A multi-feature fusion-based cross-generator generated image discrimination method and system

Abstract

The application relates to the field of electronic information and discloses a cross-generator generated image identification method and system based on multi-feature fusion, which comprises the following steps: uniformly adjusting the size of a to-be-detected image and reserving an original input image; dividing the preprocessed image into multiple image blocks, calculating the texture richness of the image blocks, sorting and recombining the image blocks according to the texture richness to generate a rearranged image; performing high-pass filtering on the rearranged image, modeling the frequency domain response and weighting the multi-spectrum attention to extract frequency domain texture enhancement features; extracting global semantic representation features of the original input image by using a pre-trained visual model; dynamically calculating weights and weighting and fusing the frequency domain texture enhancement features and the global semantic representation features; classifying according to the fused feature representation and outputting a discrimination result. The system comprises corresponding modules for executing the above method. The application comprehensively utilizes bottom-layer local texture abnormalities and high-layer semantic information, and improves the accuracy of cross-generator identification.

Description

Technical Field

[0001] This invention relates to the field of electronic information technology, specifically to a cross-generator image identification method and system based on multi-feature fusion. Background Technology

[0002] With the rapid development of deep learning technology, image generation techniques such as generative adversarial networks and diffusion models have become increasingly mature, producing images with a high degree of visual realism. While these technologies have wide applications in artistic creation and image editing, they also bring potential risks such as the spread of misinformation, copyright disputes, and privacy concerns. Therefore, effectively verifying the authenticity of images has become particularly important.

[0003] Existing generative image detection methods typically rely on the inherent data distribution of a specific generative model or specific artifact features for identification. These methods achieve good detection results when faced with generated images from the same source as their training set, but in real-world open scenarios, the images to be detected often originate from unknown generators. Different image generation models have significant differences in their generation mechanisms and spectral statistics, which leads to insufficient generalization ability and a significant drop in detection accuracy when faced with unseen generator output images.

[0004] Furthermore, existing detection networks often directly utilize convolutional neural networks to extract features from the entire image. Since artifacts in generated images are typically hidden within local high-frequency texture details, and deep neural networks tend to focus on the high-level semantic structure of the image during training, subtle local texture anomalies are easily smoothed out or ignored during feature extraction. Simultaneously, the normal global semantic layout in the original image can interfere with the perception of local artifacts, making it difficult for single-view feature extraction to simultaneously consider both local detail artifacts and global semantic consistency, thus limiting the model's ability to detect complex forged images. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a cross-generator generated image identification method and system based on multi-feature fusion, which solves the problems of insufficient generalization ability of existing generated image detection methods under cross-generator conditions, and the difficulty of feature extraction from a single perspective in simultaneously taking into account local abnormal texture perception and global semantic features.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of this invention provides a method for cross-generator generated image identification based on multi-feature fusion, comprising the following steps: The input image to be detected is uniformly resized to obtain a preprocessed image, while retaining the original input image; The preprocessed image is locally divided, the texture richness of each image patch is calculated, and the image patches are sorted and recombined according to the texture richness to generate a rearranged image. A high-pass filtering operation is performed on the rearranged image, and frequency domain response modeling and multi-spectral attention weighting are performed on the filtered features. The features are then extracted to obtain frequency domain texture enhancement features. The pre-trained visual model is used to extract features from the original input image to obtain global semantic representation features; The frequency domain texture enhancement features and the global semantic representation features are weighted and fused to obtain a fused feature representation. Based on the fused feature representation, classification is performed, and the result of the judgment that the image to be detected belongs to a real image or a generated image is output.

[0007] Preferably, the step of locally segmenting the preprocessed image, calculating the texture richness of each image patch, sorting and recombining the image patches according to the texture richness to generate a rearranged image specifically includes: The preprocessed image is divided into multiple image blocks of the same size; Each color image patch is converted into a grayscale image, and the information entropy is calculated based on the grayscale distribution to obtain the texture richness of each image patch; The image blocks are sorted from largest to smallest according to their entropy values; The sorted image blocks are reorganized into strong texture regions and weak texture regions to generate a rearranged image.

