Remote sensing image steganography method and device based on deep learning and storage medium

Through the improved remote sensing image steganography network, the FCDensenet architecture and frequency domain transformation module are used to solve the problem of generating high-quality images and decryption quality in remote sensing image steganography, and high-quality image hiding and decryption effects are achieved.

CN119094668APending Publication Date: 2024-12-06CHINA UNIV OF GEOSCIENCES (WUHAN)
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411220800.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-02
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The prior art cannot generate high-quality remote sensing image steganography images, and decrypted image color texture abnormalities are prone to occur during decryption.

Method used

Adopt the improved remote sensing image steganography network based on the FCDensenet convolutional neural network architecture, including preprocessing network, image frequency domain transformation module, encrypted network and decryption network. By introducing attention mechanism and frequency domain transformation module, the hiding and decryption of images are improved. quality.

Benefits of technology

Generate high-quality secret images and effectively decrypt secret image information in practical applications, improve the PSNR and SSIM index values ​​of the image, weaken the contour of the secret image, and ensure that the decrypted image is similar to the original secret image.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119094668A_ABST
    Figure CN119094668A_ABST
Patent Text Reader

Abstract

The invention relates to the field of image steganography, and discloses a remote sensing image steganography method and device based on deep learning and a storage medium, and the method comprises the steps: constructing an improved remote sensing image steganography network; the improved remote sensing image steganography network is based on an FCDensenet convolutional neural network architecture. The system comprises a preprocessing network, an image frequency domain transformation module, an encryption network, a decryption network and an image recovery network. Training an improved remote sensing image steganography network; and performing remote sensing image steganography by using the trained network. According to the method, the secret image information contained in the high-quality secret-containing image can be effectively decrypted in practical application while the high-quality secret-containing image is generated.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of image steganography, and in particular to a remote sensing image steganography method, device and storage medium based on deep learning. Background Art

[0002] The application of communication technology in various fields such as industrial communication, monitoring, health care and military data communication has become more and more common, and it has become easier and simpler to obtain digital data or information during the transmission process. In the process of collecting, transmitting and storing remote sensing images, there is a risk of information being stolen, tampered with or leaked. Most remote sensing images have accurate geographic coordinate information. By analyzing remote sensing images, users can obtain information about disasters, natural resources and even the location of some military facilities. Therefore, studying how to protect sensitive information in remote sensing images and ensure information security is of great significance to safeguarding national security, commercial interests and personal privacy.

[0003] In recent years, with the rapid development of machine learning, image steganography methods based on deep learning have achieved remarkable results in the field of images. Deep learning models, especially convolutional neural networks (CNN) and generative adversarial networks (GAN), have become important methods for image processing and analysis. Among these methods, the CNN network model has a significant effect on image reconstruction, can learn the complex features and representations of images, and can generate realistic steganographic images. However, remote sensing images have more complex features than natural images, which makes neural network training unstable and difficult to generate high-quality steganographic images. For example, the artifact outline of the secret remote sensing image can be seen in the secret image. In addition, when decrypting images, it is easy to encounter abnormal color and texture of the decrypted image. Summary of the invention

[0004] The purpose of the present invention is to propose a remote sensing image steganography method, device and storage medium based on deep learning, so as to solve the technical problems that the prior art cannot generate high-quality secret images and the decrypted images contain abnormal information during decryption.

[0005] The present invention provides a remote sensing image steganography method based on deep learning, comprising the following steps:

[0006] S1. Construct an improved remote sensing image steganography network; the improved remote sensing image steganography network is based on the FCDensenet convolutional neural network architecture, specifically including: a preprocessing network, an image frequency domain transformation module, an encryption network, a decryption network and an image restoration network;

[0007] S2, training the improved remote sensing image steganography network;

[0008] S3. Use the trained network to perform remote sensing image steganography.

[0009] A storage medium stores instructions and data for implementing a remote sensing image steganography method based on deep learning.

[0010] A remote sensing image steganography device based on deep learning comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a remote sensing image steganography method based on deep learning.

