Image storing and retrieving method and system based on pixel-level reward mechanism

Abstract

This invention discloses an image-hiding method and system based on a pixel-level reward mechanism. When hiding a secret image, a color carrier image and a grayscale secret image are input into a trained image-hiding network to generate a hidden image. When extracting the secret image, the hidden image is input into the trained image-hiding network to obtain a reconstructed secret image. The training of the image-hiding network includes: constructing the image-hiding network using an image cascaded hiding network, a reconstruction network based on a U-Net++ structure, and a referee network; constructing the total loss function of the image-hiding network based on the loss function of the image cascaded hiding network, the loss function of the reconstruction network based on the U-Net++ structure, the discrimination result of the referee network, and the pixel-level reward matrix; optimizing the image-hiding network with the goal of minimizing the total loss function, resulting in a trained image-hiding network. Advantages: This invention has high concealment, high security, and high computational efficiency.

Description

Technical Field

[0001] This invention relates to a method and system for image-to-image hiding based on a pixel-level reward mechanism, belonging to the field of image processing technology. Background Technology

[0002] Image hiding is a technique that uses embedding algorithms to subtly conceal a secret image within a cover image, which is then extracted by the recipient using a decryption algorithm. Unlike cryptography, image hiding not only ensures the security of the secret information itself but also enhances its security during transmission. In recent years, image hiding technology has been applied in many fields, such as secure data communication and copyright protection.

[0003] In 2017, Baluja published the first deep learning steganography algorithm for image-to-image hiding at NIPS [Baluja S. Hiding images in plain sight: Deep steganography[C]. In Proceedings of the Neural Information Processing Systems. Cambridge: MIT Press, 2017: 2069-2079.]. Since then, numerous steganography models utilizing various deep learning networks have emerged. A good image-to-image hiding technique needs to solve two major challenges: concealment and security; that is, the secret image should not be detectable by the human eye or steganography analysis models. However, current technologies struggle to achieve a balance between these two aspects. Furthermore, reducing the computational complexity of the model is also a significant challenge in the field of image-to-image hiding. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method and system for hiding images based on a pixel-level reward mechanism, which has high hiding capacity and high concealment.

[0005] To address the aforementioned technical problems, this invention provides a method for embedding images within images based on a pixel-level reward mechanism, comprising:

[0006] When hiding secret images, the color carrier image and the grayscale secret image are input into the trained image-hiding network to generate a secret image;

[0007] When extracting secret images, the secret image is input into a trained image-hidden network to obtain the reconstructed secret image;

[0008] The training of the image-hidden network includes:

[0009] The image-to-image-hidden-image network is constructed using an image-cascaded hidden network, a reconstruction network based on the U-Net++ architecture, and a referee network. The image-cascaded hidden network generates a hidden image based on a color carrier image and a grayscale secret image. The reconstruction network based on the U-Net++ architecture reconstructs the hidden image to obtain a reconstructed secret image. The referee network judges the hidden image output by the image-cascaded hidden network during the training of the image-to-image-hidden-image network, obtaining the judgment result and a pixel-level reward matrix.

[0010] The total loss function of the image-hidden image network is constructed based on the loss function of the image cascaded hidden network, the loss function of the reconstruction network based on the U-Net++ structure, the discrimination result of the judge network, and the pixel-level reward matrix.

[0011] The graph-hidden-graph network is optimized with the goal of minimizing the total loss function. Training is completed when the loss decreases and remains stable, resulting in a well-trained graph-hidden-graph network. During the optimization process, the weights of the referee network are fixed.

[0012] Furthermore, the step of inputting the color carrier image and the grayscale secret image into the trained image-hidden network includes:

[0013] The carrier image / secret image is preprocessed using a resolution diversification operation to obtain three pairs of carrier image / secret image pairs with different resolutions: original size, half size, and quarter size.

[0014] Three pairs of carrier images / secret images with different resolutions are input into the image-hidden image network.

[0015] Furthermore, the image cascaded hidden network includes: a low-resolution semantic branch, a medium-resolution detail branch, a high-resolution detail branch, a first cascaded feature fusion module, a second cascaded feature fusion module, and an upsampling operation group;

[0016] The input to the low-resolution semantic branch is a quarter-size carrier image / secret image pair;

[0017] The input to the medium resolution detail branch is a half-size carrier image / secret image pair;

[0018] The input to the high-resolution detail branch is a carrier image / secret image pair of the original size;

[0019] The outputs of the low-resolution semantic branch and the medium-resolution detail branch serve as the inputs to the first fusion module of the cascaded features.

[0020] The output of the high-resolution detail branch of the first cascaded feature fusion module serves as the input of the second cascaded feature fusion module; the output of the second cascaded feature fusion module is connected to the upsampling operation group.

[0021] The output of the upsampling operation group is a dense image.

[0022] Furthermore, the low-resolution semantic branch includes: a carrier branch hidden probability guidance module, a carrier branch first convolution operation group, a carrier branch second convolution operation group, a carrier branch third convolution operation group, a carrier branch fourth convolution operation group, a carrier branch fifth convolution operation group, a secret branch first convolution operation group, a secret branch second convolution operation group, a secret branch third convolution operation group, a secret branch fourth convolution operation group, a secret branch fifth operation group, a deconvolution first operation group, a deconvolution second operation group, and a deconvolution third operation group;

[0023] The input to the carrier branch hiding probability guidance module is a quarter-size carrier image;

[0024] The input to the first convolution operation group of the carrier branch is the output of the carrier branch hidden probability guidance module;

[0025] The input to the second convolution operation group of the carrier branch is the output of the first convolution operation group of the carrier branch.

[0026] The input to the third convolution operation group of the carrier branch is the output of the second convolution operation group of the carrier branch.

[0027] The input to the fourth convolution operation group of the carrier branch is the output of the third convolution operation group of the carrier branch.

[0028] The input to the fifth convolution operation group of the carrier branch is the output of the fourth convolution operation group of the carrier branch.

[0029] The input to the first convolution operation group of the secret branch is a secret image of one-quarter size;

[0030] The input to the second convolution operation group of the secret branch is the output of the first convolution operation group of the secret branch.

