A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization
By employing a multi-feature fusion and progressive block optimization approach, a parallel multi-branch feature extraction network is constructed and progressively trained. This addresses the issues of insufficient feature representation and suboptimal training strategies in existing technologies, enabling efficient detection of AI-generated images.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI JIAOTONG UNIV
- Filing Date
- 2026-03-17
- Publication Date
- 2026-05-29
AI Technical Summary
Existing AI-generated image detection methods suffer from insufficient feature representation in complex scenarios, lack of optimized training strategies, weak generalization ability, and difficulty in effectively identifying forgery traces from diverse generative models.
A multi-feature fusion and progressive block optimization approach is adopted. High-level semantic, residual artifact and local texture features are extracted through a parallel multi-branch feature extraction network. A progressive block collaborative training strategy is used for feature fusion. Each branch is trained independently until convergence and then the optimal weights are loaded for fully connected fusion.
It significantly improves the robustness and generalization ability of AI-generated images, enhances the accuracy and discriminative power of feature extraction, and ensures the accuracy and reliability of classification.
Smart Images
Figure CN122116017A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image detection technology, and in particular to a lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization. Background Technology
[0002] With the rapid development of AI-generated content (AIGC) technologies such as Generative Adversarial Networks (GANs) and Diffusion Models, AI-generated images have reached a level of visual quality that is virtually indistinguishable from real images. While these technologies have shown great potential in fields such as artistic creation and the entertainment industry, they have also been maliciously used to create fake news, fabricate evidence, and produce deepfake content, posing a serious threat to the authenticity of information. Therefore, how to effectively detect and identify AI-generated forged images has become a critical problem that urgently needs to be solved in the fields of computer vision and information security.
[0003] Existing AI-generated image detection methods are mainly divided into single-feature-based methods and end-to-end deep learning-based methods. The former typically relies only on appearance features in the RGB space or only utilizes residual features in the frequency and noise domains. Although simple to implement, it is difficult to cope with diverse generative models. The latter usually uses large convolutional neural networks or Transformers for end-to-end training. Although the accuracy is high, it suffers from high computational complexity, large model size, and overfitting to training data. For example, patent application number CN202210229571.7, entitled "A Deep Forgery Detection Method Based on Frequency Domain Filtering Residuals," utilizes frequency domain filtering residual features, while patent application number CN202410345964.3, entitled "A Generalizable Deep Forgery Image Detection Method and System Based on Noise Perception," introduces abnormal noise distribution clues. While these technologies have made some breakthroughs in spatial high-frequency or noise statistics, they are generally limited to surface statistics in a single domain (space or frequency), and are insufficient in mining deep forgery traces such as cross-domain collaboration and semantic consistency, resulting in a sharp drop in robustness when facing cross-generative architectures, strongly compressed or mixed tampered samples.
[0004] Therefore, existing technologies have not yet formed a solution with sufficient feature representation, optimized training strategy and strong generalization ability when dealing with AI-generated image detection in complex scenarios. There is an urgent need to propose a new detection method that can combine multi-dimensional features and adopt progressive optimization training. Summary of the Invention
[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization, comprising: S1: Preprocess the input image to generate the original RGB image, SRM artifact feature map, and LBP-HOG texture feature map; S2: Construct a parallel multi-branch feature extraction network including an RGB semantic branch, an SRM artifact branch, and an LBP-HOG texture branch, and extract high-level semantic features, residual artifact features, and local texture features through the parallel multi-branch feature extraction network; S3: The parallel multi-branch feature extraction network is trained using a progressive block-based collaborative training strategy. In the first stage, each branch feature extractor is trained independently until convergence. In the second stage, each branch feature extractor is loaded with the optimal weights based on the validation set, and the fully connected fusion layer is trained. S4: Input the high-level semantic features, residual artifact features, and local texture features output from each branch into the fully connected fusion layer for weighted combination to obtain combined features. The combined features are then used to calculate the image forgery probability using the Sigmoid activation function to achieve real and fake image classification.
