AIGC image detection method and system based on deep fusion network

Abstract

The invention discloses an AIGC image detection method and system based on a deep fusion network, and relates to the technical field of image processing, and the AIGC image detection method comprises the steps: setting a classification label for initial image data, and obtaining label image data containing the classification label; establishing a feature extraction model based on a time delay neural network, and inputting the label image data into the feature extraction model to obtain visual features of the label image data; fusing the visual features with predefined image element features based on a deep fusion network to form image feature representation; and detecting an artificial intelligence image from the label image data by using the artificial intelligence image detection model to obtain an artificial intelligence image detection result. According to the method, collaborative extraction of global and local features is provided through a multi-channel fusion mode, feature distribution of the AIGC image in different areas can be captured more comprehensively, and therefore the detail anomaly and asymmetry phenomenon of the AIGC image can be recognized.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an AIGC image detection method and system based on a deep fusion network. Background Art

[0002] AIGC image detection (AI-Generated Content Image Detection) refers to the process of using computer algorithms and models to identify and distinguish AI-generated images from real images. With the widespread application of artificial intelligence in the field of image generation, AIGC image detection has become particularly important. For example, AIGC images AI-generated images sometimes appear blurred, discontinuous, or have unnatural transitions in edge processing, especially complex structures (such as fingers, eyes, and facial details), which may have problems such as asymmetry and abnormal shapes. As a result, the light and shadow of AIGC images often do not conform to the laws of real physics, especially in light exposure and shadow formation.

[0003] There may be significant differences in details and textures between AIGC images and real images, such as blurred edges, abnormal details, unnatural noise distribution, etc. If the model lacks multi-level feature extraction capabilities, it will be difficult for the model to accurately capture these subtle differences, resulting in low detection accuracy. Multi-level feature extraction usually covers the global structure and local details of the image. In the absence of this capability, the model may not be able to simultaneously obtain overall features (such as light and shadow effects, structure) and local features (such as texture, edges), making it difficult to identify the comprehensive features of AIGC images and prone to misjudgment.

[0004] In the prior art, images generated by different generative models differ in style and features, but they all lack multi-level feature extraction, which makes it difficult for the model to adapt to these diverse features. Some AIGC images may not be detected, resulting in a high missed detection rate. Multi-level feature extraction can enhance the model's adaptability to noise and external interference. If the model cannot extract features in layers, the detection results may be affected by image quality (such as resolution, clarity) or external noise, resulting in reduced detection reliability.

[0005] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0006] In response to the problems in the related technology, the present invention proposes an AIGC image detection method and system based on a deep fusion network to overcome the above-mentioned technical problems existing in the existing related technology.

[0007] To this end, the specific technical solution adopted by the present invention is as follows:

[0008] According to one aspect of the present invention, an AIGC image detection method based on a deep fusion network is provided, and the AIGC image detection method includes:

[0009] S1, extracting initial image data from a central database, and setting classification labels for the initial image data to obtain labeled image data containing the classification labels;

[0010] S2. Establishing a feature extraction model based on a time-delay neural network, and inputting the label image data into the feature extraction model to obtain visual features of the label image data;

[0011] S3, based on the deep fusion network, the visual features are fused with the predefined image meta-features to form an image feature representation, and an adaptive weight adjustment strategy is configured in the deep fusion network to assign weights to the image feature representation;

[0012] S4. An artificial intelligence image detection model is established based on the image feature representation after the weights are assigned. The artificial intelligence image detection model is used to detect artificial intelligence images (equivalent to AIGC images) from the labeled image data to obtain artificial intelligence image detection results.

[0013] Preferably, the initial image data is extracted from the central database, and a classification label is set for the initial image data, and the label image data containing the classification label is obtained, including:

[0014] S11, extracting initial image data from a central database, and performing data conversion and data standardization on the initial image data in sequence to obtain optimized image data;

[0015] S12. Formulate a label classification system for optimized image data, use a pre-trained image classification model to generate classification labels for the optimized image data, and obtain labeled image data containing the classification labels;

[0016] S13. Randomly extract label image data to review the label classification results, and define a balanced loss function based on the label classification results to optimize the image classification model to balance the amount of label image data in each type of label.

[0017] Preferably, a feature extraction model is established based on a time-delay neural network, and the label image data is input into the feature extraction model to obtain visual features of the label image data including:

[0018] S21, using Fourier transform technology to convert the label image data into spectral components, and combining additional component factors to separate the spectral components into low-frequency components, high-frequency components and residual components;

[0019] S22, establishing a single-channel feature extraction model based on a time-delay neural network, and inputting the low-frequency component, the high-frequency component and the residual component into the single-channel feature extraction model for feature extraction to obtain the initial visual features of the label image data;

[0020] S23, judging the adaptability of the single-channel feature extraction model based on the initial visual features of the label image data, and based on the adaptability judgment result, fusing the single-channel feature extraction model with the group convolutional network and the local feature extraction model respectively to obtain a multi-channel feature extraction model;

[0021] S24, inputting the label image data into a multi-channel feature extraction model, and outputting multi-level visual features of the label image data through the multi-channel feature extraction model.

[0022] Preferably, the label image data is converted into spectral components using Fourier transform technology, and the spectral components are separated into low-frequency components, high-frequency components and residual components by combining additional component factors, including:

[0023] S211, converting the time domain of the label image data into the spectrum domain by using Fourier transform technology, and extracting the spectrum component of the filter from the label image data;

[0024] S212, separating low-frequency components and high-frequency components from the spectrum components, and filtering the label image data in the horizontal direction and the vertical direction of the image respectively to obtain a low-resolution spectrum graph and a high-resolution spectrum graph;

[0025] S213, based on the restriction conditions of the low-resolution spectrum graph and the high-resolution spectrum graph, respectively introduce additional component factors into the low-resolution spectrum graph and the high-resolution spectrum graph and define a binary mask;

[0026] S214, expanding residual components from the low-resolution spectrum graph and the high-resolution spectrum graph based on the binary mask to obtain low-frequency components, high-frequency components and residual components based on the label image data.

[0027] Preferably, the single-channel feature extraction model is fused with the group convolutional network and the local feature extraction model respectively to obtain a multi-channel feature extraction model including:

[0028] S231, filtering scale-invariant features by combining the Euclidean distance between edge points in the label image data and predefined feature points, and extracting semantic features in the label image data by using a group convolutional network;

[0029] S232, fusing the scale-invariant features, semantic features, and initial visual features of the label image data to obtain one branch of the multi-channel feature extraction model;

[0030] S233, performing importance analysis on the regions of interest in the label image data, and calculating the contribution of each region of interest based on the importance analysis result;

[0031] S234, sorting the contribution degrees, selecting local features with contribution degrees within a preset range, and fusing the local features with the initial visual features to obtain another branch of the multi-channel feature extraction model.