[0031] The input to the third convolution operation group of the secret branch is the output of the second convolution operation group of the secret branch.

[0032] The input to the fourth convolution operation group of the secret branch is the output of the third convolution operation group of the secret branch.

[0033] The input to the fifth convolution operation group of the secret branch is the output of the fourth convolution operation group of the secret branch.

[0034] The outputs of the fifth convolution operation group of the carrier branch and the fifth convolution operation group of the secret branch are concatenated and then input into the first deconvolution operation group;

[0035] The outputs of the fourth convolution operation group of the carrier branch, the fourth convolution operation group of the secret branch, and the first deconvolution operation group are concatenated and then input into the second deconvolution operation group to form a skip structure.

[0036] The outputs of the carrier branch third convolution operation group, the secret branch third convolution operation group, and the deconvolution second operation group are concatenated and then input into the deconvolution third operation group to form a jump structure.

[0037] One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

[0038] Furthermore, the carrier branch hiding probability guidance module is represented as follows:

[0039] p(m,n,d,θ)=count{(k,a),(l,b)∈(N x ×N y )×(N x ×N y )}

[0040]

[0041]

[0042]

[0043] Where p(m,n,d,θ) is the gray-level co-occurrence matrix of the quarter-size carrier image I, m and n are two different gray levels, d is the distance between two pixels in the quarter-size carrier image, θ is the angle between two pixels in the quarter-size carrier image, count{·} represents the total number of elements in the calculation set, (k,a), (l,b) are two pixels of the quarter-size carrier image, and N x and N y x represents the width and height of the quarter-size carrier image. t This is the entropy image of the quarter-size carrier image, where BN(·) is the batch normalization operation, σ1(·) is the ReLU activation function, and F1(·) is the 3×3 convolution transformation function. This is a pixel-by-pixel multiplication operation. For pixel-by-pixel addition, x e The output of the vector branch hiding probability guide module.

[0044] Furthermore, the medium-resolution detail branch includes: a first convolution operation group for the carrier branch, a second convolution operation group for the carrier branch, a first convolution operation group for the secret branch, a second convolution operation group for the secret branch, and a channel stitching operation;

[0045] The input to the first convolution operation group of the carrier branch is a carrier image of half size;

[0046] The input to the second convolution operation group of the carrier branch is the output of the first convolution operation group of the carrier branch.

[0047] The input to the first convolution operation group of the secret branch is a secret image of half size;

[0048] The input to the second convolution operation group of the secret branch is the output of the first convolution operation group of the secret branch.

[0049] The outputs of the carrier branch second convolution operation group and the secret branch second convolution operation group are the inputs of the channel splicing operation;

[0050] One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

[0051] Furthermore, the high-resolution detail branch includes: a first convolution operation group for the carrier branch, a second convolution operation group for the carrier branch, a first convolution operation group for the secret branch, a second convolution operation group for the secret branch, and a channel splicing operation.

[0052] The input to the first convolution operation group of the carrier branch is the original-size carrier image;

[0053] The input to the second convolution operation group of the carrier branch is the output of the first convolution operation group of the carrier branch.

[0054] The input to the first convolution operation group of the secret branch is the original-size secret image;

[0055] The input to the second convolution operation group of the secret branch is the output of the first convolution operation group of the secret branch.

[0056] The outputs of the carrier branch second convolution operation group and the secret branch second convolution operation group are the inputs of the channel splicing operation;

[0057] One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

[0058] Furthermore, the first fusion module of the cascaded features is represented as follows:

[0059]

[0060] Where f1 is the output of the low-resolution semantic branch, f2 is the output of the medium-resolution detail branch, Up(·) is the 2×2 upsampling operation, F2(·) is the 1×1 convolution transformation function, σ2(·) is the LeakyReLU activation function, and f3 is the output of the first fusion module of the cascaded features.

[0061] Furthermore, the second fusion module of the cascaded features is represented as follows:

[0062]

[0063] Where f4 is the output of the high-resolution detail branch, and f5 is the output of the second fusion module of the cascaded features. Further, the upsampling operation group is represented as:

[0064] c'=σ3(F3(BN(σ1(F3(BN(σ1(F3(f5))))))))

[0065] Where F3(·) is the 4×4 deconvolution transform function, σ3(·) is the Tanh activation function, and c' is the dense image.

[0066] Furthermore, the reconstructed network based on the U-Net++ structure is a skip connection structure, including: convolution operation group 1, convolution operation group 2, convolution operation group 3, convolution operation group 4, deconvolution operation group 1, convolution operation group 5, deconvolution operation group 2, convolution operation group 6, deconvolution operation group 3, convolution operation group 7, deconvolution operation group 4, convolution operation group 8, deconvolution operation group 5, convolution operation group 9, deconvolution operation group 6, and convolution operation group 10;

[0067] The outputs of the first deconvolution operation group and the first convolution operation group are concatenated and then input into the fifth convolution operation group.

[0068] The outputs of the deconvolution operation group 2 and the convolution operation group 2 are concatenated and then input into the convolution operation group 6.

[0069] The outputs of the deconvolution operation group three, the convolution operation group one, and the convolution operation group five are concatenated and then input into the convolution operation group seven to form a jump connection structure.

[0070] The outputs of the deconvolution operation group four and the convolution operation group three are concatenated and then input into the convolution operation group eight.

[0071] The outputs of the deconvolution operation group five, the convolution operation group two, and the convolution operation group six are concatenated and then input into the convolution operation group nine to form a jump connection structure.

[0072] The outputs of the deconvolution operation group six, the convolution operation group one, the convolution operation group five, and the convolution operation group seven are concatenated and then input into the convolution operation group ten to form a jump connection structure.

[0073] The output of the convolution operation group 10 is the reconstructed secret image;

[0074] One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

[0075] Furthermore, the referee network is a XuNet steganalysis network, obtained through the following training steps:

[0076] Collect multiple image samples and determine the label for each sample image; the label is the probability that the corresponding sample image contains a secret image.

[0077] The XuNet steganalysis network is trained using each image sample as input and the label of each image sample as output, resulting in the referee network.

