A method and system for reversible image information hiding based on convolutional neural networks

Abstract

This application provides a method and system for reversible information hiding of images based on convolutional neural networks, relating to the field of reversible information hiding technology. This method, during reversible information hiding, divides the original image into independent branch images and point images, predicts the branch images and point images separately, and generates a reversible information-hiding image based on the difference between the original image and the predicted image. Furthermore, during the prediction process, image features of different receptive fields are extracted and channel-wise superimposed, which can more fully utilize the correlation between pixels, improve the prediction accuracy of target pixels, and thus significantly increase the embedding capacity.

Description

Technical Field

[0001] This application relates to the field of reversible information hiding technology, and in particular to an image reversible information hiding method and system based on convolutional neural networks. Background Technology

[0002] The statements in this section are merely background information related to this application and do not necessarily constitute prior art.

[0003] Reversible information hiding can embed information into the original image and recover the original image and information without loss after decryption and extraction. Based on these two advantages, reversible information hiding is widely used in military, medical, and super-resolution processing fields.

[0004] Currently, there are many methods for reversible information hiding (RDH). One type mainly improves the embedding method, such as difference expansion, histogram shifting, and error prediction expansion. Another type mainly improves the prediction method by continuously improving the prediction performance to enhance the prediction accuracy, such as difference predictors, median edge direction predictors, gradient adaptive predictors, and bilinear interpolation predictors.

[0005] The inventors discovered that in the process of reversible information hiding, traditional pixel prediction methods use linear methods and only utilize one or a few neighboring pixels to predict the target pixel of the image. They cannot utilize the correlation of more neighboring pixels, resulting in low prediction accuracy and thus a large space for reversible information hiding. Summary of the Invention

[0006] To address the issues of low target pixel prediction accuracy and low embedding capacity in existing methods, this application provides an image reversible information hiding method and system based on convolutional neural networks.

[0007] The specific technical solution adopted is as follows:

[0008] In a first aspect, embodiments of this application provide an image reversible information hiding method based on a convolutional neural network, comprising:

[0009] The original image is divided into mutually independent cross set images and point set images;

[0010] The image features of different receptive fields of the cross set image are extracted using a point set image predictor and channel superposition is performed to obtain a predicted point set image; a first difference information between the point set image and the predicted point set image is determined, secret information is embedded into the first difference information, and added to the predicted point set image to obtain a first secret-carrying image.

[0011] Image features of different receptive fields of the first secret image are extracted using a cross set image predictor and channels are superimposed to obtain a predicted cross set image; a second difference information between the cross set image and the predicted cross set image is determined, secret information is embedded into the second difference information, and the second secret image is obtained by adding it to the predicted cross set image.

[0012] The first hidden image and the second hidden image are fused to obtain a reversible information hiding image.

[0013] In one possible implementation, the point set image predictor and the fork set image predictor have the same structure and both employ convolutional neural networks; image features of the target image are extracted using convolutional kernels of different sizes of the convolutional neural network, and the receptive field is expanded according to the superposition and combination of image features; the target image includes fork set images and point set images.

[0014] In one possible implementation, the convolutional neural network includes a backbone prediction network and an auxiliary prediction network. In the backbone prediction network, the target image is processed by a convolutional network with an Inception structure, which outputs image features with receptive fields of various sizes. The image features are then stacked in channel dimension, and the target pixels are predicted using the image features in different channels to obtain the backbone prediction image. In the auxiliary prediction network, the target image is processed by a convolutional kernel of a specific size to predict the target pixels, and the dimensions are reduced to obtain the auxiliary prediction image. The predicted image is determined based on the backbone prediction image and the auxiliary prediction image.

[0015] In one possible implementation, a channel attention mechanism is added after the backbone prediction network overlays the image features by channel dimension to adjust the weights of the image features on different channels.

[0016] In one possible implementation, the channel attention mechanism includes ECA-Net; ECA-Net uses non-dimensionality-reduced GAP to aggregate convolutional features and then adaptively selects a one-dimensional convolutional kernel size to learn channel attention.