[0032] Preferably, performing importance analysis on the regions of interest in the label image data, and calculating the contribution of each region of interest based on the importance analysis result includes:

[0033] S2331, extracting several regions of interest from the label image data, converting the regions of interest into feature vectors, and calculating the marginal contribution of each feature vector;

[0034] S2332, setting a Shapley value for the marginal contribution of each eigenvector, multiplying the marginal contribution of the region of interest by the Shapley value, and obtaining a contribution value of each region of interest;

[0035] S2333. Sort the contribution value of each area of ​​interest from large to small, and select contribution values ​​within a preset range from the sorted list and overlay them on the heat map for visual display.

[0036] Preferably, the expression of the scale-invariant feature of the label image data is:

[0037]

[0038] Wherein, z(a, b) represents the modulus of the scale-invariant feature point of the label image data; (a, b) represents the feature point in the label image data; l(a, b) represents the gradient of the feature point in the label image data.

[0039] Preferably, an artificial intelligence image detection model is established in combination with the image feature representation after the weights are assigned, and the artificial intelligence image detection model is used to detect the artificial intelligence image from the label image data, and the artificial intelligence image detection result obtained includes:

[0040] S41, establishing an artificial intelligence image detection model based on the image feature representation, and formulating a priority detection order of the artificial intelligence image detection model according to the weights of the assigned image feature representations;

[0041] S42, inputting the label image data into the artificial intelligence image detection model, preferentially detecting the label image data corresponding to the image feature representation with high weight according to the priority detection order, and outputting the detection probability value of the label image data;

[0042] S43. Compare the detection probability value with a preset threshold. If the detection probability value is greater than the preset threshold, it indicates that the current label image data is an artificial intelligence image; if the detection probability value is less than the preset threshold, it indicates that the current label image data is a real image.

[0043] Preferably, formulating the priority detection order of the artificial intelligence image detection model according to the weights represented by the assigned image features includes:

[0044] S411, defining a priority detection order based on the weights represented by the image features, and assigning corresponding label image data to the artificial intelligence image detection model according to the priority detection order;

[0045] S412. Count the detection volume of the artificial intelligence image detection model, and set a high priority queue and a low priority queue to balance the detection tasks of each queue of the artificial intelligence image detection model under the condition that the detection tasks are fault-tolerant and schedulable;

[0046] Among them, the high priority queue is used to schedule real-time detection tasks, the low priority queue is used to schedule non-real-time detection tasks, and the high priority queue can extract detection tasks from the low priority queue.

[0047] According to another aspect of the present invention, there is also provided an AIGC image detection system based on a deep fusion network, the AIGC image detection system comprising an image optimization module, a visual feature extraction module, a feature fusion module and an image detection module;

[0048] An image optimization module is used to extract initial image data from a central database and set classification labels for the initial image data to obtain labeled image data containing the classification labels;

[0049] A visual feature extraction module is used to establish a feature extraction model based on a time-delay neural network, and input the label image data into the feature extraction model to obtain the visual features of the label image data;

[0050] A feature fusion module is used to fuse visual features with predefined image meta-features based on a deep fusion network to form an image feature representation, and configure an adaptive weight adjustment strategy in the deep fusion network to assign weights to the image feature representation;

[0051] The image detection module is used to establish an artificial intelligence image detection model in combination with the image feature representation after the weights are assigned, and use the artificial intelligence image detection model to detect artificial intelligence images from the labeled image data to obtain artificial intelligence image detection results.

[0052] The beneficial effects of the present invention are:

[0053] 1. The present invention generates classification labels for optimized image data through an image classification model and establishes a label classification system, thereby ensuring the accurate classification of label image data, helping to improve the classification effect of subsequent detection models, and reducing detection errors caused by label errors. The label classification results are reviewed by randomly sampling label image data, and the classification model results are verified to ensure the accuracy and consistency of the classification results, so as to balance the amount of image data in various types of labels and avoid category imbalance problems in the model during training, thereby improving the ability to recognize images and enhancing the generalization and robustness of the detection model.

[0054] 2. The present invention provides collaborative extraction of global and local features through a multi-channel fusion method, which can more comprehensively capture the feature distribution of AIGC images in different areas, thereby identifying the detail anomalies and asymmetry phenomena of AIGC images, and utilizes the common characteristics of AIGC images such as blurred edges, abnormal details, unnatural light and shadow, and uneven noise distribution. Through multi-dimensional spectral component extraction, fusion of group convolutional networks and local feature extraction models, multi-channel contribution analysis and other technologies, the detection of AIGC images is more accurate and comprehensive, further effectively improving the accuracy of AIGC image detection.

[0055] 3. The present invention formulates a priority detection order by assigning weighted image features, so that the model can give priority to processing high-weight image features. This strategy gives priority to the detection of important feature information, ensures detection accuracy while reducing the consumption of computing resources, and sets high-priority and low-priority queues to maintain balanced scheduling of detection tasks, ensuring that the model can still process low-priority tasks after high-priority tasks are completed, avoiding the decrease in detection accuracy due to improper resource allocation. In terms of fault tolerance and schedulability of detection tasks, the detection load is balanced through priority queues to reduce the bottleneck problem caused by high-priority feature tasks, thereby ensuring the operational stability of the detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0057] Figure 1 is a flow chart of an AIGC image detection method based on a deep fusion network according to an embodiment of the present invention;

[0058] Figure 2 It is a principle block diagram of an AIGC image detection system based on a deep fusion network according to an embodiment of the present invention.

[0059] In the figure:

[0060] 1. Image optimization module; 2. Visual feature extraction module; 3. Feature fusion module; 4. Image detection module. DETAILED DESCRIPTION

[0061] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments. They can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and advantages of the present invention.

[0062] According to an embodiment of the present invention, a method and system for AIGC image detection based on a deep fusion network are provided.

[0063] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to an AIGC image detection method based on a deep fusion network according to an embodiment of the present invention, the AIGC image detection method includes:

[0064] S1. Extracting initial image data from a central database and setting classification labels for the initial image data to obtain labeled image data containing the classification labels.

