A generative image detection method without neural network assistance

Abstract

The present invention provides a neural network-free generated image detection method, which relates to the field of image analysis technology. The method measures the correlation between high-frequency signals and texture complexity in an image and the sensitivity of the image to standard uniformly distributed noise perturbations by establishing a neural network-free training indicator. The method comprises the following steps: constructing a neural network-free training indicator based on high-frequency-texture correlation and a neural network-free training indicator based on the image's sensitivity to standard uniformly distributed noise perturbations; weightedly fusing the two indicators to obtain a fused indicator; setting a generated image discrimination threshold and a generated image discrimination rule; performing a weighted search optimization on the fused indicator using Bayesian optimization; and obtaining a weighted fused generated image detection indicator after reaching the maximum number of Bayesian optimization rounds. The method then inputs the image to be detected into the weighted fused generated image detection indicator to obtain a detection result of whether the image to be detected is a generated image.

Description

Technical Field

[0001] The present invention relates to the technical field of image analysis, and in particular to a generated image detection method without the assistance of a neural network. Background Art

[0002] In today's digital age, rapid advances in artificial intelligence (AI) have revolutionized numerous fields, most notably image generation. Thanks to innovations in deep learning methods like generative adversarial networks (GANs) and diffusion models, image generation has made the leap from simple imitation to highly realistic images. These technologies can now create incredibly convincing images of human faces, so realistic that they are virtually indistinguishable to the naked eye.

[0003] However, while this widespread use of technology brings convenience, it also quietly breeds a series of serious security risks. Faced with this challenge, research institutions and companies around the world have invested heavily in developing more advanced generation technologies. However, despite the rapid advancements in generation technology, the corresponding generation face image recognition technology lags behind, and a reliable and effective solution has yet to be established.

[0004] Existing generative face image detection technologies all use deep learning techniques, based on manually designed artificial neural networks. This approach is not only time-consuming and labor-intensive, but also requires designers to have extensive experience in deep learning model design and a deep understanding of generative images, making the design of generative face image detection extremely difficult. Designing a high-performance generative face image detection model requires a significant amount of training and debugging time and computing power, making the widespread application of deep learning technology in generative face image detection difficult. Second, existing detection methods based on artificial neural networks are computationally intensive. Even with powerful hardware computing power, it still takes several seconds to detect a single image in actual application scenarios. Furthermore, these neural networks occupy a significant amount of storage space. These shortcomings make it difficult for these detection methods based on artificial neural networks to be widely used for generative image detection on low-computing devices or edge devices.

[0005] In recent years, with the continuous development of the field of generative image identification, training-free generative image identification techniques have gradually demonstrated promising performance, without the need for extensive manual network design and training and tuning. However, these methods also require the assistance of neural networks when detecting generated images in applications, resulting in high memory and time overhead in these applications. Furthermore, they are generally difficult to generalize to a wide variety of advanced generative model-generated image detection. This poses a significant challenge to the application of training-free generative image detection techniques in real-world production scenarios. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a generated image detection method without the assistance of a neural network. By establishing a training-free index that does not require the assistance of a neural network, the correlation between high-frequency signals and texture complexity in the image is measured, and the sensitivity of the image to standard uniformly distributed noise disturbance is measured. The scores of the training-free index that does not require the assistance of a neural network are then fused by weighted fusion to obtain a fusion index and a weight vector. The weight vector is searched and optimized using Bayesian optimization technology to achieve identification of the generated image. The method comprises the following steps:

[0007] Step 1: Obtain labeled images and construct a generated image detection dataset, preprocess the labeled images in the generated image detection dataset to obtain images to be detected; the labels include information indicating whether the images are generated images or real images;

[0008] The specific method of the pretreatment is:

[0009] According to the generated image detection data set, all images in the channel The expectation and variance of the pixel values on the generated image detection dataset are used to generate the image in the channel The upper position is The pixel value of the pixel Perform mean variance normalization to obtain standardized pixel values ;