[0017] In one possible implementation, the secret information is embedded into the difference information, which includes first interpolation information and second difference information, in the following manner:

[0018] Use the pixel v in the region i,j To predict target pixel u i,j The predicted pixel can be represented as: u' i,j =CNN(v i,j );

[0019] By predicting pixel u' i,j and u i,j The difference between the two is d. i,j:d i,j =u' i,j -u i,j ;

[0020] Information embedding process: D i,j =2d i,j +b;

[0021] Among them, D i,j For the expanded prediction error, b represents the payload to be embedded, and then D i,j The changes are made using the histogram translation method; after information embedding, u i,j Calculated as U i,j :U i,j =D i,j +u' i,j ;

[0022] Using the predicted value u' i,j and the modified pixel value U i,j The modified prediction error value is calculated by the following formula: D i,j =U i,j -u' i,j ;

[0023] The value of the secret information is: b = D i,j mod2;

[0024] The value of the original pixel is:

[0025] In one possible implementation, information is extracted from the reversibly hidden image in the following manner:

[0026] The reversible information hiding image is divided into mutually independent secret point set images and secret fork set images; based on the secret point set images and fork set image predictors, the predicted secret fork set images are obtained; a second secret difference image is determined between the secret fork set images and the predicted secret fork set images, and secret information and fork set images are extracted from the second secret difference image;

[0027] Based on the fork set image and the point set image predictor, obtain the predicted secret-carrying point set image; determine the first secret-carrying difference image between the secret-carrying point set image and the predicted secret-carrying point set image, and extract the secret information and the point set image from the first secret-carrying difference image;

[0028] The original image is obtained by fusing the fork set image and the point set image.

[0029] Secondly, embodiments of this application provide an image reversible information hiding system based on a convolutional neural network, comprising:

[0030] The image segmentation module is used to divide the original image into mutually independent cross set images and point set images;

[0031] The first secret image acquisition module is used to extract image features of different receptive fields of the cross set image using a point set image predictor, and perform channel superposition to obtain a predicted point set image; determine the first difference information between the point set image and the predicted point set image, embed secret information into the first difference information, and add it to the predicted point set image to obtain the first secret image.

[0032] The second secret image acquisition module uses a cross set image predictor to extract image features of different receptive fields of the first secret image and performs channel superposition to obtain a predicted cross set image; determines the second difference information between the cross set image and the predicted cross set image, embeds the secret information into the second difference information, and adds it to the predicted cross set image to obtain the second secret image;

[0033] The fusion module is used to fuse the first hidden image and the second hidden image to obtain a reversible hidden image.

[0034] Thirdly, embodiments of this application provide a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the image reversible information hiding method based on convolutional neural networks as described in the first aspect and any possible implementation of the first aspect.

[0035] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the image reversible information hiding method based on a convolutional neural network as described in the first aspect and any possible implementation thereof.

[0036] The beneficial effects of this application are:

[0037] 1. In performing reversible information hiding, this application divides the original image into independent fork set images and point set images, and predicts the fork set images and point set images respectively. Based on the difference information between the original image and the predicted image, a reversible information hiding image is generated. In addition, during the prediction process, image features of different receptive fields of the image are extracted and channels are superimposed. This can make fuller use of the correlation between pixels, improve the prediction accuracy of the target pixel, and thus greatly improve the embedding capacity.

[0038] 2. Under the same image, the error prediction through the convolutional deep network not only expands the range of the receptive field in space and makes better use of the correlation between pixels, but also adds an attention mechanism to the channels based on the convolutional neural network to enhance the correlation between pixels between channels, improve the prediction performance of the convolutional neural network, and increase the information embedding capacity. Attached Figure Description

[0039] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0040] Figure 1 This is a flowchart illustrating the image reversible information hiding method based on a convolutional neural network provided in an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the information embedding process provided in the embodiments of this application;

[0042] Figure 3 This is a schematic diagram of the network structure of the convolutional neural network predictor provided in the embodiments of this application;

[0043] Figure 4 This is a schematic diagram of the ECA-Net network structure provided in the embodiments of this application;

[0044] Figure 5 This is a schematic diagram of the information extraction process provided in the embodiments of this application;

[0045] Figure 6 This is the first error distribution histogram of the convolutional neural network provided in the embodiments of this application;

[0046] Figure 7 This is the second error distribution histogram of the convolutional neural network provided in the embodiments of this application;

[0047] Figure 8 This is a conventional MED error distribution histogram provided in the embodiments of this application;

[0048] Figure 9 This is a schematic diagram of the structure of the image reversible information hiding system based on a convolutional neural network provided in the embodiments of this application;

[0049] Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0050] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0051] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0052] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0053] Example 1

[0054] like Figure 1 As shown, based on the problems of low target pixel prediction accuracy and low embedding capacity of traditional predictors, this application provides an image reversible information hiding method based on convolutional neural networks, which specifically includes the following steps:

[0055] S101: Divide the original image into mutually independent fork set images and point set images.

