, and obtain a second steganalytic loss based on the comparison result. Specifically, the calculation method of the second steganalytic loss can be expressed as follows:
[0122]
[0123] Among them, L rev Let z represent the second steganalysis loss, and z ~ p(z) represent the auxiliary training variables that follow a normal distribution of the additional training information. This represents the nth image used for reconstructing the training image.
[0124] S607. Perform discrete wavelet transform on the carrier training image and the steganalysis training image respectively to obtain the low-frequency sub-band of the carrier training image and the low-frequency sub-band of the steganalysis training image.
[0125] In this embodiment of the disclosure, the hidden domain image enhancement device can perform DWT transformation on the carrier training image and the steganalysis training image respectively to obtain the low-frequency subband of the carrier training image and the low-frequency subband of the steganalysis training image.
[0126] S608. Based on the low-frequency sub-band of the carrier training image and the low-frequency sub-band of the steganalysis training image, the third steganalysis loss is calculated.
[0127] The third steganalysis loss in this embodiment can be understood as a loss introduced to ensure the stealth performance of the image steganalysis model. Specifically, it can be understood as a loss introduced to ensure that the difference between the low-frequency part of the carrier image before and after embedding the secret image is small during the steganalysis process. Since the secret image is hidden in the high-frequency sub-band of the carrier image during the steganalysis process, the low-frequency sub-band should be as similar as possible before and after steganalysis.
[0128] In this embodiment of the disclosure, the hidden domain image enhancement device can compare the low-frequency subbands of the carrier training image and the steganalysis training image after obtaining them, and obtain a third steganalysis loss based on the comparison result. Specifically, the calculation method of the third steganalysis loss can be expressed as follows:
[0129]
[0130] Among them, L freq H represents the third steganalysis loss, and H represents the DWT transform. The low-frequency subband representing the training images of the carrier. This represents the low-frequency subband of the steganography training image.
[0131] S609. Based on the first steganalysis loss, the second steganalysis loss, the third steganalysis loss, and the preset steganalysis loss weights, the parameters of the image steganalysis network are adjusted to obtain the trained image steganalysis model.
[0132] The steganalysis loss weight in this embodiment can be understood as a pre-set parameter that measures the importance of each steganalysis loss, and its specific value can be set as needed.
[0133] In this embodiment of the disclosure, the hidden domain image enhancement device can, after calculating the first steganalysis loss, the second steganalysis loss, and the third steganalysis loss, determine the overall steganalysis loss by combining it with preset steganalysis loss weights. Then, based on this overall steganalysis loss, the parameters of the image steganalysis network are adjusted to minimize the overall steganalysis loss. Finally, an image steganalysis model is obtained based on the parameter-adjusted image steganalysis network. The overall steganalysis loss can be expressed as follows:
[0134] L hid (Θ H )=λ C L con +λ R L rev +λ F L freq
[0135] Among them, L hid (Θ H ) represents the overall steganalytic loss, λ C , λ R and λF This represents the weight of the steganography loss.
[0136] This embodiment of the disclosure acquires a set of user training images and carrier training images for training. The user training images include a first training image and a second training image with the same image content. The image intensity of the first training image is less than that of the second training image. The first training image and the carrier training image are input into the forward branch of a preset image steganalysis network to obtain steganalysis training images and additional training information. An auxiliary training variable with the same normal distribution as the additional training information is determined. The steganalysis training images and the auxiliary training variable are input into the backward branch of the image steganalysis network to obtain a recovered training image. Based on the steganalysis training images and the carrier training images, a first steganalysis loss is calculated. Based on the first training image and the recovered training image, a second steganalysis loss is calculated. The steganalysis loss is calculated by performing discrete wavelet transform on both the carrier training image and the steganalysis training image to obtain the low-frequency sub-bands of the carrier training image and the steganalysis training image, respectively. Based on these low-frequency sub-bands, a third steganalysis loss is calculated. The parameters of the image steganalysis network are adjusted based on the first, second, and third steganalysis losses and preset steganalysis loss weights to obtain a trained image steganalysis model. This model can train an image steganalysis model with good concealment and revelation performance, reducing the difference between the user image after steganography and restoration and the original user image. It also reduces the difference between the carrier image before and after steganography, lowering the probability of attackers launching targeted attacks after discovering the image is a steganalysis image, thus protecting user privacy.