[0010] Step 2: Construct a non-neural network-assisted training-free indicator based on high-frequency-texture correlation, calculate the texture complexity vector and the high-frequency signal fluctuation vector of the image to be detected, and calculate the correlation coefficient between the texture complexity vector and the high-frequency signal fluctuation vector of the image to be detected as the indicator score of the non-neural network-assisted training-free indicator based on high-frequency-texture correlation;

[0011] Step 2.1: The image to be detected Split into The size is Image blocks , get the image block vector ,in, is the number of channels of the image block, is the width and height of the image block;

[0012] Step 2.2: Calculate the image patch vector The texture complexity of each image block in the image to be detected is obtained The texture complexity vector ;

[0013] Image patch vector Middle Image blocks Texture complexity As shown in the following formula:

[0014] ;

[0015] in, is the texture complexity function, For image blocks No. The image on each channel, For image blocks The row index of the pixel in , For image blocks The column index of the pixel in , For image blocks No. The coordinates on the channel are The pixel value of the pixel point, For image blocks No. The coordinates on the channel are The pixel value of the pixel point, For image blocks No. The coordinates on the channel are The pixel value of the pixel point;

[0016] Will Image blocks in The texture complexity of each channel is accumulated to obtain the image to be detected The texture complexity vector ;

[0017] Step 2.3: Calculate the image patch vector The standard deviation of each image block in the image to be detected is obtained The high-frequency signal fluctuation vector ;

[0018] Calculate the image block vector separately The standard deviation of each image block in the image to be detected is obtained The high-frequency signal fluctuation vector , as shown in the following formula:

[0019] ;

[0020] in, To calculate the variance operation;

[0021] Step 2.4: Calculate the image to be detected The texture complexity vector and high-frequency signal volatility vector The Pearson correlation coefficient between them is used to obtain the index score of the non-neural network assisted training-free index based on the high-frequency-texture correlation of the image block. ;

[0022] Step 3: Construct a non-neural network-assisted, training-free indicator based on the image's sensitivity to standard uniformly distributed noise perturbations. Use standard uniformly distributed noise to perturb the image to be tested. Calculate the cosine similarity between the image to be tested and the image to be tested after being perturbed by standard uniformly distributed noise as the indicator score of the non-neural network-assisted, training-free indicator based on the image's sensitivity to standard uniformly distributed noise perturbations.

[0023] Step 3.1: The image to be detected Flattened to a one-dimensional vector , initialize a vector with one dimension Standard uniformly distributed noise of the same shape ;

[0024] Step 3.2: Use standard uniform noise Perturb the image to be detected , the perturbation is from a one-dimensional vector Subtract standard uniformly distributed noise from the pixel value Calculate the value of the one-dimensional vector and the standard uniformly distributed noise The cosine similarity of the perturbed image to be detected is used to obtain the index score of the non-neural network-assisted training-free index based on the image's sensitivity to standard uniformly distributed noise perturbations. , as shown in the following formula:

[0025] ;

[0026] in, is the cosine similarity operation, is a one-dimensional vector The transpose of is the 2-norm operation;

[0027] Step 4: Perform weighted sum fusion on the index scores of the non-neural network-assisted training-free index based on high-frequency-texture correlation and the non-neural network-assisted training-free index based on the image sensitivity to standard uniformly distributed noise disturbance to obtain a fusion index. Set the generated image discrimination threshold and generated image discrimination rule based on the fusion index, use Bayesian optimization to perform weight search optimization on the fusion index, and add the weight vector and objective function score of each iteration round of the Bayesian optimization process to the search record set;

[0028] Step 4.1: Scoring the metric for the non-neural network-assisted training-free metric based on high-frequency-texture correlation and the index score of the non-neural network-assisted training-free index based on the sensitivity of the image to standard uniformly distributed noise perturbations Perform weighted sum fusion to obtain the fusion index and its weight vector , as shown in the following formula:

[0029] ;

[0030] ;

[0031] in, and are all weight parameters;

[0032] Step 4.2: Based on fusion indicators Set the generated image discrimination threshold and generate image discrimination rules , as shown in the following formula:

[0033] ;

[0034] in, is a binary indicator function, that is, when the weighted fusion index Less than or equal to the generated image discrimination threshold hour, , determine that the image to be detected is a generated image, otherwise , determine that the image to be detected is a real image;

[0035] Step 4.3: Use Bayesian optimization to search and optimize the weight vector of the fusion indicator on the generated image detection dataset. Create a search record set to store the weight vector and objective function score of each iteration of the Bayesian optimization process.

[0036] Based on the labels of the images in the generated image detection dataset and the image prediction categories obtained by the fusion indicator detection in the current round of search optimization, the number of true positives TP, the number of false positives FP, and the number of false negatives FN are counted, and the average precision AP of the detection is calculated, where the number of true positives TP refers to the number of generated images predicted as generated images, the number of false positives FP refers to the number of real images predicted as generated images, and the number of false negatives FN refers to the number of generated images predicted as real images;

[0037] The average precision AP is used as the objective function in the Bayesian optimization search, and the weight vector is searched by Bayesian Perform iterative search and record the current generation in each iteration round The weight vector and the objective function score , store it in the search record collection ;

[0038] Step 5: After completing the Bayesian optimization, obtain the weighted fusion generated image detection index, input the image to be detected into the weighted fusion generated image detection index, and obtain the detection result of whether it is a generated image;

[0039] Step 5.1: After completing Bayesian optimization, select the search record set The weight vector with the largest objective function score , the weight vector with the largest objective function score Applied to the fusion index, the weighted fusion generated image detection index is obtained;

[0040] After reaching the maximum number of Bayesian optimization rounds, the weight vector with the maximum objective function score is As shown in the following formula:

[0041] ;

[0042] in, is the maximum number of Bayesian optimization rounds, For the The weight vector of the iteration round, For the The objective function score of the iteration round, To find the maximum index function;

[0043] The maximum weight vector of the objective function score Applied to the weighted fusion index, the generated image detection index of weighted fusion is obtained;

[0044] Step 5.2: Generate image discrimination threshold according to the setting and generate image discrimination rules , the image to be detected The generated image detection index of weighted fusion is applied to calculate and obtain the detection result of whether the image to be detected is a generated image or a real image.

[0045] The beneficial effect of adopting the above technical solution is that: the present invention provides a generated image detection method without the assistance of a neural network. For a given generated image data set, by constructing two training-free detection indicators without the assistance of a neural network, and using the searched weight combination vector to weight and fuse each indicator, the method does not rely on the deployment and use of neural networks, and can achieve a high detection rate in the generated images of various generation models. At the same time, it only requires extremely low memory and time overhead, and can realize effective generated image identification on edge devices or low computing power. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1A flowchart of a method for generating image detection without neural network assistance provided by an embodiment of the present invention;

[0047] Figure 2 A schematic diagram of a neural network-free method for generating image detection provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] Below in conjunction with accompanying drawing and embodiment, specific embodiment of the present invention is described in further detail.The following examples are used to illustrate the present invention, but are not used to limit the scope of the invention.On the contrary, the purpose of providing these embodiments is to make the disclosure of the application be understood more thoroughly and comprehensively.