[0006] Preferably, in step S1, the input image is preprocessed to generate an RGB original image, an SRM artifact feature map, and an LBP-HOG texture feature map, including: S11: Normalize the size of the input image to obtain the original RGB image; S12: Convert the original RGB image to a grayscale image, and perform a convolution operation between the grayscale image and a pre-defined fixed-weight SRM convolution kernel to generate the SRM artifact feature map, as shown below: ,in For the first One SRM convolution kernel, The grayscale image; S13: Perform local binary mode transformation and histogram of oriented gradients transformation on the grayscale image to obtain LBP feature map and HOG feature map respectively. Concatenate the LBP feature map and the HOG feature map in the channel dimension to generate the dual-channel LBP-HOG texture feature map.
[0007] Preferably, in step S13, generating the dual-channel LBP-HOG texture feature map includes: S131: By comparing the grayscale values of the center pixel and its 8-neighbor pixels in the grayscale image, an 8-bit binary code is generated as an LBP feature, as shown below: in, For the grayscale image, For indicator functions, 8-neighbor offset set ; S132: Calculate the gradient values of the grayscale image in the x and y directions using the Sobel operator. and gradient value Combined with the gradient value and the gradient value The gradient magnitude and gradient direction of each pixel are calculated, and the gradient direction is uniformly quantized into 9 angular intervals to obtain the HOG feature map, as shown below: The formula for calculating the gradient magnitude M is: The gradient direction The calculation formula is: ; S133: Concatenate the LBP feature map and the HOG feature map in the channel dimension to form a dual-channel initial texture feature map, and perform normalization and standardization processing on the initial texture feature map to obtain the LBP-HOG texture feature map.
[0008] Preferably, in step S2, constructing the parallel multi-branch feature extraction network includes: The RGB semantic branch uses a pre-trained EfficientNet-B0 as the backbone network, removes the original classifier module of the EfficientNet-B0, inputs the three-channel RGB original image into the backbone network, and outputs the high-level semantic features after feature extraction. The SRM artifact branch uses a pre-trained EfficientNet-B0 as the backbone network. The number of input channels of the first convolutional layer of the backbone network is modified to the preset number of SRM channels to adapt to the SRM artifact feature map. The classifier module is removed, and the SRM artifact feature map is preprocessed to obtain a preprocessed SRM artifact feature map. The preprocessed SRM artifact feature map is input into the backbone network, and the residual artifact features are output. The LBP-HOG texture branch uses a pre-trained EfficientNet-B0 as the backbone network. The number of input channels of the first convolutional layer of the backbone is modified to the preset number of texture channels to adapt to the LBP-HOG texture feature map. The classifier module is removed, and the LBP-HOG texture feature map is input into the backbone network to output the local texture features.
[0009] Preferably, the SRM artifact feature map is preprocessed to obtain a preprocessed SRM artifact feature map, including: Min-Max normalization is performed on each channel of the SRM artifact feature map to obtain a normalized SRM artifact feature map. The pixel values in each channel are mapped to the [0, 1] interval. The normalization calculation formula is as follows: in, The original pixel value within the channel. This is the minimum pixel value for that channel. This represents the maximum pixel value for that channel. These are the normalized pixel values; The normalized SRM artifact feature map is standardized using single-channel statistics from the ImageNet dataset for each channel. The standardization formula is as follows: ,in, These are the standardized pixel values.
[0010] Preferably, in step S3, the parallel multi-branch feature extraction network is trained using a progressive block-based collaborative training strategy, including: In the first stage, the network parameters of the RGB semantic branch, the SRM artifact branch, and the LBP-HOG texture branch are initialized and optimized independently. Each branch is trained using the binary cross-entropy loss function and the Adam optimizer. A learning rate scheduler and an early stopping mechanism are introduced. When the validation loss does not decrease for several consecutive rounds, training is stopped and the optimal model of each branch is saved. In the second stage, the optimal model parameters of the branch are loaded and frozen. The training set and validation set samples are forward propagated through the three branches to obtain three training feature vectors. The training feature vectors are then concatenated to form a concatenated feature vector. Only one fully connected fusion layer with an input dimension of 3 and an output dimension of 1 is trained. The weight parameters of the fully connected fusion layer are optimized until the validation loss converges.