[0065] The initial image data is extracted from the central database, and a classification label is set for the initial image data, and the obtained labeled image data containing the classification label includes:

[0066] S11, extracting initial image data from a central database, and performing data conversion and data standardization processing on the initial image data in sequence to obtain optimized image data.

[0067] It should be noted that the initial image data is extracted from the central database, and the initial image data is sequentially converted and standardized to obtain the optimized image data including:

[0068] The required initial image data is retrieved and loaded from the central database, and the initial image data includes image files in various formats.

[0069] Format conversion: Convert image data to a target format, such as converting RAW format to PNG or JPEG, to ensure data consistency.

[0070] Color space conversion: Unify the color space of image data, such as converting from RGB to grayscale (in suitable detection tasks) or YCbCr color space, so as to focus on specific color information during processing.

[0071] Resolution adjustment: Unify the image resolution according to the requirements of the detection task, such as scaling it to a fixed size or processing it at a specific resolution, to facilitate subsequent feature extraction and model input.

[0072] Image enhancement: By adjusting brightness, contrast, sharpening and other enhancement operations, the image quality is improved to make its features clearer, which is conducive to subsequent standardization and detection processing.

[0073] After data conversion and standardization, the generated optimized image data has a unified format, resolution and numerical range, which reduces the diversity and noise interference of the input data and improves the quality of the model input data. The optimized image data provides a reliable foundation for subsequent feature extraction and detection, and can significantly improve the performance and accuracy of the detection model.

[0074] S12. Formulate a label classification system for optimized image data, use a pre-trained image classification model to generate classification labels for the optimized image data, and obtain labeled image data containing the classification labels;

[0075] S13. Randomly extract label image data to review the label classification results, and define a balanced loss function based on the label classification results to optimize the image classification model to balance the amount of label image data in each type of label.

[0076] It should be noted that a label classification system for optimizing image data is formulated, and a pre-trained image classification model is used to generate classification labels for the optimized image data. The obtained labeled image data containing classification labels include:

[0077] According to the characteristics of image data and expected application requirements, formulate and optimize the label classification system of image data to ensure that the label classification system covers all possible image types to improve the applicability of the classification model.

[0078] Use a pre-trained image classification model (such as residual network, Inception network, etc.) to perform classification prediction on the optimized image data and generate classification labels, which are then attached to the optimized image data to form "labeled image data" containing classification labels.

[0079] Randomly extract label image data to review the label classification results, and define a balanced loss function based on the label classification results to optimize the image classification model to balance the amount of label image data in each type of label, including:

[0080] Randomly extract a certain proportion of samples from the generated labeled image data and manually review their classification labels to ensure the accuracy of the labels, for example:

[0081] Small-scale data sets (number of samples < 10,000): Usually 20%-30% of the samples are extracted for review. When the amount of data is small, a higher proportion can be extracted to more accurately verify the label classification quality of the model.

[0082] Medium-sized datasets (sample size between 10,000 and 100,000): 10%-15% of the samples can be sampled for review. For medium-sized datasets, sampling more than 10% of the samples can effectively detect label errors and correct them.

[0083] Large-scale data sets (sample size > 100,000): Generally, 5%-10% of the samples are taken. For large data sets, selecting 5% to 10% can reduce the cost of verification while maintaining high accuracy.

[0084] Minority categories with specific labels: For minority categories (samples account for less than 5%), the review ratio can be appropriately increased, such as 30%-50%, to ensure the accuracy of minority categories and reduce the errors caused by category imbalance.

[0085] For misclassified samples, readjust their labels or correct the label classification system to ensure the labeling quality of the sample data.

[0086] Define a balanced loss function to optimize the classification model:

[0087] Based on the review results, the distribution of various labels in the labeled image data is analyzed. For unbalanced label categories (for example, there are fewer samples in the minority category), a balanced loss function is defined to improve the performance of the model on the minority category.

[0088] Balanced loss functions such as weighted cross entropy and focal loss are used to give higher weights to minority categories, ensuring that the bias towards label categories with large amounts of data is reduced during model training. For example, the expression of the weighted cross entropy loss function is:

[0089]

[0090] In the formula, N represents the number of samples, y i represents the actual label, represents the predicted probability, w yi Represents the label y i The weight of .

[0091] The adjusted labeled data is reused for training or fine-tuning the image classification model, so that the model can further optimize the classification performance under the constraint of the balanced loss function.

[0092] S2. Establish a feature extraction model based on a time-delay neural network, and input the label image data into the feature extraction model to obtain the visual features of the label image data.

[0093] Among them, a feature extraction model is established based on a time-delay neural network, and the label image data is input into the feature extraction model, and the visual features of the label image data obtained include:

[0094] S21. Convert the label image data into spectral components using Fourier transform technology, and separate the spectral components into low-frequency components, high-frequency components and residual components by combining additional component factors.

[0095] Among them, the Fourier transform technology is used to convert the label image data into spectral components, and the spectral components are separated into low-frequency components, high-frequency components and residual components by combining additional component factors, including:

[0096] S211 , using Fourier transform technology to convert the time domain of the label image data into the spectrum domain, and extracting the spectrum components of the filter from the label image data.

[0097] It should be noted that Fourier transform includes:

[0098] The label image data is converted from the spatial (time) domain to the spectral domain to analyze its frequency components. The image in the spatial domain represents the spatial distribution of pixels, while in the spectral domain, the image is decomposed into various frequency components (low frequency and high frequency).

[0099] The fast Fourier transform (FFT) algorithm is used to perform a two-dimensional Fourier transform on each image to obtain its spectrum representation. The Fourier transform regards the brightness value of each pixel as a sine wave and analyzes the amplitude and frequency of different waves. The fast Fourier transform (FFT) algorithm is used to perform a Fourier transform on the two-dimensional image data, including:

[0100] The two-dimensional FFT transforms an image from the spatial domain (pixel values ​​in a spatial coordinate system) to the frequency domain (a combination of frequency and phase). The output is a matrix of complex values, where each complex number represents the amplitude (absolute value) and phase (angle) of a specific frequency component.

[0101] Low-frequency components: In the central area of ​​the spectrum graph, they represent the overall outline and basic structural information of the image. These components correspond to slower changes in brightness or color (smooth changes over a large range).

[0102] High-frequency components: distributed in the edge area of ​​the spectrum graph, representing the details and edge information of the image. These components correspond to rapidly changing brightness or color changes (such as fine edges and textures).

