Image tampering detection method based on cross feedback optimization network
By adopting a cross-feedback-based optimization network method in biomedical image tamper detection, multi-scale control tamper features are extracted and fused with residual features, the problem of insufficient detection accuracy in the prior art is solved, and higher detection accuracy and robustness are achieved.
Patent Information
- Application Number
- CN202411947745.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-27
- Publication Date
- 2025-04-25
AI Technical Summary
The prior art has the problem of insufficient detection accuracy in biomedical image tamper detection, especially in stitching type image tampering, which makes it difficult to effectively detect.
An image tamper detection method based on cross-feedback optimization network is adopted. By adjusting the brightness and contrast of the medical images to be detected, a control tamper feature extraction network is established, including spliced branches and non-split branches, multi-scale convolution is performed, and multi-scale control tamper features are extracted, and tampered area segmentation and classification prediction are carried out through the control interaction network and the multi-scale residual features.
It significantly improves the accuracy and robustness of medical image stitching tampering detection, reduces the impact of noise and interference on the detection results, enhances the discriminant ability of the model, and avoids false alarms caused by background feature interference.
Smart Images
Figure CN119887672A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of image tampering detection, and in particular to an image tampering detection method based on a cross feedback optimization network. Background Art
[0002] Image manipulation in biomedicine has become a widely-discussed issue. Unethical researchers maliciously manipulate images to falsify experimental results and cover up unfavorable experimental phenomena through technical means such as splicing. This behavior brings significant harm, including reducing the integrity and credibility of biomedical literature, causing economic losses, and misleading academic research in this field.
[0003] However, existing methods cannot directly detect spliced biomedical academic images. Among various image tampering behaviors, image splicing tampering is very challenging because the source of the spliced components is unknown and cannot be solved by comparison, and there are no regional artifacts caused by similar removal operations. When detecting the splicing of biomedical academic images, existing methods for natural images usually have the problem of insufficient detection accuracy.
[0004] For natural image tampering detection, existing methods are mainly divided into the following categories: Feature extraction methods based on deep learning, such as RRU, Transforensics, PSCC and HiFi, mainly improve the detection effect through attention mechanism, multi-scale feature fusion and other technologies. Detection methods based on noise features, such as RGB-N, IF-OSN, Noiseprint, etc., focus on analyzing and extracting noise patterns in images. Detection methods based on edge artifacts, such as GSR-Net, MVSS, EMT-Net, etc., improve detection accuracy through edge supervision. Methods based on JPEG compression artifacts and frequency domain features, such as CAT-Net, ObjectFormer, etc., focus on the extraction of compression traces and frequency domain features.
[0005] However, these methods perform poorly in the biomedical field. The difference between medical academic images and natural images at the semantic level leads to an additional reliance on tampering trace features for detection. Existing studies usually use arbitrary methods based on natural image prior design, which are prone to miss subtle clues when applied to biomedical images, and confuse splicing clues with physical interference factors. In the scenario of biomedical image tampering, splicing usually occurs between different images taken with the same device and under the same conditions, which makes it impossible for methods that rely on camera noise to obtain effective clues. At the same time, the features of digital artifacts and physical interference factors in the basic network are prone to produce too close representations, and the feature differences learned by existing methods are not obvious and easy to confuse. This leads to a high false alarm rate for methods that focus on edge artifacts. In addition, the commonly used post-processing operations lead to excessive differences in the significance of tampering traces in different regions, and weak clues are easily ignored, resulting in rough and incomplete detection / localization results. Therefore, better methods are needed for tampering detection of medical academic images. Summary of the invention
[0006] In view of the above problems, the present invention provides an image tampering detection method based on a cross-feedback optimization network, which solves the technical problem in the prior art that image features are poorly effective for medical image tampering.