[0011] Preferably, each branch is trained using a binary cross-entropy loss function, including: in For the Sigmoid function, For real labels, This is the branch prediction value.
[0012] Preferably, the weighted fusion calculation method of the fully connected fusion layer is as follows: ,in, To concatenate feature vectors, For the weights of the fusion layer, For bias terms, To ultimately falsify the score.
[0013] Preferably, the method further includes full model inference during the testing phase, including: Load the parameters of the branch-optimal model and the fully connected fusion layer; For the test sample, features are extracted in parallel through three branches. The high-level semantic features, the residual artifact features and the local texture features are then concatenated and input into the fully connected fusion layer. The output of the fully connected fusion layer is calculated by the Sigmoid function to obtain the final forgery probability. If the probability is greater than the probability threshold, it is determined to be an AI-generated image; otherwise, it is a real image.
[0014] Preferably, the output of the fully connected fusion layer is used to calculate the final forgery probability using the Sigmoid function, including: in, For test samples The probability of forgery, For the Sigmoid function, For fully connected fusion layer functions, , , These are feature extraction functions for the three branches.
[0015] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention generates RGB original image, SRM artifact feature map and LBP-HOG texture feature map by preprocessing the input image, and constructs a parallel multi-branch feature extraction network containing RGB semantic branch, SRM artifact branch and LBP-HOG texture branch, so as to achieve comprehensive representation of the image from three different dimensions: semantic content, forgery residue traces and local texture details. This solves the problem that single modality features are difficult to cope with diverse generation models, thus significantly improving the detection robustness and generalization ability of AI generated images.
[0016] (2) This invention extracts high-level semantic features through the RGB semantic branch, extracts residual artifact features through the SRM artifact branch, and extracts local texture features through the LBP-HOG texture branch, thereby achieving parallel extraction and decoupling of multi-dimensional features. This allows each branch to focus on learning the forgery traces of a specific modality, avoiding mutual interference between different features, thus effectively improving the accuracy and discriminative power of feature extraction from each branch, and providing high-quality input features for subsequent fusion.
[0017] (3) This invention adopts a progressive block-based collaborative training strategy. In the first stage, each branch feature extractor is trained independently until convergence. In the second stage, the optimal weights of each branch are loaded and the adaptive fusion layer is trained. This decouples the optimization process of feature learning and feature fusion, avoiding gradient dominance and optimization conflict problems that may occur in end-to-end joint training. This improves the model convergence stability and enables the adaptive fusion layer to learn the optimal combination weights of each branch feature, further improving the accuracy and reliability of the final classification. Attached Figure Description
[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.
[0019] Figure 1 This is a flowchart of a lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to the present invention. Figure 2 This is a schematic diagram of the parallel multi-branch network architecture of the present invention; Figure 3 This is a flowchart illustrating the progressive block training strategy of the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. Obviously, the described embodiments are only some, not all, of the embodiments described in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application.
[0021] Those skilled in the art will understand that, unless otherwise stated, the singular forms “a” and “an” used herein, and “the”, may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.
[0022] First Embodiment Please see Figure 1 As shown, this embodiment provides a lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization, including the following steps: S1: Preprocess the input image to generate the original RGB image, SRM artifact feature map, and LBP-HOG texture feature map.
