A large-capacity image hiding method and system based on a deep convolutional neural network
Patent Information
- Application Number
- CN202410028376.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-08
- Publication Date
- 2026-10-30
- Estimated Expiration
- 2044-01-08
AI Technical Summary
虽然基于深度学习的图像隐藏方法取得了一定的成就，但是在视觉不可见性和隐藏容量方面仍有待提高
[0038]本发明提出的一种基于深度卷积神经网络的大容量图像隐藏方法，包含了两个预处理模块，一个图像隐藏网络和提取网络，实现1幅载体图像中同时有效隐藏和提取2幅大尺寸秘密图像。在秘密图像被隐藏前，预处理模块对载体图像和秘密图像分别进行处理，提高隐藏特征提取效率同时减少网络参数数量。为了更好地提取和融合多尺度、多区域的特征，图像隐藏网络和提取网络中引入U型残差注意力模块、并行双通道注意力机制模块、混连的空洞卷积模块和基于Pixelshuffle的金字塔池化模块，保证图像质量的同时提高隐藏容量。实验结果表明，本发明提出的方法相对隐藏容量可达800％，同时具有较好的视觉质量。与相同隐藏容量的方法相比，本发明方法在视觉质量上有显著提高，载体图像的平均PSNR和SSIM分别提高了2.46dB和1.52％。多个数据集的测试结果表明本发明方法具有良好的泛化能力。
Smart Images
Figure CN117880431B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image hiding technology, and relates to a high-capacity image hiding method and system based on deep convolutional neural networks. Background Technology
[0002] With the rapid development of information technology and the widespread adoption of smart devices, the amount of information disseminated on the internet has exploded. While people enjoy the convenience brought by digital information, information security is seriously threatened. Image hiding technology is one of the technical means to protect information security.
[0003] Early image information hiding methods based on the spatial and transform domains always left traces of modification in the embedded image after embedding secret information. This made the embedded image vulnerable to statistical image steganalysis, and also suffered from capacity limitations and image quality issues. However, with the rapid development and application of deep learning technology in various fields, including information security, these methods have become more sophisticated.
[0004] In recent years, researchers have proposed solutions for hiding and extracting secret information using convolutional networks. Although deep learning-based image hiding methods have achieved some success, there is still room for improvement in terms of visual invisibility and hiding capacity. Summary of the Invention
[0005] The purpose of this invention is to solve the problems in the prior art and provide a high-capacity image hiding method and system based on deep convolutional neural networks.
[0006] To achieve the above objectives, the present invention employs the following technical solution:
[0007] This invention proposes a high-capacity image hiding method based on deep convolutional neural networks, comprising the following steps:
[0008] Constructing image hiding networks and image extraction networks;
[0009] Image preprocessing module for acquiring secret images and image preprocessing module for carrier images;
[0010] Based on the image preprocessing module of the secret image and the image preprocessing module of the carrier image, the preprocessed feature map is obtained;
[0011] Weight files are obtained based on preprocessed feature maps, image hiding networks, and image extraction networks to achieve image hiding.
[0012] Preferably, the image hiding network consists of a main network, a parallel dual attention mechanism, and a convolutional layer 1 connected in series.
[0013] Preferably, the parallel dual attention mechanism consists of a spatial attention mechanism and a channel attention mechanism connected in parallel;
[0014] The weights W of the channel attention mechanism c (P) is represented as follows:
[0015] W c (P)=σ(L(Avg c (p)+Max c (p)))
[0016] Among them, Avg c For average pooling downsampling in the channel attention mechanism, Max c For max pooling downsampling in the channel attention mechanism, L is a fully connected layer and σ is the sigmoid activation function;
[0017] Weights of spatial attention mechanisms s (P) is represented as follows:
[0018] W s (P)=σ(Conv(Avg s (p)+Max s (p)))
[0019] Among them, Avg s Max represents the average pooling downsampling in the spatial attention mechanism. s This represents max pooling downsampling in the spatial attention mechanism, and Conv represents convolution.
[0020] Feature map P and channel attention mechanism weights W c (P), weights W of the spatial attention mechanism s (P) Weighted multiplication yields the final output feature map.
[0021] Preferably, the main network consists of 8 RSAU modules, 2 hybrid dilated convolutional modules, and a pyramid pooling module with Pixelshuffle introduced, connected in series and skipped.