[0049] A method for generating image detection without the assistance of a neural network in this embodiment, such as Figure 1 As shown, the following steps are included:

[0050] Step 1: Obtain labeled images and construct a generated image detection dataset, preprocess the labeled images in the generated image detection dataset to obtain images to be detected; the labels include information indicating whether the images are generated images or real images;

[0051] The specific method of the pretreatment is:

[0052] According to the generated image detection data set, all images in the channel The expectation and variance of the pixel values on the generated image detection dataset are used to generate the image in the channel The upper position is The pixel value of the pixel Row mean variance normalization to obtain standardized pixel values , as shown in the following formula:

[0053] ;

[0054] in, To generate image detection dataset, all images in the channel The expected pixel value on To generate image detection dataset, all images in the channel The variance of the pixel values on , is the pixel position information;

[0055] Step 2: Construct a non-neural network-assisted training-free indicator based on high-frequency-texture correlation, calculate the texture complexity vector and the high-frequency signal fluctuation vector of the image to be detected, and calculate the correlation coefficient between the texture complexity vector and the high-frequency signal fluctuation vector of the image to be detected as the indicator score of the non-neural network-assisted training-free indicator based on high-frequency-texture correlation;

[0056] Step 2.1: The image to be detected Split into The size is Image blocks , get the image block vector , as shown in the following formula:

[0057] ;

[0058] in, is the number of channels of the image block, are the width and height of the image block, is the image block vector Middle image blocks, is the set of real numbers;

[0059] Step 2.2: Calculate the image patch vector The texture complexity of each image block in the image to be detected is obtained The texture complexity vector ;

[0060] Image patch vector Middle Image blocks Texture complexity As shown in the following formula:

[0061] ;

[0062] in, is the texture complexity function, For image blocks No. The image on each channel, For image blocks The row index of the pixel in , For image blocks The column index of the pixel in , For image blocks No. The coordinates on the channel are The pixel value of the pixel point, For image blocks No. The coordinates on the channel are The pixel value of the pixel point, For image blocks No. The coordinates on the channel are The pixel value of the pixel point;

[0063] Will Image blocks in The texture complexity of each channel is accumulated to obtain the image to be detected The texture complexity vector , as shown in the following formula:

[0064] ;

[0065] Step 2.3: Calculate the image patch vector The standard deviation of each image block in the image to be detected is obtained The high-frequency signal fluctuation vector ;

[0066] Calculate the image block vector separately The standard deviation of each image block in the image to be detected is obtained The high-frequency signal fluctuation vector , as shown in the following formula:

[0067] ;

[0068] in, To calculate the variance operation;

[0069] Step 2.4: Calculate the image to be detected The texture complexity vector and high-frequency signal volatility vector The Pearson correlation coefficient between them is used to obtain the index score of the non-neural network assisted training-free index based on the high-frequency-texture correlation of the image block. ;

[0070] Indicator score of non-neural network assisted training-free indicator based on high-frequency-texture correlation of image patches As shown in the following formula:

[0071] ;

[0072] in, To calculate the Pearson correlation coefficient operation, To calculate the mean operation;

[0073] Step 3: Construct a non-neural network-assisted, training-free indicator based on the image's sensitivity to standard uniformly distributed noise perturbations. Use standard uniformly distributed noise to perturb the image to be tested. Calculate the cosine similarity between the image to be tested and the image to be tested after being perturbed by standard uniformly distributed noise as the indicator score of the non-neural network-assisted, training-free indicator based on the image's sensitivity to standard uniformly distributed noise perturbations.

[0074] Step 3.1: The image to be detected Flattened to a one-dimensional vector , initialize a vector with one dimension Standard uniformly distributed noise of the same shape ;

[0075] Step 3.2: Use standard uniform noise Perturb the image to be detected , the perturbation is from a one-dimensional vector Subtract standard uniformly distributed noise from the pixel value Calculate the value of the one-dimensional vector and the standard uniformly distributed noise The cosine similarity of the perturbed image to be detected is used to obtain the index score of the non-neural network-assisted training-free index based on the image's sensitivity to standard uniformly distributed noise perturbations. , as shown in the following formula:

[0076] ;

[0077] in, is the cosine similarity operation, is a one-dimensional vector The transpose of is the 2-norm operation;