[0103] In the spectrogram, the absolute value (amplitude) of the complex value represents the intensity (or energy) of a specific frequency, while the phase value represents the phase information of these frequency components. By selectively retaining or removing the spectral information of specific areas (such as low-frequency or high-frequency areas), the structural information and detail information of the image can be extracted respectively for subsequent steps such as feature extraction or classification.

[0104] Suppose a 256×256 grayscale image is subjected to Fourier transform:

[0105] After performing a two-dimensional Fourier transform, a 256×256 complex matrix is ​​obtained. The low-frequency part in the center of the matrix represents the overall contour information of the image, and the high-frequency part at the edge of the matrix represents the details of the image. When visualizing the spectrum, the amplitude can be scaled by logarithmic transformation to more clearly observe different frequency components.

[0106] Spectral component extraction:

[0107] In the spectrum diagram, the low-frequency components of the image are concentrated in the center of the spectrum, representing the overall structure and brightness changes of the image, while the high-frequency components are located around it, representing the details and edge information of the image.

[0108] Bandpass filters are used to extract components of specific frequencies in the spectrum domain. The design of the filter determines the extracted frequency band. Different bandwidths can be selected according to the application to extract the desired low-frequency and high-frequency components.

[0109] S212, separating low-frequency components and high-frequency components from the spectrum components, and filtering the label image data in the horizontal direction and the vertical direction of the image respectively to obtain a low-resolution spectrum graph and a high-resolution spectrum graph;

[0110] S213, based on the restriction conditions of the low-resolution spectrum graph and the high-resolution spectrum graph, additional component factors are introduced into the low-resolution spectrum graph and the high-resolution spectrum graph respectively, and then a binary mask is defined.

[0111] It should be noted that the low-frequency and high-frequency restrictions are set according to the actual application requirements. For example, if the low-frequency component must account for 80% of the total spectrum energy and the high-frequency component for 20%, this restriction can be set by the spectrum energy distribution. If the energy ratio of the low frequency and the high frequency does not meet the preset conditions, the balance of the spectrum energy can be controlled by adjusting the additional component factor so that the low-frequency and high-frequency components reach the expected values ​​respectively, where the binary mask includes:

[0112] A binary mask matrix is ​​generated based on the constraints to identify low-frequency and high-frequency areas in the image. The "1" in the mask indicates areas in the spectrum that meet the low-frequency or high-frequency conditions, while "0" indicates areas that do not meet the conditions. Additional component factors are introduced to the low-frequency and high-frequency maps so that the value of each pixel position meets the constraints, and then the results are binarized to generate a binary mask.

[0113] S214, expanding residual components from the low-resolution spectrum graph and the high-resolution spectrum graph based on the binary mask to obtain low-frequency components, high-frequency components and residual components based on the label image data.

[0114] S22. A single-channel feature extraction model is established based on a time-delay neural network, and low-frequency components, high-frequency components and residual components are input into the single-channel feature extraction model for feature extraction to obtain initial visual features of the label image data.

[0115] It should be noted that a single-channel feature extraction model is established based on a time-delay neural network, and the low-frequency component, high-frequency component and residual component are input into the single-channel feature extraction model for feature extraction, and the initial visual features of the label image data obtained include:

[0116] Step 1: Build a single-channel time-delay neural network. The time-delay neural network can capture data features under different time delays and is suitable for extracting features from frequency and residual information. Design multiple layers of one-dimensional convolutional layers and pooling layers. The convolutional layer can effectively extract features under different time delays, and the pooling layer can gradually reduce the size of the feature map to extract more representative features. Use the ReLU activation function to improve the nonlinear fitting ability of the model, so that it can better capture the complex features in low-frequency, high-frequency and residual components.

[0117] Step 2: Convert the low-frequency component, high-frequency component, and residual component into a format suitable for single-channel input. Assuming that the data shape of each component is a two-dimensional matrix (for example, 256×256), they need to be standardized and converted into a single-channel format to ensure the consistency of the input. In practical applications, you can choose to concatenate the three components (low-frequency, high-frequency, and residual) into a single input data stream, or input them one by one into TDNN for feature extraction so that the model can process each type of spectral information in turn.

[0118] Step 3: Concatenate or fuse the low-frequency, high-frequency and residual feature representations in the model, and form a joint feature vector in a weighted manner. This can comprehensively consider the information of each component, making the final feature expression more comprehensive. After feature fusion, the features are input into the fully connected layer to further extract high-dimensional features to generate the final initial visual features.

[0119] S23. The adaptability of the single-channel feature extraction model is judged based on the initial visual features of the label image data, and based on the adaptability judgment result, the single-channel feature extraction model is respectively fused with the group convolutional network and the local feature extraction model to obtain a multi-channel feature extraction model.

[0120] Among them, the single-channel feature extraction model is fused with the group convolutional network and the local feature extraction model respectively, and the multi-channel feature extraction model is obtained, including:

[0121] S231. Filter scale-invariant features by combining the Euclidean distance between edge points in the label image data and predefined feature points, and use a group convolutional network to extract semantic features in the label image data.

[0122] It should be noted that the scale-invariant features selected by combining the Euclidean distance between the edge points in the label image data and the predefined feature points include:

[0123] Step 1: Use the edge detection algorithm (equivalent to the Canny algorithm) to detect edge points in the image. The edge points are concentrated in the contours and mutation areas of the image. For example, for a 256×256 image, the Canny edge detection may output a set of edge point coordinates such as (x1, y1), (x2, y2), etc.

[0124] The key point detection algorithm (equivalent to Scale-Invariant Feature Transform) is used to identify predefined feature points in the image. Feature points represent points that remain stable under different scales and rotations, usually in the corners, textures, or high-variance areas of the image. The SIFT algorithm outputs a set of coordinates of the feature point as well as the magnitude and direction of the feature point (usually expressed as gradient magnitude and angle).

[0125] Step 2: Calculate the Euclidean distance between each edge point and feature point, and set the distance threshold for screening: Set a distance threshold, for example, 5 pixels, and screen out edge point-feature point pairs that meet this distance threshold. The selected point pairs are relatively stable at a certain scale and have scale invariance.

[0126] Step 3: Combine the information of mode and direction to refine the screening of scale-invariant features:

[0127] Magnitude analysis: The magnitude of the magnitude indicates the gradient amplitude of the feature point, which can measure the significance of the feature. Usually, points with larger gradient amplitudes are selected to ensure that the feature points have significant local differences. The screening threshold M is defined as agnitude >M thresh , for example, setting M thresh =10, only retain feature points with gradient magnitude greater than 10.