[0008] Preferably, performing a high-pass filtering operation on the rearranged image specifically includes: The rearranged image is filtered using a high-pass convolution kernel to enhance high-frequency texture structures and weaken low-frequency smoothing information. The filtered feature map is divided into two regions, representing weak texture regions and strong texture regions, according to the spatial height direction. Preliminary feature extraction is performed on the divided regions using convolutional layers and batch normalization layers.

[0009] Preferably, the step of performing frequency domain response modeling and multi-spectral attention weighting on the filtered features and extracting features to obtain frequency domain texture enhancement features specifically includes: The feature map after preliminary feature extraction is subjected to spectral mapping by discrete cosine transform to obtain the spectral response and spectral description vector of the corresponding frequency coordinates; The spectral description vector is processed by combining linear transformation parameters with the ReLU activation function and the Sigmoid function to calculate the channel weights. The pooled representation of the feature map after preliminary feature extraction is multiplied by the channel weights to obtain the frequency domain enhanced features; The frequency domain enhancement features are input into a local texture feature extraction network for extraction to obtain frequency domain texture enhancement features.

[0010] Preferably, the step of extracting features from the original input image using a pre-trained visual model to obtain global semantic representation features specifically includes: The original input image is normalized by combining the mean and variance of the current data distribution with the mean and variance required by the pre-trained model. The normalized input image is fed into the pre-trained visual model to perform forward propagation calculation, thereby obtaining global semantic representation features that reflect the overall semantic structure and contextual information of the image.

[0011] Preferably, the weighted fusion of the frequency domain texture enhancement features and the global semantic representation features to obtain the fused feature representation specifically includes: Feature alignment and dimension matching are performed on the frequency domain texture enhancement features and the global semantic representation features; Calculate the computational weights of the frequency domain texture enhancement features and the global semantic representation features, respectively; The frequency domain texture enhancement features and the global semantic representation features are weighted and summed according to the calculated weights to obtain the fused feature representation.

[0012] Preferably, the calculation of the computational weights of the frequency domain texture enhancement features and the global semantic representation features specifically includes: The input features are mapped and calculated by using the first and second layer linear transformation parameters, combined with the ReLU activation function and the Sigmoid function. The calculated weights corresponding to the frequency domain texture enhancement features and the calculated weights corresponding to the global semantic representation features are output respectively.

[0013] Preferably, the step of classifying based on the fused feature representation and outputting a judgment result indicating whether the image to be detected belongs to a real image or a generated image specifically includes: The fused feature representation is input into a multilayer perceptron classifier to perform feature dimensionality reduction and classification mapping calculations, and the probability distribution mapping results are obtained. Based on the probability distribution mapping result, the result of the judgment that the image to be detected belongs to a real image or a generated image is output.

[0014] Preferably, the method further includes an end-to-end training step: A supervised training set is constructed using real image samples and generated image samples; The prediction error between the calculated discrimination result and the true label of the sample; Based on the prediction error, the parameters in the network are updated synchronously using the backpropagation algorithm.

[0015] A second aspect of the present invention provides a cross-generator generated image identification system based on multi-feature fusion, comprising: The preprocessing module is used to resize the input image to be detected, obtain the preprocessed image, and retain the original input image; The image patch rearrangement module is used to locally divide the preprocessed image, calculate the texture richness of each image patch, sort and recombine the image patches according to the texture richness, and generate a rearranged image. The frequency domain texture enhancement module is used to perform high-pass filtering on the rearranged image, perform frequency domain response modeling and multi-spectral attention weighting on the filtered features, and extract features to obtain frequency domain texture enhancement features; The global semantic extraction module is used to extract features from the original input image using a pre-trained visual model to obtain global semantic representation features. An adaptive feature fusion module is used to perform weighted fusion of the frequency domain texture enhancement features and the global semantic representation features to obtain a fused feature representation; The classification and discrimination module is used to classify the image based on the fused feature representation and output the discrimination result of whether the image to be detected belongs to a real image or a generated image.