[0022] Preferably, two dilated convolutional submodules are cascaded and skipped to form a hybrid dilated convolutional module. The output of the hybrid dilated convolution is represented as follows:
[0023] MDCM=DCM2(DCM1(X)+X)+DCM1(X)+X
[0024] Where X is the input of the hybrid dilated convolution, MDCM is the output of the hybrid dilated convolution, DCM1 is the output of dilated convolution module 1, and DCM2 is the output of dilated convolution module 2.
[0025] The dilated convolutional submodule is composed of convolutional layer 2, batch normalization 1, ReLU activation function 1, convolutional layer 3, batch normalization 1, ReLU activation function 1, convolutional layer 4, batch normalization 1, and ReLU activation function 1 connected in series.
[0026] Preferably, the pyramid pooling module structure that introduces Pixelshuffle is as follows: the pyramid pooling scale is set to 1×1, 2×2, 4×4 and 8×8, and the internal processing order of the pyramid pooling module is adjusted to max pooling, Pixelshuffle upsampling and convolution, and then concatenated.
[0027] The RSAU module consists of three parts: downsampling encoding, upsampling decoding, parallel dual attention mechanism, and skip connections.
[0028] Preferably, the image extraction network is composed of extraction network 1 and extraction network 2 connected in parallel; both extraction network 1 and extraction network 2 are composed of a pixel shuffle layer, a parallel dual attention mechanism, and a convolutional layer 1 connected in series.
[0029] The image preprocessing module for the carrier image is preprocessing module 1, which consists of convolutional layer 2, batch normalization layer 1 and ReLU activation function 1 connected in series. The image preprocessing module for the secret image is preprocessing module 2, which consists of max pooling layer and average pooling layer connected in parallel. The outputs of the max pooling layer and average pooling layer are concatenated and then connected in series with convolutional layer 2, batch normalization layer 1 and ReLU activation function 1.
[0030] This invention proposes a high-capacity image hiding system based on a deep convolutional neural network, comprising:
[0031] A network structure building module is used to construct an image hiding network and an image extraction network.
[0032] A preprocessing module is used to acquire the image preprocessing module of the secret image and the image preprocessing module of the carrier image;
[0033] The image preprocessing module is used to obtain a preprocessed feature map based on the image preprocessing module of the secret image and the image preprocessing module of the carrier image.
[0034] The weight file acquisition module is used to acquire a weight file based on the preprocessed feature map, image hiding network and image extraction network to achieve image hiding.
[0035] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a large-capacity image hiding method based on a deep convolutional neural network.
[0036] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a large-capacity image hiding method based on a deep convolutional neural network.
[0037] Compared with the prior art, the present invention has the following beneficial effects:
[0038] This invention proposes a high-capacity image hiding method based on deep convolutional neural networks, comprising two preprocessing modules, an image hiding network, and an extraction network. This method effectively hides and extracts two large-sized secret images simultaneously from a single carrier image. Before the secret images are hidden, the preprocessing module processes the carrier image and the secret images separately, improving the efficiency of hidden feature extraction while reducing the number of network parameters. To better extract and fuse multi-scale and multi-region features, a U-shaped residual attention module, a parallel dual-channel attention mechanism module, a hybrid dilated convolution module, and a pixelshuffle-based pyramid pooling module are introduced into the image hiding and extraction networks, ensuring image quality while increasing hiding capacity. Experimental results show that the proposed method achieves a relative hiding capacity of up to 800% while maintaining good visual quality. Compared with methods of the same hiding capacity, the proposed method significantly improves visual quality, with the average PSNR and SSIM of the carrier image increasing by 2.46 dB and 1.52%, respectively. Test results on multiple datasets demonstrate that the proposed method has good generalization ability.
[0039] This invention proposes a high-capacity image hiding system based on a deep convolutional neural network. The system is divided into a network structure construction module, a preprocessing module, an image preprocessing module, and a weight file acquisition module to achieve image hiding. The modular approach ensures that each module is independent, facilitating unified management of all modules. Attached Figure Description
[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0041] Figure 1 This is a flowchart of the large-capacity image hiding method based on deep convolutional neural networks according to the present invention.
[0042] Figure 2 This is an algorithmic framework diagram of the large-capacity image hiding method based on deep convolutional neural networks of the present invention.
[0043] Figure 3 This is a structural diagram of the preprocessing module of the present invention.
