A social network channel-oriented image distortion generation method and medium

Abstract

The application relates to a social network channel-oriented image distortion generation method and medium, which comprises the following steps: inputting a to-be-processed image into a pre-trained real channel noise simulator, outputting a distortion image of a simulated real channel processing result, and training steps comprising the following steps: acquiring original clean images and channel degradation images to constitute a data set; inputting the original clean images into a conditional feature extraction network to extract channel conditional features and predict image compression quality parameters; inputting the channel conditional features and random latent variables into a cascaded noise generation network to generate high-frequency sensitive degradation disturbance before compression, and fusing the high-frequency sensitive degradation disturbance with the original clean images to obtain a pre-compression distortion image, and then obtaining a distortion image after real channel processing based on the image compression quality parameters and the pre-compression distortion image; and constructing a loss function to iteratively train the real channel noise simulator until the training is completed. Compared with the prior art, the application has the advantages of improving stability and the like.

Description

Technical Field

[0001] This invention relates to the fields of multimedia content security and image processing, and in particular to a method and medium for generating image distortion for social network channels. Background Technology

[0002] With the widespread use of social networking platforms, instant messaging software, and cloud-based image services, images typically undergo complex real-world channel processing during actual transmission. For example, during uploading, forwarding, downloading, or re-encoding, images may be affected by a combination of factors, including size scaling, resampling, color space conversion, JPEG compression re-encoding, quantization truncation, high-frequency information loss, and platform-adaptive compression strategies. These processing steps are often not single, fixed, or completely transparent image degradation operations, but rather exhibit platform-dependent, content-dependent, and nonlinear coupling characteristics. In existing technologies, common image noise simulation methods typically use Gaussian noise with fixed parameters, salt-and-pepper noise, blurring, random cropping, or JPEG compression with a fixed quality factor to approximate real-world channel distortion. While these methods are simple to implement, they struggle to accurately reflect the complex degradation characteristics in the real image transmission chain. On one hand, real-world platforms may employ different compression qualities, resampling strategies, and encoding parameters for different image content; on the other hand, there is a coupling relationship between spatial domain degradation and frequency domain compression degradation, making it difficult for a single noise model to simultaneously describe smooth degradation, color shift, blockiness, high-frequency truncation, and local texture damage.

[0003] Furthermore, while some methods can learn image degradation maps through neural networks, they typically treat the real-world channel as a black-box end-to-end transformation, lacking structured modeling of spatial continuous distortion and frequency domain compression distortion. This results in insufficient model interpretability and limited generalization ability across different platforms, compression strategies, or image content. For tasks such as digital watermarking, image steganography, image restoration, image forensics, content authentication, and robust visual model training, if the noise model used during training is inconsistent with the real-world channel distortion distribution, the model's performance in actual propagation environments will significantly degrade. Therefore, there is an urgent need for an image noise simulator capable of jointly modeling spatial degradation, frequency domain compression, high-frequency loss, and content adaptive compression behavior in real-world channels, tailored to real image propagation links. This simulator should be able to learn the real-world channel distortion patterns from the original image and its corresponding channel degradation image, and generate simulated degradation images that are more consistent with the results of real-world channel processing. This would provide a more reliable real-world channel simulation environment for subsequent applications such as robust image training, multimedia content security, image restoration, digital watermarking, and image forensics. Summary of the Invention

[0004] The purpose of this invention is to provide an image distortion generation method and medium for social network channels that simulates and closely resembles real social channels and has high distortion restoration accuracy.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for generating image distortion for social network channels includes the following steps: The image to be processed is input into a pre-trained realistic channel noise simulator, and the output is a distorted image simulating the processing result of a realistic channel. The realistic channel noise simulator includes a conditional feature extraction network and a cascaded noise generation network, and its training steps include: The original clean image and the corresponding channel-degraded image after passing through the platform channel are obtained to form a dataset, wherein the channel-degraded image is used as the ground truth label; The original clean image is input into the conditional feature extraction network to extract the channel conditional features related to the actual channel distortion, and to predict the image compression quality parameters corresponding to the actual channel. The channel condition features and random latent variables are input into a cascaded noise generation network to generate a high-frequency sensitive degradation perturbation before compression. The high-frequency sensitive degradation perturbation before compression is then fused with the original clean image to obtain a pre-compressed distorted image. Finally, based on the image compression quality parameters and the pre-compressed distorted image, a distorted image after actual channel processing is obtained. Based on the difference between the distorted image and the channel degradation image, a loss function is constructed to iteratively train the real channel noise simulator until training is complete.

[0006] Furthermore, the step of obtaining the channel condition features includes: For the first in the same batch A clean original image Channel-level high-frequency residual extraction is performed to obtain high-frequency residual features, which are represented as follows: In the formula, For the first A clean original image Extracted high-frequency residual features, This is a high-frequency residual extraction module; The first A clean original image High-frequency residual characteristics Concatenation is performed along the channel dimension to obtain the fused input features, represented as follows: In the formula, To fuse input features, This indicates a channel-dimensional splicing operation; The fused input features The input is fed into a deep feature fusion network to obtain the first... A clean original image The corresponding channel condition characteristics are expressed as follows: In the formula, For the first A clean original image Corresponding channel condition characteristics, This represents a deep feature fusion network.

[0007] Furthermore, the step of obtaining the image compression quality parameters includes: Global feature aggregation is performed on the aforementioned channel condition features to obtain a global channel representation, which is expressed as: In the formula, For global channel characterization, This indicates a global feature aggregation operation. For the first A clean original image Corresponding channel condition characteristics; The global channel characterization is input into the compression quality parameter prediction module to obtain normalized quality parameters, which are expressed as: In the formula, To normalize the quality parameters, This indicates the compression quality parameter prediction module. These are the learnable parameters for the compression quality parameter prediction module; Let the preset compression quality range be The normalized quality parameters are mapped to the compression quality range to obtain the image compression quality parameters, which are expressed as follows: In the formula, For the first A clean original image The corresponding image compression quality parameters, , These represent the upper and lower limits of compression quality.