[0016] This invention provides a method and system for cross-generator generated image identification based on multi-feature fusion. It has the following beneficial effects: 1. This invention involves locally segmenting the input image and then sorting and recombining the image patches based on the calculated texture richness. This spatial rearrangement disrupts the original semantic layout of the image, concentrating regions with similar texture intensity, thereby effectively reducing the interference of the original global structure on the judgment of local details. Based on this, feature extraction enhances the detection network's ability to perceive subtle local texture differences and spatial artifacts in the generated image.

[0017] 2. This invention utilizes high-pass filtering and a multi-spectral channel attention mechanism to enhance the frequency domain features of the rearranged image. High-pass filtering preserves high-frequency edge details and suppresses low-frequency smoothing of the background, while multi-spectral attention weighting assigns adaptive learning weights to different frequency components. This method directly models the inherent differences in spectral distribution between the real and generated images, enabling subsequent networks to extract more discriminative frequency domain features and improving the accuracy of identifying hidden generation artifacts.

[0018] 3. This invention employs a dual-branch network structure and introduces an adaptive feature fusion mechanism. The network acquires frequency-domain texture enhancement features through a local texture branch, while simultaneously utilizing a pre-trained visual model to extract global semantic representation features of the image through a global semantic branch. The adaptive feature fusion module dynamically calculates the weights of these two types of features and performs joint modeling, enabling the classification network to comprehensively utilize both low-level texture anomalies and high-level semantic information. This joint feature discrimination method does not depend on the data distribution of a specific generative model, effectively improving the model's generalization ability across generator conditions and open scenarios. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method steps of the present invention; Figure 2 This is a system module architecture diagram of the present invention. Detailed Implementation

[0020] The technical solutions in 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see the appendix Figure 1 This invention provides a method for cross-generator generated image identification based on multi-feature fusion, comprising the following steps: S1. Input the image to be detected into the preprocessing module, adjust the size of the input image uniformly, and retain the original image content as the input for the subsequent global semantic branch.

[0022] S2. Input the preprocessed image into the image block rearrangement module to perform local division of the image to obtain multiple image blocks; calculate the texture richness of each image block and sort the image blocks according to the texture richness; then recombine the sorted image blocks according to the preset arrangement to generate a rearranged image, and input the rearranged image into the local texture branch.

[0023] S3. Input the rearranged image into the frequency domain texture enhancement module, perform high-pass filtering on the rearranged image to highlight high-frequency texture information in the image; further perform frequency domain response modeling and multi-spectral attention weighting on the enhanced features to obtain frequency domain enhanced features, and input the frequency domain enhanced features into the local texture feature extraction network to obtain frequency domain texture enhanced features.

[0024] S4. Input the original input image into the global semantic branch and use the pre-trained visual model to extract the global semantic representation features of the image.

[0025] S5. Input the frequency domain enhancement features and global semantic features into the adaptive feature fusion module, perform joint modeling and weighted fusion of different branch features, and obtain the fused feature representation for the identification task.

[0026] S6. Input the fused features into the classification and discrimination module, and output the discrimination result of whether the image to be detected belongs to the real image or the generated image.

[0027] Please see the appendix Figure 2 Based on the above method, the present invention provides a cross-generator generated image identification system based on multi-feature fusion, comprising: The preprocessing module is used to uniformly resize the input image and retain the original image content as input for the subsequent global semantic branch; The image patch rearrangement module is used to locally divide the image to obtain multiple image patches, calculate the texture richness of each image patch, sort the image patches according to the texture richness, recombine the sorted image patches according to a preset arrangement to generate a rearranged image, and input the rearranged image into the local texture branch. The frequency domain texture enhancement module is used to perform high-pass filtering on the rearranged image, perform frequency domain response modeling and multi-spectral attention weighting on the enhanced features to obtain frequency domain enhanced features, and input the frequency domain enhanced features into the local texture feature extraction network to obtain frequency domain texture enhanced features. The global semantic extraction module is used to receive the original input image and extract the global semantic representation features of the image using a pre-trained visual model; The adaptive feature fusion module is used to receive frequency domain enhanced features and global semantic features, jointly model and weightedly fuse different branch features to obtain a fused feature representation for the discrimination task; The classification and discrimination module is used to receive fused features and output the discrimination result of whether the image to be detected belongs to the real image or the generated image.