[0044] Figure 4 This is a process diagram of the image hiding network and extraction network of the present invention.
[0045] Figure 5 This is a structural diagram of the RASU of the present invention.
[0046] Figure 6 This is a block diagram of the parallel dual attention mechanism of the present invention.
[0047] Figure 7 This is a structural diagram of the hybrid dilated convolution and dilated convolution module of the present invention.
[0048] Figure 8 This is a diagram of the improved pyramid pooling module of the present invention.
[0049] Figure 9 This is the result of the generalization test of the present invention.
[0050] Figure 10 This is a diagram of the high-capacity image hiding system based on a deep convolutional neural network according to the present invention. Detailed Implementation
[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0052] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0053] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0054] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper," "lower," "horizontal," or "inner" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, terms such as "first" and "second" are only used to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0055] Furthermore, the use of the term "horizontal" does not imply that the component must be absolutely horizontal, but rather that it can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted.
[0056] In the description of the embodiments of the present invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.
[0057] The present invention will now be described in further detail with reference to the accompanying drawings:
[0058] This invention proposes a high-capacity image hiding method based on deep convolutional neural networks, such as... Figure 1 As shown, it includes the following steps:
[0059] S1. Construct an image hiding network and an image extraction network;
[0060] S2, Image preprocessing module for acquiring secret images and image preprocessing module for carrier images;
[0061] S3. An image preprocessing module based on the secret image and an image preprocessing module based on the carrier image, to obtain the preprocessed feature map;
[0062] S4. Obtain weight files based on preprocessed feature maps, image hiding networks, and image extraction networks to achieve image hiding.
[0063] Preprocessing module 1 consists of convolutional layer 2, batch normalization layer 1, and ReLU activation function 1 connected in series. Preprocessing module 2 consists of max pooling layer and average pooling layer connected in parallel. The outputs of max pooling layer and average pooling layer are concatenated and then connected in series with convolutional layer 2, batch normalization layer 1, and ReLU activation function 1.
[0064] The image extraction network consists of extraction network 1 and extraction network 2 connected in parallel; both extraction network 1 and extraction network 2 are composed of a pixel shuffle layer, a parallel dual attention mechanism, and a convolutional layer 1 connected in series.
[0065] The image hiding network consists of a main network, a parallel dual attention mechanism, and a convolutional layer 1 connected in series.
[0066] The parallel dual attention mechanism consists of a spatial attention mechanism and a channel attention mechanism connected in parallel.
[0067] The weights W of the channel attention mechanism c (P) is represented as follows:
[0068] W c (P)=σ(L(Avg c (p)+Max c (p)))
[0069] Weights of spatial attention mechanisms s (P) is represented as follows:
[0070] W s (P)=σ(Conv(Avg s (p)+Max s (p)))
[0071] Among them, Avg c For average pooling downsampling in the channel attention mechanism, Max c In the channel attention mechanism, max pooling downsampling is used, where L is a fully connected layer and σ is the sigmoid activation function; Avg s Max represents the average pooling downsampling in the spatial attention mechanism. s This represents max pooling downsampling in the spatial attention mechanism, and Conv represents convolution.
[0072] Feature map P and channel attention mechanism weights W c (P), weights W of the spatial attention mechanism s (P) Weighted multiplication yields the final output feature map.
[0073] The main network consists of 8 RSAU modules, 2 hybrid dilated convolutional modules, a pyramid pooling module with Pixelshuffle introduced, and skip connections.
[0074] Two dilated convolutional submodules are cascaded and skipped to form a hybrid dilated convolutional module. The output of the hybrid dilated convolution is represented as follows:
[0075] MDCM=DCM2(DCM1(X)+X)+DCM1(X)+X
[0076] Where X is the input of the hybrid dilated convolution, MDCM is the output of the hybrid dilated convolution, DCM1 is the output of dilated convolution module 1, and DCM2 is the output of dilated convolution module 2.
[0077] The dilated convolutional submodule consists of convolutional layer 2, batch normalization 1, ReLU activation function 1, convolutional layer 3, batch normalization 1, ReLU activation function 1, convolutional layer 4, batch normalization 1, and ReLU activation function 1 connected in series.
[0078] The pyramid pooling module structure with Pixelshuffle is as follows: the pyramid pooling scale is set to 1×1, 2×2, 4×4 and 8×8, and the internal processing order of the pyramid pooling module is adjusted to max pooling, Pixelshuffle upsampling and convolution, and finally concatenated.