[0056] In practice, the original image is segmented into non-overlapping image blocks of the same size. Each image block is then divided into independent cross sets and point sets, and changes in pixels within one set do not affect the other.

[0057] S102: Use a point set image predictor to extract image features of different receptive fields of the cross set image, and perform channel superposition to obtain a predicted point set image; determine the first difference information between the point set image and the predicted point set image, embed secret information into the first difference information, and add it to the predicted point set image to obtain a first secret-carrying image.

[0058] S103: Use a cross set image predictor to extract image features of different receptive fields of the first secret image and perform channel superposition to obtain a predicted cross set image; determine the second difference information between the cross set image and the predicted cross set image, embed the secret information into the second difference information, and add it to the predicted cross set image to obtain a second secret image.

[0059] S104: Fuse the first hidden image and the second hidden image to obtain a reversible information hiding image.

[0060] In specific implementation, such as Figure 2 As shown, the information embedding process is summarized as follows:

[0061] (1) First, divide the image I into a chessboard-shaped set of "points" P1 and a set of "forks" F1.

[0062] (2) Using a convolutional neural network, the first-order "fork" set F1 is used to predict the "point" set P1. The difference between the predicted "point" set and the original "point" set is obtained by subtracting the difference information M1. The secret information S1 is embedded into the difference information M1. Then, the difference information M1 is added to the predicted "point" set image to obtain the secret-carrying "point" set P1'. In the second stage, the secret-carrying "point" set P1' is used to predict the "fork" set F1. The difference between the predicted "fork" set and the original "fork" set is obtained by subtracting the difference information M2. The secret information S2 is embedded into the difference information M2 using error prediction expansion technology. Then, the difference information M2 is added to the predicted "fork" set image F1'.

[0063] (3) After embedding secret information in two stages and adding them together, the final secret image Im is obtained.

[0064] In this way, when performing reversible information hiding, the original image is divided into mutually independent fork set images and point set images. The fork set images and point set images are predicted separately. The reversible information hiding image is generated based on the difference information between the original image and the predicted image. In addition, during the prediction process, image features of different receptive fields of the image are extracted and channel superimposed. The prediction accuracy of the target pixel is improved by using a larger range of surrounding pixels through nonlinear transformation, thereby significantly improving the embedding capacity.

[0065] As an optional implementation, in order to improve the prediction accuracy of the convolutional neural network for the target pixel, the point set image predictor and the fork set image predictor have the same structure and both adopt a convolutional neural network; the image features of the target image are extracted by using convolutional kernels of different sizes of the convolutional neural network, and the receptive field is expanded according to the superposition and combination of image features. The target image includes a fork set image and a first carrier image.

[0066] Optionally, the convolutional neural network includes a backbone prediction network and an auxiliary prediction network. In the backbone prediction network, the target image is processed by a convolutional network with an Inception structure to output image features with receptive fields of various sizes. The image features are then stacked in channel dimension, and the target pixels are predicted using the image features in different channels to obtain the backbone prediction image. In the auxiliary prediction network, the target image is processed by a convolutional kernel of a specific size to predict the target pixels, and the dimension is reduced to obtain the auxiliary prediction image. The predicted image is determined based on the backbone prediction image and the auxiliary prediction image.

[0067] In specific implementation, this embodiment uses two predictors, a backbone prediction and an auxiliary prediction, to predict the target pixel. Taking the prediction of the fork set image using the point set image as an example, the specific steps include:

[0068] (1) The target image is first processed by convolution kernels of size 5*5 and 3*3 to obtain a receptive field of 7*7, and the image features are initially extracted.

[0069] (2) Backbone Prediction Network: The extracted image features are fed into a convolutional network with an Inception structure. Dimensionality is reduced by four convolutions with 1x1 kernels and 32 channels each, resulting in a more compact network structure. The dimensionality-reduced image features are then combined with convolutions of 3x3, 5x5, and 7x7 kernels, yielding receptive fields of 5x5, 7x7, and 9x9, respectively.

[0070] (3) The four receptive field image features with different sizes output by the Inception structure are spliced ​​and superimposed in the second dimension (32+32+32+32=128) to obtain 128-dimensional image features, that is, the number of image channels is 128.

[0071] (4) As an optional implementation, in order to make full use of the image features after being superimposed in the channel dimension through different receptive fields, a channel attention mechanism is added to adjust the weights of image features in different channels. Optionally, this embodiment introduces ECA-Net to strengthen the weights of feature images in important channels, and similarly, reduces the weights of channels with weak image feature representation.