[0023] Preferably, in step S1, the input image is preprocessed to generate an RGB original image, an SRM artifact feature map, and an LBP-HOG texture feature map, including: S11: Normalize the size of the input image to obtain the original RGB image. Specifically, in this embodiment, the input image is received. Resize to using bilinear interpolation ; S12: Convert the original RGB image to a grayscale image, and perform a convolution operation between the grayscale image and a pre-defined fixed-weight SRM convolution kernel to generate an SRM artifact feature map, as shown below: ,in For the first One SRM convolution kernel, To convert the image to grayscale, specifically in this embodiment, 30 fixed [values / values] are used. SRM convolutional kernels extract artifact features. The weights of these kernels are based on the definitions given by Jessica Fridrich and Jan Kodovsky in their paper "Rich Models for Steganalysis of Digital Images". The convolutions produce a 30-channel response map, which is then normalized using Min-Max. The range is then standardized using ImageNet single-channel statistics (mean=0.485, std=0.229); S13: Perform local binary mode transformation and histogram of oriented gradients transformation on the grayscale image to obtain LBP feature map and HOG feature map respectively. Concatenate the LBP feature map and HOG feature map in the channel dimension to generate a dual-channel LBP-HOG texture feature map.
[0024] Preferably, in step S13, generating a dual-channel LBP-HOG texture feature map includes: S131: By comparing the grayscale values of the center pixel and its 8-neighbor pixels in the grayscale image, an 8-bit binary code is generated as an LBP feature, as shown below: in, To convert the image to grayscale, For indicator functions, 8-neighbor offset set ; S132: Calculate the gradient values of the grayscale image in the x and y directions using the Sobel operator. and gradient value Combined with gradient values and gradient value The gradient magnitude and gradient direction of each pixel are calculated, and the gradient direction is uniformly quantized into 9 angular intervals to obtain the HOG feature map, as shown below: The formula for calculating the gradient magnitude M is: gradient direction The calculation formula is: Specifically, in this embodiment, LBP generates an 8-bit binary code by comparing the center pixel with its 8 neighboring pixels. HOG calculates the gradient direction and magnitude using the Sobel operator and quantizes it into 9 directional intervals. Finally, the LBP and HOG images are concatenated into a 2-channel texture feature map, which is then normalized using Min-Max. The range is then standardized using ImageNet single-channel statistics (mean=0.485, std=0.229); S133: Concatenate the LBP feature map and the HOG feature map in the channel dimension to form a dual-channel initial texture feature map. Perform normalization and standardization processing on the initial texture feature map to obtain the LBP-HOG texture feature map.
[0025] S2: Construct a parallel multi-branch feature extraction network including an RGB semantic branch, an SRM artifact branch, and an LBP-HOG texture branch. Extract high-level semantic features, residual artifact features, and local texture features through the parallel multi-branch feature extraction network.
[0026] Please see Figure 2 As shown, in step S2, a parallel multi-branch feature extraction network is constructed, including: The RGB semantic branch uses a pre-trained EfficientNet-B0 as the backbone network. The original classifier module of EfficientNet-B0 is removed, and the three-channel RGB raw image is input into the backbone network. After feature extraction, high-level semantic features are output. Specifically, in this embodiment, the dropout layer of the classifier module is removed, and a 3-channel input is used to output 1000-dimensional semantic features. ; The SRM artifact branch uses a pre-trained EfficientNet-B0 as the backbone network. The number of input channels in the first convolutional layer of the backbone network is modified to the preset number of SRM channels to adapt to the SRM artifact feature maps. The classifier module is removed, and the SRM artifact feature maps are preprocessed to obtain preprocessed SRM artifact feature maps. These preprocessed SRM artifact feature maps are then input into the backbone network, outputting residual artifact features. Specifically, in this embodiment, the input channels of the first convolutional layer are changed to 30 (using the single-channel parameters of the original first-layer weights for weight initialization), the dropout layer of the classifier module is removed, and 1000-dimensional artifact features are output. ; The LBP-HOG texture branch uses a pre-trained EfficientNet-B0 as the backbone network. The number of input channels in the first convolutional layer of the backbone is modified to a preset number of texture channels to adapt to the LBP-HOG texture feature map. The classifier module is removed, and the LBP-HOG texture feature map is input into the backbone network to output local texture features. Specifically, in this embodiment, the input channels of the first convolutional layer are changed to 2 (the single-channel parameters of the original first-layer weights are copied to initialize the weights), the dropout layer of the classifier module is removed, and 1000-dimensional texture features are output. Each branch ends with a linear layer with an output dimension of 1, which serves as the training input for the adaptive fusion layer.