[0079] The RSAU module consists of three parts: downsampling encoding, upsampling decoding, parallel dual attention mechanism, and skip connections.
[0080] This invention proposes a large-capacity image hiding method based on multi-feature fusion, the overall framework of which is as follows: Figure 2 As shown. To improve network training efficiency and increase steganography capacity, both the carrier image and the secret image are preprocessed before embedding the secret image into the carrier image. Then, an image hiding network is used for steganography to obtain the hidden image. Two extraction networks are used to extract the secret image from the hidden image. The preprocessing module, image hiding network, extraction network structure, and loss function will be described in detail below.
[0081] The specific steps are as follows:
[0082] 1) Preprocessing module
[0083] Before embedding the secret image into the carrier image, both the carrier and secret images are preprocessed separately, i.e., the RGB image is converted into a feature image, and then input into the image hiding network. This improves the training speed of the network and increases the steganalysis capacity. Figure 3As shown, secret image preprocessing involves concatenating two secret images along the channel dimension and then performing max pooling and average pooling respectively. Max pooling extracts high-frequency features from the secret image, while average pooling extracts low-frequency features. Fusing these two types of features increases the efficiency of secret image feature extraction and significantly improves the network's representational power. The resulting feature maps are then concatenated along with the carrier image and fed into a deep convolutional neural network consisting of 3×3 convolutions, BatchNorm layers, and ReLU activation functions.
[0084] 2) Constructing image hiding networks and image extraction networks
[0085] The image hiding network consists of a main network, a parallel dual attention mechanism, and a convolutional layer 1 connected in series. The main network consists of 8 RSAU modules, 2 hybrid dilated convolutional modules, a pyramid pooling module with pixel shuffle introduced, and skip connections connected in series.
[0086] The image extraction network consists of extraction network 1 and extraction network 2 connected in parallel; both extraction network 1 and extraction network 2 are composed of a pixel shuffle layer, a parallel dual attention mechanism, and a convolutional layer 1 connected in series.
[0087] The image preprocessing module for the carrier image is preprocessing module 1, which consists of convolutional layer 2, batch normalization layer 1, and ReLU activation function 1 connected in series. The image preprocessing module for the secret image is preprocessing module 2, which consists of max pooling layer and average pooling layer connected in parallel. The outputs of the max pooling layer and average pooling layer are concatenated and then connected in series with convolutional layer 2, batch normalization layer 1, and ReLU activation function 1.
[0088] Among them, the structures of image hiding networks and image extraction networks are as follows: Figure 4 As shown, Figure 4 In (a), they all contain the same main network structure, such as Figure 4 In (b), the main network is primarily composed of RSAU modules. RSAU introduces a parallel attention mechanism into the residual network to acquire features at different scales and fuse them effectively. Its structure is as follows: Figure 5 As shown, Figure 5 In the diagram, L represents the layer number of the RSAU module. The RSAU module consists of three parts: downsampling encoding, upsampling decoding, and skip connections. The left side represents the downsampling encoding part, which obtains multi-scale features and a larger receptive field. The right side represents the upsampling decoding part, which concatenates the high-level features from the encoding part along the channel dimension via skip connections. Finally, a parallel dual-channel attention mechanism is fed into the end of the module. The RSAU module can fuse low-level and high-level features, taking into account both local feature information and global contextual information.