[0078] Step 4: Perform weighted sum fusion on the index scores of the non-neural network-assisted training-free index based on high-frequency-texture correlation and the non-neural network-assisted training-free index based on the image sensitivity to standard uniformly distributed noise disturbance to obtain a fusion index. Set the generated image discrimination threshold and generated image discrimination rule based on the fusion index, use Bayesian optimization to perform weight search optimization on the fusion index, and add the weight vector and objective function score of each iteration round of the Bayesian optimization process to the search record set;

[0079] Step 4.1: Scoring the metric for the non-neural network-assisted training-free metric based on high-frequency-texture correlation and the index score of the non-neural network-assisted training-free index based on the sensitivity of the image to standard uniformly distributed noise perturbations Perform weighted sum fusion to obtain the fusion index and its weight vector , as shown in the following formula:

[0080] ;

[0081] ;

[0082] in, and are all weight parameters;

[0083] Step 4.2: Based on fusion indicators Set the generated image discrimination threshold and generate image discrimination rules , as shown in the following formula:

[0084] ;

[0085] in, is a binary indicator function, that is, when the weighted fusion index Less than or equal to the generated image discrimination threshold hour, , determine that the image to be detected is a generated image, otherwise , determine that the image to be detected is a real image;

[0086] In this embodiment, the image discrimination threshold is set to generate ;

[0087] Step 4.3: Use Bayesian optimization to search and optimize the weight vector of the fusion indicator on the generated image detection dataset. Create a search record set to store the weight vector and objective function score of each iteration of the Bayesian optimization process.

[0088] Based on the labels of the images in the generated image detection dataset and the image prediction categories obtained by the fusion indicator detection in the current round of search optimization, the number of true positives TP, the number of false positives FP, and the number of false negatives FN are counted, and the average precision AP of the detection is calculated, where the number of true positives TP refers to the number of generated images predicted as generated images, the number of false positives FP refers to the number of real images predicted as generated images, and the number of false negatives FN refers to the number of generated images predicted as real images. The average precision AP is shown in the following formula:

[0089] ;

[0090] ;

[0091] in, is the accuracy, is the recall rate;

[0092] The average precision AP is used as the objective function in the Bayesian optimization search, and the weight vector is searched by Bayesian Perform iterative search and record the current iteration in each iteration round The weight vector and the objective function score , store it in the search record collection , search the record collection As shown in the following formula:

[0093] ;

[0094] in, is the maximum number of Bayesian optimization rounds, is the weight vector of the first iteration round, is the objective function score of the first iteration round, For the The weight vector of the iteration round, For the The objective function score of the iteration round;

[0095] Step 5: After completing the Bayesian optimization, obtain the weighted fusion generated image detection index, input the image to be detected into the weighted fusion generated image detection index, and obtain the detection result of whether it is a generated image;

[0096] Step 5.1: After completing Bayesian optimization, select the search record set The weight vector with the largest objective function score , the weight vector with the largest objective function score Applied to the fusion index, the weighted fusion generated image detection index is obtained;

[0097] After reaching the maximum number of Bayesian optimization rounds, the weight vector with the maximum objective function score is As shown in the following formula:

[0098] ;

[0099] in, To find the maximum index function;

[0100] The maximum weight vector of the objective function score Applied to the fusion index, the weighted fusion generated image detection index is obtained;

[0101] Step 5.2: Generate image discrimination threshold according to the setting and generate image discrimination rules , the image to be detected Apply the generated image detection index of weighted fusion to calculate and obtain the detection result of whether the image to be detected is a generated image or a real image, such as Figure 2 shown.

[0102] In order to further verify the detection performance and computational efficiency of this method, this implementation adopts UniversalFake Detection to generate image detection classification dataset to perform image classification detection simulation test, in which the memory and time overhead are performed on a single RTX 4090 GPU, and the weight vector The Bayesian optimization of the proposed method is performed on the data subset of the LDM200 cfg generation model of this dataset. The test results are shown in Table 1. The test results of the traditional method are shown in Table 2. The traditional method used in the test is a detection method assisted by a neural network.