[0137] Figure 7 This is a flowchart illustrating a method for training an image enhancement model according to an embodiment of this disclosure. Figure 7 As shown, based on the above embodiments, the image enhancement model can be trained using the following method.
[0138] S701. After the image steganography model is trained, the steganography training image is input into a preset image enhancement network and subjected to denoising, deblurring and super-resolution processing respectively to obtain the wavelet sub-band of the first enhanced training image, the wavelet sub-band of the second enhanced training image and the wavelet sub-band of the third enhanced training image.
[0139] In this embodiment of the disclosure, the hidden domain image enhancement device can freeze the parameters in the image steganalysis model after the image steganalysis model has been trained, and train the image enhancement model based on the steganalysis training image output by the image steganalysis model. Specifically, the steganalysis training image can be input into a preset image enhancement network and subjected to denoising, deblurring and super-resolution processing respectively to obtain the wavelet sub-band of the first enhanced training image after denoising, the wavelet sub-band of the second enhanced training image after deblurring and the wavelet sub-band of the third enhanced training image after super-resolution processing.
[0140] In one exemplary embodiment of this disclosure, the hidden domain image enhancement device can obtain three steganalysis training images generated by the image steganalysis model based on three different first training images corresponding to the same second training image. The three different first training images corresponding to the same second training image can be images generated by adding noise, blurring, and reducing resolution, respectively, from the second training image. The hidden domain image enhancement device can determine the type of enhancement processing to be performed on the first training image according to the processing method when generating the first training image, and then perform the corresponding type of enhancement processing to obtain the wavelet subband of the first enhanced training image, the wavelet subband of the second enhanced training image, and the wavelet subband of the third enhanced training image.
[0141] S702. Input the wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image, together with the auxiliary training variables, into the image steganography model to obtain the enhanced first target training image, the second target training image, and the third target training image corresponding to the first training image.
[0142] In this embodiment of the present disclosure, the hidden domain image enhancement device can, after obtaining the wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image, input the wavelet sub-band of the first enhanced training image and auxiliary training variables into the image steganalysis model. The image steganalysis model extracts and reconstructs the enhanced first target training image corresponding to the first training image. The wavelet sub-band of the second enhanced training image and auxiliary training variables are then input into the image steganalysis model to obtain the output enhanced second target training image. The wavelet sub-band of the third enhanced training image and auxiliary training variables are then input into the image steganalysis model to obtain the output enhanced third target training image.
[0143] S703. Perform inverse wavelet transform on the wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image, respectively, to obtain the first enhanced training image, the second enhanced training image, and the third enhanced training image.
[0144] In this embodiment of the present disclosure, the hidden domain image enhancement device can perform inverse DWT transformation on each wavelet sub-band after obtaining the wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image.
[0145] S704. Based on the first target training image, the second target training image, the third target training image, and the second training image, calculate the first enhancement loss, the second enhancement loss, and the third enhancement loss respectively.
[0146] The first enhancement loss, the second enhancement loss, and the third enhancement loss in the embodiments of this disclosure can be understood as losses introduced to ensure that the difference between the enhanced secret image and the real high-quality image is small.
[0147] In this embodiment of the disclosure, the hidden domain image enhancement device can, after obtaining the first target training image, the second target training image, and the third target training image, compare them respectively with the corresponding second training image, and obtain a first enhancement loss, a second enhancement loss, and a third enhancement loss based on the comparison results. Specifically, the calculation methods for the first enhancement loss, the second enhancement loss, and the third enhancement loss can be expressed as follows:
[0148]
[0149]
[0150]
[0151] Among them, L N1 L represents the first enhanced loss. B1 L represents the second enhanced loss. s1 Represents the third enhanced loss, The image represents the first target training image. The training image represents the second objective. The training image represents the third objective. This represents the corresponding second training image.
[0152] S705. Based on the first enhanced training image, the second enhanced training image, and the third enhanced training image and the carrier training image, the fourth enhancement loss, the fifth enhancement loss, and the sixth enhancement loss are calculated respectively.