[0089] In the image hiding process, the input consists of the carrier image and the secret image, whose features are obtained through a preprocessing module. The left side of the main network in the image hiding network is the downsampling encoding process, which consists of six stages. The first four stages are composed of RSAU modules with layer parameters L of 7, 6, 5, and 4, respectively. The last two stages are composed of a hybrid dilated convolutional module and a pyramid pooling module based on PixelShuffle. Skip connections are used between RSAU modules to maximize the retention of more image features and improve feature utilization. A pyramid module based on PixelShuffle is used in the center of the main network to better fuse feature information from different regions and improve the network's ability to acquire global multi-scale information. The right side of the main network is the upsampling decoding process, where the structure of each module is the same as the symmetrical position on the left side, and the connections between each level of RSAU module are as follows: Figure 4 As shown in b, the RSAU module performs upsampling via pixel shuffle after concatenation along the channel dimension, and all upsampled results are concatenated for output. A parallel dual-attention mechanism is added after each module in the encoding and decoding stages to increase the weights beneficial to steganalytic features and decrease the weights of features unimportant to steganalytic features. Finally, the feature tensors output by each module in the decoding stage are upsampled to the same size as the carrier image via pixel shuffle and then concatenated along the channel dimension to obtain the final feature tensor output by the main network. This is then passed through the parallel dual-attention mechanism and convolution again to obtain the dense image. The image hiding process can be represented as follows:
[0090] C'=H(C,S1,S2)=Conv(PDAM(MN(Pre1(C)+Pre2(S1+S2))))
[0091] Where C and C' represent the carrier image and the hidden image, respectively; S1 and S2 represent secret image 1 and secret image 2, respectively; and H represents the image hiding network. Pre1 and Pre2 represent the preprocessing modules, respectively; and MN, PDAM, and Conv represent the main network, the parallel dual attention mechanism, and the convolutional layer, respectively.
[0092] During the extraction process, the input secret image is fed into two extraction networks with identical structures. Taking one extraction network as an example, the secret image extraction process is illustrated. First, the secret image is passed through the main network to obtain a series of feature tensors. Then, these feature tensors are sequentially passed through pixel shuffle upsampling and a parallel dual attention mechanism to obtain one of the secret images. Although the two extraction networks have the same structure, their output targets differ during training, resulting in different parameters after training. The secret information extraction process can be represented as follows:
[0093] S1'=R1(C')=Conv(PDAM(UP(MN(C'))))
[0094] S2'=R2(C')=Conv(PDAM(UP(MN(C'))))
[0095] S1' and S2' represent the extracted secret image 1 and secret image 2, respectively. R1 and R2 represent the two extraction networks. MN, PDAM, Conv, and UP represent the main network, parallel dual attention mechanism, convolutional layer, and pixel shuffle upsampling, respectively.
[0096] The parallel dual attention mechanism is described in detail below:
[0097] To improve feature extraction efficiency and feature utilization, this invention proposes a parallel dual-attention mechanism module, which connects spatial attention and channel attention mechanisms in parallel. Unlike existing serial attention mechanism structures, the parallel dual-attention mechanism takes into account both spatial and channel dimensions. They can learn the features of the original input feature map directly without affecting each other, thus enabling more accurate calculation of the feature weights in the hidden stage.
[0098] The parallel dual attention mechanism module is shown in Figure 6. To improve feature extraction efficiency and aggregate feature information from different spaces, Figure 6 The upper channel attention mechanism module first performs max pooling downsampling and average pooling downsampling on the input feature P in the spatial dimension, changing the size of the feature tensor from H×W×C to 1×1×C. Then, the two feature tensors are input into the fully connected layer to obtain the feature weights in the two spaces. After adding them together, the result is passed through the sigmoid activation function to finally obtain the channel attention weight W. c (P). The process is described below:
[0099] W c (P)=σ(L(Avg c (p)+Max c (p)))
[0100] Among them, Avg c Max c The first two sections represent average pooling downsampling and max pooling downsampling in the channel attention mechanism, respectively. L is a fully connected layer, and σ is the sigmoid activation function. Below is the spatial attention mechanism, which compresses the input feature P along the channel dimension using both max pooling and average pooling, generating two H×W×1 feature tensors. These two feature tensors are then concatenated and fed into a 5×5 convolutional layer to optimize the weights, finally yielding the spatial attention weights. The process is shown below:
[0101] W s (P)=σ(Conv(Avgs (p)+Max s (p)))
[0102] Among them, Avg s Max s represents average pooling downsampling and max pooling downsampling in the spatial attention mechanism, and Conv represents convolution.
[0103] Finally, the feature map P and the two weights W c (P), W s (P) Weighted multiplication yields the final output feature map.