[0011] The beneficial effects provided by the present invention are:

[0012] 1. The improved attention mechanism preprocessing network introduced in the present invention can effectively extract the contour information of the carrier image and the secret image, which is conducive to the hiding of the secret image; the use of the multi-layer perceptron layer (Mlp) introduced therein greatly improves the values ​​of image PSNR, SSIM and other indicators while ensuring the hiding effect; the final addition operation of the network forms a residual connection, so that the image after the preprocessing network retains its original image information when its contour information is extracted for the subsequent network processing.

[0013] 2. The introduced improved frequency domain transformation module can effectively weaken the outline of the secret image, allowing it to be better hidden in the carrier image and improve the quality of the decrypted image.

[0014] 3. Introduce the pre-decryption recovery network and the post-decryption recovery network before the decryption network to restore the pixel loss caused by image rounding in practical applications, ensuring that the final decrypted image is similar to the original secret image;

[0015] In general, the method of the present invention can not only generate a high-quality secret image, but also effectively decrypt the secret image information contained therein in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic flow chart of the method of the present invention;

[0017] Figure 2 A remote sensing image steganography model diagram established for the present invention;

[0018] Figure 3 This is the effect diagram generated by the remote sensing image steganography network of the present invention;

[0019] Figure 4 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0020] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0021] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.

[0022] Please refer to Figure 1 , Figure 1 It is a schematic flow diagram of the method of the present invention;

[0023] The present invention provides a remote sensing image steganography method based on deep learning, comprising the following steps:

[0024] S1. Construct an improved remote sensing image steganography network; the improved remote sensing image steganography network is based on the FCDensenet convolutional neural network architecture, specifically including: a preprocessing network, an image frequency domain transformation module, an encryption network, a decryption network and an image restoration network;

[0025] Please refer to Figure 2 , Figure 2 It is a remote sensing image steganography model diagram established by the present invention;

[0026] Among them, Pre-processing Network is a preprocessing network, which is used to extract the contour information of the carrier image and the secret image, and promote the feature fusion of the image, so that the secret image can be better hidden in the carrier image; D represents the image frequency domain transformation module, which is used to weaken the contour information of the secret image; Encoding Network is an encryption network, which is used to generate a secret image from the preprocessed carrier image and the frequency domain transformed secret image; PreRecovery is a pre-decryption recovery network, which is used to recover the secret image and generate a pre-decryption recovery image; Decoding Network is a decryption network, which extracts the decrypted image from the pre-decryption recovery image; PostRecovery is a post-decryption recovery network, which recovers the decrypted image to obtain the decrypted recovery image, and ID is a frequency domain inverse transformation module, after which the final extracted image reveal is obtained.

[0027] It should be noted that the preprocessing network is used to allow the entire model to learn the contour information of the secret image secret and the carrier image cover, and obtain the preprocessed secret image secret_pre and the preprocessed carrier image cover_pre.

[0028] Specifically, the preprocessing network in the present invention consists of two parts: a channel attention module and a spatial attention module. The channel attention module consists of: 1 global maximum pooling layer, 1 global average pooling layer, 2 multi-layer perceptron layers and 1 Sigmod activation function; the spatial attention module consists of: 1 global maximum pooling layer, 1 global average pooling layer, 1 convolution layer and 1 Sigmod activation function.

[0029] As an embodiment, the structure of the preprocessing network is as follows:

[0030] The first layer: a channel attention module, which is used to extract features between image channels. At the same time, a global maximum pooling layer and a global average pooling layer are used to perform pooling operations on the input original tensor x to obtain the average pooling result (x_avg) and the maximum pooling result (x_max). The purpose of this is to capture the information of different spatial positions in the feature map and compress it into a global representation of the channel dimension.

[0031] Global maximum pooling focuses on the position with the strongest response in the feature map, while global average pooling considers the average of all positions. Then, the results of x_avg and x_max after passing through two multi-layer perceptron layers mlp are added to obtain the combined attention weight accumulation channel_att_sum, learn the complex nonlinear relationship between different channels, activate channel_att_sum through the Sigmod function, and increase the dimension to obtain the channel attention weight scale, and then multiply the original input x by scale to obtain the final channel attention feature channel_att, so that the model can pay more attention to important channels and suppress unimportant channels.