[0027] Preferably, the SRM artifact feature map is preprocessed to obtain a preprocessed SRM artifact feature map, including: Min-Max normalization is performed on each channel of the SRM artifact feature map to obtain a normalized SRM artifact feature map. The pixel values in each channel are mapped to the [0, 1] interval. The normalization calculation formula is as follows: in, The original pixel value within the channel. This is the minimum pixel value for that channel. This represents the maximum pixel value for that channel. These are the normalized pixel values; For the normalized SRM artifact feature map, standardization was performed on each channel using single-channel statistics from the ImageNet dataset. The standardization calculation formula is as follows: ,in, These are the standardized pixel values.
[0028] S3: A progressive block-based collaborative training strategy is adopted to train the parallel multi-branch feature extraction network. In the first stage, each branch feature extractor is trained independently until convergence. In the second stage, each branch feature extractor is loaded with the optimal weights based on the validation set, and the fully connected fusion layer is trained.
[0029] Please see Figure 3 As shown, in step S3, a progressive block-based collaborative training strategy is used to train the parallel multi-branch feature extraction network, including: In the first stage, the network parameters of the RGB semantic branch, SRM artifact branch, and LBP-HOG texture branch are initialized and optimized independently. Each branch is trained using a binary cross-entropy loss function and the Adam optimizer, with a learning rate scheduler and an early stopping mechanism introduced. Training stops and the optimal model of each branch is saved when the validation loss does not decrease for several consecutive rounds. Specifically, in this embodiment, the Adam optimizer is used with a learning rate of 0.0001. A learning rate scheduler is introduced, and when the validation loss does not decrease for two consecutive rounds, the learning rate is multiplied by 0.5. The early stopping patience value is set to 5 rounds. If the validation loss does not improve for 5 consecutive rounds, the current optimal model is saved and training of that branch is stopped. Each branch is optimized independently to avoid a dominant branch (such as the RGB branch) from dominating the gradient in the early training, causing underfitting of other branches. At the same time, independent training allows each branch to fully learn the optimal representation of its specific modality.
[0030] Preferably, each branch is trained using a binary cross-entropy loss function, including: in For the Sigmoid function, For real labels, For the branch prediction value, specifically, in this embodiment, for each branch Define the binary cross-entropy loss function:
[0031] In the second stage, the optimal model parameters of the branch are loaded and frozen. The training set and validation set samples are forward propagated through the three branches to obtain three training feature vectors. The training feature vectors are then concatenated to form a concatenated feature vector. Only one fully connected fusion layer with an input dimension of 3 and an output dimension of 1 is trained. The weight parameters of the fully connected fusion layer are optimized until the validation loss converges.
[0032] Preferably, the weighted fusion calculation method for the fully connected fusion layer is as follows: ,in, To concatenate feature vectors, For the weights of the fusion layer, For bias terms, To ultimately falsify the score.
[0033] Better yet, load the optimal checkpoint for each branch based on the validation set and freeze the parameters of all branch feature extractors; Perform forward propagation on the entire training set, and cache the three-branch output for each sample. Construct the feature matrix and label vector Similarly, construct validation set features. ; Initialize fusion layer parameters Zero initialization is used to ensure that the weights of each branch are similar; Training the fusion layer solely based on cached features still uses the binary cross-entropy loss function, but the optimization objective is now: A high learning rate of 0.01 and the Adam optimizer were used, along with ReduceLROnPlateau scheduling, until the loss was verified to converge.
[0034] The technical advantage of this block training strategy lies in decoupling the optimization process of feature learning and feature fusion, ensuring that the fusion layer learns weights in a fixed and optimal feature space, and avoiding optimization conflicts between the feature extractor and the fusion unit in end-to-end training.
[0035] S4: The high-level semantic features, residual artifact features, and local texture features output from each branch are input into a fully connected fusion layer for weighted combination to obtain combined features. The combined features are then used to calculate the image forgery probability using a Sigmoid activation function, thus achieving real vs. fake image classification. Specifically, in this embodiment, on the test sample... Features are extracted in parallel through three branches, weighted and summed through a fusion layer, and the forgery probability is obtained through the Sigmoid function. ,like If the image is not generated by AI, it is considered an AI-generated image; otherwise, it is considered a real image.