[0128] Direction analysis: The direction of each feature point represents the gradient direction, which is used for screening based on rotation invariance. On the basis of scale invariance, direction invariance is further increased to ensure that the image can still match the feature point when it is rotated. Select direction range θ min≤θ≤θ max to match the target direction, for example, if horizontal edge points are required, the direction range can be set to [85°, 95°][85°, 95°].

[0129] Among them, the expression of the scale-invariant feature of the label image data is:

[0130]

[0131]

[0132] Wherein, z(a, b) represents the modulus of the scale-invariant feature point of the label image data; (a, b) represents the feature point in the label image data; l(a, b) represents the gradient of the feature point in the label image data; θ(a, b) represents the direction of the scale-invariant feature point of the label image data.

[0133] Step 4: The edge point-feature point pairs selected by distance, modulus and direction are used as the final scale-invariant feature set. For each pair of edge point-feature point pairs that meet the conditions, their coordinates, modulus and direction information are retained to form a feature descriptor, which is used for subsequent feature extraction and matching.

[0134] It should be noted that the semantic features extracted from the labeled image data using the group convolutional network include:

[0135] Step 1: Group convolution is a group convolution strategy that divides the input channels into different groups, and each group of channels performs convolution operations independently. Compared with traditional convolution, group convolution can significantly reduce computational complexity while enhancing feature diversity, and is suitable for processing high-dimensional image feature data.

[0136] Step 2: Use the initial visual features (such as low-frequency, high-frequency and residual components) obtained by the single-channel feature extraction model and the selected scale-invariant feature set to concatenate the feature maps of different components to form a multi-channel input feature matrix. For example, for each 256×256 image, the concatenated feature map matrix may be (256, 256, 3), where each channel represents a different feature component (such as low-frequency features, high-frequency features and residual features).

[0137] Step 3: Divide the number of channels of the input feature map into multiple groups to reduce computational overhead. Assuming that the number of channels of the initial feature map is 3, it can be set to 3 groups to form a 1-channel group convolution and build a multi-layer group convolution network. The group convolution of each layer operates independently on each group of feature maps to capture information between different channels.

[0138] The first layer of group convolution: using a 3×3 convolution kernel, convolution operations are performed independently in each group to extract preliminary edge and detail features.

[0139] Subsequent layer group convolution: The convolution kernel size is increased layer by layer (such as 5×5) to capture more complex semantic patterns, and the outputs of each group are merged to form a comprehensive feature map.

[0140] Pooling layer: After each convolution operation, a pooling layer (such as maximum pooling or average pooling) is inserted to gradually reduce the size of the feature map and retain the main features. The pooling layer helps to perform multi-scale analysis of the image and thus extract stable semantic features.

[0141] Step 4: Activation function and batch normalization:

[0142] Activation function: After each layer of group convolution, the ReLU activation function is used to introduce nonlinearity and make the model more expressive.

[0143] Batch normalization: Batch normalization is applied after each convolution layer to ensure the stability of feature data and accelerate network convergence. Batch normalization can eliminate data bias and ensure that the features extracted by the network remain consistent across groups.

[0144] Step 5. After the last layer of group convolution, the output results of each group of convolutions are combined to form a final feature vector, which represents the semantic information in the image and can characterize the content and structure of the image. The output feature map is reduced in dimension or globally averaged pooled to further obtain the semantic feature vector of the image. This vector contains the semantic information of the image and helps to classify or identify specific image patterns.

[0145] S232, fusing the scale-invariant features, semantic features, and initial visual features of the label image data to obtain one branch of the multi-channel feature extraction model;

[0146] S233: Perform importance analysis on the regions of interest in the label image data, and calculate the contribution of each region of interest based on the importance analysis result.

[0147] Among them, the importance analysis of the interest region in the label image data is performed, and based on the importance analysis result, the contribution of each interest region is calculated, including:

[0148] S2331, extracting several regions of interest from the label image data, converting the regions of interest into feature vectors, and calculating the marginal contribution of each feature vector;

[0149] S2332. Develop a Shapley value for the marginal contribution of each eigenvector, multiply the marginal contribution of the region of interest by the Shapley value, and obtain the contribution value of each region of interest.

[0150] It should be noted that the Shapley value is formulated for the marginal contribution of each eigenvector, and the marginal contribution of the region of interest is multiplied by the Shapley value to obtain the contribution value of each region of interest, including:

[0151] Step 1: Define a baseline contribution value, which can be the overall feature performance of the model without the features of the region of interest, such as classification accuracy or feature difference index, and calculate its marginal contribution by adding or removing the features of a region of interest one by one. Assume that the baseline contribution value is V baseline , after adding or removing a region of interest i, we get a new contribution V withi , then the marginal contribution can be expressed as:

[0152] ΔV i =V withi -V baseline ;

[0153] In the formula, the marginal contribution ΔV i Represents the unique contribution of the interest region i to the overall feature.

[0154] Step 2: Shapley value is an allocation method introduced from cooperative game theory to fairly measure the contribution of each feature in the overall feature combination. Calculating the Shapley value of each region of interest includes:

[0155] For each region of interest i, calculate its marginal contribution in all possible feature subsets. Assuming there are N regions of interest, the marginal contribution in all combinations can be expressed as the average marginal contribution of each region of interest in all other feature combinations.

[0156] The Shapley value calculates the average marginal contribution of region of interest i in all possible combinations, ensuring that its importance is measured in different combination scenarios.

[0157] Step 3: The final contribution value is represented by the product of the marginal contribution and the Shapley value, which is used to determine the comprehensive contribution of the region of interest to the overall characteristics. For region of interest i, its contribution value Ci can be expressed as:

[0158] Ci=ΔVi×φi;

[0159] Where ΔVi represents the marginal contribution of the region of interest i, and φi represents its Shapley value.

[0160] By multiplying the marginal contribution and the Shapley value, the comprehensive contribution of each interest area in different combinations can be quantified, thereby ensuring the fairness and stability of the evaluation. For example, assuming there are three interest areas A, B and C, the calculation process is as follows:

[0161] Calculate the marginal contribution of each region of interest relative to the baseline contribution, for example:

[0162] ΔV A =0.2;

[0163] ΔV B =0.3;

[0164] ΔV C =0.15;

[0165] By calculating the marginal contribution of each subset combination, we can obtain the respective Shapley values, for example:

[0166] φA=0.25;

[0167] φB=0.35;

[0168] φC=0.2;

[0169] Finally, the contribution value of each area of ​​interest is:

[0170] C A =0.2×0.25=0.05;

[0171] C B =0.3×0.35=0.105;

[0172] C C =0.15×0.2=0.03.