[0103] Table 1. Image classification simulation test results of this method in the Universal Fake Detection dataset.

[0104]

[0105] Table 2: The Universal Fake Detection dataset uses traditional methods to generate simulated image classification test results.

[0106]

[0107] According to the test results in Tables 1 and 2, the method of the present invention has stronger performance than existing methods and requires only minimal memory usage and time overhead, allowing the method to be deployed on edge devices and low-computing power devices to achieve efficient and accurate generation of image detection and classification.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for generating image detection without the aid of a neural network, characterized in that: The following steps are involved: Step 1: Obtain labeled images and construct a generated image detection dataset, preprocess the labeled images in the generated image detection dataset to obtain images to be detected; the labels include information indicating whether the images are generated images or real images; Step 2: Construct a non-neural network-assisted training-free indicator based on high-frequency-texture correlation, calculate the texture complexity vector and the high-frequency signal fluctuation vector of the image to be detected, and calculate the correlation coefficient between the texture complexity vector and the high-frequency signal fluctuation vector of the image to be detected as the indicator score of the non-neural network-assisted training-free indicator based on high-frequency-texture correlation; Step 3: Construct a non-neural network-assisted, training-free indicator based on the image's sensitivity to standard uniformly distributed noise perturbations. Use standard uniformly distributed noise to perturb the image to be tested. Calculate the cosine similarity between the image to be tested and the image to be tested after being perturbed by standard uniformly distributed noise as the indicator score of the non-neural network-assisted, training-free indicator based on the image's sensitivity to standard uniformly distributed noise perturbations. Step 4: Perform weighted sum fusion on the index scores of the non-neural network-assisted training-free index based on high-frequency-texture correlation and the non-neural network-assisted training-free index based on the image sensitivity to standard uniformly distributed noise disturbance to obtain a fusion index. Set the generated image discrimination threshold and generated image discrimination rule based on the fusion index, use Bayesian optimization to perform weight search optimization on the fusion index, and add the weight vector and objective function score of each iteration round of the Bayesian optimization process to the search record set; Step 5: After completing the Bayesian optimization, obtain the weighted fusion generated image detection index, input the image to be detected into the weighted fusion generated image detection index, and obtain the detection result of whether it is a generated image.

2. The method for generating image detection without the aid of a neural network according to claim 1, characterized in that: The specific method of the pretreatment in step 1 is: According to the generated image detection data set, all images in the channel The expectation and variance of the pixel values on the generated image detection dataset are used to generate the image in the channel The upper position is The pixel value of the pixel Perform mean variance normalization to obtain standardized pixel values .

3. The method for generating image detection without the aid of a neural network according to claim 2, characterized in that: The step 2 includes: Step 2.1: The image to be detected Split into The size is Image blocks , get the image block vector ,in, is the number of channels of the image block, is the width and height of the image block; Step 2.2: Calculate the image patch vector The texture complexity of each image block in the image to be detected is obtained The texture complexity vector ; Step 2.3: Calculate the image patch vector The standard deviation of each image block in the image to be detected is obtained The high-frequency signal fluctuation vector ; Step 2.4: Calculate the image to be detected The texture complexity vector and high-frequency signal volatility vector The Pearson correlation coefficient between them is used to obtain the index score of the non-neural network assisted training-free index based on the high-frequency-texture correlation of the image block. .

4. The method for generating image detection without the aid of a neural network according to claim 3, characterized in that: The step 2.2 is specifically as follows: Image patch vector Middle Image blocks Texture complexity As shown in the following formula: ； in, is the texture complexity function, For image blocks No. The image on each channel, For image blocks The row index of the pixel in , For image blocks The column index of the pixel in , For image blocks No. The coordinates on the channel are The pixel value of the pixel point, For image blocks No. The coordinates on the channel are The pixel value of the pixel point, For image blocks No. The coordinates on the channel are The pixel value of the pixel point; Will Image blocks in The texture complexity of each channel is accumulated to obtain the image to be detected The texture complexity vector .