[0007] The present invention provides an image tampering detection method based on a cross feedback optimization network, comprising the following steps:
[0008] Step S1, adjusting the brightness and contrast of the medical image to be detected to obtain a standardized image;
[0009] Step S2, establishing a control tampering feature extraction network, wherein the control tampering feature extraction network includes a splicing branch and a non-splicing branch, wherein the splicing branch and the non-splicing branch respectively perform multi-scale convolution on the standardized image to obtain multi-scale control tampering features;
[0010] Step S3, using a residual convolution network to perform multi-scale convolution on the standardized image to obtain multi-scale residual features;
[0011] Step S4: using a control interaction network to fuse and decode the multi-scale control tampering features and the multi-scale residual features to obtain a predicted tampering region segmentation result; performing classification prediction on the tampering region segmentation result to obtain the tampering category of the medical image to be detected.
[0012] Preferably, the step S2 specifically includes:
[0013] Step S2-1, the stitching branch performs multi-scale convolution on the input standardized image, including: repeatedly passing the input standardized image through the downsampling convolution block three times in sequence to obtain stitching texture features of the first, second and third scales respectively; and using the stitching texture features of the first, second and third scales as multi-scale stitching texture features;
[0014] Step S2-2, the non-stitching branch performs multi-scale convolution on the input standardized image, including: repeatedly passing the input standardized image through the downsampling convolution block three times in sequence to obtain non-stitching texture features of the first, second and third scales respectively; and using the non-stitching texture features of the first, second and third scales as multi-scale non-stitching texture features;
[0015] Step S2-3: using the multi-scale splicing texture feature and the multi-scale non-splicing texture feature together as the multi-scale comparison tampering feature.
[0016] Preferably, in step S2-1 and step S2-2, the processing process of the downsampling convolution block includes:
[0017] The input feature map of the downsampling convolution block passes through the downsampling convolution block with a convolution kernel size of 3×3 and a step size of 2, the batch normalization layer, the ReLU activation layer, the ordinary convolution block with a convolution kernel size of 3×3 and a step size of 1, the batch normalization layer, and the ReLU activation layer in turn, and finally obtains the output feature map;
[0018] The output convolution block includes a layer of 1×1 convolution blocks and a Sigmoid activation layer connected in sequence.
[0019] Preferably, the image tampering detection method based on the cross feedback optimization network further includes: training the control tampering feature extraction network, specifically including:
[0020] Acquire medical image tampering training data, wherein the medical image tampering training data includes multiple groups of paired medical images and tampered area annotated images;
[0021] Preprocessing the medical image, and obtaining a supervision target based on the preprocessed image and the tampered region annotated image;
[0022] An optimization loss function is determined based on the supervision objective, and the control tampering feature extraction network is trained by minimizing the mean absolute error loss function as the optimization objective.
[0023] Preferably, the step of obtaining the supervision target specifically includes:
[0024] Performing brightness and contrast adjustment preprocessing on the medical images in the medical image tampering training data;
[0025] The preprocessed image is processed by Sobel operator to obtain a gradient image, and the gradient image is intercepted based on the tampered area annotated image to obtain the spliced gradient supervision target g f and the non-spliced gradient supervision target g b ;
[0026] Using the splicing gradient supervision target as the supervision target of the splicing branch in the control tampering feature extraction network, and using the non-splicing gradient supervision target as the supervision target of the non-splicing branch in the control tampering feature extraction network;
[0027] g f and g b The expression is:
[0028]
[0029] Where x represents the input image, y represents the tampered region annotation image corresponding to the input image, expand(·) represents the channel dimension to expand the truth matrix to 3 dimensions to ensure consistency with the input image, Sobel(·) represents the Sobel operator, and y rev It represents the result of reversing y from 0 to 1;
[0030] Determine the mean absolute error loss function MAE as the optimization target, the expression is:
[0031]
[0032] Among them, y i represents the i-th spliced and unspliced gradient supervision target, represents the splicing and non-splicing targets calculated by the splicing and non-splicing branches, and n represents the total number of medical images used for training;
[0033] The training is performed with minimizing the mean absolute error loss function as the optimization goal, and finally a trained control tampering feature extraction network is obtained.