[0032] The multi-layer perceptron layer Mlp consists of a flattening layer Flatten, which is used to flatten the multi-dimensional input tensor tensor into one dimension; a fully connected layer Linear, which reduces the feature dimension of the input data from the number of channels of the input feature to the specified number of channels (calculated by channels / / reduction_ratio, where reduction_ratio: dimensionality reduction ratio factor, which controls the number of channels of the output feature).

[0033] A nonlinear activation function layer ReLU layer is used to introduce nonlinearity to help the network learn complex features; the last one is a fully connected layer Linear layer, which upscales the output image of the previous layer (the number of channels is channels / / reduction_ratio) back to the number of channels.

[0034] The second layer: a spatial attention module, which is used to extract the feature information related to the image space. The channel attention feature (channel_att) output by the channel attention module is used as the input feature to perform maximum pooling and average pooling in the channel dimension, compressing the spatial dimension of the feature map while retaining the information of the channel dimension, and splicing the compressed channel feature (x_compress) in the channel dimension. After this operation, the obtained x_compress contains the maximum and average information of each channel, which is helpful for the subsequent learning of spatial attention weights. Then, after a convolution layer with a convolution kernel size of 7*7 and a stride of 1, it captures more extensive and detailed dependencies in space, and after a Sigmod activation function, the spatial attention weight (scale2) is obtained. Then the original input channel attention feature (channel_att) is multiplied by the spatial attention weight (scale2) to obtain the final spatial attention feature (spatial_att). The final spatial_att is the feature map processed by the spatial attention mechanism, which contains important feature information related to the spatial position.

[0035] The third layer: Add the image features spatial_att output by the channel attention and spatial attention to the original input image tensor x to form a residual connection, so that the network can learn the contour information of the image while retaining the original information of the image, which is beneficial to the feature fusion of the carrier image and the secret image in the subsequent encryption network.

[0036] The preprocessed secret image secret_pre is transformed using an image frequency domain transformation module to obtain a frequency domain transformed secret image secret_pre_dwt, thereby weakening the contour of the secret image.

[0037] It should be noted that the image frequency domain transformation module structure is composed of an improved discrete wavelet transform, including two parts: a forward transform and an inverse transform.

[0038] As an embodiment, the encryption network consists of 1 input convolution layer, 11 dense blocks, 5 downsampling layers, 5 upsampling layers and 1 output convolution layer;

[0039] The decryption network consists of 6 separable convolutional blocks and 2 convolutional layers, 2 normalization layers and 2 ELU activation functions;

[0040] The image restoration network consists of two parts: a pre-decryption restoration network and a post-decryption restoration network. The pre-decryption restoration network consists of: 2 convolutional layers, 2 spatial channel reconstruction convolutional layers SCConv and a ReLU activation function; the post-decryption restoration network consists of: 5 convolutional layers, 4 dropout layers and 4 LeakyReLU layers.

[0041] The specific processing process of the image frequency domain module is as follows:

[0042] Forward transformation part:

[0043] Step 1: Perform discrete wavelet transform (DWT) on the input image. The input image is decomposed into four sub-bands: LL, HL, LH, and HH, which represent the frequency domain information in the horizontal-horizontal, horizontal-vertical, vertical-horizontal, and vertical-vertical directions respectively;

[0044] Step 2: By filling the LL, HL, LH, HH sub-bands into a tensor according to the following filling rules, the image hiding is promoted. The specific filling rules are shown in formulas (1)(2)(3)(4):

[0045] h[:,:,0::2,0::2]=x LL (1)

[0046] h[:,:,1::2,0::2]=x HL (2)

[0047] h[:,:,0::2,1::2]=x LH (3)

[0048] h[:,:,1::2,1::2]=x HH (4)

[0049] The fusion of the four sub-bands is achieved through matrix indexing and slicing operations according to the following filling rules, and the four sub-bands LL, HL, LH, and HH after two-dimensional discrete wavelet transform decomposition are filled into the h image; in the above description, h is the filled image, the size of h image is equal to the size of the original secret image, and the shape of h is four dimensions, the first dimension is the batch size, the second dimension is the number of channels, the third dimension is the height, and the fourth dimension is the width. The size of the four sub-bands LL, HL, LH, and HH is one-fourth of the size of the original secret image, and also one-fourth of the size of the h image;

[0050] Specifically, the first two ':' in formula (1) select all batch sizes and all dimension sizes of h. The first 0::2 selects every other element starting from 0 in the height dimension of h, that is, selects all even-indexed elements. The second 0::2 selects all even-indexed elements in the width dimension of h, that is, fills the subband LL into the even rows and even columns of h.