[0153] The fourth, fifth, and sixth enhancement losses in the embodiments of this disclosure can be understood as losses introduced to ensure that the difference between the enhanced steganographic image and the original carrier image is small.
[0154] In this embodiment of the disclosure, the hidden domain image enhancement device can, after obtaining the first enhanced training image, the second enhanced training image, and the third enhanced training image, compare them with their respective original carrier training images, and calculate the fourth enhancement loss, the fifth enhancement loss, and the sixth enhancement loss based on the comparison results. Specifically, the calculation methods for the fourth enhancement loss, the fifth enhancement loss, and the sixth enhancement loss can be expressed as follows:
[0155]
[0156]
[0157]
[0158] Among them, L N2 L represents the fourth enhanced loss. B2 L represents the fifth enhanced loss. S2 Represents the sixth enhanced loss, Represents the first augmented training image. Represents the second enhanced training image. Represents the third augmented training image. Representative training images for the carrier.
[0159] S706. Based on the first enhancement loss, the second enhancement loss, the third enhancement loss, the fourth enhancement loss, the fifth enhancement loss, the sixth enhancement loss, and the preset enhancement loss weights, the parameters of the image enhancement network are adjusted to obtain the trained image enhancement model.
[0160] In this embodiment of the disclosure, the hidden domain image enhancement device can calculate the first enhancement loss, the second enhancement loss, the third enhancement loss, the fourth enhancement loss, the fifth enhancement loss, and the sixth enhancement loss, and then determine the overall enhancement loss by combining it with preset enhancement loss weights. Based on this overall enhancement loss, the parameters of the image enhancement network are adjusted to minimize the overall enhancement loss. An image enhancement model is then obtained based on the parameter-adjusted image enhancement network. The overall enhancement loss can be expressed as follows:
[0161] L hid (Θ R )=λ res (λ N1 L N1 +λ B1 L B1 +λ S1 L S1 )
[0162] +λ stego (λ N2 L N2 +λ B2 LB2 +λ S2 L S2 )
[0163] Among them, L hid (Θ R ) represents the overall enhancement loss, λ res , λ stego , λ N1 , λ B1 , λ S1 , λ N2 , λ B2 and λ S2 This represents an enhanced loss weight.
[0164] In this embodiment, after the image steganalysis model is trained, the steganalysis training images are input into a preset image augmentation network for denoising, deblurring, and super-resolution processing, respectively, to obtain wavelet sub-bands of a first augmented training image, a second augmented training image, and a third augmented training image. These wavelet sub-bands, along with auxiliary training variables, are then input into the image steganalysis model to obtain augmented first, second, and third target training images corresponding to the first training image. Inverse wavelet transforms are then performed on the wavelet sub-bands of the first, second, and third augmented training images, respectively. Given the first, second, and third enhanced training images, and the second training image, respectively, a first enhancement loss, a second enhancement loss, and a third enhancement loss are calculated based on the first, second, and third target training images and the second training image. Based on the first, second, and third enhanced training images and the carrier training image, a fourth enhancement loss, a fifth enhancement loss, and a sixth enhancement loss are calculated. Based on the first, second, third, fourth, fifth, and sixth enhancement losses and preset enhancement loss weights, the parameters of the image enhancement network are adjusted to obtain the trained image enhancement model. This model can produce an image enhancement model with good image enhancement performance and minimal content difference between the images before and after enhancement.
[0165] Figure 8 This is a schematic diagram of the structure of a hidden domain image enhancement device provided in an embodiment of this disclosure. Figure 8As shown, the hidden domain image enhancement device 800 includes: an image acquisition module 810, an embedding module 820, an enhancement module 830, and an extraction module 840. The image acquisition module 810 acquires an original image uploaded by a user. The embedding module 820 embeds the original image into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganographic image. The enhancement module 830 performs image enhancement processing on the steganographic image based on a pre-trained image enhancement model to obtain a wavelet subband of the enhanced image. The extraction module 840 extracts and reconstructs the enhanced target image corresponding to the original image from the wavelet subband of the enhanced image based on the image steganography model. The image steganography model is a two-branch network structure, with a forward branch for performing the embedding operation and a backward branch for performing the extraction operation. The forward branch and the backward branch use the same parameters. The image steganography model and the image enhancement model are trained based on supervised learning.