[0034] Preferably, the step S3 specifically includes:
[0035] Res2Net is used to perform multi-scale convolution on the standardized image. The input image first passes through the Res2Net initial convolution block, and then passes through the Res2Net downsampling convolution block repeated four times to obtain residual features of the first, second, third and fourth scales, respectively, as the multi-scale residual features.
[0036] Preferably, the step S4 specifically includes:
[0037] Step S4-1, respectively connecting the spliced texture features of the first, second and third scales and the non-spliced texture features of the first, second and third scales to obtain the comparison tampering features of the first, second and third scales;
[0038] Step S4-2, inputting the control tampering features of the first, second and third scales and the residual features of the first, second and third scales into three control interaction networks for fusion processing, to obtain the fusion features of the first, second and third scales;
[0039] Step S4-3, processing the residual features of the fourth scale through an upsampling convolution block, connecting the processing result with the fusion features of the third scale, processing through an upsampling convolution block, connecting the processing result with the fusion features of the second scale, processing through an upsampling convolution block, connecting the processing result with the fusion features of the first scale, and processing through an upsampling convolution block to finally obtain a tampered area segmentation result;
[0040] Step S4-4: perform global maximum pooling processing on the tampered region segmentation result, and perform classification prediction to finally obtain the tampered category.
[0041] Preferably, in step S4-2, the processing of comparing the interactive network specifically includes:
[0042] For the input residual features, firstly, the image with the tampered area annotation is intercepted to obtain the residual features of the spliced part and the residual features of the non-splicing part;
[0043] For the residual features of the spliced part, they are input into the first shuffle fusion module together with the spliced texture features in the comparison tampering features; for the residual features of the non-spliced part, they are input into the second shuffle fusion module together with the non-spliced texture features in the comparison tampering features;
[0044] The output features of the first and second shuffle fusion modules are processed by the comparison attention module to obtain a fused feature map containing the tampered area and a fused feature map containing the untampered area;
[0045] The fused feature map containing the tampered area and the fused feature map containing the untampered area are respectively intercepted and merged using the tampered area annotation image to finally obtain the fused feature.
[0046] Compared with the prior art, the present invention has at least the following beneficial effects:
[0047] (1) The present invention uses a control tampering feature extraction network to extract multi-scale control tampering features, learns the difference features between the spliced area and the original area, and effectively strengthens the feature expression of the splicing traces. This differential feature extraction reduces the impact of noise and other interference on the detection results, enhances the model's discrimination ability, avoids false alarms caused by background feature interference, and significantly improves the accuracy and robustness of medical image splicing tampering detection.
[0048] (2) The cross-feedback mechanism of the present invention optimizes the refined performance of detection by cross-fusion between the tampering features of multiple scales, which helps to capture the fine edge features of the splicing area, making the segmentation of the splicing area more complete and accurate.
[0049] (3) The present invention fuses the multi-scale control tampering features with the multi-scale residual features through the control interaction network, obtains accurate tampering region segmentation results, and further performs classification prediction, which effectively improves the detection efficiency of the system and ensures that the model can perform stably and well on different biomedical image datasets. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] The drawings are only for the purpose of illustrating particular embodiments and are not to be construed as limiting the invention.
[0051] Figure 1 A flow chart of the image tampering detection method based on a cross feedback optimization network provided by the present invention.
[0052] Figure 2 This is an overall structural diagram of the cross-feedback optimization network provided by the present invention.
[0053] Figure 3 This is a structural diagram of the control tampering feature extraction network provided by the present invention.
[0054] Figure 4 This is a structural diagram of the downsampling convolution block provided by the present invention.
[0055] Figure 5 This is a structural diagram of the control interaction network provided by the present invention. DETAILED DESCRIPTION
[0056] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other. In addition, the present invention can also be implemented in other ways different from those described herein, and therefore, the protection scope of the present invention is not limited by the specific embodiments disclosed below.