[0008] Furthermore, the cascaded noise generation network includes a spatial distortion flow network, a frequency domain perturbation flow network, and a differentiable image compression module. The spatial distortion flow network is used to compress the original clean image. Channel condition characteristics and random latent variables As input, the output is a spatially degraded image. Continuous degradation disturbance in space and spatial stage latent variables The operational expression for the spatially distorted flow network is: In the formula, Representation spatial distortion flow network; The frequency domain perturbation flow network is used to degrade the spatial phase image. Channel condition characteristics Spatial stage latent variables and image compression quality parameters As input, output high-frequency sensitive degradation perturbation before compression. and pre-compressed distorted images The operational expression for the frequency domain perturbation flow network is as follows: In the formula, Represents a frequency domain perturbation flow network; The differentiable image compression module is used to compress the spatially degraded image. High-frequency sensitive degradation perturbation before compression and image compression quality parameters As input, the output is a distorted image simulating real-world channel processing in the current iteration round, where the operational expression of the differentiable image compression module is: In the formula, The image is distorted. This indicates a differentiable image compression module.

[0009] Furthermore, the spatial distortion flow network includes multiple cascaded conditionally reversible flow modules, and the execution steps of each conditionally reversible flow module include: Let the first The input to the conditionally reversible flow module is After passing through the reversible channel replacement layer The result was: In the formula, The input after the substitution; Will Divided along the channel dimension into first spatial feature components Second spatial characteristic components ,Right now: In the formula, Indicates concatenation of channel dimensions; Based on the second spatial feature component and channel condition characteristics Generates the first spatial feature component The scale parameters and translation parameters are determined, and the first output feature component is obtained. The first spatial feature component The scale parameters, translation parameters, and first output feature components They are represented as follows: In the formula, , These are respectively applied to the first spatial feature components. The scale parameters and translation parameters, Represents the scaling subnetwork. Represents the translation transformation subnetwork. Represents the scaling constraint function; Based on the first spatial feature component and channel condition characteristics Generates the feature components that act on the second space. The scale parameters and translation parameters are obtained, and the second output feature components are obtained. The second spatial feature component The scale parameters, translation parameters, and second output feature components They are represented as follows: In the formula, , These are the eigencomponents acting on the second space. The scale parameters and translation parameters, Represents the scaling subnetwork. Represents the translation transformation subnetwork; The first output feature component Second output feature component Perform channel splicing to obtain the first... The output of a conditionally reversible flow module is represented as: In the formula, For the first The output of the conditionally reversible flow module, i.e. the first... The input of a conditional reversible flow module.

[0010] Furthermore, the frequency domain perturbation flow network includes multiple conditionally reversible flow modules based on KAN nonlinear channel mapping, and the execution steps of each conditionally reversible flow module based on KAN nonlinear channel mapping include: Let the first The input of a conditionally reversible flow module based on KAN nonlinear channel mapping is via reversible channel replacement layer The result was: In the formula, The input after the substitution; Will Divided along the channel dimension into the first frequency domain feature components Second frequency domain characteristic components ,Right now: In the formula, Indicates concatenation of channel dimensions; Based on the first frequency domain feature component Second frequency domain characteristic components and channel condition characteristics Generate the first frequency domain output feature component. , is represented as: In the formula, Represents the scale constraint function. , These are the eigencomponents acting on the first frequency domain. The scale parameters and translation parameters, and These represent the scale parameter generation unit and translation parameter generation unit based on KAN nonlinear channel mapping, respectively. Based on the first frequency domain output feature components Channel condition characteristics The components are then spliced ​​together and processed by the scale parameter generation unit and translation parameter generation unit based on KAN nonlinear channel mapping to obtain the second frequency domain output feature components. , is represented as: In the formula, , These are the eigencomponents acting on the second frequency domain. The scale parameters and translation parameters; The first frequency domain output feature component Second frequency domain output characteristic components By concatenating along the channel dimension, we obtain the [number]. The output of a conditionally reversible flow module based on KAN nonlinear channel mapping is expressed as: In the formula, For the first The output of the conditionally reversible flow module based on KAN nonlinear channel mapping, i.e., the... The input to a conditionally reversible flow module based on KAN nonlinear channel mapping.

[0011] Furthermore, the execution steps of the conditionally reversible flow module based on KAN nonlinear channel mapping also include: performing nonlinear mapping on the channel vector at each spatial location, specifically including: After obtaining the first frequency domain output feature component At that time, the second frequency domain feature component Channel condition characteristics The concatenation result is used as input for the nonlinear mapping. After obtaining the second frequency domain output feature component At that time, the first frequency domain feature component Channel condition characteristics The concatenation result is used as input for the nonlinear mapping. ; Take the input In spatial location The channel vector is represented as: In the formula, For spatial location The channel vector at that location; right A nonlinear transformation based on a learnable spline function is performed to obtain the output channel vector, which is expressed as: In the formula, This is the output channel vector. This represents a channel-dimensional nonlinear mapping composed of learnable basis functions; Based on all spatial locations The corresponding output channel vector yields the output features of the nonlinear mapping. The output features The scale parameter , or translation parameters , The output features Used to obtain the first frequency domain output feature component through radiometric transformation Or the second frequency domain output characteristic components To obtain the first The output of a conditionally reversible flow module based on KAN nonlinear channel mapping .

[0012] Furthermore, the loss function is: in: Intermediate state monitoring loss for: Final image pixel domain loss for: Frequency domain high-frequency subband loss for: Perceived consistency loss for: Predicted loss of compressive mass parameters for: In the formula, For the total loss, , , , , , These are the weighting coefficients. For spatial stage degradation images, This is the smoothed intermediate degraded image. This indicates a downsampling operation. Indicates an upsampling operation. For channel degradation images, The image is distorted. This indicates the extraction of horizontal and vertical high-frequency subbands from the wavelet transform. This indicates the extraction of diagonal high-frequency subbands from the wavelet transform. This represents the perceptual distance calculated based on a pre-trained perceptual feature extraction network. and These represent the distorted images. and channel degradation image The image obtained by mapping the numerical range [0,1] to the numerical range [-1,1]. Predicted values ​​for image compression quality parameters. This represents the true value of the image compression quality parameter.