[0166] Optionally, the image steganography model includes a first discrete wavelet transform layer, a first wavelet lifting layer, and an inverse wavelet transform layer. The embedding module 820 includes: a first wavelet transform unit, used to input the original image and the carrier image into the first discrete wavelet transform layer to obtain wavelet sub-bands of the original image and wavelet sub-bands of the carrier image; a first wavelet lifting unit, used to input the wavelet sub-bands of the original image and the carrier image into the first wavelet lifting layer, and perform iterative prediction and update processing on the wavelet sub-bands of the original image and the carrier image to obtain wavelet sub-bands of the steganography image; and a first inverse transform unit, used to input the wavelet sub-bands of the steganography image into the inverse wavelet transform layer to obtain the steganography image.
[0167] Optionally, the first wavelet lifting layer includes a predictor and an updater. The first wavelet lifting unit includes: a prediction subunit, used to input the wavelet sub-band of the carrier image into the predictor to obtain a first prediction result; a first generation subunit, used to iteratively generate a new wavelet sub-band of the original image based on the wavelet sub-band of the original image and the first prediction result; an update subunit, used to input the wavelet sub-band of the original image into the updater to obtain a first update result; a second generation subunit, used to iteratively generate a new wavelet sub-band of the carrier image based on the wavelet sub-band of the carrier image and the first update result; and a determination subunit, used to determine the wavelet sub-band of the carrier image after iteration as the wavelet sub-band of the steganographic image.
[0168] Optionally, the image enhancement model includes a second discrete wavelet transform layer, a second wavelet lifting layer, an enhancement processing layer, and a first inverse wavelet lifting layer. The enhancement module 830 includes: a second wavelet transform unit, used to input the steganalyte image into the second discrete wavelet transform layer to obtain a first high-frequency sub-band and a first low-frequency sub-band; a second wavelet lifting unit, used to input the first high-frequency sub-band and the first low-frequency sub-band into the second wavelet lifting layer, and perform iterative prediction and update processing on the first high-frequency sub-band and the first low-frequency sub-band to obtain a second high-frequency sub-band and a second low-frequency sub-band; an enhancement unit, used to input the second high-frequency sub-band into the enhancement processing layer to obtain an enhanced third high-frequency sub-band; and a first execution unit, used for... The second low-frequency sub-band is input into the second discrete wavelet transform layer, and the above processing procedure is executed until the number of executions reaches a preset number, to obtain the first target low-frequency sub-band and the enhanced first target high-frequency sub-band; the first inverse wavelet lifting unit is used to input the first target low-frequency sub-band and the first target high-frequency sub-band into the first inverse wavelet lifting layer to obtain the second target low-frequency sub-band; the second execution unit is used to input the second target low-frequency sub-band and the enhanced second target high-frequency sub-band generated in the previous execution into the first inverse wavelet lifting layer to obtain the third target low-frequency sub-band, until the enhanced target high-frequency sub-band generated in the first execution is input into the first inverse wavelet lifting layer to obtain the wavelet sub-band of the enhanced image.
[0169] Optionally, the image enhancement processing includes denoising, deblurring, and super-resolution processing. The hidden domain image enhancement device further includes: a type acquisition module for acquiring the enhancement type of the image enhancement processing input by the user; the enhancement processing layer includes a first encoder, a second encoder, a third encoder, a first decoder, a second decoder, and a third decoder constructed based on a nonlinear inactivation-free network; the enhancement unit includes: a first encoding subunit for inputting the second high-frequency subband into the first encoder to obtain a first encoding result; and a first decoding subunit for inputting the first encoding result into the first encoder if the enhancement type is denoising. A first decoder obtains the enhanced third high-frequency subband; a second encoding subunit is configured to input the first encoding result into the second encoder to obtain a second encoding result if the enhancement type is not denoising processing; a second decoding subunit is configured to input the second encoding result into the second decoder to obtain the enhanced third high-frequency subband if the enhancement type is deblurring processing; a third encoding subunit is configured to input the second encoding result into the third encoder to obtain a third encoding result if the enhancement type is super-resolution processing; and a third decoding subunit is configured to input the third encoding result into the third decoder to obtain the enhanced third high-frequency subband.