[0078] Furthermore, the pixel-level reward matrix is ​​represented as follows:

[0079]

[0080]

[0081] Where σ1(·) is the ReLU activation function, F k For the k-th channel of the feature map output by the last convolutional layer of the referee network, α k R represents the weight of the k-th feature map. ij Let be the pixel-level reward matrix, where i and j represent the row and column positions of image pixels, respectively; H' and W' are the height and width of the k-th feature map, and z' is the prediction result of the referee network. F represents k The value of the (i,j)th pixel.

[0082] Furthermore, the total loss function L is:

[0083] L=βL c +ηL s

[0084] Among them, L c For the loss of the image cascaded hidden network, L s η is the loss of the reconstruction network based on the U-Net++ structure, and β and η are the weights used to control the loss of the image cascaded hidden network and the loss of the reconstruction network based on the U-Net++ structure.

[0085] The image cascaded hidden network loss L c The calculation formula is:

[0086] L c =μL v +ωL a

[0087]

[0088]

[0089] Among them, L v L represents quality loss and is used to measure the visual quality of dense images. a Representing security loss, used to measure the security of a dense image, μ and ω are weights used to control quality loss and security loss, c is the number of pixels in the carrier image, L is the total number of pixels in the image, and c' is the number of pixels in the dense image. c μ c' σc and c' are the average values ​​of c and c', representing the brightness of the carrier image and the dense image, respectively. K1 is a constant less than or equal to 1, M is a user-defined scale, and σc is the average value of c and c'. c σ c' σ represents the standard deviations of c and c', respectively, and σ' represents the contrast between the carrier image and the dense image. cc' Let c be the covariance of c and c', representing the structural similarity between the carrier image and the dense image, K2 be a constant less than or equal to 1, G be the Gaussian filter parameters, α and γ be hyperparameters used to control the weights, H and W be the height and width of the image, z be the true label of the image, and z' be the predicted value of the referee network.

[0090] The reconstruction network loss L based on the U-Net++ structure s The calculation formula is:

[0091]

[0092] Where s is the secret image pixel, s' is the reconstructed secret image pixel, and μ s μ s' σ represents the average values ​​of s and s', respectively, indicating the brightness of the secret image and the reconstructed secret image. s σ s' σ represents the standard deviations of s and s', respectively, and σ' represents the contrast between the secret image and the reconstructed secret image. ss' Let s be the covariance of s and s', representing the structural similarity between the secret image and the reconstructed secret image.

[0093] A pixel-level reward mechanism-based image-hiding system includes:

[0094] The generation unit is used to input the color carrier image and the grayscale secret image into the trained image-hiding network to generate a secret image during secret image hiding.

[0095] The reconstruction unit is used to input the secret image into the trained image-hidden network during secret image extraction to obtain the reconstructed secret image;

[0096] The training of the image-hidden network includes:

[0097] The image-to-image-hidden-image network is constructed using an image-cascaded hidden network, a reconstruction network based on the U-Net++ architecture, and a referee network. The image-cascaded hidden network generates a hidden image based on a color carrier image and a grayscale secret image. The reconstruction network based on the U-Net++ architecture reconstructs the hidden image to obtain a reconstructed secret image. The referee network judges the hidden image output by the image-cascaded hidden network during the training of the image-to-image-hidden-image network, obtaining the judgment result and a pixel-level reward matrix.

[0098] The total loss function of the image-hidden image network is constructed based on the loss function of the image cascaded hidden network, the loss function of the reconstruction network based on the U-Net++ structure, the discrimination result of the judge network, and the pixel-level reward matrix.

[0099] The graph-hidden-graph network is optimized with the goal of minimizing the total loss function. Training is completed when the loss decreases and remains stable, resulting in a well-trained graph-hidden-graph network. During the optimization process, the weights of the referee network are fixed.

[0100] The beneficial effects achieved by this invention are as follows:

[0101] This invention designs an image cascaded hiding network that improves the visual quality of dense images while reducing computational complexity;

[0102] A hidden probability guidance module was designed to guide the hiding of secret information with prior knowledge, which improves the visual quality of dense images while reducing the exploration cost of neural networks.

[0103] A pixel-level reward mechanism was designed to provide rewards in real time based on the hiding effect of the secret image, guiding the secret image to be hidden in a safer image area and improving the security of the secret image. Attached Figure Description

[0104] Figure 1 This is a flowchart of an image-to-image method based on a pixel-level reward mechanism, as shown in one embodiment.

[0105] Figure 2 This is a schematic diagram of an image cascaded hidden network structure according to one embodiment;

[0106] Figure 3 This is a schematic diagram of the network structure of a low-resolution semantic branch in one embodiment;

[0107] Figure 4 This is a schematic diagram of the network structure of a medium-resolution detail branch in one embodiment;

[0108] Figure 5 This is a schematic diagram of a network structure with high-resolution detail branches in one embodiment;

[0109] Figure 6 This is a schematic diagram of a reconstructed network based on the U-Net++ architecture, as shown in one embodiment. Detailed Implementation

[0110] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0111] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0112] The application principle of the present invention will be described in detail below with reference to the accompanying drawings.

[0113] Example 1

[0114] This invention provides a method for hiding images using images based on a pixel-level reward mechanism, such as... Figure 1 As shown, the specific steps include the following:

[0115] S10, Input a color carrier image and a grayscale secret image into the image cascaded hiding network to generate a secret image;

[0116] In the specific implementation process, a resolution diversification operation is used to preprocess the carrier image / secret image pair, resulting in three pairs of carrier image / secret image pairs with different resolutions: original size, half size, and quarter size. These three pairs of carrier image / secret image pairs with different resolutions are then input into the image cascaded hiding network.

[0117] In one specific embodiment of the present invention, such as Figure 2As shown, the image cascaded hidden network includes: low-resolution semantic branch LSB, medium-resolution detail branch MDB, high-resolution detail branch HDB, cascaded feature first fusion module CFF_1, cascaded feature second fusion module CFF_2, and upsampling operation group UPG.