5. The method for generating image detection without the aid of a neural network according to claim 4, characterized in that: The step 2.3 is specifically as follows: Calculate the image block vector separately The standard deviation of each image block in the image to be detected is obtained The high-frequency signal fluctuation vector , as shown in the following formula: ； in, To calculate the variance operation.

6. The method for generating image detection without the aid of a neural network according to claim 5, characterized in that: The step 3 includes: Step 3.1: The image to be detected Flattened to a one-dimensional vector , initialize a vector with one dimension Standard uniformly distributed noise of the same shape ; Step 3.2: Use standard uniform noise Perturb the image to be detected , the perturbation is from a one-dimensional vector Subtract standard uniformly distributed noise from the pixel value Calculate the value of the one-dimensional vector and the standard uniformly distributed noise The cosine similarity of the perturbed image to be detected is used to obtain the index score of the non-neural network-assisted training-free index based on the image's sensitivity to standard uniformly distributed noise perturbations. , as shown in the following formula: ； in, is the cosine similarity operation, is a one-dimensional vector The transpose of is the 2-norm operation.

7. The method for generating image detection without the aid of a neural network according to claim 6, characterized in that: The step 4 comprises: Step 4.1: Scoring the metric for the non-neural network-assisted training-free metric based on high-frequency-texture correlation and the index score of the non-neural network-assisted training-free index based on the sensitivity of the image to standard uniformly distributed noise perturbations Perform weighted sum fusion to obtain the fusion index and its weight vector , as shown in the following formula: ； ； in, and are all weight parameters; Step 4.2: Based on fusion indicators Set the generated image discrimination threshold and generate image discrimination rules , as shown in the following formula: ； in, is a binary indicator function, that is, when the weighted fusion index Less than or equal to the generated image discrimination threshold hour, , determine that the image to be detected is a generated image, otherwise , determine that the image to be detected is a real image; Step 4.3: On the generated image detection dataset, use Bayesian optimization to search and optimize the weight vector of the fusion indicator, and create a search record set to store the weight vector and objective function score of each iteration round in the Bayesian optimization process.

8. The method for generating image detection without the aid of a neural network according to claim 7, characterized in that: The step 4.3 is specifically as follows: Based on the labels of the images in the generated image detection dataset and the image prediction categories obtained by the fusion indicator detection in the current round of search optimization, the number of true positives TP, the number of false positives FP, and the number of false negatives FN are counted, and the average precision AP of the detection is calculated, where the number of true positives TP refers to the number of generated images predicted as generated images, the number of false positives FP refers to the number of real images predicted as generated images, and the number of false negatives FN refers to the number of generated images predicted as real images; The average precision AP is used as the objective function in the Bayesian optimization search, and the weight vector is searched by Bayesian Perform iterative search and record the current generation in each iteration round The weight vector and the objective function score , store it in the search record collection .

9. The method for generating image detection without the aid of a neural network according to claim 8, characterized in that: The step 5 comprises: Step 5.1: After completing Bayesian optimization, select the search record set The weight vector with the largest objective function score , the weight vector with the largest objective function score Applied to the fusion index, the weighted fusion generated image detection index is obtained; Step 5.2: Generate image discrimination threshold according to the setting and generate image discrimination rules , the image to be detected The generated image detection index of weighted fusion is applied to calculate and obtain the detection result of whether the image to be detected is a generated image or a real image.

10. The method for generating image detection without the aid of a neural network according to claim 9, characterized in that: The step 5.1 is specifically as follows: After reaching the maximum number of Bayesian optimization rounds, the weight vector with the maximum objective function score is As shown in the following formula: ； in, is the maximum number of Bayesian optimization rounds, For the The weight vector of the iteration round, For the The objective function score of the iteration round, To find the maximum index function; The maximum weight vector of the objective function score Applied to the weighted fusion index, the generated image detection index of weighted fusion is obtained.

Images