[0028] In its implementation, the system divides the input image into multiple equally sized image blocks using an image block rearrangement module, and calculates the texture richness of each block. Based on this texture richness, the system sorts and reorganizes the image blocks, spatially redistributing regions with strong textures to those with weak textures. This spatial redistribution reduces the impact of the original semantic layout on the classification process and enhances the perceptibility of local texture differences in subsequent feature extraction.

[0029] The system performs high-pass filtering on the rearranged image in the frequency domain texture enhancement module, preserving high-frequency information and suppressing low-frequency background components. Subsequently, the system performs frequency domain mapping on the high-pass filtered features and models different frequency components using a multi-spectral attention mechanism. The model automatically learns frequency components that are crucial for detecting generated images, assigning higher weights to relevant frequency responses to improve the ability to perceive frequency domain anomalies in the generated image. Since real and generated images differ in texture formation mechanisms and spectral statistics, the frequency domain texture enhancement module amplifies the expression of these differences in the feature space, enabling subsequent convolutional feature extraction networks to obtain discriminative input.

[0030] The system employs a pre-trained visual model to extract global features from the original input image, obtaining a global semantic representation that reflects the overall semantic structure, object relationships, and contextual information of the image. The local texture branch focuses on fine-grained artifacts, while the global semantic branch focuses on high-level structure and overall consistency. The system inputs these two types of features into an adaptive feature fusion module. Through feature alignment, weight allocation, and joint expression, the model simultaneously utilizes local low-level texture information and global high-level semantic information within the same discriminative framework. Finally, the classification and discrimination module outputs the final discrimination result.

[0031] In step S1, the image to be detected is input to the preprocessing module, where the input image is uniformly resized, and the original image content is retained as input for the subsequent global semantic branch. In specific implementation, this preprocessing process may further include the following sub-steps: S101. The preprocessing module receives the input image to be detected. Since the images to be detected from different sources or different generation models have different spatial resolutions, they need to be normalized to meet the fixed input size requirements of the subsequent feature extraction network.

[0032] S102. The preprocessing module performs uniform size adjustment on the input image to be detected. For image size adjustment, those skilled in the art can use conventional interpolation algorithms, such as bilinear interpolation or bicubic interpolation to resample image pixels. Changing the spatial resolution of the image through interpolation calculations to achieve a preset uniform size standard is a well-known technique in the art and will not be elaborated further here.

[0033] S103. The preprocessing module outputs the uniformly resized image to the subsequent image patch rearrangement module for local texture branch processing. Simultaneously, the preprocessing module retains the original image content as input for the subsequent global semantic branch. The global semantic branch primarily utilizes a pre-trained visual model to extract the global semantic representation features of the image. Retaining the original image content ensures that the overall semantic structure, object relationships, and contextual information of the image are not lost due to unnecessary image transformations, providing a complete data foundation for subsequent global semantic extraction.

[0034] In step S2, the preprocessed image is input into the image block rearrangement module to locally divide the image into multiple image blocks; the texture richness of each image block is calculated, and the image blocks are sorted according to their texture richness; then, the sorted image blocks are recombine according to a preset arrangement to generate a rearranged image, which is then input into the local texture branch. In specific implementation, this step may further include the following sub-steps: S201. The image block rearrangement module receives the preprocessed image and divides the input image into multiple image blocks of the same size. Through spatial partitioning, the entire image is discretized into multiple independent local regions, providing a computational basis for subsequent measurement of the internal statistical characteristics of each local region.