[0118] The input to the low-resolution semantic branch is a quarter-size carrier image / secret image pair;

[0119] The input to the medium resolution detail branch is a half-size carrier image / secret image pair;

[0120] The input to the high-resolution detail branch is a carrier image / secret image pair of the original size;

[0121] The outputs of the low-resolution semantic branch and the medium-resolution detail branch serve as the inputs to the first fusion module of the cascaded features.

[0122] The output of the high-resolution detail branch of the first cascaded feature fusion module serves as the input of the second cascaded feature fusion module; the output of the second cascaded feature fusion module is connected to the upsampling operation group.

[0123] The output of the upsampling operation group is a dense image.

[0124] like Figure 3 As shown, the low-resolution semantic branch includes the carrier branch hidden probability guidance module HPGM, the carrier branch first convolution operation group L_c_conv_1, the carrier branch second convolution operation group L_c_conv_2, the carrier branch third convolution operation group L_c_conv_3, the carrier branch fourth convolution operation group L_c_conv_4, the carrier branch fifth convolution operation group L_c_conv_5, the secret branch first convolution operation group L_s_conv_1, the secret branch second convolution operation group L_s_conv_2, the secret branch third convolution operation group L_s_conv_3, the secret branch fourth convolution operation group L_s_conv_4, the secret branch fifth operation group L_s_conv_5, the deconvolution first operation group L_convT_1, the deconvolution second operation group L_convT_2, and the deconvolution third operation group L_convT_3;

[0125] The outputs of the fifth convolution operation group of the carrier branch and the fifth convolution operation group of the secret branch are concatenated and then input into the first deconvolution operation group;

[0126] The outputs of the fourth convolution operation group of the carrier branch, the fourth convolution operation group of the secret branch, and the first deconvolution operation group are concatenated and then input into the second deconvolution operation group to form a skip structure.

[0127] The outputs of the carrier branch third convolution operation group, the secret branch third convolution operation group, and the deconvolution second operation group are concatenated and then input into the deconvolution third operation group to form a jump structure.

[0128] One convolution operation group includes a convolutional layer Conv, an activation layer LeakyReLU, and a batch normalization layer BN arranged in sequence; one deconvolution operation group includes a deconvolutional layer ConvT, an activation layer LeakyReLU, and a batch normalization layer BN arranged in sequence.

[0129] The carrier branch hiding probability guidance module is represented as follows:

[0130] p(m,n,d,θ)=count{(k,a),(l,b)∈(N x ×N y )×(N x ×N y )}

[0131]

[0132]

[0133]

[0134] Where p(m,n,d,θ) is the gray-level co-occurrence matrix of the quarter-size carrier image I, m and n are two different gray levels, d is the distance between two pixels in the quarter-size carrier image, θ is the angle between two pixels in the quarter-size carrier image, count{·} represents the total number of elements in the calculation set, (k,a), (l,b) are two pixels of the quarter-size carrier image, and N x and N y x represents the width and height of the quarter-size carrier image. t This is the entropy image of the quarter-size carrier image, where BN(·) is the batch normalization operation, σ1(·) is the ReLU activation function, and F1(·) is the 3×3 convolution transformation function. This is a pixel-by-pixel multiplication operation. For pixel-by-pixel addition, x e The output of the vector branch hiding probability guide module.

[0135] like Figure 4 As shown, the medium-resolution detail branch includes the first convolution operation group M_c_conv_1 of the carrier branch, the second convolution operation group M_c_conv_2 of the carrier branch, the first convolution operation group M_s_conv_1 of the secret branch, and the second convolution operation group M_s_conv_2 of the secret branch; and the channel concatenation operation Concat.

[0136] The outputs of the second convolution operation group of the carrier branch and the second convolution operation group of the secret branch serve as the inputs for the channel splicing operation.

[0137] like Figure 5 As shown, the high-resolution detail branch includes the first convolution operation group H_c_conv_1 of the carrier branch, the second convolution operation group H_c_conv_2 of the carrier branch, the first convolution operation group H_s_conv_1 of the secret branch, the second convolution operation group H_s_conv_2 of the secret branch, and the channel concatenation operation Concat.

[0138] The outputs of the second convolution operation group of the carrier branch and the second convolution operation group of the secret branch serve as the inputs for the channel splicing operation.

[0139] The first fusion module of the cascaded features is represented as follows:

[0140]

[0141] Where f1 is the output of the low-resolution semantic branch, f2 is the output of the medium-resolution detail branch, Up(·) is the 2×2 upsampling operation, F2(·) is the 1×1 convolution transformation function, σ2(·) is the LeakyReLU activation function, and f3 is the output of the first fusion module of the cascaded features.

[0142] The second fusion module of the cascaded features is represented as follows:

[0143]

[0144] Here, f4 is the output of the high-resolution detail branch, and f5 is the output of the second fusion module of the cascaded features.

[0145] The upsampling operation group is represented as follows:

[0146] c'=σ3(F3(BN(σ1(F3(BN(σ1(F3(f5))))))))

[0147] Where F3(·) is the 4×4 deconvolution transform function, σ3(·) is the Tanh activation function, and c' is the dense image.

[0148] As can be seen, the image cascade network takes cascaded image pairs of different resolutions as input. Low-resolution image pairs (1 / 4 size) are input into a complex semantic network to extract coarse-grained semantic information, while medium- and high-resolution image pairs (1 / 2 size and original size) are input into a lightweight detail network to refine fine-grained detail information. Although the semantic network consists of complex upsampling and downsampling layers, the input resolution is low, allowing for the extraction of rich image information with low computational complexity, thus improving the visual quality of the hidden image. Furthermore, the hiding probability guidance module adaptively enhances the semantic information in high-frequency regions of the carrier image, constraining the secret image to be hidden in textured areas, providing the correct direction for hiding the secret image, and effectively reducing the exploration cost of the hiding network.