[0057] The present invention achieves synchronous learning of artificially tampered and non-tampered interference areas by introducing the true value information of the tampered area into the training target of the decoupled tampering feature extractor. A dual-branch structure is used to process the features of the tampered and non-tampered areas respectively, and the interaction is achieved through the comparative attention mechanism to highlight the difference between the real tampering traces and interference factors. The dual-dimensional feedback structure is used to optimize the feature representation and segmentation results, effectively eliminating the residual interference in each branch, thereby obtaining a more accurate detection effect.
[0058] In order to illustrate the effectiveness of the method proposed by the present invention, the above technical solution of the present invention is described in detail below through a specific embodiment. Figure 1 As shown, a method for detecting image tampering based on a cross-feedback optimization network is disclosed, and the specific implementation steps are as follows:
[0059] Step S1: adjusting the brightness and contrast of the medical image to be detected to obtain a standardized image.
[0060] The present invention preprocesses the input biomedical image to be detected, and the preprocessing may include brightness and contrast normalization, denoising, and size adjustment, etc., to form a standardized image.
[0061] Step S2, establishing a control tampering feature extraction network, wherein the control tampering feature extraction network includes a splicing branch and a non-splicing branch, wherein the splicing branch and the non-splicing branch respectively perform multi-scale convolution on the standardized image to obtain multi-scale control tampering features;
[0062] The overall structure diagram of the cross-feedback optimization network provided by the present invention is as follows: Figure 2 As shown, for the input standardized image, the comparison tampering feature extraction network is used to extract features. The structure of the comparison tampering feature extraction network is as follows Figure 3 shown.
[0063] For the trained control tampering feature extraction network, the input standardized image is subjected to multi-scale convolution of the splicing branch and multi-scale convolution of the non-splicing branch. The multi-scale convolution process of the splicing branch includes: the input standardized image is repeatedly passed through the downsampling convolution block three times in sequence to obtain the splicing texture features of the first, second and third scales respectively, and then passed through the output convolution block to obtain the output splicing target; the splicing texture features of the first, second and third scales are used as multi-scale splicing texture features.
[0064] The multi-scale convolution process of the non-stitching branch includes: repeatedly passing the input standardized image through the downsampling convolution block three times in sequence to obtain the non-stitching texture features of the first, second and third scales respectively, and then passing through the output convolution block to obtain the output non-stitching target; the non-stitching texture features of the first, second and third scales are used as multi-scale non-stitching texture features. The multi-scale splicing texture features and the multi-scale non-stitching texture features are used together as the multi-scale comparison tampering features.
[0065] The processing process of the downsampling convolution block is as follows Figure 4 As shown in the figure, it consists of two groups of convolution blocks-batch normalization layers-ReLU activation layers. The input feature map of the downsampling convolution block passes through the downsampling convolution block with a convolution kernel size of 3×3 and a step size of 2, the batch normalization layer, the ReLU activation layer, the ordinary convolution block with a convolution kernel size of 3×3 and a step size of 1, the batch normalization layer, and the ReLU activation layer in turn, and finally the output feature map is obtained.
[0066] The output convolution block contains a 1×1 convolution block and a sigmoid activation layer to produce a single channel output and map it to the [0, 1] range.
[0067] The present invention also includes the step of training the control tampering feature extraction network, specifically including:
[0068] (1) Obtaining medical image tampering training data, wherein the medical image tampering training data includes multiple sets of paired medical images and tampering region annotated images, wherein the tampering region annotated images are binary images that annotate pixel-level tampering locations of the medical images.
[0069] (2) Preprocessing the medical image to obtain a standardized image, and obtaining a supervision target based on the standardized image and the tampered area annotated image.
[0070] In this step, the gradient of the input normalized image is extracted and the spliced and non-spliced images are separated. The gradient map describes the direction and magnitude of the image brightness change, and using it as the supervision of the spliced and non-spliced tampering-related feature extractors can guide them to capture the tampering-related information of the image.