[0072] (5) Use 5*5 and 3*3 convolution kernels to predict the target pixels on the important channel feature images extracted by ECA-Net to obtain the backbone prediction "fork" set image, and calculate the backbone MSELoss1 with the target "fork" set image.

[0073] (6) Auxiliary prediction network: The feature image output by (1) is input to the convolutional pixel prediction with a size of 5*5. After 1*1 convolutional dimensionality reduction, the auxiliary prediction "fork" set image is obtained, and the auxiliary MSELoss2 is calculated with the target "fork" set image.

[0074] (7) Finally, the loss between the predicted "fork" set image and the target "fork" set image is:

[0075] MSELoss = MSELoss1 + λ × MSELoss2; (λ is a constant, λ = 0.2)

[0076] By training a convolutional neural network predictor, the strong correlation between pixels is utilized to predict the values ​​of the other half of the target pixels from half of the pixels. Different combinations of convolutional kernel sizes (3x3, 5x5, 7x7) are used to obtain receptive fields of 5x5, 7x7, 9x9, and 11x11 pixels. The convolutional neural network structure is as follows. Figure 3 As shown in the image.

[0077] To improve the performance of convolutional neural networks, this embodiment uses the ECA-Net channel attention network structure as follows: Figure 4 As shown, a weight is calculated using the channel attention network, and this weight is then used in conjunction with the feature map to correct and change the channel features of the feature map.

[0078] Here, ECA-Net uses non-dimensionality-reduced GAP aggregation convolutional features, first applying an adaptive convolutional kernel of size K, then performing one-dimensional convolution, and finally using a sigmoid function to learn channel attention. To ensure efficiency and effectiveness, local cross-channel interaction is used with the frequency band matrix W. k Let's learn about channel attention. k The specific representation is as follows:

[0079]

[0080] Where K is the kernel size and C is the number of channels.

[0081] output = σ(C1D) k (y));

[0082] Where y is the feature image before attention is applied, C1D represents one-dimensional convolution, σ represents the Sigmoid activation function, and output is the output image after attention is applied.

[0083] The following table lists the detailed parameters of a convolutional neural network:

[0084] Table 1 Detailed parameters of the convolutional neural network

[0085]

[0086]

[0087] Traditional error prediction techniques predict target pixels using several adjacent pixels, failing to adequately consider the correlation between adjacent pixels. This embodiment uses pixels within a pixel region to predict the target pixel, assuming the target pixel is u. i,j Use the pixels v around the target pixel i,j .

[0088] All pixels in the original image are divided into two sets: a "fork" set and a "point" set. In the first stage, the fork set is used for prediction, and the point set is used for information embedding; in the second stage, the point set is used for prediction, and the fork set is used for information embedding. Therefore, secret information can be embedded into the difference information, which includes first interpolation information and second difference information, in the following way:

[0089] Use the pixel v in the region i,j To predict target pixel ui,j The predicted pixel can be represented as: u' i,j =CNN(v i,j );

[0090] By predicting pixel u' i,j and u i,j The difference between the two is d. i,j :d i,j =u' i,j -u i,j ;

[0091] Information embedding process: D i,j =2d i,j +b;

[0092] Among them, D i,j For the expanded prediction error, b represents the payload to be embedded, and then D i,j The changes are made using the histogram translation method; after information embedding, u i,j Calculated as U i,j :U i,j =D i,j +u' i,j ;

[0093] Using the predicted value u' i,j and the modified pixel value U i,j The modified prediction error value is calculated by the following formula: D i,j =U i,j -u' i,j ;

[0094] The value of the secret information is: b = D i,j mod2;

[0095] The value of the original pixel is:

[0096] As an optional implementation, information is extracted from reversibly hidden images in the following manner:

[0097] The reversible information hiding image is divided into mutually independent secret point set images and secret fork set images; based on the secret point set images and fork set image predictors, the predicted secret fork set images are obtained; a second secret difference image is determined between the secret fork set images and the predicted secret fork set images, and secret information and fork set images are extracted from the second secret difference image;

[0098] Based on the fork set image and the point set image predictor, obtain the predicted secret-carrying point set image; determine the first secret-carrying difference image between the secret-carrying point set image and the predicted secret-carrying point set image, and extract the secret information and the point set image from the first secret-carrying difference image;

[0099] The original image is obtained by fusing the fork set image and the point set image.