[0173] S2333. Sort the contribution value of each area of ​​interest from large to small, and select contribution values ​​within a preset range from the sorted list and overlay them on the heat map for visual display.

[0174] S234, sorting the contribution degrees, selecting local features with contribution degrees within a preset range, and fusing the local features with the initial visual features to obtain another branch of the multi-channel feature extraction model.

[0175] S24, inputting the label image data into a multi-channel feature extraction model, and outputting multi-level visual features of the label image data through the multi-channel feature extraction model.

[0176] It should be noted that the fusion features of the two branches of the multi-channel feature extraction model include:

[0177] The first branch covers the overall structure, edge and pattern information of the image, which helps to capture the generation pattern and overall semantic characteristics of AIGC images.

[0178] The second branch focuses on capturing local features of important areas of interest, which helps to identify detail differences and abnormal features in local areas of the image.

[0179] Assume that feature extraction is performed on different categories of AIGC images:

[0180] Generated Image 1: The first branch features show that the image has highly repetitive edge patterns, and the second branch presents irregular noise features in key interest areas.

[0181] Generated image 2: The first branch feature shows that the overall texture of the image is relatively smooth, and there is no obvious mutation in the local area features, which is consistent with the feature pattern generated by VAE.

[0182] S3. Based on the deep fusion network, the visual features are fused with the predefined image meta-features to form an image feature representation, and an adaptive weight adjustment strategy is configured in the deep fusion network to assign weights to the image feature representation.

[0183] It should be noted that the deep fusion network is a neural network structure that deeply integrates multiple feature sources (such as visual features, image meta-features, text descriptions, contextual information, etc.). Its core idea is to fuse information from different feature sources into a unified feature representation through multi-layer nonlinear transformations and adaptive weight adjustments to improve the model's ability to understand and process complex data. Among them, the deep fusion network is used to fuse visual features with predefined image meta-features to form an image feature representation, and an adaptive weight adjustment strategy is configured in the deep fusion network to assign weights to the image feature representation, including:

[0184] Combine visual features with image meta-features using the feature fusion layer of a deep fusion network. Features can be fused via concatenation, weighted averaging, or fully connected layers:

[0185] Concatenation: Concatenate visual features with image meta-features in the channel dimension to form a high-dimensional feature representation.

[0186] Weighted average: Assign initial weights to the visual features and image meta-features respectively, and calculate the weighted average features.

[0187] Fully connected layer: The concatenated features are linearly combined through the fully connected layer.

[0188] Adaptive weight adjustment module:

[0189] The core of the adaptive weight adjustment strategy is to assign dynamically adjusted weights to different feature sources, so that the model can adaptively adjust the importance of features according to different image contents. The main methods for implementing adaptive weights include attention mechanism and gating mechanism.

[0190] Attention Mechanism:

[0191] The self-attention mechanism is used to calculate the weights of visual features and image meta-features. The weight vectors α and β are generated based on the input features to represent the importance of visual features and image meta-features. The formula of the weight vector generated by the self-attention mechanism is as follows:

[0192] α=softmax(Wv·V+bv);

[0193] β = softmax(Wm·M+bm);

[0194] Where V represents visual features, M represents image element features, Wv and Wm represent weight matrices, bv and bm represent bias terms, and the weights are distributed between 0 and 1 after passing through the softmax function to ensure that the sum is 1.

[0195] The gate unit is used to control the weight adjustment of each feature source. The gate mechanism dynamically adjusts the proportion of different features in the final feature representation by learning the correlation between different features. The gate unit outputs a weight value for weighting visual features and image meta-features.

[0196] Feature fusion: Use adaptive weights to perform a weighted combination of visual features and image meta-features to generate the final image feature representation F:

[0197] F = α·V + β·M;

[0198] Where α and β represent the adaptive weights of visual features and image meta-features, respectively, and V and M represent the standardized visual features and image meta-features.

[0199] The deep fusion network is trained through supervised learning, and the weight coefficients of the features are gradually optimized using the adaptive weight adjustment module. Back propagation optimizes the weight parameters and attention modules in the network, so that the adaptive weights can be automatically adjusted according to task requirements.

[0200] Feature weight regularization: To prevent feature weight imbalance, regularization methods are used to constrain the weights to prevent the weight of a single feature source from being too high. For example, the visual features of an image are 512-dimensional vectors, and the image meta-features are 128-dimensional vectors.

[0201] Feature concatenation: concatenate the 512-dimensional visual features and the 128-dimensional meta-features to form a 640-dimensional input.

[0202] Attention mechanism generates weights: The visual feature weight α=0.7 and the meta-feature weight β=0.3 are calculated through the attention module.

[0203] Feature fusion: Weighting visual features and meta-features to generate the final image feature representation:

[0204] F = 0.7·V + 0.3·M;

[0205] The fused image feature representation is obtained for subsequent image classification or feature matching tasks.

[0206] S4. An artificial intelligence image detection model is established based on the image feature representation after the weights are assigned. The artificial intelligence image detection model is used to detect artificial intelligence images from the labeled image data to obtain artificial intelligence image detection results.

[0207] Among them, an artificial intelligence image detection model is established in combination with the image feature representation after the weight is assigned, and the artificial intelligence image detection model is used to detect the artificial intelligence image from the label image data, and the artificial intelligence image detection results obtained include:

[0208] S41. Establish an artificial intelligence image detection model based on the image feature representation, and formulate a priority detection order of the artificial intelligence image detection model according to the weights of the assigned image feature representation.

[0209] Among them, the priority detection order of the artificial intelligence image detection model is formulated according to the weights represented by the assigned image features, including:

[0210] S411, defining a priority detection order based on the weights represented by the image features, and assigning corresponding label image data to the artificial intelligence image detection model according to the priority detection order;

[0211] S412. Count the detection volume of the artificial intelligence image detection model, and set high-priority queues and low-priority queues to balance the detection tasks of each queue of the artificial intelligence image detection model while ensuring that the detection tasks are fault-tolerant and schedulable.

[0212] It should be noted that in the deep fusion network mentioned above, the feature representation of each image has been extracted and calculated, and each feature has been weighted according to its contribution (through an adaptive weight adjustment strategy). These weight values ​​reflect the importance of the image in AI detection. Usually, a high-weighted image indicates that the image has more significant artificial features, while a low-weighted image tends to have real image features.