[0051] In formula (2), all batch sizes and all dimension sizes of h are selected. The first 1::2 selects every other element starting from 1 in the height dimension of h, that is, selects all elements with odd indexes. The second 0::2 selects all elements with even indexes in the width dimension of h, that is, fills the subband HL into the odd rows and even columns of h.

[0052] Formula (3) is to fill the sub-band LH into the even rows and odd columns of h, and formula (4) is to fill the sub-band HH into the odd rows and odd columns of h;

[0053] Inverse transformation part:

[0054] Step 1: Extract the LL, HL, LH, and HH sub-band information from the decrypted image tensor according to the filling rule, that is, extract LL from the even rows and even columns of the decrypted image, extract HL from the odd rows and even columns, extract LH from the even rows and odd columns, and extract HH ​​from the odd rows and odd columns;

[0055] Step 2: Using the extracted LL, HL, LH, and HH sub-band information, perform inverse discrete wavelet transform to obtain the actual secret image.

[0056] Among them, an encryption network is used to encrypt the preprocessed carrier image cover_pre and the frequency domain transformed secret image secret_pre_dwt into a secret image steg containing secret image feature information.

[0057] It should be noted that the encrypted network consists of 1 input convolution layer, 11 dense blocks, 5 downsampling layers, 5 upsampling layers and 1 output convolution layer;

[0058] As an embodiment, the encrypted network structure is:

[0059] First layer: Use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to convolve the input 6-channel image to complete shallow feature extraction and obtain a shallow feature image.

[0060] The second to sixth layers are composed of 5 dense blocks (DB) and 5 downsampling layers (TD) for downsampling to extract deep image information.

[0061] Layer 7: A dense block (DB) is used to modify the number of image channels output by the sixth layer to facilitate subsequent upsampling operations.

[0062] Layer 8-Layer 12: It consists of 5 dense blocks (DB) and 5 upsampling layers (TU) for upsampling to enhance the spatial resolution of the feature map so as to better capture details and boundary information in the steganography task, thereby improving the accuracy and detail capture capability of the secret image.

[0063] Thirteenth layer: Use a convolution layer with a convolution kernel size of 1×1 and a stride of 1 to convolve the upsampled feature map to ensure that the output is a three-channel secret image.

[0064] As an embodiment, the process of encrypting the secret image steg generated by the network is as follows:

[0065] S111: Get the carrier natural image cover and the secret remote sensing image secret;

[0066] S112: using the preprocessing network to process the carrier natural image and the secret remote sensing image of S201 to obtain the preprocessed carrier image cover_pre and the preprocessed secret image secret_pre, so that the network can learn the contour information of the two images, which is beneficial to the feature fusion of the encryption network and the feature extraction of the decryption network;

[0067] S113: The secret image pre-processed in S202 is subjected to an image frequency domain transformation module to obtain a frequency domain transformed secret image secret_pre_dwt;

[0068] S114: connect the carrier image cover_pre output by S202 and the frequency domain transformed secret image secret_pre_dwt output by S203 in the channel dimension, and then input them into the encryption network for encryption;

[0069] S115: Use a convolution layer with a convolution kernel size of 3×3 and a step size of 1 to convolve the input 6-channel image to complete shallow feature extraction and obtain a shallow feature image.

[0070] S116: Downsampling is performed using 5 dense blocks (DB) and 5 downsampling layers (TD) to extract deep image information.

[0071] S117: A dense block (DB) is used to modify the number of image channels output by the sixth layer to facilitate subsequent upsampling operations.

[0072] S118: It consists of 5 dense blocks (DB) and 5 upsampling layers (TU) for upsampling to enhance the spatial resolution of the feature map so as to better capture details and boundary information in the steganographic task, thereby improving the accuracy and detail capture capability of the secret image.