[0035] S202, the image patch rearrangement module calculates the texture richness of each image patch. This invention employs a global histogram entropy-based texture richness calculation method. Specifically, the process involves converting color image patches to grayscale images and then calculating their information entropy based on the grayscale distribution. Information entropy can serve as an effective descriptor of local texture strength, used to specifically calculate the texture richness contained in each image patch. For the conversion of color image patches to grayscale images, those skilled in the art can use a conventional pixel channel weighting algorithm; the channel conversion operation is a well-known technique in the field and will not be elaborated upon here. The specific formulas for calculating information entropy and probability distribution are as follows: ; ; In the formula, For the first The image patch in the first Pixel count at each gray level; The gray level; A very small constant introduced to prevent taking the logarithm of zero; The entropy response of the image patch; For the first The first image patch The probability distribution of each gray level.

[0036] S203. The image patch rearrangement module sorts the image patches according to their texture richness. The image patch rearrangement module uses the calculated entropy response results of each image patch to sort the image patches from largest to smallest entropy value, thus obtaining the sorting result.

[0037] S204. The image patch rearrangement module recombines the sorted image patches according to a preset arrangement to generate a rearranged image. Based on the aforementioned sorting results, the image patch rearrangement module reorganizes the image patches into strong texture regions and weak texture regions. This arrangement redistributes the strong texture regions and weak texture regions in space, weakening the influence of the original semantic layout on the classification process and enhancing the perceptibility of local texture differences in the subsequent feature extraction process.

[0038] S205. The image block rearrangement module outputs the generated rearranged image and inputs it into the local texture branch as the data basis for the subsequent frequency domain texture enhancement module.

[0039] In step S3, the rearranged image is input into the frequency domain texture enhancement module, where a high-pass filtering operation is performed to highlight high-frequency texture information. Further, frequency domain response modeling and multi-spectral attention weighting are applied to the enhanced features to obtain frequency domain enhanced features. These enhanced features are then input into the local texture feature extraction network to obtain frequency domain texture enhanced features. In specific implementations, this step may further include the following sub-steps: S301, the frequency domain texture enhancement module receives the rearranged image and performs filtering processing on the rearranged image using a high-pass convolution kernel. This high-pass filtering operation retains high-frequency information with discriminative value, weakens low-frequency smoothing information, and suppresses low-frequency background components, thereby enhancing the edge structure and texture details in the image and highlighting the high-frequency texture structure. For the specific convolution calculation implementation of the high-pass filtering operation, those skilled in the art can use conventional edge detection operators or high-pass filters for processing. The two-dimensional convolution filtering operation is a well-known technique in the field and will not be described in detail here.

[0040] S302, the frequency domain texture enhancement module divides the feature map after high-pass filtering into two regions, upper and lower, along the spatial height direction. These two regions represent the weak texture region and strong texture region constructed by the previous image patch rearrangement module, respectively. The frequency domain texture enhancement module then performs preliminary feature extraction on the divided regions using convolutional layers and batch normalization layers.

[0041] The S303 frequency domain texture enhancement module introduces a multi-spectral channel attention mechanism in each of the two branches extracting features from two regions. The module performs spectral mapping on the feature map using discrete cosine transform, performs frequency domain response modeling and multi-spectral attention weighting, enabling the network to model the texture statistical characteristics of different regions differently. Through this mechanism, the model automatically learns frequency components that are crucial for detecting generated images and assigns higher weights to relevant frequency responses to improve its ability to detect frequency domain anomalies in the generated image. The specific multi-spectral attention weighting and feature enhancement formulas are as follows: ; ; ; In the formula, The spatial coordinate representation of the input features after preliminary processing; These are discrete cosine basis functions; These are discrete cosine basis functions; , For the corresponding different frequency coordinates; The height of the feature map; The width of the feature map; The calculated spectral response; This is a frequency spectrum description vector; These are the parameters for the first-level linear transformation; These are the parameters for the second-level linear transformation; For the Sigmoid function; It is the ReLU activation function; This is the final channel weight; The pooled representation of the input features; This is for the final frequency domain enhancement features.