[0100] In specific implementation, such as Figure 5 As shown, information extraction is the inverse transformation of the embedding process. The information extraction process is as follows:

[0101] (1) At the information extraction end, the image is divided into a set of "points" and a set of "forks" in the shape of a chessboard.

[0102] (2) Step 1: By using the set of "points" carrying secrets and the convolutional neural network model in the training process, the "fork" set of the second stage is predicted and embedded. The difference between the predicted "fork" set and the set of "forks" carrying secrets is obtained to obtain the difference image M2. At the same time, the secret information S2 and the original difference image are extracted using the difference image M2. Step 2: The extraction process is the same as the first step.

[0103] (3) After extracting the secret information from the secret image Im in step (2), the secret information and the original image I are obtained.

[0104] Under the same image, the error prediction through convolutional deep network not only expands the receptive field in space and makes better use of the correlation between pixels, but also adds an attention mechanism to the channels based on the convolutional neural network to enhance the correlation between pixels between channels, improve the prediction performance of the convolutional neural network, and increase the information embedding capacity.

[0105] By selecting MSE, mean, and variance as evaluation criteria for predictor performance, it can be concluded that the prediction performance based on convolutional neural networks is far superior to that of traditional methods.

[0106] The prediction performance of convolutional neural network error prediction and traditional MED predictor on Lena images is compared, and the comparison results are shown in the table below:

[0107] Table 2 Comparison of Predictive Performance

[0108] Convolutional Neural Network Predictor Traditional MED predictor MSE 7.8721 50.7138 mean -0.016 -0.0384 variance 7.8718 50.7125

[0109] And, as Figure 6-8 As shown, the peak points of the two error predictions of the convolutional neural network are much higher than the peak points of the difference image between the error prediction image and the target pixel of the MED predictor on the Lena image.

[0110] Example 2

[0111] Please see Figure 9 , Figure 9 This application provides a crowd evacuation congestion propagation prediction system, such as... Figure 9As shown in the illustration, this application also provides an image reversible information hiding system 900 based on a convolutional neural network, comprising:

[0112] The image segmentation module 910 is used to divide the original image into mutually independent cross set images and point set images;

[0113] The first secret image acquisition module 920 is used to extract image features of different receptive fields of the cross set image using a point set image predictor, and perform channel superposition to obtain a predicted point set image; determine the first difference information between the point set image and the predicted point set image, embed secret information into the first difference information, and add it to the predicted point set image to obtain the first secret image.

[0114] The second secret image acquisition module 930 uses a cross set image predictor to extract image features of different receptive fields of the first secret image and performs channel superposition to obtain a predicted cross set image; determines the second difference information between the cross set image and the predicted cross set image, embeds the secret information into the second difference information, and adds it to the predicted cross set image to obtain the second secret image.

[0115] The fusion module 940 is used to fuse the first hidden image and the second hidden image to obtain a reversible information hiding image.

[0116] Example 3

[0117] Please see Figure 10 , Figure 10 This is a schematic diagram of a computer device according to an embodiment of this application. Figure 10 As shown, the computer device 1000 includes a processor 1010, a memory 1020, and a bus 1030.

[0118] The memory 1020 stores machine-readable instructions executable by the processor 1010. When the computer device 1000 is running, the processor 1010 communicates with the memory 1020 via the bus 1030. When the machine-readable instructions are executed by the processor 1010, they can perform the operations described above. Figures 1 to 5 The steps of the image reversible information hiding method based on convolutional neural networks in the method embodiment shown are specifically implemented in the method embodiment and will not be repeated here.

[0119] Example 4

[0120] Based on the same inventive concept, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the image reversible information hiding method based on a convolutional neural network described in the above method embodiments.

[0121] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0122] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for reversible information hiding in images based on convolutional neural networks, characterized in that, include: The original image is divided into mutually independent cross set images and point set images; The image features of different receptive fields of the fork image are extracted using a point set image predictor and then channel overlay is performed to obtain the predicted point set image. Determine the first difference information between the point set image and the predicted point set image, embed the secret information into the first difference information, and add it to the predicted point set image to obtain the first secret-carrying image; The image features of different receptive fields of the first carrier image are extracted using a cross set image predictor and then channel overlay is performed to obtain the predicted cross set image. Determine the second difference information between the fork image and the predicted fork image, embed the secret information into the second difference information, and add it to the predicted fork image to obtain the second secret image; The first hidden image and the second hidden image are fused to obtain a reversible information hiding image; The point set image predictor and the fork set image predictor have the same structure and both use convolutional neural networks. The image features of the target image are extracted by using convolutional kernels of different sizes of convolutional neural networks. The receptive field is expanded by superimposing and combining the image features. The target image includes fork set images and point set images. The convolutional neural network includes a backbone prediction network and an auxiliary prediction network. In the backbone prediction network, the target image is processed by a convolutional network with an Inception structure, which outputs image features with receptive fields of various sizes. The image features are then stacked in channel dimension, and the target pixels are predicted using the image features in different channels to obtain the backbone prediction image. In the auxiliary prediction network, the target image is processed by a convolutional kernel of a specific size to predict the target pixels, and the dimensions are reduced to obtain the auxiliary prediction image. The predicted image is determined based on the backbone prediction image and the auxiliary prediction image.