[0170] Optionally, the first wavelet lifting unit is specifically used to: input the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image into the first wavelet lifting layer, and perform iterative prediction and update processing on the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image to obtain the wavelet sub-bands of the steganographic image and additional information; the image steganography model further includes a second inverse wavelet lifting layer, and the extraction module 840 includes: a first determining unit, used to determine an auxiliary variable that has the same normal distribution as the additional information; a second inverse wavelet lifting unit, used to input the wavelet sub-bands of the enhanced image and the auxiliary variable into the second inverse wavelet lifting layer to obtain the wavelet sub-bands of the target image; and a second inverse transform unit, used to input the wavelet sub-bands of the target image into the inverse wavelet transform layer to obtain the target image.
[0171] Optionally, the hidden domain image enhancement device further includes an image steganalysis model training module. The image steganalysis model training module includes: a training data acquisition unit, used to acquire a set of user training images and carrier training images for training, wherein the user training images include a first training image and a second training image with the same image content, and the image intensity of the first training image is less than that of the second training image; a steganalysis training unit, used to input the first training image and the carrier training image into the forward branch of a preset image steganalysis network to obtain steganalysis training images and additional training information; a second determination unit, used to determine auxiliary training variables that have the same normal distribution as the additional training information; and a recovery training unit, used to input the steganalysis training images and the auxiliary training variables into the reverse branch of the image steganalysis network to obtain a recovery training image. The system comprises: a first steganalysis loss calculation unit for calculating a first steganalysis loss based on the steganalysis training image and the carrier training image; a second steganalysis loss calculation unit for calculating a second steganalysis loss based on the first training image and the recovered training image; a third wavelet transform unit for performing discrete wavelet transforms on the carrier training image and the steganalysis training image respectively to obtain low-frequency subbands of the carrier training image and the steganalysis training image; a third steganalysis loss calculation unit for calculating a third steganalysis loss based on the low-frequency subbands of the carrier training image and the steganalysis training image; and a first parameter adjustment unit for adjusting the parameters of the image steganalysis network based on the first steganalysis loss, the second steganalysis loss, the third steganalysis loss, and preset steganalysis loss weights to obtain a trained image steganalysis model.
[0172] Optionally, the hidden domain image enhancement device further includes an image enhancement model training module, which includes: an enhancement training unit, used to input the steganalysis training image into a preset image enhancement network after the image steganalysis model has been trained, and perform denoising, deblurring, and super-resolution processing respectively to obtain wavelet sub-bands of a first enhanced training image, a second enhanced training image, and a third enhanced training image; an extraction unit, used to input the wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image, together with the auxiliary training variables, into the image steganalysis model to obtain the enhanced first target training image, the second target training image, and the third target training image corresponding to the first training image; and a third inverse transform unit, used to process the wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image... Wavelet subbands and wavelet subbands of the third enhanced training image are subjected to inverse wavelet transforms to obtain the first enhanced training image, the second enhanced training image, and the third enhanced training image; a first enhancement loss calculation unit is used to calculate the first enhancement loss, the second enhancement loss, and the third enhancement loss based on the first target training image, the second target training image, the third target training image, and the second training image, respectively; a second enhancement loss calculation unit is used to calculate the fourth enhancement loss, the fifth enhancement loss, and the sixth enhancement loss based on the first enhanced training image, the second enhanced training image, the third enhanced training image, and the third enhanced training image and the carrier training image, respectively; a second parameter adjustment unit is used to adjust the parameters of the image enhancement network based on the first enhancement loss, the second enhancement loss, the third enhancement loss, the fourth enhancement loss, the fifth enhancement loss, the sixth enhancement loss, and preset enhancement loss weights to obtain the trained image enhancement model.
[0173] The hidden domain image enhancement device provided in this embodiment can perform the method described in any of the above embodiments, and its execution method and beneficial effects are similar, so they will not be described again here.
[0174] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure.
[0175] like Figure 9 As shown, the computer device may include a processor 910 and a memory 920 storing computer program instructions.
[0176] Specifically, the processor 910 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.