[0042] S304. The frequency domain texture enhancement module inputs the calculated frequency domain enhancement features into the local texture feature extraction network for feature extraction, resulting in frequency domain texture enhancement features. Since real images and generated images differ in texture formation mechanisms and spectral statistics, the frequency domain texture enhancement module effectively amplifies the expression of this difference in the feature space through the aforementioned high-pass filtering and spectral mapping operations. This allows the local texture feature extraction network to obtain discriminative input, improving the model's ability to detect hidden artifacts in the generated image.

[0043] In step S4, the original input image is input into the global semantic branch, and the global semantic representation features of the image are extracted using a pre-trained visual model. In specific implementations, this step may further include the following sub-steps: S401, the global semantic extraction module receives the original input image retained during the preprocessing stage. The local texture branch focuses on fine-grained artifacts, while the global semantic branch focuses on high-level structure and overall consistency; the two focus on different levels of information. The input is the original image without local spatial rearrangement and frequency domain high-pass filtering, which preserves the complete scene layout and object dependencies in the image, providing a complete input for subsequent extraction of high-level structural information.

[0044] S402. The global semantic extraction module normalizes the original input image to align the statistical distribution of the input data with the input space required by the pre-trained model. For image normalization and linear transformation of the underlying pixel scale, those skilled in the art can use conventional data normalization methods; the data mapping calculation is a well-known technique in the field and will not be elaborated upon here. The specific calculation formula for this normalization operation is as follows: ; In the formula, The input image before normalization; The input image is normalized; The mean value required for the pre-trained model; The variance required for the pre-trained model; This represents the mean of the current data distribution. This represents the variance of the current data distribution.

[0045] S403, the global semantic extraction module inputs the normalized input image into the pre-trained visual model, performs forward propagation calculations, and obtains global semantic representation features. These global semantic representation features reflect the overall semantic structure, object relationships, and contextual information of the image, compensating for the shortcomings of local texture branches in global semantic perception. The feature extraction calculation formula is as follows: ; In the formula, The input image is normalized; For the feature mapping function of the pre-trained visual model; This is the obtained global semantic feature representation. The global semantic extraction module outputs this global semantic feature representation to the subsequent adaptive feature fusion module as the data input for cross-modal feature joint modeling.

[0046] In step S5, the frequency domain enhanced features and global semantic features are input into the adaptive feature fusion module to jointly model and weightedly fuse features from different branches, obtaining a fused feature representation for the discrimination task. In specific implementation, this step may further include the following sub-steps: S501, the adaptive feature fusion module receives the frequency domain texture enhancement features output from the local texture branch and the global semantic features output from the global semantic branch. Since the local low-level texture information and the global high-level semantic information are located in different feature spaces and have different dimensions, the adaptive feature fusion module performs dimension matching processing on these two types of features through feature alignment operations, establishing a unified computational basis for subsequent weight allocation.

[0047] S502, the adaptive feature fusion module adaptively assigns weights to the two types of input features. Based on the information contained in the features themselves, the module dynamically calculates the weight coefficients of the two types of features during the fusion and discrimination process. For the specific forward computation implementation of the multilayer perceptron and linear transform, those skilled in the art can use conventional fully connected layers of a neural network for construction. The parameter matrix multiplication and nonlinear activation calculations are well-known techniques in the field and will not be elaborated upon here. The corresponding weight calculation formula is as follows: ; ; In the formula, Enhance the frequency domain texture features of the input; The input is global semantic features; It is the ReLU activation function; For the Sigmoid function; The parameters for the first-level linear transform used to calculate the frequency domain feature weights; The parameters for the second-level linear transform are used to calculate the frequency domain feature weights; The parameters for the first-level linear transformation are used to calculate the weights of the global semantic features; The parameters for the second-level linear transformation are used to calculate the weights of the global semantic features; The calculation weights for frequency domain texture enhancement features; The weights are used to calculate the global semantic features.