2. The image reversible information hiding method based on convolutional neural networks as described in claim 1, characterized in that, After the backbone prediction network superimposes the image features by channel dimension, a channel attention mechanism is added to adjust the weights of the image features on different channels.

3. The image reversible information hiding method based on convolutional neural networks as described in claim 2, characterized in that, The channel attention mechanism includes ECA-Net; ECA-Net uses non-dimensionality-reduced GAP to aggregate convolutional features and then adaptively selects a one-dimensional convolutional kernel size to learn channel attention.

4. The image reversible information hiding method based on convolutional neural networks as described in claim 1, characterized in that, Secret information is embedded into difference information, wherein the difference information includes first interpolation information and second difference information, in the following manner: Use pixels within the region To predict target pixels The predicted pixel can be represented as: ; By predicting pixels and The difference between the two is obtained : = - ; Information embedding process: ; in, For the expanded prediction error, Indicate the payload to be embedded, then The changes are made using the histogram translation method; after information embedding... becomes : ; Using predicted values and modified pixel values The revised prediction error value is calculated using the following formula: ; The value of the secret information is: ; The value of the original pixel is: .

5. The image reversible information hiding method based on convolutional neural networks as described in claim 1, characterized in that, Information extraction is performed on reversibly hidden images using the following method: The reversible information hiding image is divided into mutually independent secret point set images and secret fork set images; based on the secret point set images and fork set image predictors, the predicted secret fork set images are obtained; a second secret difference image is determined between the secret fork set images and the predicted secret fork set images, and secret information and fork set images are extracted from the second secret difference image; Based on the fork set image and the point set image predictor, obtain the predicted secret-carrying point set image; determine the first secret-carrying difference image between the secret-carrying point set image and the predicted secret-carrying point set image, and extract the secret information and the point set image from the first secret-carrying difference image; The original image is obtained by fusing the fork set image and the point set image.

6. An image reversible information hiding system based on a convolutional neural network, characterized in that, include: The image segmentation module is used to divide the original image into mutually independent cross set images and point set images; The first image acquisition module is used to extract image features of different receptive fields of the cross set image using a point set image predictor, and to perform channel overlay to obtain the predicted point set image. Determine the first difference information between the point set image and the predicted point set image, embed the secret information into the first difference information, and add it to the predicted point set image to obtain the first secret-carrying image; The second carrier image acquisition module uses a cross set image predictor to extract image features of different receptive fields of the first carrier image and performs channel overlay to obtain the predicted cross set image. Determine the second difference information between the fork image and the predicted fork image, embed the secret information into the second difference information, and add it to the predicted fork image to obtain the second secret image; The fusion module is used to fuse the first hidden image and the second hidden image to obtain a reversible hidden image; The point set image predictor and the fork set image predictor have the same structure and both use convolutional neural networks. The image features of the target image are extracted by using convolutional kernels of different sizes of convolutional neural networks. The receptive field is expanded by superimposing and combining the image features. The target image includes fork set images and point set images. The convolutional neural network includes a backbone prediction network and an auxiliary prediction network. In the backbone prediction network, the target image is processed by a convolutional network with an Inception structure, which outputs image features with receptive fields of various sizes. The image features are then stacked in channel dimension, and the target pixels are predicted using the image features in different channels to obtain the backbone prediction image. In the auxiliary prediction network, the target image is processed by a convolutional kernel of a specific size to predict the target pixels, and the dimensions are reduced to obtain the auxiliary prediction image. The predicted image is determined based on the backbone prediction image and the auxiliary prediction image.

7. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the image reversible information hiding method based on a convolutional neural network as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the image reversible information hiding method based on a convolutional neural network as described in any one of claims 1 to 5.

Images