[0073] S119: Use a convolution layer with a convolution kernel size of 1×1 and a stride of 1 to convolve the upsampled feature map to ensure that the output is a three-channel secret image steg.

[0074] Among them, the decryption network is used to decrypt the secret image to obtain the decrypted image (decode), and then the decrypted image is inversely transformed in the frequency domain to obtain the final extracted image (reveal).

[0075] It should be noted that the decryption network consists of 6 separable convolution blocks and 2 convolution layers, 2 normalization layers and 2 ELU activation functions;

[0076] As an embodiment, the decryption network structure is:

[0077] The first layer to the sixth layer: use 6 separable convolution blocks (SCR) to extract features. Using SCR through multi-layer convolution operations and nonlinear activation functions can increase network depth, reduce the number of parameters, improve computational efficiency and model performance, enhance feature extraction capabilities, and help the network learn more discriminative features.

[0078] The seventh layer: consists of 2 convolutional layers, 2 normalization layers and 2 ELU activation functions;

[0079] Layer 8: Use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to convolve the feature map to obtain a 3-channel output image.

[0080] Among them, the image restoration network restores the secret image and the decrypted image respectively.

[0081] It should be noted that the image restoration network consists of two parts: a pre-decryption restoration network and a post-decryption restoration network. The pre-decryption restoration network consists of: 2 convolutional layers, 2 spatial channel reconstruction convolutional layers SCConv and a ReLU activation function; the post-decryption restoration network consists of: 5 convolutional layers, 4 dropout layers and 4 LeakyReLU layers.

[0082] As an embodiment, the network structure restored before decryption is:

[0083] First layer: Use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to convolve the feature map to obtain a shallow feature image.

[0084] Second layer: A RELU activation function is used.

[0085] The third layer: Use the spatial channel reconstruction convolution layer (ScConv), which combines the spatial reconstruction unit (SRU) and the channel reconstruction unit (CRU) to improve the network's perception and expression capabilities of image features.

[0086] SRU normalizes the input tensor x to reduce the scale difference between different feature maps, and then separates the informative features (x_1) and the non-informative features (x_2) by thresholding. The 4-channel feature map of x_1 is split into a feature map containing the first two channels (x_11) and a feature map containing the last two channels (x_12), and the 4-channel feature map of x_2 is split into a feature map containing the first two channels (x_21) and a feature map containing the last two channels (x_22), forming 4 sub-feature sets. Then the first two channels of x_1 are split into The channel feature map x_11 is added to the last two channel feature maps x_22 of x_2 to obtain a new feature map (x_a), and then the last two channel feature maps x_12 of x_1 are added to the first two channel feature maps x_21 of x_2 to obtain a new feature map (x_b), and then x_a and x_b are spliced ​​in the channel dimension to obtain the final output tensor (x_sru), and the separated feature sets are transformed and reorganized to generate refined spatial feature outputs for further processing, thereby improving the efficiency of convolutional neural network processing features;

[0087] CRU takes the output tensor x_sru of the above SRU as input, first uses the splitting method to split x_sru into upper channel features (up) and lower channel features (low) in the channel dimension, divides the features into two parts, and performs the next step of extraction. Then, the convolution layer is used to compress up and low respectively, and the compressed up is subjected to group convolution and point convolution to further extract features, and the results of group convolution and point convolution are added to obtain the extracted features (Y1), and the compressed low is subjected to point convolution to further extract features, and the point convolution result is concatenated with the compressed low. Get the extracted feature (Y2), concatenate Y1 and Y2 in the channel dimension to get the concatenated feature (out), perform adaptive average pooling on out, and pass the Softmax activation function. Then split the activated out into two tensors out1 and out2 in the channel dimension. Then add out1 and out2 element by element, and return the added result (x_cru). x_cru aggregates the information of the front and back parts of out, effectively integrating the features of the first 2 channels and the last 2 channels of out, which is conducive to further processing of the features in the next task.

[0088] The fourth layer: Use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to convolve the feature map to obtain a further feature image.