[0213] Priority detection order definition:

[0214] According to the weight of the image features, all image data are sorted from high weight to low weight. High-weight images are processed first because they are more likely to be AI-generated images. These sorted image data are assigned to the AI ​​image detection model and input into the detection model according to the preset priority order.

[0215] During operation, the number of images processed by the AI ​​image detection model is counted in real time, and the detection volume in each priority queue is recorded to monitor the load and model processing speed.

[0216] To set up high and low priority queues:

[0217] High-priority queue: Used for real-time detection tasks to process image data corresponding to high-weight image feature representations. Since these images are highly likely to be generated by artificial intelligence, they need to be processed preferentially.

[0218] Low-priority queue: Used for non-real-time detection tasks to process image data corresponding to low-weight image feature representations, which may be images with a lower probability of being generated by artificial intelligence or real images.

[0219] Queue balancing: To ensure the balance of each queue, a dynamic scheduling strategy is adopted. For example, when the number of tasks in the high-priority queue is insufficient, tasks can be extracted from the low-priority queue for processing, so as to make full use of system resources and ensure queue balance.

[0220] Among them, the high-priority queue is used to schedule real-time detection tasks, the low-priority queue is used to schedule non-real-time detection tasks, and the high-priority queue can extract detection tasks from the low-priority queue.

[0221] S42. Input the labeled image data into the artificial intelligence image detection model, and preferentially detect the labeled image data corresponding to the high-weight image feature representations according to the priority detection order, and output the detection probability value of the labeled image data;

[0222] S43. Compare the detection probability value with a preset threshold. If the detection probability value is greater than the preset threshold, it means that the current labeled image data is an artificial intelligence image; if the detection probability value is less than the preset threshold, it means that the current labeled image data is a real image.

[0223] It should be noted that the labeled image data assigned to the high-priority and low-priority queues is input into the artificial intelligence image detection model in the order of priority; the images in the high-priority queue will enter the detection model for processing first; the image data in the low-priority queue will be detected only when the tasks in the high-priority queue are processed or the resources are idle.

[0224] The model outputs a detection probability value for each image, indicating the probability that the current image is recognized as an image generated by artificial intelligence. A detection probability threshold T is preset to determine whether the image is generated by artificial intelligence. If the detection probability value P is greater than T, then the image is highly likely to be an image generated by artificial intelligence; otherwise, it is determined to be a real image. For example:

[0225] If P≥T: Then the current labeled image data is marked as "image generated by artificial intelligence".

[0226] If P<T: Then the current labeled image data is marked as "real image".

[0227] According to another embodiment of the present invention, there is provided an AIGC image detection system based on a deep fusion network, the AIGC image detection system comprising an image optimization module, a visual feature extraction module, a feature fusion module and an image detection module;

[0228] The image optimization module 1 is used to extract the initial image data from the central database and set classification labels for the initial image data to obtain label image data containing the classification labels;

[0229] Visual feature extraction module 2, used to establish a feature extraction model based on a time-delay neural network, and input the label image data into the feature extraction model to obtain the visual features of the label image data;

[0230] A feature fusion module 3 is used to fuse the visual features with predefined image element features based on a deep fusion network to form an image feature representation, and configure an adaptive weight adjustment strategy in the deep fusion network to assign weights to the image feature representation;

[0231] The image detection module 4 is used to establish an artificial intelligence image detection model in combination with the image feature representation after the weights are assigned, and use the artificial intelligence image detection model to detect artificial intelligence images from the label image data to obtain artificial intelligence image detection results.

[0232] In summary, with the help of the above technical solution of the present invention, the present invention generates classification labels for optimized image data through an image classification model, and establishes a label classification system, ensuring the accurate classification of label image data, which helps to improve the classification effect of subsequent detection models, reduce detection errors caused by label errors, and verify the label classification results by randomly extracting label image data, verifying the classification model results, ensuring the accuracy and consistency of the classification results, balancing the amount of image data in various labels, so that the model avoids the problem of category imbalance during training, thereby improving the recognition ability of images and enhancing the generalization ability and robustness of the detection model. The present invention provides collaborative extraction of global and local features through a multi-channel fusion method, which can more comprehensively capture the feature distribution of AIGC images in different regions, thereby identifying the abnormal details and asymmetry of AIGC images, and utilizing the common characteristics of edge blur, abnormal details, unnatural light and shadow, and uneven noise distribution in AIGC images. Through multi-dimensional spectral component extraction, fusion of group convolutional networks and local feature extraction models, multi-channel contribution analysis and other technologies, the detection of AIGC images is more accurate and comprehensive, and the accuracy of AIGC image detection is further effectively improved. The present invention formulates a priority detection order by assigning weighted image feature representations, so that the model can give priority to processing high-weight image features. This strategy gives priority to detecting important feature information, ensuring detection accuracy while reducing the consumption of computing resources, and setting high-priority and low-priority queues to maintain balanced scheduling of detection tasks, ensuring that the model can still process low-priority tasks after high-priority tasks are completed, avoiding the decrease in detection accuracy due to improper resource allocation. In terms of fault tolerance and schedulability of detection tasks, the detection load is balanced through priority queues, reducing the bottleneck problem caused by high-priority feature tasks, thereby ensuring the operational stability of the detection model.

[0233] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An AIGC image detection method based on a deep fusion network, characterized in that: The AIGC image detection method includes: S1, extracting initial image data from a central database, and setting classification labels for the initial image data to obtain labeled image data containing the classification labels; S2. Establishing a feature extraction model based on a time-delay neural network, and inputting the label image data into the feature extraction model to obtain visual features of the label image data; S3, based on the deep fusion network, the visual features are fused with the predefined image meta-features to form an image feature representation, and an adaptive weight adjustment strategy is configured in the deep fusion network to assign weights to the image feature representation; S4. An artificial intelligence image detection model is established based on the image feature representation after the weights are assigned. The artificial intelligence image detection model is used to detect artificial intelligence images from the labeled image data to obtain artificial intelligence image detection results.

2. According to claim 1, the AIGC image detection method based on deep fusion network is characterized in that: The extracting of the initial image data from the central database and setting a classification label for the initial image data to obtain the label image data containing the classification label includes: S11, extracting initial image data from a central database, and performing data conversion and data standardization on the initial image data in sequence to obtain optimized image data; S12. Formulate a label classification system for optimized image data, use a pre-trained image classification model to generate classification labels for the optimized image data, and obtain labeled image data containing the classification labels; S13. Randomly extract label image data to review the label classification results, and define a balanced loss function based on the label classification results to optimize the image classification model to balance the amount of label image data in each type of label.