[0177] Memory 920 may include a large-capacity storage for information or instructions. For example, and not limitingly, memory 920 may include a hard disk drive (HDD), a floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 920 may include removable or non-removable (or fixed) media. Where appropriate, memory 920 may be internal or external to the integrated gateway device. In a particular embodiment, memory 920 is a non-volatile solid-state memory. In a particular embodiment, memory 920 includes read-only memory (ROM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (Electrically Programmable ROM, EPROM), an electrically erasable programmable PROM (EEPROM), an electrically alterable ROM (EAROM), or flash memory, or a combination of two or more of these.
[0178] The processor 910 reads and executes computer program instructions stored in the memory 920 to perform the steps of the hidden domain image enhancement method provided in the embodiments of this disclosure.
[0179] In one example, the computer device may also include a transceiver 930 and a bus 940. Wherein, as... Figure 9 As shown, the processor 910, memory 920 and transceiver 930 are connected via bus 940 and communicate with each other.
[0180] Bus 940 may include hardware, software, or both. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industrial Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a MicroChannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, bus 940 may include one or more buses. Although specific buses are described and illustrated in the embodiments of this application, this application considers any suitable bus or interconnection.
[0181] This disclosure also provides a computer-readable storage medium that can store a computer program that, when executed by a processor, enables the processor to implement the hidden domain image enhancement method provided in this disclosure.
[0182] The aforementioned storage medium may, for example, include a memory 920 containing computer program instructions, which can be executed by the processor 910 of the hidden domain image enhancement device to complete the hidden domain image enhancement method provided in the embodiments of this disclosure. Optionally, the storage medium may be a non-transitory computer-readable storage medium, such as a ROM, random access memory (RAM), compact disc ROM (CD-ROM), magnetic tape, floppy disk, and optical data storage device. The aforementioned computer program may be written in any combination of one or more programming languages to perform the operations of the embodiments of this disclosure. The programming languages include object-oriented programming languages such as Java and C++, as well as conventional procedural programming languages such as C or similar languages. The program code may be executed entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.
[0183] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0184] The above description is merely a specific embodiment of this disclosure, enabling those skilled in the art to understand or implement it. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method of hiding domain image enhancement, characterized by, include: Get the original image uploaded by the user; The original image is embedded into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganographic image; The steganographic image is enhanced using a pre-trained image enhancement model to obtain wavelet subbands of the enhanced image. Based on the image steganography model, the enhanced target image corresponding to the original image is extracted and reconstructed from the wavelet subband of the enhanced image. The image steganography model is a two-branch network structure, with the forward branch used to perform the embedding operation and the backward branch used to perform the extraction operation. The forward branch and the backward branch use the same parameters. The image steganography model and the image enhancement model are obtained based on supervised learning training. The image enhancement model includes a second discrete wavelet transform layer, a second wavelet lifting layer, an enhancement processing layer, and a first inverse wavelet lifting layer. The pre-trained image enhancement model performs image enhancement processing on the stegana to obtain wavelet subbands of the enhanced image, including: The steg image is input into the second discrete wavelet transform layer to obtain the first high-frequency sub-band and the first low-frequency sub-band. The first high-frequency sub-band and the first low-frequency sub-band are input into the second wavelet lifting layer, and the first high-frequency sub-band and the first low-frequency sub-band are iteratively predicted and updated to obtain the second high-frequency sub-band and the second low-frequency sub-band. The second high-frequency subband is input into the enhancement processing layer to obtain the enhanced third high-frequency subband; The second low-frequency sub-band is input into the second discrete wavelet transform layer, and the above prediction and update process is executed until the number of executions reaches the preset number, so as to obtain the first target low-frequency sub-band and the enhanced first target high-frequency sub-band. The first target low-frequency sub-band and the first target high-frequency sub-band are input into the first inverse wavelet lifting layer to obtain the second target low-frequency sub-band. The second target low-frequency sub-band and the enhanced second target high-frequency sub-band generated in the previous execution are input into the first inverse wavelet lifting layer to obtain the third target low-frequency sub-band, until the enhanced target high-frequency sub-band generated in the first execution is input into the first inverse wavelet lifting layer to obtain the wavelet sub-band of the enhanced image.