[0013] Furthermore, the trained real-world channel noise simulator is integrated into the training process of the downstream visual model. During training, the parameters of the real-world channel noise simulator are frozen and backpropagated to the downstream visual model via gradients from the simulator's output. The downstream visual model then outputs the reconstructed image under clean transmission conditions. Input a realistic channel noise simulator to obtain a reconstructed image output under simulated realistic channel disturbance conditions. And construct a robust training loss for the downstream model: In the formula, Loss due to robust training The channel covariance matrix, It is the Frobenius norm. The total number of feature points, For batch size, , These represent the height and width of the latent space noise disturbance characteristics, respectively. To obtain the centered channel feature matrix , To characterize latent space noise disturbances, by... Expanding along the batch and spatial dimensions yields the channel feature matrix. , This indicates a latent space encoder.

[0014] The present invention also provides a storage medium storing a program thereon, which, when executed, implements an image distortion generation method for social network channels as described above.

[0015] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention learns platform-specific channel condition features from the original clean image through a condition feature extraction network and predicts the image compression quality parameters. Then, it decouples three types of distortions—spatial color shift, nonlinear truncation, and JPEG quantization—using a two-level serial stream network in the spatial and frequency domains. It models the disturbances in a hierarchical and independent manner, resulting in smaller errors between the simulated distorted image and the real platform channel degradation image, and highly restoring the composite degradation effect of online transmission.

[0016] (2) This invention constructs a learnable noise generation network for real-world image transmission links by learning the correspondence between the original image and the real-world channel degradation image. This network can simulate complex degradation processes such as uploading to social networking platforms, compression recoding, scaling resampling, color shifting, quantization truncation, block artifacts, and high-frequency information loss. Compared with traditional fixed noise or fixed quality factor JPEG compression methods, this invention decomposes real-world channel degradation into two stages: spatial continuous distortion and frequency domain compression degradation. It also improves the realism and interpretability of noise simulation by jointly modeling the spatial distortion stream network, the frequency domain compression stream network, and the differentiable image compression module.

[0017] (3) The spatial distortion flow network of this invention independently models two types of spatial domain perturbations: spatial color shift and pixel nonlinear truncation. The frequency domain perturbation flow network generates quantized frequency domain perturbations based on the predicted image compression quality parameters, i.e., high-frequency sensitive degradation perturbations before compression. The three types of distortion components are modeled independently in layers and superimposed step by step to generate simulated distorted images. This decoupling method can accurately fit the degradation law of real images on social platforms and improve simulation accuracy; each distortion component is independently adjustable to adapt to the differentiated channels of multiple social platforms; gradients do not cross conflict, and training converges faster and more stably; some distortions can be enabled separately as needed to adapt to various downstream tasks such as steganography and watermarking; the independent modules facilitate iterative optimization and have stronger generalization ability.

[0018] (4) This invention introduces a high-frequency residual condition perception and compression quality parameter prediction mechanism, enabling the real-world channel noise simulator to adaptively generate channel condition features and compression control parameters based on the input image content. This real-world channel noise simulator can independently generate degraded images that closely resemble the processing results of a real platform, and can also be used as a differentiable real-world channel module to access tasks such as digital watermarking, image steganography, image restoration, image forensics, content authentication, and robust visual model training, thereby improving the stability, robustness, and generalization ability of downstream models in actual propagation scenarios. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the overall network structure of the present invention; Figure 3 These are the noise simulation results at each stage of this invention; Figure 4 The following are comparison results of the application of the present invention to image steganography, wherein (a) is the secret image after passing through the Facebook channel, (b) is the real OSN noise, (c) is the original secret image, (d) is the recovered secret image after training, and (e) is the recovered secret image after noise-free training. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0021] Example 1 This embodiment provides an image distortion generation method for social network channels, which improves the realism, interpretability, and generalization ability of image distortion simulation in real-world channels. It addresses the problem that existing fixed-noise models struggle to accurately simulate complex degradation processes such as social platform uploads, image recoding, compression quality changes, scaling and resampling, color shifts, and high-frequency losses, thereby enhancing the model's noise resistance in real-world channels. Specifically, it combines... Figure 2 The overall network structure shown is as follows: Figure 1 As shown, the method includes the following steps: (a) Collect a dataset consisting of channel degradation images and original clean images that have passed through the corresponding platform channel. Input the original clean images into a conditional feature extraction network to extract conditional features related to the actual channel distortion and predict the image compression quality parameters corresponding to the actual channel. The conditional feature extraction network obtains channel conditional features to guide subsequent noise generation through high-frequency residual extraction and deep feature fusion.

[0022] In this embodiment, the channel-degraded image refers to the image after channel degradation by the corresponding platform, which serves as the truth label; the original clean image refers to the original clear image without noise and without channel degradation.

[0023] The conditional feature extraction network is shown in Table 1.

[0024] Table 1: Conditional Feature Extraction Network Structure Table Among them, C cond The number of channels is a characteristic of the channel conditions; in one embodiment, it is taken as C. cond=64. The conditional feature extraction network first performs high-frequency residual filtering on the three color channels (RGB) of the input image, resulting in 30 high-frequency residual channels for each channel. Then, these 90 high-frequency residual channels are concatenated with the original RGB image to obtain a 93-channel joint feature. After convolutional fusion and residual dense feature extraction, this joint feature outputs the channel conditional feature h. cond Meanwhile, the compression quality parameter prediction module predicts h. cond Global average pooling and fully connected mapping are performed to output normalized quality parameters, which are then mapped to a preset compression quality range to obtain the predicted image compression quality parameters (QF).