[0036] Preferably, the method further includes full model inference during the testing phase, including: Load the parameters of the branch-optimal model and the fully connected fusion layer; For the test samples, features are extracted in parallel through three branches. The high-level semantic features, residual artifact features and local texture features are then concatenated and input into the fully connected fusion layer. The output of the fully connected fusion layer is used to calculate the final forgery probability using the Sigmoid function. If the probability is greater than the probability threshold, it is determined to be an AI-generated image; otherwise, it is a real image.
[0037] Preferably, the output of the fully connected fusion layer is used to calculate the final forgery probability using the Sigmoid function, including: in, For test samples The probability of forgery, For the Sigmoid function, For fully connected fusion layer functions, , , These are feature extraction functions for the three branches.
[0038] This embodiment achieves a significant improvement in detection accuracy compared to a single RGB branch baseline on the dataset by complementing multi-dimensional features of semantics, artifacts, and texture, combined with a phased optimization training mechanism, and also demonstrates stronger generalization ability for unseen generative models.
[0039] Second Embodiment Based on the same concept, this embodiment also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, which, when executed by the processor, cause the processor to perform the steps of a precise measurement method for automated scanning of building components as described in the embodiment.
[0040] Based on the same concept, the present invention also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of a precise measurement method for automated scanning of building components as described in any one embodiment.
[0041] It is understood that, for the aforementioned lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization, if all components are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer server or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.
[0042] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.
[0043] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization, characterized in that, Includes the following steps: S1: Preprocess the input image to generate the original RGB image, SRM artifact feature map, and LBP-HOG texture feature map; S2: Construct a parallel multi-branch feature extraction network including an RGB semantic branch, an SRM artifact branch, and an LBP-HOG texture branch, and extract high-level semantic features, residual artifact features, and local texture features through the parallel multi-branch feature extraction network; S3: The parallel multi-branch feature extraction network is trained using a progressive block-based collaborative training strategy. In the first stage, each branch feature extractor is trained independently until convergence. In the second stage, the feature extractors of each branch load the optimal weights based on the validation set and train the fully connected fusion layer. S4: Input the high-level semantic features, residual artifact features, and local texture features output from each branch into the fully connected fusion layer for weighted combination to obtain combined features. The combined features are then used to calculate the image forgery probability using the Sigmoid activation function to achieve real and fake image classification.
2. The lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 1, characterized in that, In step S1, the input image is preprocessed to generate the original RGB image, SRM artifact feature map, and LBP-HOG texture feature map, including: S11: Normalize the size of the input image to obtain the original RGB image; S12: Convert the original RGB image to a grayscale image, and perform a convolution operation between the grayscale image and a pre-defined fixed-weight SRM convolution kernel to generate the SRM artifact feature map, as shown below: ， in For the first One SRM convolution kernel, The grayscale image; S13: Perform local binary mode transformation and histogram of oriented gradients transformation on the grayscale image to obtain LBP feature map and HOG feature map respectively. Concatenate the LBP feature map and the HOG feature map in the channel dimension to generate the dual-channel LBP-HOG texture feature map.
3. The lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 2, characterized in that, In step S13, the dual-channel LBP-HOG texture feature map is generated, including: S131: By comparing the grayscale values of the center pixel and its 8-neighbor pixels in the grayscale image, an 8-bit binary code is generated as an LBP feature, as shown below: in, For the grayscale image, For indicator functions, 8-neighbor offset set ; S132: Calculate the gradient values of the grayscale image in the x and y directions using the Sobel operator. and gradient value Combined with the gradient value and the gradient value The gradient magnitude and gradient direction of each pixel are calculated, and the gradient direction is uniformly quantized into 9 angular intervals to obtain the HOG feature map, as shown below: The formula for calculating the gradient magnitude M is: The gradient direction The calculation formula is: ； S133: Concatenate the LBP feature map and the HOG feature map in the channel dimension to form a dual-channel initial texture feature map, and perform normalization and standardization processing on the initial texture feature map to obtain the LBP-HOG texture feature map.