3. The AIGC image detection method based on deep fusion network according to claim 1 is characterized in that: The feature extraction model is established based on the time-delay neural network, and the label image data is input into the feature extraction model to obtain the visual features of the label image data, including: S21, using Fourier transform technology to convert the label image data into spectral components, and combining additional component factors to separate the spectral components into low-frequency components, high-frequency components and residual components; S22, establishing a single-channel feature extraction model based on a time-delay neural network, and inputting the low-frequency component, the high-frequency component and the residual component into the single-channel feature extraction model for feature extraction to obtain the initial visual features of the label image data; S23, judging the adaptability of the single-channel feature extraction model based on the initial visual features of the label image data, and based on the adaptability judgment result, fusing the single-channel feature extraction model with the group convolutional network and the local feature extraction model respectively to obtain a multi-channel feature extraction model; S24, inputting the label image data into a multi-channel feature extraction model, and outputting multi-level visual features of the label image data through the multi-channel feature extraction model.

4. The AIGC image detection method based on deep fusion network according to claim 3 is characterized in that: The method of converting the label image data into spectral components by using Fourier transform technology and separating the spectral components into low-frequency components, high-frequency components and residual components by combining additional component factors includes: S211, converting the time domain of the label image data into the spectrum domain by using Fourier transform technology, and extracting the spectrum component of the filter from the label image data; S212, separating low-frequency components and high-frequency components from the spectrum components, and filtering the label image data in the horizontal direction and the vertical direction of the image respectively to obtain a low-resolution spectrum graph and a high-resolution spectrum graph; S213, based on the restriction conditions of the low-resolution spectrum graph and the high-resolution spectrum graph, respectively introduce additional component factors into the low-resolution spectrum graph and the high-resolution spectrum graph and define a binary mask; S214, expanding residual components from the low-resolution spectrum graph and the high-resolution spectrum graph based on the binary mask to obtain low-frequency components, high-frequency components and residual components based on the label image data.

5. The AIGC image detection method based on deep fusion network according to claim 4 is characterized in that: The single-channel feature extraction model is respectively integrated with the group convolutional network and the local feature extraction model to obtain a multi-channel feature extraction model, including: S231, filtering scale-invariant features by combining the Euclidean distance between edge points in the label image data and predefined feature points, and extracting semantic features in the label image data by using a group convolutional network; S232, fusing the scale-invariant features, semantic features, and initial visual features of the label image data to obtain one branch of the multi-channel feature extraction model; S233, performing importance analysis on the regions of interest in the label image data, and calculating the contribution of each region of interest based on the importance analysis result; S234, sorting the contribution degrees, selecting local features with contribution degrees within a preset range, and fusing the local features with the initial visual features to obtain another branch of the multi-channel feature extraction model.

6. The AIGC image detection method based on deep fusion network according to claim 5 is characterized in that: The performing importance analysis on the regions of interest in the label image data and calculating the contribution of each region of interest based on the importance analysis result includes: S2331, extracting several regions of interest from the label image data, converting the regions of interest into feature vectors, and calculating the marginal contribution of each feature vector; S2332, setting a Shapley value for the marginal contribution of each eigenvector, multiplying the marginal contribution of the region of interest by the Shapley value, and obtaining a contribution value of each region of interest; S2333. Sort the contribution value of each area of ​​interest from large to small, and select contribution values ​​within a preset range from the sorted list and overlay them on the heat map for visual display.

7. The AIGC image detection method based on deep fusion network according to claim 6 is characterized in that: The expression of the scale-invariant feature of the label image data is: Wherein, z(a, b) represents the modulus of the scale-invariant feature point of the label image data; (a, b) represents the feature point in the label image data; l(a, b) represents the gradient of the feature point in the label image data.

8. The AIGC image detection method based on deep fusion network according to claim 1 is characterized in that: The artificial intelligence image detection model is established by combining the image feature representation after the weights are assigned, and the artificial intelligence image detection model is used to detect the artificial intelligence image from the label image data to obtain the artificial intelligence image detection result, including: S41, establishing an artificial intelligence image detection model based on the image feature representation, and formulating a priority detection order of the artificial intelligence image detection model according to the weights of the assigned image feature representations; S42, inputting the label image data into the artificial intelligence image detection model, preferentially detecting the label image data corresponding to the image feature representation with high weight according to the priority detection order, and outputting the detection probability value of the label image data; S43. Compare the detection probability value with a preset threshold. If the detection probability value is greater than the preset threshold, it indicates that the current label image data is an artificial intelligence image; if the detection probability value is less than the preset threshold, it indicates that the current label image data is a real image.

9. The AIGC image detection method based on deep fusion network according to claim 8 is characterized in that: The priority detection order of the artificial intelligence image detection model is formulated according to the weights represented by the assigned image features, including: S411, defining a priority detection order based on the weights represented by the image features, and assigning corresponding label image data to the artificial intelligence image detection model according to the priority detection order; S412. Count the detection volume of the artificial intelligence image detection model, and set a high priority queue and a low priority queue to balance the detection tasks of each queue of the artificial intelligence image detection model under the condition that the detection tasks are fault-tolerant and schedulable; Among them, the high priority queue is used to schedule real-time detection tasks, the low priority queue is used to schedule non-real-time detection tasks, and the high priority queue can extract detection tasks from the low priority queue.

10. An AIGC image detection system based on a deep fusion network, used to implement the AIGC image detection method based on a deep fusion network described in any one of claims 1 to 9, characterized in that: The AIGC image detection system includes an image optimization module, a visual feature extraction module, a feature fusion module and an image detection module; The image optimization module is used to extract initial image data from the central database and set classification labels for the initial image data to obtain labeled image data containing the classification labels; The visual feature extraction module is used to establish a feature extraction model based on a time-delay neural network, and input the label image data into the feature extraction model to obtain the visual features of the label image data; The feature fusion module is used to fuse the visual features with the predefined image element features based on the deep fusion network to form an image feature representation, and configure an adaptive weight adjustment strategy in the deep fusion network to assign weights to the image feature representation; The image detection module is used to establish an artificial intelligence image detection model in combination with the image feature representation after the weights are assigned, and use the artificial intelligence image detection model to detect artificial intelligence images from the labeled image data to obtain artificial intelligence image detection results.

Images