[0025] The specific process of step (a) is as follows: (a1) Let the input original clean image be ,in Includes multiple color channels; This indicates the high-frequency residual extraction module, which then processes the input image. Channel-level high-frequency residual extraction is performed to obtain high-frequency residual features: , in, This represents the original clean image as input. The residual information obtained after high-frequency residual filtering of each color channel is used to characterize image components that are sensitive to texture, edge and high-frequency details during the real channel degradation process.

[0026] (a2) Let This indicates a channel-dimensional splicing operation, allowing To represent a deep feature fusion network, the input image is... High-frequency residual characteristics The concatenation is performed along the channel dimension and then input into a deep feature fusion network to obtain channel condition features: , in, This represents channel condition characteristics related to actual channel degradation. This represents the learnable parameters of a deep feature fusion network.

[0027] (a3) Let This indicates a global feature aggregation operation. The compression quality parameter prediction module then considers channel condition characteristics. Perform global aggregation to obtain the global channel representation: .

[0028] Furthermore, the global channel representation Input the compression quality parameter prediction module to obtain the normalized quality parameters: , in, ∈[0,1], This represents the learnable parameters of the compression quality parameter prediction module.

[0029] (a4) Set the preset compression mass range to Then the normalized quality parameters will be... Mapping to a preset compression quality range yields the predicted compression quality parameters: .

[0030] For the first in the same batch Input image First, the high-frequency residual features are obtained through the high-frequency residual extraction module: .

[0031] Then, the input image High-frequency residual characteristics By concatenating the data along the channel dimension, the fused input features are obtained: .

[0032] Furthermore, the input features will be fused. Inputting the deep feature fusion network yields the first... Channel condition features corresponding to each input image: .

[0033] Again Perform global feature aggregation and image compression quality parameter prediction to obtain the first... Compression quality parameters of the predicted image corresponding to each input image: .

[0034] Therefore, the conditional feature extraction network can generate corresponding channel conditional features for different input images. and predicted image compression quality parameters To adapt to the image content-related adaptive compression behavior in real-world platform channels.

[0035] (b) The random latent variables and channel condition features are input into a cascaded noise generation network, which then passes through a spatial distortion flow network and a frequency domain perturbation flow network to generate a pre-compression high-frequency sensitive degradation perturbation. The spatial distortion flow network is used to output the spatial stage latent variables that characterize the spatial continuous distortion, and the frequency domain perturbation flow network is used to generate the pre-compression high-frequency sensitive degradation perturbation based on the spatial stage latent variables. The pre-compression high-frequency sensitive degradation perturbation is fused with the original clean image to obtain a pre-compressed distorted image, and the final distorted image after simulating the real channel is obtained by passing it through a differentiable image compression module according to the predicted image compression quality parameters.

[0036] The spatial distortion flow network structure and the frequency domain compressed flow network are shown in Table 2 and Table 3, respectively.

[0037] Table 2: Spatial Distortion Flow Network Structure Table Where, N s N represents the number of conditionally reversible flow modules in a spatially distorted flow network, which is taken as N in practical applications. s =6. Spatial distortion flow networks first map random latent variables to the wavelet domain using discrete wavelet transform; then, based on channel condition characteristics h... cond Guided by this approach, inverse flow mapping is performed through multiple conditionally reversible flow modules to obtain spatial stage latent variables. Then, spatial perturbation components are obtained through inverse wavelet transform, and these perturbation components are added to the input image to generate an intermediate distorted image. This module is mainly used to simulate spatial continuous distortions such as scaling, resampling, color shifting, and smoothing degradation in real-world channels.

[0038] Table 3: Frequency Domain Compressed Flow Network Structure Table Where, N f N represents the number of KAN conditionally reversible flow modules in a frequency-domain compressed flow network. f =4. The frequency-domain compressed streaming network receives intermediate latent variables from the output of the spatially distorted streaming network, rather than resampling new random variables, thus maintaining the continuity between the spatial degradation stage and the frequency-domain degradation stage. The KAN channel-dimensional nonlinear mapping unit in the frequency-domain compressed streaming network is used to perform nonlinear mapping on the multi-channel feature vectors at each spatial location without unfolding the spatial dimension, thereby enhancing the model's ability to fit high-frequency texture variations, compression-sensitive regions, and local nonlinear degradation.

[0039] The structure of the KAN channel-dimensional nonlinear mapping unit is shown in Table 4 below.

[0040] Table 4: Structure of KAN channel-dimensional nonlinear mapping unit. The execution steps of the cascaded reversible flow mapping network include: first, sampling random latent variables with the same size as the input image, and performing discrete wavelet transform on the random latent variables to obtain low-frequency and high-frequency components of the latent variables; then, inputting the low-frequency and high-frequency components of the latent variables into the spatial distortion flow network, performing inverse reversible flow mapping under the constraints of channel condition characteristics to generate spatial perturbation components, and superimposing the spatial perturbation components onto the input image to obtain an intermediate degraded image; simultaneously, the spatial distortion flow network outputs spatial stage latent variables; subsequently, inputting the spatial stage latent variables into the frequency domain compression flow network, continuing inverse reversible flow mapping under the constraints of channel condition characteristics and predicted compression quality parameters to generate pre-compression high-frequency sensitive degradation perturbations; finally, superimposing the pre-compression high-frequency sensitive degradation perturbations onto the input image to obtain a pre-compressed distorted image, and performing compression recoding processing on the pre-compressed distorted image through a differentiable image compression unit to obtain the final distorted image after simulating real-world channel processing. Specifically, the detailed process of step (b) is as follows: (b1) Let the spatially distorted flow network be Frequency domain compressed stream network is The microimage compression module is ;in, The learnable parameters of spatially distorted flow networks are represented. Let represent the learnable parameters of the frequency domain compressed stream network. Let the input image be... The channel condition characteristics are The random latent variable is The predicted image compression quality parameters are: First, , and Input spatial distortion stream network to obtain spatial stage degradation image Continuous degradation disturbance in space and spatial stage latent variables : .