[0104] The hybrid dilated convolutional modules are described in detail below:
[0105] Because the feature map obtained after several downsampling iterations is relatively small, it is easy to lose a significant amount of feature information during image hiding and extraction. Therefore, this invention utilizes dilated convolution to compensate for this loss, such as... Figure 7 As shown. To avoid mesh effects (refer to Hybrid Dilated Convolution (HDC)), the dilation coefficient d of each dilated convolutional module is set to 1, 2, and 3 respectively, as follows. Figure 7 As shown in (b) above. Considering the influence of the dilation coefficient on the feature information of large and small targets in the image, the algorithm of this invention concatenates and skips the two dilated convolution sub-modules to form a hybrid dilated convolution module, as shown in (b). Figure 7 As shown in (a) above, the output of the hybrid dilated convolution is represented as follows:
[0106] MDCM=DCM2(DCM1(X)+X)+DCM1(X)+X
[0107] Where X is the input of the hybrid dilated convolution, MDCM is the output of the hybrid dilated convolution, DCM1 is the output of dilated convolution module 1, and DCM2 is the output of dilated convolution module 2. The hybrid dilated convolution increases the receptive field to obtain information from a larger region while reducing the number of parameters, making the network more lightweight. The cascaded and skip connections effectively integrate features from both shallow and deep layers.
[0108] The specific description of introducing the pyramid pooling module of Pixelshuffle is as follows:
[0109] To fuse global feature information throughout the process and avoid the loss of contextual information, this invention introduces Pixelshuffle as an upsampling method in the pooling module and borrows the idea of pyramid pooling for feature fusion, such as... Figure 8As shown. Pixelshuffle is introduced because it effectively preserves image information, improves image resolution, maintains consistency with the image structure, and incurs no additional computational burden. Furthermore, the algorithm in this invention modifies the original pyramid pooling scale to 1×1, 2×2, 4×4, and 8×8. To ensure consistency with the input feature map size of this module, the internal processing order is adjusted to max pooling, pixelshuffle upsampling, and convolution, followed by concatenation. The appropriate pooling scale and the introduction of pixelshuffle upsampling better concatenate feature vectors from different receptive fields along the channel dimension without increasing computational cost, thus producing higher-quality images.
[0110] 3) Training the network
[0111] 3.1) Determine the objective function
[0112] During the training of the image hiding network and the image extraction network, the total loss L total Includes hidden loss L H and extraction loss L R They are represented by mean square error (MSE), which is defined as follows.
[0113]
[0114] In the above formula, I and I' represent two images, i and j represent the horizontal and vertical coordinates of the images, and H and W represent the height and width of the images. The total loss L... total Hidden losses and L H and extraction loss L R The definitions are as follows.
[0115] L total =L H +θL R
[0116]
[0117]
[0118] Total loss L total In the definition of L, θ is a hyperparameter. H and L RIn the definition, N represents the number of groups into which the carrier image and secret image are divided in each batch during training, C represents the carrier image, and C' represents the secret image. C and C' are the original secret images, and S1' and S2' are the extracted secret images. The hiding loss is used to calculate the error between the carrier image and the secret image, thus constraining the parameter updates of the hiding network. The extraction loss is used to calculate the error between the secret image and the extracted secret image, thus constraining the parameter updates of both the hiding and extraction networks.
[0119] 3.2) Training the network
[0120] The training set was input into the hidden network for training. The Adam optimizer was used during the training process. The initial learning rate of the network model was 0.001, the initial value of the hyperparameter θ was 0.75, and the number of training epochs was 1000.
[0121] 4) Save the model and test the image hiding network.
[0122] During the training of a large-capacity image hiding network based on multi-feature fusion, the weights of the network model are continuously updated and the weight file is saved. The test set is input into the trained hiding network for testing, and the saved weight file is loaded to obtain the image hiding and extraction results.
[0123] This invention uses the COCO dataset for model training, validation, and testing. This dataset contains 328,000 images, from which 30,000 are randomly selected for training and 3,000 for validation testing. The secret image is cropped to 512×512 pixels, and the carrier image is cropped to 256×256 pixels; both images are randomly generated. The experimental hardware platform consists of an Intel Xeon E5-2620 v4 CPU, an NVIDIA TITAN XP 10G GPU, and Ubuntu 20.04 operating system. The deep learning framework PyTorch version is 1.9.1, and the Python version is 3.9. The network model was trained with an initial learning rate of 0.001, 1000 epochs, and an initial hyperparameter θ of 0.75. The optimizer chosen was Adam.
[0124] This invention employs an adversarial training method, similar to GANs, where the hiding network and the extraction network compete against each other. The hiding network aims to obtain a hidden image that most closely resembles the carrier image, while the extraction network aims to obtain an extracted image that most closely resembles the secret image. The hidden image output by the hiding network is fed into the extraction network, and after calculating the corresponding hiding and extraction losses, backpropagation is performed, thus forming a closed loop between the hiding and extraction networks. This adversarial training method reduces the number of features fused into the carrier image, resulting in an extracted image that more closely resembles the secret image.