4. The lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 1, characterized in that, In step S2, the parallel multi-branch feature extraction network is constructed, including: The RGB semantic branch uses a pre-trained EfficientNet-B0 as the backbone network, removes the original classifier module of the EfficientNet-B0, inputs the three-channel RGB original image into the backbone network, and outputs the high-level semantic features after feature extraction. The SRM artifact branch uses a pre-trained EfficientNet-B0 as the backbone network. The number of input channels of the first convolutional layer of the backbone network is modified to the preset number of SRM channels to adapt to the SRM artifact feature map. The classifier module is removed, and the SRM artifact feature map is preprocessed to obtain a preprocessed SRM artifact feature map. The preprocessed SRM artifact feature map is input into the backbone network, and the residual artifact features are output. The LBP-HOG texture branch uses a pre-trained EfficientNet-B0 as the backbone network. The number of input channels of the first convolutional layer of the backbone is modified to the preset number of texture channels to adapt to the LBP-HOG texture feature map. The classifier module is removed, and the LBP-HOG texture feature map is input into the backbone network to output the local texture features.
5. A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 4, characterized in that, The SRM artifact feature map is preprocessed to obtain a preprocessed SRM artifact feature map, including: Min-Max normalization is performed on each channel of the SRM artifact feature map to obtain a normalized SRM artifact feature map. The pixel values in each channel are mapped to the [0, 1] interval. The normalization calculation formula is as follows: in, The original pixel value within the channel. This is the minimum pixel value for that channel. This represents the maximum pixel value for that channel. These are the normalized pixel values; The normalized SRM artifact feature map is standardized using single-channel statistics from the ImageNet dataset for each channel. The standardization formula is as follows: ， in, These are the standardized pixel values.
6. A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 1, characterized in that, In step S3, the parallel multi-branch feature extraction network is trained using a progressive block-based collaborative training strategy, including: In the first stage, the network parameters of the RGB semantic branch, the SRM artifact branch, and the LBP-HOG texture branch are initialized and optimized independently. Each branch is trained using the binary cross-entropy loss function and the Adam optimizer. A learning rate scheduler and an early stopping mechanism are introduced. When the validation loss does not decrease for several consecutive rounds, training is stopped and the optimal model of each branch is saved. In the second stage, the optimal model parameters of the branch are loaded and frozen. The training set and validation set samples are forward propagated through the three branches to obtain three training feature vectors. The training feature vectors are then concatenated to form a concatenated feature vector. Only one fully connected fusion layer with an input dimension of 3 and an output dimension of 1 is trained. The weight parameters of the fully connected fusion layer are optimized until the validation loss converges.
7. A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 6, characterized in that, Each branch is trained using the binary cross-entropy loss function, including: in For the Sigmoid function, For real labels, This is the branch prediction value.
8. A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 6, characterized in that, The weighted fusion calculation method for the fully connected fusion layer is as follows: ， in, To concatenate feature vectors, For the weights of the fusion layer, For bias terms, To ultimately falsify the score.
9. A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 1, characterized in that, It also includes full model inference during the testing phase, including: Load the parameters of the branch-optimal model and the fully connected fusion layer; For the test sample, features are extracted in parallel through three branches. The high-level semantic features, the residual artifact features and the local texture features are then concatenated and input into the fully connected fusion layer. The output of the fully connected fusion layer is calculated by the Sigmoid function to obtain the final forgery probability. If the probability is greater than the probability threshold, it is determined to be an AI-generated image; otherwise, it is a real image.
10. A lightweight AI-generated image detection method based on multi-feature fusion and progressive block optimization according to claim 1, characterized in that, The output of the fully connected fusion layer is used to calculate the final forgery probability using the Sigmoid function, including: in, For test samples The probability of forgery, For the Sigmoid function, For fully connected fusion layer functions, , , These are feature extraction functions for the three branches.