[0071] First, the standardized image is processed by the Sobel operator to obtain a gradient image, and then the gradient image is intercepted based on the tampered area labeled image to obtain the spliced gradient supervision target G f and the unstitched gradient supervision target G b , the expression is:
[0072]
[0073] Where x represents the input image, y represents the tampered region annotation image corresponding to the input image, expand(·) represents the channel dimension to expand the truth matrix to 3 dimensions to ensure consistency with the input image, Sobel(·) represents the Sobel operator, and y rev Represents the result of reversing y from 0 to 1.
[0074] Using the splicing gradient supervision target as the supervision target of the splicing branch in the control tampering feature extraction network, and using the non-splicing gradient supervision target as the supervision target of the non-splicing branch in the control tampering feature extraction network, can guide them to capture the tampering-related information of the image.
[0075] (3) Determine the mean absolute error (MAE) loss function as the optimization objective, expressed as:
[0076]
[0077] Among them, y i represents the i-th spliced and unspliced gradient supervision target, It indicates that the splicing and non-splicing branches calculate the splicing and non-splicing targets, and n indicates the total number of medical images used for training.
[0078] The training process of the present invention can adopt an error back propagation algorithm such as a gradient descent method, and train with minimizing the mean absolute error loss function as the optimization goal, and finally obtain a trained control tampering feature extraction network.
[0079] Step S3, using a residual convolution network to perform multi-scale convolution on the standardized image to obtain multi-scale residual features;
[0080] The present invention adopts Res2Net to perform multi-scale convolution on the standardized image. The input image first passes through the Res2Net initial convolution block, and then passes through the Res2Net downsampling convolution block repeated four times to obtain residual features of the first, second, third and fourth scales respectively as multi-scale residual features.
[0081] In this step, the present invention improves the ability to capture medical image details and features and enhances the feature detection effect by acquiring residual features of different scales.
[0082] Step S4: using a control interaction network to fuse and decode the multi-scale control tampering features and the multi-scale residual features to obtain a predicted tampering region segmentation result; performing classification prediction on the tampering region segmentation result to obtain the tampering category of the medical image to be detected.
[0083] In this step, firstly, the spliced texture features of the first, second and third scales and the non-spliced texture features of the first, second and third scales are connected respectively to obtain the comparison tampering features of the first, second and third scales;
[0084] The control tampering features of the first, second and third scales and the residual features of the first, second and third scales are respectively input into three control interaction networks for fusion processing to obtain the fusion features of the first, second and third scales.
[0085] The present invention provides a control interaction network for fusing control tampering features and residual features. The structure of the control interaction network is as follows: Figure 5 As shown in Figure 2, for the input residual features, the image of the tampered area is first intercepted to obtain the residual features of the spliced part and the residual features of the non-splicing part.
[0086] For the residual features of the spliced part, they are input into the first shuffled fusion module together with the spliced texture features in the control tampering features; for the residual features of the non-spliced part, they are input into the second shuffled fusion module together with the non-spliced texture features in the control tampering features; the output features of the first and second shuffled fusion modules are processed by a comparative attention module (CATM) to obtain a fused feature map including the tampered area and a fused feature map including the untampered area; the fused feature map including the tampered area and the fused feature map including the untampered area are respectively intercepted and merged using the tampered area annotated image to finally obtain the fused features.
[0087] The multi-scale fusion features are decoded to obtain the predicted tampered area segmentation results. The specific steps of feature decoding include:
[0088] The residual features of the fourth scale are processed by the upsampling convolution block, and the processing results are connected with the fused features of the third scale. After being processed by the upsampling convolution block, the processing results are connected with the fused features of the second scale. After being processed by the upsampling convolution block, the processing results are connected with the fused features of the first scale. After being processed by the upsampling convolution block, finally the tampered area segmentation result is obtained.
[0089] The tampered region segmentation result is subjected to global maximum pooling processing, and classification prediction is performed to finally obtain the tampered category.
[0090] It can be understood that the residual convolutional network and the control interaction network provided by the present invention can complete training by updating model parameters using methods such as gradient descent, and the present invention does not limit the training method of the residual convolutional network and the control interaction network.