[0149] S20, the secret image is input into a reconstruction network based on the U-Net++ structure to obtain the reconstructed secret image;

[0150] In one specific embodiment of this invention, the reconstructed network based on the U-Net++ structure is a skip connection structure, such as... Figure 6 The diagram includes: Convolution operation group 1, Conv_1, Convolution operation group 2, Conv_2, Convolution operation group 3, Conv_3, Convolution operation group 4, Conv_4, Deconvolution operation group 1, ConvT_1, Convolution operation group 5, Conv_5, Deconvolution operation group 2, ConvT_2, Convolution operation group 6, Conv_6, Deconvolution operation group 3, ConvT_3, Convolution operation group 7, Conv_7, Deconvolution operation group 4, ConvT_4, Convolution operation group 8, Conv_8, Deconvolution operation group 5, ConvT_5, Convolution operation group 9, Conv_9, Deconvolution operation group 6, ConvT_6, and Convolution operation group 10, Conv_10.

[0151] The outputs of the deconvolution operation group 1 and the convolution operation group 1 are concatenated and then input into the convolution operation group 5.

[0152] The outputs of the deconvolution operation group 2 and the convolution operation group 2 are concatenated and then input into the convolution operation group 6.

[0153] The outputs of the deconvolution operation group 3, the convolution operation group 1 and the convolution operation group 5 are concatenated and then input into the convolution operation group 7 to form a jump connection structure.

[0154] The outputs of the deconvolution operation group 4 and the convolution operation group 3 are concatenated and then input into the convolution operation group 8.

[0155] The outputs of the deconvolution operation group 5, the convolution operation group 2 and the convolution operation group 6 are concatenated and then input into the convolution operation group 9 to form a jump connection structure.

[0156] The outputs of the deconvolution operation group 6, convolution operation group 1, convolution operation group 5, and convolution operation group 7 are concatenated and then input into the convolution operation group 10 to form a jump connection structure.

[0157] The output of the convolution operation group 10 is the reconstructed secret image;

[0158] One convolution operation group includes a convolutional layer Conv, an activation layer LeakyReLU, and a batch normalization layer BN arranged in sequence; one deconvolution operation group includes a deconvolutional layer ConvT, an activation layer LeakyReLU or Tanh, and a batch normalization layer BN arranged in sequence.

[0159] S30, a pre-trained referee network is used to classify the dense image, obtaining a classification result and a pixel-level reward matrix. The classification result includes identifying the dense image as a carrier image or a dense image;

[0160] After obtaining the encrypted image, to improve its security and determine whether the embedding position of the secret image is reasonable, the steganalysis network XuNet is selected as the judge network, which can effectively distinguish between the encrypted image and the carrier image. Therefore, in specific implementation, the judge network can use the XuNet network to learn from encrypted images generated by image-hidden-image algorithms in different spatial domains, such as Baluja and UDH. The encrypted image is then input into the judge network for discrimination, outputting a binary classification result, that is, identifying the watermarked image as either the carrier image or the watermarked image, along with a pixel-level reward matrix.

[0161] In one specific embodiment of the present invention, the referee network is obtained through the following training steps:

[0162] Collect multiple image samples and determine the label of each sample image; the label includes the probability that the corresponding sample image contains a secret image;

[0163] XuNet is trained using each image sample as input and the label of each image sample as output to obtain the referee network.

[0164] The pixel-level reward matrix is ​​represented as follows:

[0165]

[0166]

[0167] Among them, F kFor the k-th channel of the feature map output by the last convolutional layer of the referee network, α k R represents the weight of the k-th feature map. ij Let be the pixel-level reward matrix, where i and j represent the row and column positions of the image pixels, respectively, H' and W' are the height and width of the k-th feature map, and z' is the prediction result of the referee network.

[0168] S40, a hybrid loss function is designed based on the quality of the hidden image, the quality of the reconstructed secret image, the judgment result of the referee network, and the pixel-level reward matrix. This hybrid loss function is then used as the total loss function of the image-hidden network. The image-hidden network is optimized with the goal of minimizing the total loss function. Training is considered complete when the loss decreases and remains stable. The image-hidden network includes an image cascaded hiding network, a reconstruction network based on the U-Net++ structure, and a referee network. During the optimization process, the weights of the referee network are fixed.

[0169] In one specific embodiment of the present invention, the total loss function of the graph-hidden graph network is:

[0170] L=βL c +ηL s

[0171] Among them, L c For the loss of the image cascaded hidden network, L s η represents the loss of the reconstruction network based on the U-Net++ structure, and β and η are the weights used to control the loss of the image cascaded hidden network and the loss of the reconstruction network based on the U-Net++ structure.

[0172] The image cascaded hidden network loss L c The calculation formula is:

[0173] L c =μL v +ωL a

[0174]

[0175]

[0176] Among them, L v L represents quality loss and is used to measure the visual quality of dense images. a Representing security loss, used to measure the security of a dense image, μ and ω are weights used to control quality loss and security loss, where c is the number of pixels in the carrier image, and c = {c i |1,2,...,L}, where L is the total number of pixels in the image, and c' is the number of pixels in the dense image, c'={c i '|1,2,...,L},μc μ c' σc and c' are the average values ​​of c and c', respectively, representing the brightness of the carrier image and the dense image. K1 is a constant less than or equal to 1, M is a custom scale, and here it is set to 5. c σ c' σ represents the standard deviations of c and c', respectively, and also represents the contrast between the carrier image and the dense image. cc' Let c be the covariance of c and c', which also represents the structural similarity between the carrier image and the dense image. K2 is a constant less than or equal to 1. G is the Gaussian filter parameter. α and γ are hyperparameters used to control the weights. H and W are the height and width of the image. z represents the true label of the image. z' represents the predicted value of the referee network.

[0177] The reconstruction network loss L based on the U-Net++ structure s The calculation formula is:

[0178]

[0179] Where s is the secret image pixel, s' is the reconstructed secret image pixel, and μ s μ s' σs and s' are the average values ​​of s and s', respectively, representing the brightness of the secret image and the reconstructed secret image. K1 is a constant less than or equal to 1, M is a custom scale, which is set to 5 in this case. s σ s' σ represents the standard deviations of s and s', respectively, and also represents the contrast between the secret image and the reconstructed secret image. ss' K is the covariance of s and s', which also represents the structural similarity between the secret image and the reconstructed secret image. K2 is a constant less than or equal to 1, and G is the Gaussian filter parameter.

[0180] In practical applications, a color carrier image and a grayscale secret image are input into a trained image cascaded hiding network to obtain a hidden image. Then, the hidden image is input into a trained reconstruction network based on the U-Net++ structure to extract the secret image hidden in the hidden image.