[0048] S503, the adaptive feature fusion module calculates the weight coefficients. and The frequency domain texture enhancement features and global semantic features are weighted and summed, and then joint representation processing is performed. Through this joint modeling operation, the model can simultaneously utilize local low-level texture information and global high-level semantic information within the same discriminative framework, dynamically adjust the role of different features in the discrimination process, and output a fused feature representation for the discrimination task.

[0049] In step S6, the fused features are input into the classification and discrimination module, which outputs a discrimination result indicating whether the image to be detected belongs to a real image or a generated image. In specific implementations, this step may further include the following sub-steps: S601, the classification and discrimination module receives the fused feature representation for the discrimination task output by the adaptive feature fusion module. The classification and discrimination module inputs the fused features into the multilayer perceptron classifier to perform feature dimensionality reduction and classification mapping calculation.

[0050] For the specific network structure construction and probability mapping calculation of the multilayer perceptron classifier, those skilled in the art can use conventional fully connected network layers combined with normalized exponential or logistic functions to achieve it. Its nonlinear classification mapping operation is a well-known technology in the field and will not be described in detail here.

[0051] S602. The classification and discrimination module outputs the category result of whether the image to be detected belongs to a real image or a generated image based on the probability distribution mapping result calculated by the multilayer perceptron classifier.

[0052] S603. During the training phase, a supervised training set is constructed using real image samples and generated image samples. The prediction error between the discrimination result output by the classification module and the true label of the sample is calculated, and the entire detection network is trained end-to-end based on this prediction error. The network parameters in the local texture branch, global semantic branch, and adaptive feature fusion module are updated synchronously through the backpropagation algorithm, enabling the model to have the ability to classify by comprehensively utilizing multiple feature information within the same discrimination framework.

[0053] S604. During the inference phase, only the image to be detected needs to be input. After forward propagation calculation through the aforementioned multi-branch network, the classification and discrimination module can output the corresponding identification result. This network structure and classification mechanism do not depend on the data distribution of a specific generative model and can adapt to real images from different sources and output images from different types of generative models, thereby realizing the generative image detection task in open scenes and under cross-generator conditions.

[0054] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A cross-generator generated image identification method based on multi-feature fusion, characterized in that, Includes the following steps: The input image to be detected is uniformly resized to obtain a preprocessed image, while retaining the original input image; The preprocessed image is locally divided, the texture richness of each image patch is calculated, and the image patches are sorted and recombined according to the texture richness to generate a rearranged image. A high-pass filtering operation is performed on the rearranged image, and frequency domain response modeling and multi-spectral attention weighting are performed on the filtered features. The features are then extracted to obtain frequency domain texture enhancement features. The pre-trained visual model is used to extract features from the original input image to obtain global semantic representation features; The frequency domain texture enhancement features and the global semantic representation features are weighted and fused to obtain a fused feature representation. Based on the fused feature representation, classification is performed, and the result of the judgment that the image to be detected belongs to a real image or a generated image is output.

2. The cross-generator generated image identification method based on multi-feature fusion according to claim 1, characterized in that, The process of locally segmenting the preprocessed image, calculating the texture richness of each image patch, sorting and recombining the image patches based on their texture richness to generate a rearranged image specifically includes: The preprocessed image is divided into multiple image blocks of the same size; Each color image patch is converted into a grayscale image, and the information entropy is calculated based on the grayscale distribution to obtain the texture richness of each image patch; The image blocks are sorted from largest to smallest according to their entropy values; The sorted image blocks are reorganized into strong texture regions and weak texture regions to generate a rearranged image.

3. The cross-generator generated image identification method based on multi-feature fusion according to claim 1, characterized in that, The high-pass filtering operation performed on the rearranged image specifically includes: The rearranged image is filtered using a high-pass convolution kernel to enhance high-frequency texture structures and weaken low-frequency smoothing information. The filtered feature map is divided into two regions, representing weak texture regions and strong texture regions, according to the spatial height direction. Preliminary feature extraction is performed on the divided regions using convolutional layers and batch normalization layers.