[0041] Among them, spatial distortion stream network This is used to simulate spatial continuous degradation in real-world channels, which includes at least one of scaling, resampling, color shifting, and smoothing degradation.

[0042] (b2) Spatial Distortion Flow Network It includes multiple cascaded conditional reversible flow modules. Each conditional reversible flow module includes a conditional affine coupling layer and a reversible channel permutation layer. The conditional affine coupling layer fuses the channel condition features with the input features in the channel dimension and generates corresponding scaling transformation parameters and translation transformation parameters to simulate spatial continuous distortions such as scaling, resampling, color shifting and smoothing degradation in real channels.

[0043] Order No. The input characteristics of the conditionally reversible flow module are: After passing through the reversible channel replacement layer The result was: .

[0044] Will Divided along the channel dimension into first spatial feature components Second spatial characteristic components ,Right now: .

[0045] in, This indicates concatenation of channel dimensions.

[0046] (b3) Order , Represents the scaling subnetwork. , Represents the translation transformation subnetwork. Represents the scale constraint function. Based on the second spatial feature components. and channel condition characteristics The scale parameter and translation parameter acting on the first spatial feature component are generated, and the first output feature component is obtained. : , , .

[0047] Furthermore, based on the first output feature component and channel condition characteristics The scale and translation parameters acting on the second spatial feature components are generated, and the second output feature components are obtained. : , , .

[0048] Will and Perform channel splicing to obtain the first... Output characteristics of a conditionally reversible flow module: .

[0049] (b4) After multiple conditionally reversible flow modules are cascaded, the spatially distorted flow network outputs a spatially degraded image. and spatial stage latent variables .Will Channel condition characteristics Spatial stage latent variables and predicted image compression quality parameters Input frequency domain compressed stream network Generate high-frequency sensitive degradation perturbations before compression : .

[0050] Based on the high-frequency sensitive degradation perturbation before compression Obtain a pre-compressed distorted image : .

[0051] Among them, frequency domain compressed stream network This method is used to simulate high-frequency truncation, frequency-domain nonlinear attenuation, and pre-compression artifact tendencies in real-world channels. To constrain the spatial distortion flow network to primarily learn continuous spatial degradation features and avoid directly fitting high-frequency quantization artifacts generated in the subsequent compression stage, it can output images based on the actual channel conditions. Construct intermediate degradation monitoring targets Specifically, the output image of the real channel. The image is downsampled and then upsampled back to its original size to obtain a smoothed intermediate degraded image: , in, This indicates a downsampling operation. This indicates an upsampling operation. Through downsampling and upsampling, the intermediate degraded image... It preserves continuous spatial degradation information such as scaling, resampling, color shift, and smoothing degradation in the real channel output image, while reducing high-frequency compression degradation information such as JPEG compression, quantization noise, and block artifacts.

[0052] (b5) Frequency domain compressed stream network It includes multiple conditionally reversible flow modules based on KAN nonlinear channel mapping. Each conditionally reversible flow module based on KAN nonlinear channel mapping includes a conditional affine coupling layer, a reversible channel permutation layer, and a channel-dimensional nonlinear mapping unit. The execution principle of this conditional affine coupling layer and reversible channel permutation layer is the same as that of the conditional affine coupling layer and reversible channel permutation layer in the spatial distortion flow network. The channel-dimensional nonlinear mapping unit is used to approximate the channel information of the input features with a nonlinear function to simulate frequency domain discrete distortions such as recoding, high-frequency truncation, and block effects in real channels.

[0053] For the A frequency-domain conditionally reversible current module, whose input characteristics are... via reversible channel replacement layer The result was: .

[0054] Will Divided along the channel dimension into the first frequency domain feature components Second frequency domain characteristic components : .

[0055] make and Let represent the scale parameter generation unit and translation parameter generation unit based on KAN nonlinear channel mapping, respectively. Then: , , .

[0056] In the formula, The first frequency domain output feature components; Furthermore, and The concatenation is then passed again through a parameter generation unit based on KAN nonlinear channel mapping to obtain: , , .

[0057] In the formula, The second frequency domain output feature components; Will and By concatenating along the channel dimension, we obtain the [number]. Output characteristics of a frequency-domain conditionally reversible current module: .

[0058] (b6) The KAN nonlinear channel mapping is used to perform a nonlinear mapping of the channel vector at each spatial location. Let the input features be... ,in Indicates the number of channels. and These represent the height and width of the feature map, respectively. When generating the first frequency domain output feature component, the input feature... The second frequency domain characteristic component Channel condition characteristics The concatenation result, when generating the second frequency domain output feature component, is compared with the input feature. Output characteristic components in the first frequency domain Channel condition characteristics The splicing result.

[0059] Extract input features In spatial location Channel vector at: .

[0060] KAN nonlinear channel mapping unit A nonlinear transformation based on a learnable spline function is performed to obtain the output channel vector: .

[0061] in, This represents a channel-dimensional nonlinear mapping consisting of learnable basis functions. For all spatial locations... After performing the above mapping, the output features are obtained. Wherein, when the KAN nonlinear channel mapping unit is a scale parameter generation unit... At that time, output features For scale parameter features; when the KAN nonlinear channel mapping unit is a translation parameter generation unit At that time, output features These are translation parameter features. The scale parameter features and translation parameter features are used to perform an affine transformation on the corresponding frequency domain feature components to obtain the first frequency domain output feature component. Or the second frequency domain output characteristic components The first frequency domain output characteristic component and the second frequency domain output characteristic component are concatenated to form the output of the corresponding conditionally reversible current module. .

[0062] The KAN nonlinear channel mapping unit does not expand the spatial dimension of the entire feature map, but instead performs channel-dimensional nonlinear function approximation independently at each spatial location. This enhances the fitting ability to high-frequency sensitive perturbations and local nonlinear degradation before compression while controlling computational and memory overhead.