[0181] In summary, the image-to-image method based on a pixel-level reward mechanism in this embodiment of the invention inputs a color carrier image and a grayscale secret image into an image cascaded hiding network to generate a hidden image; the hidden image is then input into a reconstruction network based on the U-Net++ structure to obtain a reconstructed secret image; a pre-trained referee network is used to determine whether the hidden image contains a secret image and to allocate pixel-level rewards based on the hiding effect. This method has high concealment, high security, and high computational efficiency.

[0182] To verify the effectiveness of this invention, the proposed image-hidden image model was first trained on the public dataset PASCAL-VOC2012 and then tested on the public dataset LFW. The experimental results regarding image quality are shown in Table 1. Among them, the Baluja model [Shumeet Baluja. Hiding images in plain sight: Deep steganography. In NIPS, pages 2069-2079, 2017.] is the first image-hiding model, the UDH model [Chaoning Zhang, Philipp Benz, Adil Karjauv, Geng Sun and In So Kweon. UDH: Universal deep hiding for steganography, watermarking, and light field messaging. In NIPS, pages 10223-10234, 2020.] is the first image-hiding model that does not rely on the texture of the carrier image, and the ISGAN model [Ru Zhang, Shiqi Dong and Jianyi Liu. Invisible steganography via generative adversarial networks. Multimedia Tools and Applications, 78(7): 8559-8575, 2019.] is the best model for generating and reconstructing hidden images by embedding grayscale secret images.The experimental results regarding security are shown in Table 2. Among them, XuNet [Guanshuo Xu, Han-Zhou Wu and Yun-Qing Shi. Structural design of convolutional neural networks for steganalysis. IEEE Signal Processing Letters, 23(5):708-712, 2016.], YeNet [Jian Ye, Jiangqun Ni and Yang Yi. Deep learning hierarchical representations for image steganalysis. IEEE Transactions on Information Forensics and Security, 12(11):2545-2557, 2017.] and Yedroudj-Net [Mehdi Yedroudj, Frédéric Comby and Marc Chaumont. Yedroudj-net: An efficient CNN for spatial steganalysis. In ICASSP, pages 2092-2096, 2018.] are the three steganalysis models with better performance. The experimental results regarding computational complexity are shown in Table 3.

[0183] Table 1

[0184]

[0185] Table 2

[0186]

[0187] Table 3

[0188] Image-based image model <![CDATA[Floating-point operations per second (×10 6 )↓]]> Baluja 29125.51 UDH 10976.46 ISGAN 54084.83 This invention 1594.06

[0189] Example 2

[0190] Based on the same inventive concept as Embodiment 1, this embodiment of the invention provides an image-hiding system based on a pixel-level reward mechanism, comprising:

[0191] The generation unit is used to input a carrier image and a secret image into the image cascaded hiding network to obtain a hidden image that hides a secret image.

[0192] The reconstruction unit is used to input the secret image into a reconstruction network based on the U-Net++ structure to obtain the reconstructed secret image;

[0193] The discrimination unit is used to discriminate the dense image using a pre-trained referee network to obtain a discrimination result and a pixel-level reward matrix. The discrimination result includes identifying the dense image as a carrier image or a dense image.

[0194] The training unit is used to design a hybrid loss function based on the quality of the hidden image, the quality of the reconstructed secret image, the judgment result of the referee network, and the pixel-level reward matrix. This hybrid loss function is then used as the total loss function of the image-hidden network. The image-hidden network is optimized with the goal of minimizing the total loss function. Training is considered complete when the loss decreases and remains stable. The image-hidden network includes an image cascaded hiding network, a reconstruction network based on the U-Net++ structure, and a referee network. During the optimization process, the weights of the referee network are fixed.

[0195] The steganography unit is used to generate a hidden image from a carrier image and a secret image using a trained image-hidden-image network, and then reconstruct the secret image from the hidden image.

[0196] For specific limitations regarding the image-to-image system based on a pixel-level reward mechanism, please refer to the limitations of the image-to-image method based on a pixel-level reward mechanism mentioned above, which will not be repeated here. Each module in the aforementioned image-to-image system based on a pixel-level reward mechanism can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0197] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0198] It should be noted that the terms "first," "second," and "third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific order of objects. It is understood that "first," "second," and "third" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first," "second," and "third" can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in an order other than those illustrated or described herein.

[0199] The terms "comprising" and "having," and any variations thereof, in this application are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to such processes, methods, products, or devices.

[0200] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for hiding images using images based on a pixel-level reward mechanism, characterized in that, include: When hiding secret images, the color carrier image and the grayscale secret image are input into the trained image-hiding network to generate a secret image; When extracting secret images, the secret image is input into a trained image-hidden network to obtain the reconstructed secret image; The training of the image-hidden network includes: The image-to-image-hidden-image network is constructed using an image-cascaded hidden network, a reconstruction network based on the U-Net++ architecture, and a referee network. The image-cascaded hidden network generates a hidden image based on a color carrier image and a grayscale secret image. The reconstruction network based on the U-Net++ architecture reconstructs the hidden image to obtain a reconstructed secret image. The referee network judges the hidden image output by the image-cascaded hidden network during the training of the image-to-image-hidden-image network, obtaining the judgment result and a pixel-level reward matrix. The total loss function of the image-hidden image network is constructed based on the loss function of the image cascaded hidden network, the loss function of the reconstruction network based on the U-Net++ structure, the discrimination result of the judge network, and the pixel-level reward matrix. The graph-hidden-graph network is optimized with the goal of minimizing the total loss function. Training is completed when the loss decreases and remains stable, resulting in a well-trained graph-hidden-graph network. During the optimization process, the weights of the referee network are fixed.

2. The image-hiding method based on a pixel-level reward mechanism according to claim 1, characterized in that, The step of inputting the color carrier image and the grayscale secret image into the trained image-hidden network includes: The carrier image / secret image is preprocessed using a resolution diversification operation to obtain three pairs of carrier image / secret image pairs with different resolutions: original size, half size, and quarter size. Three pairs of carrier images / secret images with different resolutions are input into the image-hidden image network.