4. The cross-generator generated image identification method based on multi-feature fusion according to claim 1, characterized in that, The process of performing frequency domain response modeling and multi-spectral attention weighting on the filtered features, and extracting features to obtain frequency domain texture enhancement features, specifically includes: The feature map after preliminary feature extraction is subjected to spectral mapping by discrete cosine transform to obtain the spectral response and spectral description vector of the corresponding frequency coordinates; The spectral description vector is processed by combining linear transformation parameters with the ReLU activation function and the Sigmoid function to calculate the channel weights. The pooled representation of the feature map after preliminary feature extraction is multiplied by the channel weights to obtain the frequency domain enhanced features; The frequency domain enhancement features are input into a local texture feature extraction network for extraction to obtain frequency domain texture enhancement features.

5. The cross-generator generated image identification method based on multi-feature fusion according to claim 1, characterized in that, The step of extracting features from the original input image using a pre-trained visual model to obtain global semantic representation features specifically includes: The original input image is normalized by combining the mean and variance of the current data distribution with the mean and variance required by the pre-trained model. The normalized input image is fed into the pre-trained visual model to perform forward propagation calculation, thereby obtaining global semantic representation features that reflect the overall semantic structure and contextual information of the image.

6. The cross-generator generated image identification method based on multi-feature fusion according to claim 1, characterized in that, The weighted fusion of the frequency domain texture enhancement features and the global semantic representation features to obtain the fused feature representation specifically includes: Feature alignment and dimension matching are performed on the frequency domain texture enhancement features and the global semantic representation features; Calculate the computational weights of the frequency domain texture enhancement features and the global semantic representation features, respectively; The frequency domain texture enhancement features and the global semantic representation features are weighted and summed according to the calculated weights to obtain the fused feature representation.

7. The cross-generator generated image identification method based on multi-feature fusion according to claim 6, characterized in that, The calculation of the computational weights of the frequency domain texture enhancement features and the global semantic representation features specifically includes: The input features are mapped and calculated by using the first and second layer linear transformation parameters, combined with the ReLU activation function and the Sigmoid function. The calculated weights corresponding to the frequency domain texture enhancement features and the calculated weights corresponding to the global semantic representation features are output respectively.

8. The cross-generator generated image identification method based on multi-feature fusion according to claim 1, characterized in that, The classification based on the fused feature representation, and the output of the judgment result that the image to be detected belongs to a real image or a generated image, specifically includes: The fused feature representation is input into a multilayer perceptron classifier to perform feature dimensionality reduction and classification mapping calculations, and the probability distribution mapping results are obtained. Based on the probability distribution mapping result, the result of the judgment that the image to be detected belongs to a real image or a generated image is output.

9. The cross-generator generated image identification method based on multi-feature fusion according to claim 1, characterized in that, The method also includes an end-to-end training step: A supervised training set is constructed using real image samples and generated image samples; The prediction error between the calculated discrimination result and the true label of the sample; Based on the prediction error, the parameters in the network are updated synchronously using the backpropagation algorithm.

10. A cross-generator generated image identification system based on multi-feature fusion, comprising the cross-generator generated image identification method based on multi-feature fusion according to any one of claims 1-9, characterized in that, include: The preprocessing module is used to resize the input image to be detected, obtain the preprocessed image, and retain the original input image; The image patch rearrangement module is used to locally divide the preprocessed image, calculate the texture richness of each image patch, sort and recombine the image patches according to the texture richness, and generate a rearranged image. The frequency domain texture enhancement module is used to perform high-pass filtering on the rearranged image, perform frequency domain response modeling and multi-spectral attention weighting on the filtered features, and extract features to obtain frequency domain texture enhancement features; The global semantic extraction module is used to extract features from the original input image using a pre-trained visual model to obtain global semantic representation features. An adaptive feature fusion module is used to perform weighted fusion of the frequency domain texture enhancement features and the global semantic representation features to obtain a fused feature representation; The classification and discrimination module is used to classify the image based on the fused feature representation and output the discrimination result of whether the image to be detected belongs to a real image or a generated image.

Images