[0089] The fifth layer: Use the spatial channel reconstruction convolution layer (ScConv), which combines the spatial reconstruction unit (SRU) and the channel reconstruction unit (CRU) to improve the network's perception and expression capabilities of image features.

[0090] The sixth layer: Use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to convolve the feature map to obtain the final output image.

[0091] The network structure restored after decryption is:

[0092] First layer: Use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to convolve the input x0 feature map to obtain a shallow feature image, then activate it through the LeakyReLU activation function, call the dropout function to reduce overfitting, and obtain x1.

[0093] Second layer: concatenate the input x0 and x1 in the channel dimension to get x 01 , use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to perform feature mapping on the feature map x 01 Perform convolution, then use the LeakyReLU activation function to operate, call the dropout function to reduce overfitting, and get x2.

[0094] The third layer: concatenate the input x0, x1 and x2 in the channel dimension to get x 012 , use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to perform feature mapping on the feature map x 012 Perform convolution, then use the LeakyReLU activation function to operate, call the dropout function to reduce overfitting, and get x3.

[0095] Fourth layer: concatenate the input x0, x1, x2 and x3 in the channel dimension to get x 0123 , use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to perform feature mapping on the feature map x 0123 Perform convolution, then use the LeakyReLU activation function to operate, call the dropout function to reduce overfitting, and get x4.

[0096] The fifth layer: input x0, x1, x 2、 x3 and x4 are concatenated in the channel dimension to obtain x 01234 , use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to perform feature mapping on the feature map x 01234 Perform convolution, then use the LeakyReLU activation function to operate, call the dropout function to reduce overfitting, and get x5 for output.

[0097] As an embodiment, the specific process of the decryption network generating the decrypted image decode and generating the final extracted image reveal is as follows:

[0098] S121: In the lossy process, the pre-decryption recovery network needs to be used to perform the first step of restoring the steg generated in S209 to obtain the restored image steg_restoration.

[0099] S122: Use 6 separable convolution blocks (SCR) to process steg_restoration for feature extraction.

[0100] S123: Enhanced using 2 convolutional layers, 2 normalization layers, and 2 ELU activation functions.

[0101] S124: Use a convolution layer with a convolution kernel size of 3×3 and a stride of 1 to convolve the feature map to obtain a 3-channel output decrypted image decode.

[0102] S125: After decrypting the decrypted image decode in S123, the network is restored to obtain decode_restoration.

[0103] S126: Perform frequency domain inverse transform on the decode_restoration image of S124 to obtain the final extracted image reveal.

[0104] S2, training the improved remote sensing image steganography network;

[0105] It should be noted that the training method in step S2 is as follows:

[0106] S21, perform the first round of training under lossless conditions, alternately iteratively training the preprocessing network, the encryption network, and the decryption network;

[0107] S22. After the preprocessing network, the encryption network, and the decryption network converge, the image restoration network is trained under lossy conditions to restore the secret image generated by the trained encryption network and the decrypted image generated by the trained decryption network, respectively.

[0108] As an example, the loss calculation of the entire network can be expressed as formula (5):

[0109] loss=loss l1_steg_cover +loss l1_secret_reveal +loss l1_decode_secret_pre_dwt +loss l1_steg_steg_restoration +loss mov_steg_cover +loss mov_secret_reveal +loss mov_decode_secret_pre_dwt +loss mov_steg_steg_restoration (5)

[0110] In formula (5), loss l1_steg_cover and lossmov_steg_cover are the mean of the L1 loss and variance of the encrypted image generated by the encryption network and the input carrier image, respectively. l1_secret_reveal and loss mov_secret_reveal are the mean of the L1 loss and variance of the extracted image generated by the inverse transform in the frequency domain and the input secret image, respectively. l1_decode_secret_pre_dwt and loss mov_decpde_secret_pre_dwt are the mean of the L1 loss and variance of the decrypted image generated by the decryption network and the secret image transformed in the frequency domain, respectively. l1_steg_steg_restoration and loss mov_steg_steg_restoration They are respectively the mean of the L1 loss and variance of the encrypted image generated by the encryption network and the restored image after being processed by the recovery network before decryption.

[0111] S3. Use the trained network to perform remote sensing image steganography.