3. The image-hiding method based on a pixel-level reward mechanism according to claim 2, characterized in that, The image cascaded hidden network includes: a low-resolution semantic branch, a medium-resolution detail branch, a high-resolution detail branch, a first cascaded feature fusion module, a second cascaded feature fusion module, and an upsampling operation group. The input to the low-resolution semantic branch is a quarter-size carrier image / secret image pair; The input to the medium resolution detail branch is a half-size carrier image / secret image pair; The input to the high-resolution detail branch is a carrier image / secret image pair of the original size; The outputs of the low-resolution semantic branch and the medium-resolution detail branch serve as the inputs to the first fusion module of the cascaded features. The output of the high-resolution detail branch of the first cascaded feature fusion module serves as the input of the second cascaded feature fusion module; the output of the second cascaded feature fusion module is connected to the upsampling operation group. The output of the upsampling operation group is a dense image.

4. The image-hiding method based on a pixel-level reward mechanism according to claim 3, characterized in that, The low-resolution semantic branch includes: a carrier branch hidden probability guidance module, a carrier branch first convolution operation group, a carrier branch second convolution operation group, a carrier branch third convolution operation group, a carrier branch fourth convolution operation group, a carrier branch fifth convolution operation group, a secret branch first convolution operation group, a secret branch second convolution operation group, a secret branch third convolution operation group, a secret branch fourth convolution operation group, a secret branch fifth operation group, a deconvolution first operation group, a deconvolution second operation group, and a deconvolution third operation group; The input to the carrier branch hiding probability guidance module is a quarter-size carrier image; The input to the first convolution operation group of the carrier branch is the output of the carrier branch hidden probability guidance module; The input to the second convolution operation group of the carrier branch is the output of the first convolution operation group of the carrier branch. The input to the third convolution operation group of the carrier branch is the output of the second convolution operation group of the carrier branch. The input to the fourth convolution operation group of the carrier branch is the output of the third convolution operation group of the carrier branch. The input to the fifth convolution operation group of the carrier branch is the output of the fourth convolution operation group of the carrier branch. The input to the first convolution operation group of the secret branch is a secret image of one-quarter size; The input to the second convolution operation group of the secret branch is the output of the first convolution operation group of the secret branch. The input to the third convolution operation group of the secret branch is the output of the second convolution operation group of the secret branch. The input to the fourth convolution operation group of the secret branch is the output of the third convolution operation group of the secret branch. The input to the fifth convolution operation group of the secret branch is the output of the fourth convolution operation group of the secret branch. The outputs of the fifth convolution operation group of the carrier branch and the fifth convolution operation group of the secret branch are concatenated and then input into the first deconvolution operation group; The outputs of the fourth convolution operation group of the carrier branch, the fourth convolution operation group of the secret branch, and the first deconvolution operation group are concatenated and then input into the second deconvolution operation group to form a skip structure. The outputs of the carrier branch third convolution operation group, the secret branch third convolution operation group, and the deconvolution second operation group are concatenated and then input into the deconvolution third operation group to form a jump structure. One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

5. The image-hiding method based on a pixel-level reward mechanism according to claim 4, characterized in that, The carrier branch hiding probability guidance module is represented as follows: p(m,n,d,θ)=count{(k,a),(l,b)∈(N x ×N y )×(N x ×N y )} Where p(m,n,d,θ) is the gray-level co-occurrence matrix of the quarter-size carrier image I, m and n are two different gray levels, d is the distance between two pixels in the quarter-size carrier image, θ is the angle between two pixels in the quarter-size carrier image, count{·} represents the total number of elements in the calculation set, (k,a), (l,b) are two pixels of the quarter-size carrier image, and N x and N y x represents the width and height of the quarter-size carrier image. t This is the entropy image of the quarter-size carrier image, where BN(·) is the batch normalization operation, σ1(·) is the ReLU activation function, and F1(·) is the 3×3 convolution transformation function. This is a pixel-by-pixel multiplication operation. For pixel-by-pixel addition, x e The output of the vector branch hiding probability guide module.

6. The image-hiding method based on a pixel-level reward mechanism according to claim 3, characterized in that, The medium-resolution detail branch includes: a first convolution operation group for the carrier branch, a second convolution operation group for the carrier branch, a first convolution operation group for the secret branch, a second convolution operation group for the secret branch; and channel stitching operation. The input to the first convolution operation group of the carrier branch is a carrier image of half size; The input to the second convolution operation group of the carrier branch is the output of the first convolution operation group of the carrier branch. The input to the first convolution operation group of the secret branch is a secret image of half size; The input to the second convolution operation group of the secret branch is the output of the first convolution operation group of the secret branch. The outputs of the carrier branch second convolution operation group and the secret branch second convolution operation group are the inputs of the channel splicing operation; One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

7. The image-hiding method based on a pixel-level reward mechanism according to claim 3, characterized in that, The high-resolution detail branch includes: carrier branch first convolution operation group, carrier branch second convolution operation group, secret branch first convolution operation group, secret branch second convolution operation group, and channel splicing operation; The input to the first convolution operation group of the carrier branch is the original-size carrier image; The input to the second convolution operation group of the carrier branch is the output of the first convolution operation group of the carrier branch. The input to the first convolution operation group of the secret branch is the original-size secret image; The input to the second convolution operation group of the secret branch is the output of the first convolution operation group of the secret branch. The outputs of the carrier branch second convolution operation group and the secret branch second convolution operation group are the inputs of the channel splicing operation; One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

8. The image-hiding method based on a pixel-level reward mechanism according to claim 3, characterized in that, The first fusion module of the cascaded features is represented as follows: Where f1 is the output of the low-resolution semantic branch, f2 is the output of the medium-resolution detail branch, Up(·) is the 2×2 upsampling operation, F2(·) is the 1×1 convolution transformation function, σ2(·) is the LeakyReLU activation function, and f3 is the output of the first fusion module of the cascaded features.

9. The image-hiding method based on a pixel-level reward mechanism according to claim 8, characterized in that, The second fusion module of the cascaded features is represented as follows: Here, f4 is the output of the high-resolution detail branch, and f5 is the output of the second fusion module of the cascaded features.