2. The method of claim 1, wherein, The image steganography model includes a first discrete wavelet transform layer, a first wavelet lifting layer, and an inverse wavelet transform layer. The pre-trained image steganography model embeds the original image into a pre-acquired carrier image to obtain a steganographic image, including: The original image and the carrier image are input into the first discrete wavelet transform layer to obtain wavelet sub-bands of the original image and wavelet sub-bands of the carrier image. The wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image are input into the first wavelet lifting layer, and the wavelet sub-bands of the original image and the carrier image are iteratively predicted and updated to obtain the wavelet sub-bands of the steganographic image. The wavelet subband of the steganalysis image is input into the inverse wavelet transform layer to obtain the steganalysis image.
3. The method of claim 2, wherein, The first wavelet lifting layer includes a predictor and an updater. The step of inputting the wavelet sub-bands of the original image and the carrier image into the first wavelet lifting layer, and performing iterative prediction and update processing on the wavelet sub-bands of the original image and the carrier image to obtain the wavelet sub-bands of the steganographic image includes: The wavelet subband of the carrier image is input into the predictor to obtain the first prediction result; Based on the wavelet sub-bands of the original image and the first prediction result, a new wavelet sub-band of the original image is iteratively generated; The wavelet subband of the original image is input into the updater to obtain the first update result; Based on the wavelet sub-bands of the carrier image and the first update result, a new wavelet sub-band of the carrier image is iteratively generated; The wavelet sub-band of the carrier image after iteration is determined as the wavelet sub-band of the steganalysis image.
4. The method of claim 1, wherein, The image enhancement processing includes denoising, deblurring, and super-resolution processing. Before the pre-trained image steganography model embeds the original image into a pre-acquired carrier image to obtain the steganographic image, the method further includes: Obtain the enhancement type for image enhancement processing as input by the user; The enhancement processing layer includes a first encoder, a second encoder, a third encoder, a first decoder, a second decoder, and a third decoder constructed based on a nonlinear inactivation-free network. The step of inputting the second high-frequency subband into the enhancement processing layer to obtain the enhanced third high-frequency subband includes: The second high-frequency sub-band is input into the first encoder to obtain the first encoding result; If the enhancement type is noise reduction processing, then the first encoding result is input into the first decoder to obtain the enhanced third high-frequency subband; If the enhancement type is not denoising, then the first encoding result is input into the second encoder to obtain the second encoding result; If the enhancement type is deblurring, the second encoding result is input into the second decoder to obtain the enhanced third high-frequency subband; If the enhancement type is super-resolution processing, then the second encoding result is input into the third encoder to obtain the third encoding result; The third encoding result is input into the third decoder to obtain the enhanced third high-frequency subband.
5. The method of claim 2, wherein, The step of inputting the wavelet sub-bands of the original image and the carrier image into the first wavelet lifting layer, and performing iterative prediction and update processing on the wavelet sub-bands of the original image and the carrier image to obtain the wavelet sub-bands of the steganographic image includes: The wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image are input into the first wavelet lifting layer. Iterative prediction and update processing is performed on the wavelet sub-bands of the original image and the wavelet sub-bands of the carrier image to obtain the wavelet sub-bands and additional information of the steganographic image. The image steganography model further includes a second inverse wavelet lifting layer. The step of extracting and reconstructing the enhanced target image corresponding to the original image from the wavelet subbands of the enhanced image based on the image steganography model includes: Identify auxiliary variables that have the same normal distribution as the additional information; The wavelet sub-bands of the enhanced image and the auxiliary variables are input into the second inverse wavelet lifting layer to obtain the wavelet sub-bands of the target image; The wavelet subband of the target image is input into the inverse wavelet transform layer to obtain the target image.
6. The method according to any one of claims 1-5, characterized in that, The image steganography model is trained based on the following steps: A set of user training images and carrier training images are obtained for training. The user training images include a first training image and a second training image with the same image content, wherein the image intensity of the first training image is less than that of the second training image. The first training image and the carrier training image are input into the forward branch of a preset image steganalysis network to obtain steganalysis training images and additional training information. Identify auxiliary training variables that have the same normal distribution as the additional training information; The steganalytic training image and the auxiliary training variables are input into the reverse branch of the image steganalytic network to obtain the recovered training image; Based on the steganalysis training image and the carrier training image, the first steganalysis loss is calculated; The second steganalysis loss is calculated based on the first training image and the recovered training image; Discrete wavelet transform is performed on the carrier training image and the steganalysis training image respectively to obtain the low-frequency sub-band of the carrier training image and the low-frequency sub-band of the steganalysis training image; The third steganalysis loss is calculated based on the low-frequency subband of the carrier training image and the low-frequency subband of the steganalysis training image. Based on the first steganalysis loss, the second steganalysis loss, the third steganalysis loss, and the preset steganalysis loss weights, the parameters of the image steganalysis network are adjusted to obtain the trained image steganalysis model.