[0063] (b7) Finally, compress the distorted image and predicted image compression quality parameters Input differentiable image compression module The final distorted image after simulating real-world channel processing in the current iteration round is obtained: .

[0064] Therefore, the overall generation process of the cascaded noise generation network can be represented as follows: , in: .

[0065] This differentiable image compression module simulates JPEG recoding, quantization truncation, block effects, and high-frequency information loss in real-world channels by performing differentiable compression on pre-compressed distorted images.

[0066] (c) Based on the difference between the generated distorted image and the channel degradation image, a loss function is constructed, and the conditional feature extraction network and the cascaded noise generation network are trained to obtain a real channel noise simulator.

[0067] The joint loss function includes intermediate state supervision loss, final image pixel domain loss, frequency domain high-frequency subband loss, perceptual consistency loss, and compression quality parameter prediction loss. The specific process of step (c) is as follows: in, The intermediate state supervision loss is used to constrain the intermediate distortion images generated by the spatial distortion flow network. The pixel domain loss of the final image is used to constrain the difference in pixel space between the final simulated degraded image and the real channel degraded image; For the high-frequency subband loss in the frequency domain, it is used to constrain the difference between the final simulated degradation image and the real channel degradation image in the high-frequency subband; The perceptual consistency loss is used to constrain the difference between the final simulated degraded image and the real channel degraded image in the perceptual feature space; The image compression quality parameter prediction loss is used to constrain the difference between the predicted image compression quality parameters and the true compression quality parameters. These are the weighting coefficients for the corresponding loss terms.

[0068] Let the input image be The real channel degradation image is The intermediate distortion image output by the spatial distortion stream network is The final simulated degraded image is The predicted image compression quality parameters are: The actual compression quality parameters are First, based on the actual channel degradation image... Constructing intermediate state supervision targets: , in, This indicates a downsampling operation. This indicates an upsampling operation. The intermediate state monitoring loss is then: , The final image pixel domain loss is defined as: , The high-frequency subband loss in the frequency domain is defined as: , Perceived consistency loss for: , The loss for predicting compressible mass parameters is defined as: , In the formula, For the total loss, , , , , , These are the weighting coefficients. This indicates the extraction of horizontal and vertical high-frequency subbands from the wavelet transform. To represent the extraction of diagonal high-frequency subbands in wavelet transform, This represents the perceptual distance calculated based on a pre-trained perceptual feature extraction network. and They respectively represent the following: and The image obtained by mapping the numerical range [0,1] to the numerical range [-1,1].

[0069] When some samples do not have valid true compression quality parameters, only those that meet the following conditions can be used. ) Valid sample calculation .

[0070] (d) After training, freeze the network parameters of the conditional feature extraction network, spatial distortion stream network, frequency domain compression stream network and differentiable image compression module, input the image to be processed into the real channel noise simulator, and output the distorted image of the simulated real channel processing result, thereby simulating the degradation process of the corresponding platform channel.

[0071] The noise generation process of the cascaded noise generation network, specifically step (d), is as follows: (d1) The image to be processed Input the conditional feature extraction network after training Channel condition characteristics are obtained. and predicted image compression quality parameters At the same time, sample random latent variables It is used to characterize random degradation factors in real-world channel processing.

[0072] (d2) The image to be processed Channel condition characteristics and random latent variables Input the trained spatial distortion stream network to generate intermediate distortion images. Continuous degradation disturbance in space and spatial stage latent variables .

[0073] , Among them, the intermediate distorted image It is used to simulate continuous spatial degradation results such as scaling, resampling, color shifting, and smoothing degradation in the corresponding platform channel.

[0074] (d3) Transform the intermediate distorted image Channel condition characteristics Spatial stage latent variables and predicted image compression quality parameters Input the trained frequency domain compressed stream network to generate a pre-compressed distorted image. and high-frequency sensitive degradation perturbation .

[0075] , Among them, the image before compression is distorted. It is used to simulate high-frequency truncation, frequency-domain nonlinear attenuation, and pre-compression artifact tendency that further arise on the basis of spatial continuous degradation.

[0076] (d4) The image before compression is distorted. Input the trained differentiable image compression module and compress it according to the predicted image compression quality parameters. Perform differentiable compression to obtain the final simulated degradation image. : .

[0077] , Among them, the final simulated degraded image Used to approximate the output result of the image to be processed after passing through the corresponding real-world platform channel.

[0078] In this embodiment, the noise simulation results obtained at each stage are as follows: Figure 3 As shown in Table 5, the simulation quality of Facebook channel noise using a real-world channel noise simulator is compared to that using a traditional simulator.

[0079] Table 5: Comparison of Facebook channel noise simulation quality with traditional simulators This embodiment demonstrates that the method can be applied to image steganography. A comparison of the effects of training a simulated distortion channel using real online social network (OSN) noise is shown in the following results. Figure 4 As shown. Figure 4Figure (a) shows the encrypted image after transmission through the Facebook channel. It can be seen that the compression and transmission distortion introduced by the OSN platform have significantly reduced the quality of the carrier image. Figure 4 Figure (b) shows a noise sample collected from a real OSN channel, which presents a complex color noise pattern, reflecting the non-additive and non-Gaussian distortion introduced by multi-stage processing such as JPEG compression, scaling, and filtering in actual social networking platforms; Figure 4 Figure (c) in the figure is the secret image to be embedded; Figure 4 Figure (d) in the figure shows the secret image recovered after training with real OSN noise. The extracted secret information has good visual quality and the details are well preserved. Figure 4 Figure (e) shows the secret image recovered under the condition of no noise training (no noise training). It can be seen that the secret information has been severely distorted and the valid content cannot be identified.