[0112] As an example, please refer to Figure 3 , Figure 3 This is an effect diagram generated by the remote sensing image steganography network of the present invention. The initial input images of the flowchart are the carrier image cover and the secret image secret. Then, cover and secret are respectively subjected to the preprocessing network to obtain the preprocessed carrier image cover_pre and the preprocessed secret image secret_pre. Secret_pre is subjected to the frequency domain transformation module to obtain the frequency domain transformed secret image secret_pre_dwt. Then, cover_pre and secret_pre_dwt are connected in the channel dimension and sent to the encryption network to obtain the secret image steg. Steg is passed through the pre-decryption restoration network to obtain the pre-decryption restored image steg_restoration. Steg_restoration is input into the decryption network to obtain the decrypted image decode. Decode is passed through the post-decryption restoration network to obtain the post-decryption restored image decode_restoration. Finally, decode_restoration is passed through the frequency domain inverse transformation module to obtain the final extracted image reveal.

[0113] See also Figure 4 , Figure 4 40 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, and the hardware device specifically includes: a remote sensing image steganography device 401 based on deep learning, a processor 402 and a storage medium 403.

[0114] A remote sensing image steganography device 401 based on deep learning: The remote sensing image steganography device 401 based on deep learning implements the remote sensing image steganography method based on deep learning.

[0115] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the remote sensing image steganography method based on deep learning.

[0116] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the remote sensing image steganography method based on deep learning.

[0117] The beneficial effects of the present invention are:

[0118] 1. The improved attention mechanism preprocessing network introduced in the present invention can effectively extract the contour information of the carrier image and the secret image, which is conducive to the hiding of the secret image; the use of the multi-layer perceptron layer (Mlp) introduced therein greatly improves the values ​​of image PSNR, SSIM and other indicators while ensuring the hiding effect; the final addition operation of the network forms a residual connection, so that the image after the preprocessing network retains its original image information when its contour information is extracted for the subsequent network processing.

[0119] 2. The introduced improved frequency domain transformation module can effectively weaken the outline of the secret image, allowing it to be better hidden in the carrier image and improve the quality of the decrypted image.

[0120] 3. Introduce the pre-decryption recovery network and the post-decryption recovery network before the decryption network to restore the pixel loss caused by image rounding in practical applications, ensuring that the final decrypted image is similar to the original secret image;

[0121] In general, the method of the present invention can not only generate a high-quality secret image, but also effectively decrypt the secret image information contained therein in practical applications.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A remote sensing image steganography method based on deep learning, characterized by: The method comprises the following steps: S1. Construct an improved remote sensing image steganography network; the improved remote sensing image steganography network is based on the FCDensenet convolutional neural network architecture, specifically including: a preprocessing network, an image frequency domain transformation module, an encryption network, a decryption network and an image restoration network; S2, training the improved remote sensing image steganography network; S3. Use the trained network to perform remote sensing image steganography.

2. A remote sensing image steganography method based on deep learning as claimed in claim 1, characterized in that: The training method in step S2 is as follows: S21, perform the first round of training under lossless conditions, alternately iteratively training the preprocessing network, the encryption network, and the decryption network; S22. After the preprocessing network, the encryption network, and the decryption network converge, the image restoration network is trained under lossy conditions to restore the secret image generated by the trained encryption network and the decrypted image generated by the trained decryption network, respectively.

3. A remote sensing image steganography method based on deep learning as claimed in claim 1, characterized in that: The preprocessing network obtains a secret image secret and a carrier image cover, and outputs a preprocessed secret image secret_pre and a preprocessed carrier image cover_pre, and its structure includes two parts: a channel attention module and a spatial attention module.

4. A remote sensing image steganography method based on deep learning as claimed in claim 3, characterized in that: The channel attention module consists of: 1 global maximum pooling layer, 1 global average pooling layer, 2 multi-layer perceptron layers and 1 Sigmod activation function; The spatial attention module consists of: 1 global maximum pooling layer, 1 global average pooling layer, 1 convolutional layer and 1 Sigmod activation function.