7. The method according to claim 6, characterized in that, The image enhancement model is trained based on the following steps: After the image steganalysis model is trained, the steganalysis training image is input into a preset image enhancement network and subjected to denoising, deblurring and super-resolution processing respectively to obtain the wavelet sub-bands of the first enhanced training image, the second enhanced training image and the third enhanced training image. The wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image are input together with the auxiliary training variables into the image steganalysis model to obtain the enhanced first target training image, the second target training image, and the third target training image corresponding to the first training image. Perform inverse wavelet transform on the wavelet sub-bands of the first enhanced training image, the second enhanced training image, and the third enhanced training image, respectively, to obtain the first enhanced training image, the second enhanced training image, and the third enhanced training image. Based on the first target training image, the second target training image, the third target training image, and the second training image, the first enhancement loss, the second enhancement loss, and the third enhancement loss are calculated respectively. Based on the first enhanced training image, the second enhanced training image, and the third enhanced training image and the carrier training image, the fourth enhancement loss, the fifth enhancement loss, and the sixth enhancement loss are calculated respectively. Based on the first enhancement loss, the second enhancement loss, the third enhancement loss, the fourth enhancement loss, the fifth enhancement loss, the sixth enhancement loss, and the preset enhancement loss weights, the parameters of the image enhancement network are adjusted to obtain the trained image enhancement model.
8. A hidden domain image enhancement device, characterized in that, include: The image acquisition module is used to acquire the original images uploaded by the user; The embedding module is used to embed the original image into a pre-acquired carrier image based on a pre-trained image steganography model to obtain a steganographic image. The enhancement module is used to perform image enhancement processing on the stegana based on a pre-trained image enhancement model to obtain wavelet subbands of the enhanced image; An extraction module is used to extract and reconstruct the enhanced target image corresponding to the original image from the wavelet subband of the enhanced image based on the image steganography model. The image steganography model is a two-branch network structure, with the forward branch used to perform the embedding operation and the backward branch used to perform the extraction operation. The forward branch and the backward branch use the same parameters. The image steganography model and the image enhancement model are obtained based on supervised learning training. The image enhancement model includes a second discrete wavelet transform layer, a second wavelet lifting layer, an enhancement processing layer, and a first inverse wavelet lifting layer. The pre-trained image enhancement model performs image enhancement processing on the stegana to obtain wavelet subbands of the enhanced image, including: The steg image is input into the second discrete wavelet transform layer to obtain the first high-frequency sub-band and the first low-frequency sub-band. The first high-frequency sub-band and the first low-frequency sub-band are input into the second wavelet lifting layer, and the first high-frequency sub-band and the first low-frequency sub-band are iteratively predicted and updated to obtain the second high-frequency sub-band and the second low-frequency sub-band. The second high-frequency subband is input into the enhancement processing layer to obtain the enhanced third high-frequency subband; The second low-frequency sub-band is input into the second discrete wavelet transform layer, and the above prediction and update process is executed until the number of executions reaches the preset number, so as to obtain the first target low-frequency sub-band and the enhanced first target high-frequency sub-band. The first target low-frequency sub-band and the first target high-frequency sub-band are input into the first inverse wavelet lifting layer to obtain the second target low-frequency sub-band. The second target low-frequency sub-band and the enhanced second target high-frequency sub-band generated in the previous execution are input into the first inverse wavelet lifting layer to obtain the third target low-frequency sub-band, until the enhanced target high-frequency sub-band generated in the first execution is input into the first inverse wavelet lifting layer to obtain the wavelet sub-band of the enhanced image.
9. A computer device, characterized in that, include: Memory; processor; And a computer program; wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method as described in any one of claims 1-7.
10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the hidden domain image enhancement method as described in any one of claims 1-7.