[0125] 4.1) Evaluation Indicators
[0126] One objective metric for measuring the effectiveness of hidden images and the quality of extracting secrets is the Peak Signal-to-Noise Ratio (PSNR), which measures the difference between pixels in two images. The PSNR value between images I and I' can be calculated using a formula, where MSE is the mean squared error.
[0127]
[0128] Where n represents the number of bits in the image.
[0129] Structural similarity (SSIM) measures the differences between two images in terms of brightness, contrast, and structure, and is defined as follows.
[0130]
[0131] Where μ represents the average pixel value of the image, σ 2 Represents the variance, and c1 and c2 are two constants.
[0132] The hidden capacity of an image is measured using relative capacity, and the calculation process is as follows:
[0133]
[0134] Where AC is the absolute capacity and SC is the size of the carrier image.
[0135] 4.2) Image hiding effect
[0136] As shown in Table 1, the average PSNR values of the dense images and extracted images generated by the algorithm of this invention in the COCO dataset are both above 35dB, indicating good image quality and visual invisibility. Test results are as follows... Figure 9 As shown, (1) is a 256*256 carrier image, (2) is a 256*256 secret image, (3) is a 512*512 secret image 1, (4) is a 512*512 extracted image 1, (5) is a 512*512 secret image 2, and (6) is a 512*512 extracted image 2.
[0137] Table 1. Relative hidden capacity and objective evaluation indicators of the algorithm of this invention.
[0138]
[0139] This invention proposes a high-capacity image hiding system based on deep convolutional neural networks, such as... Figure 10As shown, it includes a network structure building module, a preprocessing module, an image preprocessing module, and a weight file acquisition module;
[0140] The network structure building module is used to construct image hiding networks and image extraction networks;
[0141] The acquisition preprocessing module is used to acquire the image preprocessing module of the secret image and the image preprocessing module of the carrier image;
[0142] The image preprocessing module is used to obtain a preprocessed feature map based on the image preprocessing module of the secret image and the image preprocessing module of the carrier image.
[0143] The weight file acquisition module is used to acquire a weight file based on the preprocessed feature map, image hiding network, and image extraction network to achieve image hiding.
[0144] The terminal device provided in this embodiment of the invention includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.
[0145] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.
[0146] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.
[0147] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
[0148] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.
[0149] If the modules / units integrated into the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.
[0150] This invention proposes a high-capacity image hiding method based on multi-feature fusion. The network structure includes two preprocessing modules, one hiding network, and two extraction networks, enabling the simultaneous and effective hiding and extraction of two large-sized secret images from a single carrier image. Before the secret images are hidden, the preprocessing module processes the carrier image and the secret images separately, improving the efficiency of hidden feature extraction while reducing the number of network parameters. To better handle multi-scale and multi-region features, a U-shaped residual attention module, a parallel dual-channel attention mechanism module, a hybrid dilated convolution module, and a pixelshuffle-based pyramid pooling module are introduced into the hiding and extraction networks, ensuring image quality while increasing the hiding capacity. Experimental results show that the algorithm achieves a relative hiding capacity of up to 800% while maintaining good visual quality. Compared with methods with the same hiding capacity, the algorithm significantly improves visual quality, with average PSNR and SSIM of the carrier image increasing by 2.46 dB and 1.52%, respectively. Test results on multiple datasets demonstrate that the algorithm exhibits good generalization ability.
[0151] In the convolutional residual network, a U-shaped residual attention module, a parallel dual-channel attention mechanism module, a hybrid dilated convolution module, and a pixelshuffle-based pyramid pooling module are introduced to fuse multi-layer features, thereby improving hidden capacity while ensuring image quality. Therefore, the main contribution of this invention is:
[0152] 1) Adding a preprocessing module to the hidden network improves the feature extraction efficiency of the carrier image and the secret image, as well as the steganalysis capacity.
[0153] 2) A hybrid dilated convolution module is proposed to reduce network parameters while avoiding loss of more feature information, and to fuse shallow and deep information features.