[0080] contrast Figure 4 As shown in Figures (d) and (e), introducing a realistic channel noise simulator trained on real OSN channel degradation data as a simulated distortion channel during the model training phase can effectively improve the robustness of the steganography system to transmission distortion on real social network platforms, ensuring that secret information can still be accurately recovered after transmission through a real channel. Conversely, if channel distortion is ignored during training, the steganography system will completely fail in practical applications. This result verifies the crucial role of simulated distortion channel training based on real OSN noise in improving the robustness of image steganography.

[0081] Example 2 This embodiment provides a method for generating image distortion in social network channels. In this method, a trained real-world channel noise simulator is used as a differentiable real-world channel module in the downstream visual model training process. During this process, the model parameters of the noise simulator are kept frozen, and gradients are allowed to propagate back from the output of the noise simulator to the downstream visual model. The output image generated by the downstream visual model is input into the noise simulator to obtain a real-world channel perturbation image, and a robust training loss for the downstream model is constructed based on the real-world channel perturbation image. Furthermore, the robust training loss includes latent space noise decoupling loss. Let This represents the reconstructed image output by the latent space encoder and the downstream visual model under clean transmission conditions. The reconstructed image output under real-world channel perturbation conditions is The latent space noise disturbance characteristics are defined as follows: .

[0082] make Where B represents the batch size, C represents the number of latent space channels, and H and W represent the height and width of the latent space feature, respectively. Expanding along the batch and spatial dimensions yields the channel feature matrix: .

[0083] right Decentralization is performed to obtain And calculate the corresponding channel covariance matrix: .

[0084] The latent space noise decoupling loss is defined as: .

[0085] By minimizing This suppresses off-diagonal elements in the channel covariance matrix, ensuring that real-world channel noise disturbances remain low-correlation or approximately orthogonal between different latent space channels, thereby reducing channel coupling of noise disturbances in the latent space.

[0086] The rest are as in Example 1.

[0087] If the aforementioned functions 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, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0088] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0089] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0090] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0092] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0093] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for generating image distortion for social network channels, characterized in that, Includes the following steps: The image to be processed is input into a pre-trained real-world channel noise simulator, and the output is a distorted image simulating the processing result of a real-world channel. The real-world channel noise simulator includes a conditional feature extraction network and a cascaded noise generation network. The training steps include: The original clean image and the corresponding channel-degraded image after passing through the platform channel are obtained to form a dataset, wherein the channel-degraded image is used as the ground truth label; The original clean image is input into the conditional feature extraction network to extract the channel conditional features related to the actual channel distortion, and to predict the image compression quality parameters corresponding to the actual channel. The channel condition features and random latent variables are input into a cascaded noise generation network to generate a high-frequency sensitive degradation perturbation before compression. The high-frequency sensitive degradation perturbation before compression is then fused with the original clean image to obtain a pre-compressed distorted image. Finally, based on the image compression quality parameters and the pre-compressed distorted image, a distorted image after actual channel processing is obtained. Based on the difference between the distorted image and the channel degradation image, a loss function is constructed to iteratively train the real channel noise simulator until training is complete.

2. The image distortion generation method for social network channels according to claim 1, characterized in that, The steps for obtaining the channel condition features include: For the first in the same batch A clean original image Channel-level high-frequency residual extraction is performed to obtain high-frequency residual features, which are represented as follows: In the formula, For the first A clean original image Extracted high-frequency residual features, This is a high-frequency residual extraction module; The first A clean original image High-frequency residual characteristics Concatenation is performed along the channel dimension to obtain the fused input features, represented as follows: In the formula, To fuse input features, This indicates a channel-dimensional splicing operation; The fused input features The input is fed into a deep feature fusion network to obtain the first... A clean original image The corresponding channel condition characteristics are expressed as follows: In the formula, For the first A clean original image Corresponding channel condition characteristics, This represents a deep feature fusion network.

3. The image distortion generation method for social network channels according to claim 1, characterized in that, The steps for obtaining the image compression quality parameters include: Global feature aggregation is performed on the aforementioned channel condition features to obtain a global channel representation, which is expressed as: In the formula, For global channel characterization, This indicates a global feature aggregation operation. For the first A clean original image Corresponding channel condition characteristics; The global channel characterization is input into the compression quality parameter prediction module to obtain normalized quality parameters, which are expressed as: In the formula, To normalize the quality parameters, This indicates the compression quality parameter prediction module. These are the learnable parameters for the compression quality parameter prediction module; Let the preset compression quality range be The normalized quality parameters are mapped to the compression quality range to obtain the image compression quality parameters, which are expressed as follows: In the formula, For the first A clean original image The corresponding image compression quality parameters, , These represent the upper and lower limits of compression quality.

4. The image distortion generation method for social network channels according to claim 1, characterized in that, The cascaded noise generation network includes a spatial distortion flow network, a frequency domain perturbation flow network, and a differentiable image compression module. The spatial distortion flow network is used to compress the original clean image. Channel condition characteristics and random latent variables As input, the output is a spatially degraded image. Continuous degradation disturbance in space and spatial stage latent variables The operational expression for the spatially distorted flow network is: In the formula, Representation spatial distortion flow network; The frequency domain perturbation flow network is used to degrade the spatial phase image. Channel condition characteristics Spatial stage latent variables and image compression quality parameters As input, output high-frequency sensitive degradation perturbation before compression. and pre-compressed distorted images The operational expression for the frequency domain perturbation flow network is as follows: In the formula, Represents a frequency domain perturbation flow network; The differentiable image compression module is used to compress the spatially degraded image. High-frequency sensitive degradation perturbation before compression and image compression quality parameters As input, the output is a distorted image simulating real-world channel processing in the current iteration round, where the operational expression of the differentiable image compression module is: In the formula, The image is distorted. This indicates a differentiable image compression module.