10. The image-hiding method based on a pixel-level reward mechanism according to claim 9, characterized in that, The upsampling operation group is represented as follows: c'=σ3(F3(BN(σ1(F3(BN(σ1(F3(f5)))))))) Where F3(·) is the 4×4 deconvolution transform function, σ3(·) is the Tanh activation function, and c' is the dense image.

11. The image-hiding method based on a pixel-level reward mechanism according to claim 1, characterized in that, The reconstructed network based on the U-Net++ structure is a skip connection structure, including: convolution operation group 1, convolution operation group 2, convolution operation group 3, convolution operation group 4, deconvolution operation group 1, convolution operation group 5, deconvolution operation group 2, convolution operation group 6, deconvolution operation group 3, convolution operation group 7, deconvolution operation group 4, convolution operation group 8, deconvolution operation group 5, convolution operation group 9, deconvolution operation group 6, and convolution operation group 10; The outputs of the first deconvolution operation group and the first convolution operation group are concatenated and then input into the fifth convolution operation group. The outputs of the deconvolution operation group 2 and the convolution operation group 2 are concatenated and then input into the convolution operation group 6. The outputs of the deconvolution operation group three, the convolution operation group one, and the convolution operation group five are concatenated and then input into the convolution operation group seven to form a jump connection structure. The outputs of the deconvolution operation group four and the convolution operation group three are concatenated and then input into the convolution operation group eight. The outputs of the deconvolution operation group five, the convolution operation group two, and the convolution operation group six are concatenated and then input into the convolution operation group nine to form a jump connection structure. The outputs of the deconvolution operation group six, the convolution operation group one, the convolution operation group five, and the convolution operation group seven are concatenated and then input into the convolution operation group ten to form a jump connection structure. The output of the convolution operation group 10 is the reconstructed secret image; One convolution operation group includes a convolutional layer, an activation layer, and a batch normalization layer arranged in sequence; one deconvolution operation group includes a deconvolutional layer, an activation layer, and a batch normalization layer arranged in sequence.

12. The image-hiding method based on a pixel-level reward mechanism according to claim 1, characterized in that, The referee network is the XuNet steganalysis network, which is obtained through the following training steps: Collect multiple image samples and determine the label for each sample image; the label is the probability that the corresponding sample image contains a secret image. The XuNet steganalysis network is trained using each image sample as input and the label of each image sample as output, resulting in the referee network.

13. The image-hiding method based on a pixel-level reward mechanism according to claim 1, characterized in that, The pixel-level reward matrix is ​​represented as follows: Where σ1(·) is the ReLU activation function, F k For the k-th channel of the feature map output by the last convolutional layer of the referee network, α k R represents the weight of the k-th feature map. ij Let be the pixel-level reward matrix, where i and j represent the row and column positions of image pixels, respectively; H' and W' are the height and width of the k-th feature map, and z' is the prediction result of the referee network. F represents k The value of the (i,j)th pixel.

14. The image-hiding method based on a pixel-level reward mechanism according to claim 13, characterized in that, The total loss function L is: L=βL c +ηL s Among them, L c For the loss of the image cascaded hidden network, L s η is the loss of the reconstruction network based on the U-Net++ structure, and β and η are the weights used to control the loss of the image cascaded hidden network and the loss of the reconstruction network based on the U-Net++ structure. The image cascaded hidden network loss L c The calculation formula is: L c =μL v +ωL a Among them, L v L represents quality loss and is used to measure the visual quality of dense images. a Representing security loss, used to measure the security of a dense image, μ and ω are weights used to control quality loss and security loss, c is the number of pixels in the carrier image, L is the total number of pixels in the image, and c' is the number of pixels in the dense image. c μ c' σc and c' are the average values ​​of c and c', representing the brightness of the carrier image and the dense image, respectively. K1 is a constant less than or equal to 1, M is a user-defined scale, and σc is the average value of c and c'. c σ c' σ represents the standard deviations of c and c', respectively, and σ' represents the contrast between the carrier image and the dense image. cc' Let c be the covariance of c and c', representing the structural similarity between the carrier image and the dense image, K2 be a constant less than or equal to 1, G be the Gaussian filter parameters, α and γ be hyperparameters used to control the weights, H and W be the height and width of the image, z be the true label of the image, and z' be the predicted value of the referee network. The reconstruction network loss L based on the U-Net++ structure s The calculation formula is: Where s is the secret image pixel, s' is the reconstructed secret image pixel, and μ s μ s' σ represents the average values ​​of s and s', respectively, indicating the brightness of the secret image and the reconstructed secret image. s σ s' σ represents the standard deviations of s and s', respectively, and σ' represents the contrast between the secret image and the reconstructed secret image. ss' Let s be the covariance of s and s', representing the structural similarity between the secret image and the reconstructed secret image.

15. A system for hiding images using images based on a pixel-level reward mechanism, characterized in that, include: The generation unit is used to input the color carrier image and the grayscale secret image into the trained image-hiding network to generate a secret image during secret image hiding. The reconstruction unit is used to input the secret image into the trained image-hidden network during secret image extraction to obtain the reconstructed secret image; The training of the image-hidden network includes: The image-to-image-hidden-image network is constructed using an image-cascaded hidden network, a reconstruction network based on the U-Net++ architecture, and a referee network. The image-cascaded hidden network generates a hidden image based on a color carrier image and a grayscale secret image. The reconstruction network based on the U-Net++ architecture reconstructs the hidden image to obtain a reconstructed secret image. The referee network judges the hidden image output by the image-cascaded hidden network during the training of the image-to-image-hidden-image network, obtaining the judgment result and a pixel-level reward matrix. The total loss function of the image-hidden image network is constructed based on the loss function of the image cascaded hidden network, the loss function of the reconstruction network based on the U-Net++ structure, the discrimination result of the judge network, and the pixel-level reward matrix. The graph-hidden-graph network is optimized with the goal of minimizing the total loss function. Training is completed when the loss decreases and remains stable, resulting in a well-trained graph-hidden-graph network. During the optimization process, the weights of the referee network are fixed.

Images