[0154] 3) Pixelshuffle is introduced into the pooling module for upsampling. Leveraging the concept of pyramid pooling, a pyramid pooling module based on pixelshuffle is proposed. This better integrates features at different scales without increasing network computation. The design of this module significantly improves the network's hiding and extraction capabilities.
[0155] 4) A parallel dual attention mechanism module was proposed and introduced into the residual network to construct U-shaped residual attention modules (RSAU) to better extract features, thereby improving the network's steganalysis and extraction capabilities.
[0156] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A high-capacity image hiding method based on deep convolutional neural networks, characterized in that, Includes the following steps: Constructing image hiding networks and image extraction networks; Image preprocessing module for acquiring secret images and image preprocessing module for carrier images; Based on the image preprocessing module of the secret image and the image preprocessing module of the carrier image, the preprocessed feature map is obtained; Weight files are obtained based on preprocessed feature maps, image hiding networks, and image extraction networks to achieve image hiding; The image hiding network consists of a main network, a parallel dual attention mechanism, and a convolutional layer 1 connected in series; the parallel dual attention mechanism is composed of a spatial attention mechanism and a channel attention mechanism connected in parallel. Weights of channel attention mechanism The representation is as follows: The weight of spatial attention mechanism The repetition is represented as follows: Feature map Weights of the channel attention mechanism Weights of spatial attention mechanisms The final output feature map is obtained by weighted multiplication. The main network consists of 8 RSAU modules, 2 hybrid dilated convolutional modules, and a pyramid pooling module with pixel shuffle introduced, all connected in series and with skip connections. Each RSAU module comprises downsampling encoding, upsampling decoding, a parallel dual attention mechanism, and skip connections. The two dilated convolutional sub-modules are cascaded and skip-connected to form a hybrid dilated convolutional module. The output of the hybrid dilated convolution is represented as follows: The dilated convolutional submodule consists of convolutional layer 2, batch normalization 1, ReLU activation function 1, convolutional layer 3, batch normalization 1, ReLU activation function 1, convolutional layer 4, batch normalization 1, and ReLU activation function 1 connected in series; the pyramid pooling module with pixel shuffle introduced has the following structure: the pyramid pooling scale is set to 1×1, 2×2, 4×4, and 8×8, and the internal processing order of the pyramid pooling module is adjusted to max pooling, pixel shuffle upsampling, and convolution, and then concatenated. in, This refers to average pooling downsampling in the channel attention mechanism. This is a max-pooling downsampling mechanism in the channel attention mechanism. It is a fully connected layer. It is the sigmoid activation function; This represents average pooling downsampling in the spatial attention mechanism. This represents max pooling downsampling in the spatial attention mechanism. Represents convolution; For the input of the concatenated dilated convolution, The output of the concatenated dilated convolution. This is the output of dilated convolution module 1. This is the output of dilated convolution module 2.
2. The high-capacity image hiding method based on deep convolutional neural networks according to claim 1, characterized in that, The image extraction network consists of extraction network 1 and extraction network 2 connected in parallel; both extraction network 1 and extraction network 2 are composed of a pixel shuffle layer, a parallel dual attention mechanism, and a convolutional layer 1 connected in series. The image preprocessing module for the carrier image is preprocessing module 1, which consists of convolutional layer 2, batch normalization layer 1 and ReLU activation function 1 connected in series. The image preprocessing module for the secret image is preprocessing module 2, which consists of max pooling layer and average pooling layer connected in parallel. The outputs of the max pooling layer and average pooling layer are concatenated and then connected in series with convolutional layer 2, batch normalization layer 1 and ReLU activation function 1.
3. A high-capacity image hiding system based on a deep convolutional neural network, characterized in that, The large-capacity image hiding method based on deep convolutional neural networks as described in claim 1 includes: A network structure building module is used to construct an image hiding network and an image extraction network. A preprocessing module is used to acquire the image preprocessing module of the secret image and the image preprocessing module of the carrier image; The image preprocessing module is used to obtain a preprocessed feature map based on the image preprocessing module of the secret image and the image preprocessing module of the carrier image. The weight file acquisition module is used to acquire a weight file based on the preprocessed feature map, image hiding network and image extraction network to achieve image hiding.
4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes a computer program, it implements the steps of the large-capacity image hiding method based on a deep convolutional neural network as described in any one of claims 1 to 2.
5. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the large-capacity image hiding method based on a deep convolutional neural network as described in any one of claims 1 to 2.