5. The remote sensing image steganography method based on deep learning as claimed in claim 1, characterized in that: The image frequency domain transformation module transforms the preprocessed secret image secret_pre to obtain a frequency domain transformation secret image secret_pre_dwt, whose structure is composed of an improved discrete wavelet transform, including two parts: a forward transformation and an inverse transformation.

6. The remote sensing image steganography method based on deep learning as claimed in claim 1, characterized in that: The encrypted network consists of 1 input convolutional layer, 11 dense blocks, 5 downsampling layers, 5 upsampling layers and 1 output convolutional layer; The decryption network consists of 6 separable convolutional blocks and 2 convolutional layers, 2 normalization layers and 2 ELU activation functions; The image restoration network consists of two parts: a pre-decryption restoration network and a post-decryption restoration network. The pre-decryption restoration network consists of: 2 convolutional layers, 2 spatial channel reconstruction convolutional layers SCConv and a ReLU activation function; the post-decryption restoration network consists of: 5 convolutional layers, 4 dropout layers and 4 LeakyReLU layers.

7. A remote sensing image steganography method based on deep learning as claimed in claim 4, characterized in that: The multi-layer perceptron layer consists of a flatten layer, a first fully connected layer, a Linear layer, a nonlinear activation function layer, a ReLU layer, and a second fully connected layer, a Linear layer.

8. The remote sensing image steganography method based on deep learning as claimed in claim 1, characterized in that: The specific processing process of the image frequency domain module is as follows: Forward transformation part: Step 1: Perform discrete wavelet transform (DWT) on the input image. The input image is decomposed into four sub-bands: LL, HL, LH, and HH, which represent the frequency domain information in the horizontal-horizontal, horizontal-vertical, vertical-horizontal, and vertical-vertical directions respectively; Step 2: By filling the LL, HL, LH, HH sub-bands into a tensor according to the following filling rules, the image hiding is promoted. The specific filling rules are shown in formulas (1)(2)(3)(4): h[:,∶,0∷2,0∷2]=x LL (1) h[:,∶,1∷2,0∷2]=x HL (2) h[:,∶,0∷2,1∷2]=x LH (3) h[:,∶,1∷2,1∷2]=x HH (4) The fusion of the four sub-bands is achieved through matrix indexing and slicing operations according to the following filling rules, and the four sub-bands LL, HL, LH, and HH after two-dimensional discrete wavelet transform decomposition are filled into the h image; in the above description, h is the filled image, the size of h image is equal to the size of the original secret image, and the shape of h is four dimensions, the first dimension is the batch size, the second dimension is the number of channels, the third dimension is the height, and the fourth dimension is the width. The size of the four sub-bands LL, HL, LH, and HH is one-fourth of the size of the original secret image, and also one-fourth of the size of the h image; Specifically, the first two ':' in formula (1) select all batch sizes and all dimension sizes of h. The first 0::2 selects every other element starting from 0 in the height dimension of h, that is, selects all even-indexed elements. The second 0::2 selects all even-indexed elements in the width dimension of h, that is, fills the subband LL into the even rows and even columns of h. In formula (2), all batch sizes and all dimension sizes of h are selected. The first 1::2 selects every other element starting from 1 in the height dimension of h, that is, selects all elements with odd indexes. The second 0::2 selects all elements with even indexes in the width dimension of h, that is, fills the subband HL into the odd rows and even columns of h. Formula (3) is to fill the sub-band LH into the even rows and odd columns of h, and formula (4) is to fill the sub-band HH into the odd rows and odd columns of h; Inverse transformation part: Step 1: Extract the LL, HL, LH, and HH sub-band information from the decrypted image tensor according to the filling rule, that is, extract LL from the even rows and even columns of the decrypted image, extract HL from the odd rows and even columns, extract LH from the even rows and odd columns, and extract HH ​​from the odd rows and odd columns; Step 2: Using the extracted LL, HL, LH, and HH sub-band information, perform inverse discrete wavelet transform to obtain the actual secret image.

9. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a remote sensing image steganography method based on deep learning as described in any one of claims 1 to 8.

10. A remote sensing image steganography device based on deep learning, characterized in that: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a remote sensing image steganography method based on deep learning as described in any one of claims 1 to 8.