5. The image distortion generation method for social network channels according to claim 4, characterized in that, The spatial distortion flow network includes multiple cascaded conditionally reversible flow modules, and the execution steps of each conditionally reversible flow module include: Let the first The input to the conditionally reversible flow module is After passing through the reversible channel replacement layer The result was: In the formula, The input after the substitution; Will Divided along the channel dimension into first spatial feature components Second spatial characteristic components ,Right now: In the formula, Indicates concatenation of channel dimensions; Based on the second spatial feature component and channel condition characteristics Generates the first spatial feature component The scale parameters and translation parameters are determined, and the first output feature component is obtained. The first spatial feature component The scale parameters, translation parameters, and first output feature components They are represented as follows: In the formula, , These are respectively applied to the first spatial feature components. The scale parameters and translation parameters, Represents the scaling subnetwork. Represents the translation transformation subnetwork. Represents the scaling constraint function; Based on the first spatial feature component and channel condition characteristics Generates the feature components that act on the second space. The scale and translation parameters are determined, and the second output feature component is obtained. The second spatial feature component The scale parameters, translation parameters, and second output feature components They are represented as follows: In the formula, , These are the eigencomponents acting on the second space. The scale parameters and translation parameters, Represents the scaling subnetwork. Represents the translation transformation subnetwork; The first output feature component Second output feature component Perform channel splicing to obtain the first The output of a conditionally reversible flow module is represented as: In the formula, For the first The output of the conditionally reversible flow module, i.e. the first... The input of a conditional reversible flow module.

6. The image distortion generation method for social network channels according to claim 4, characterized in that, The frequency domain perturbation flow network includes multiple conditionally reversible flow modules based on KAN nonlinear channel mapping. The execution steps of each conditionally reversible flow module based on KAN nonlinear channel mapping include: Let the first The input of a conditionally reversible flow module based on KAN nonlinear channel mapping is via reversible channel replacement layer The result was: In the formula, The input after the substitution; Will Divided along the channel dimension into the first frequency domain feature components Second frequency domain characteristic components ,Right now: In the formula, Indicates concatenation of channel dimensions; Based on the first frequency domain feature component Second frequency domain characteristic components and channel condition characteristics Generate the first frequency domain output feature component. , is represented as: In the formula, Represents the scale constraint function. , These are the eigencomponents acting on the first frequency domain. The scale parameters and translation parameters, and These represent the scale parameter generation unit and translation parameter generation unit based on KAN nonlinear channel mapping, respectively. Based on the first frequency domain output feature components Channel condition characteristics The components are then spliced ​​together and processed by the scale parameter generation unit and translation parameter generation unit based on KAN nonlinear channel mapping to obtain the second frequency domain output feature component. , is represented as: In the formula, , These are the eigencomponents acting on the second frequency domain. The scale parameters and translation parameters; The first frequency domain output feature component Second frequency domain output characteristic components By concatenating along the channel dimension, we obtain the [number]. The output of a conditionally reversible flow module based on KAN nonlinear channel mapping is expressed as: In the formula, For the first The output of the conditionally reversible flow module based on KAN nonlinear channel mapping, i.e., the... The input to a conditionally reversible flow module based on KAN nonlinear channel mapping.

7. The image distortion generation method for social network channels according to claim 6, characterized in that, The execution steps of the conditionally reversible flow module based on KAN nonlinear channel mapping further include: performing nonlinear mapping on the channel vector at each spatial location, specifically including: After obtaining the first frequency domain output feature component At that time, the second frequency domain feature component Channel condition characteristics The concatenation result is used as input for the nonlinear mapping. After obtaining the second frequency domain output feature component At that time, the first frequency domain feature component Channel condition characteristics The concatenation result is used as input for the nonlinear mapping. ; Take the input In spatial location The channel vector is represented as: In the formula, Spatial location The channel vector at that location; right A nonlinear transformation based on a learnable spline function is performed to obtain the output channel vector, which is expressed as: In the formula, This is the output channel vector. This represents a channel-dimensional nonlinear mapping composed of learnable basis functions; Based on all spatial locations The corresponding output channel vector yields the output features of the nonlinear mapping. The output features The scale parameter , or translation parameters , The output features Used to obtain the first frequency domain output feature component through radiometric transformation Or the second frequency domain output characteristic components To obtain the first The output of a conditionally reversible flow module based on KAN nonlinear channel mapping .

8. The image distortion generation method for social network channels according to claim 1, characterized in that, The loss function is: in: Intermediate state monitoring loss for: Final image pixel domain loss for: Frequency domain high-frequency subband loss for: Perceived consistency loss for: Predicted loss of compressive mass parameters for: In the formula, For the total loss, , , , , , These are the weighting coefficients. For spatial stage degradation images, This is the smoothed intermediate degraded image. This indicates a downsampling operation. Indicates an upsampling operation. For channel degradation images, The image is distorted. This indicates the extraction of horizontal and vertical high-frequency subbands from the wavelet transform. This indicates the extraction of diagonal high-frequency subbands from the wavelet transform. This represents the perceptual distance calculated based on a pre-trained perceptual feature extraction network. and These represent the distorted images. and channel degradation image The image obtained by mapping the numerical range [0,1] to the numerical range [-1,1]. Predicted values ​​for image compression quality parameters. This represents the true value of the image compression quality parameter.

9. The image distortion generation method for social network channels according to claim 1, characterized in that, The trained real-world channel noise simulator is used in the training process of the downstream visual model. During training, the parameters of the real-world channel noise simulator are frozen and backpropagated from the output of the simulator to the downstream visual model via gradients. The downstream visual model then outputs the reconstructed image under clean transmission conditions. Input a realistic channel noise simulator to obtain a reconstructed image output under simulated realistic channel disturbance conditions. And construct a robust training loss for the downstream model: In the formula, Loss due to robust training The channel covariance matrix, To represent the Frobenius norm, The total number of feature points, For batch size, , These represent the height and width of the latent space noise disturbance characteristics, respectively. To obtain the centered channel feature matrix , To characterize latent space noise disturbances, by... Expanding along the batch and spatial dimensions yields the channel feature matrix. , This indicates a latent space encoder.

10. A storage medium having a program stored thereon, characterized in that, When the program is executed, it implements an image distortion generation method for social network channels as described in any one of claims 1-9.

Images