[0091] The present invention adopts a combination of contrast feature extraction module and cross feedback mechanism, and improves the detection accuracy and robustness of splicing traces through multi-level feature extraction and two-way feedback between features.
[0092] Although the specific embodiments of the present invention depict various actions or steps in a specific order, this should be understood as requiring such actions or steps to be performed in the specific order shown or in a sequential order, or requiring all illustrated actions or steps to be performed to obtain the desired results. Under certain circumstances, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be interpreted as limiting the scope of the present disclosure. Certain features described in the context of a separate embodiment can also be implemented in a single implementation in combination. On the contrary, the various features described in the context of a single implementation can also be implemented in multiple implementations individually or in any suitable sub-combination. The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or replacements that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
[0093] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.
Claims
1. An image tampering detection method based on a cross-feedback optimization network, characterized in that: The following steps are involved: Step S1, adjusting the brightness and contrast of the medical image to be detected to obtain a standardized image; Step S2, establishing a control tampering feature extraction network, wherein the control tampering feature extraction network includes a splicing branch and a non-splicing branch, wherein the splicing branch and the non-splicing branch respectively perform multi-scale convolution on the standardized image to obtain multi-scale control tampering features; Step S3, using a residual convolution network to perform multi-scale convolution on the standardized image to obtain multi-scale residual features; Step S4: using a control interaction network to fuse and decode the multi-scale control tampering features and the multi-scale residual features to obtain a predicted tampering region segmentation result; performing classification prediction on the tampering region segmentation result to obtain the tampering category of the medical image to be detected.
2. The image tampering detection method based on cross feedback optimization network according to claim 1 is characterized in that: The step S2 specifically includes: Step S2-1, the stitching branch performs multi-scale convolution on the input standardized image, including: repeatedly passing the input standardized image through the downsampling convolution block three times in sequence to obtain stitching texture features of the first, second and third scales respectively; and using the stitching texture features of the first, second and third scales as multi-scale stitching texture features; Step S2-2, the non-stitching branch performs multi-scale convolution on the input standardized image, including: repeatedly passing the input standardized image through the downsampling convolution block three times in sequence to obtain non-stitching texture features of the first, second and third scales respectively; and using the non-stitching texture features of the first, second and third scales as multi-scale non-stitching texture features; Step S2-3: using the multi-scale splicing texture feature and the multi-scale non-splicing texture feature together as the multi-scale comparison tampering feature.
3. The image tampering detection method based on cross feedback optimization network according to claim 2 is characterized in that: In step S2-1 and step S2-2, the processing process of the downsampling convolution block includes: The input feature map of the downsampling convolution block passes through the downsampling convolution block with a convolution kernel size of 3×3 and a step size of 2, the batch normalization layer, the ReLU activation layer, the ordinary convolution block with a convolution kernel size of 3×3 and a step size of 1, the batch normalization layer, and the ReLU activation layer in turn, and finally obtains the output feature map; The output convolution block includes a layer of 1×1 convolution blocks and a Sigmoid activation layer connected in sequence.
4. The image tampering detection method based on cross feedback optimization network according to claim 3 is characterized in that: Also includes: Training the control tampering feature extraction network specifically includes: Acquire medical image tampering training data, wherein the medical image tampering training data includes multiple groups of paired medical images and tampered area annotated images; Preprocessing the medical image, and obtaining a supervision target based on the preprocessed image and the tampered region annotated image; An optimization loss function is determined based on the supervision objective, and the control tampering feature extraction network is trained by minimizing the mean absolute error loss function as the optimization objective.
5. The image tampering detection method based on cross feedback optimization network according to claim 4 is characterized in that: The steps of obtaining the supervision target specifically include: Performing brightness and contrast adjustment preprocessing on the medical images in the medical image tampering training data; The preprocessed image is processed by Sobel operator to obtain a gradient image, and the gradient image is intercepted based on the tampered area annotated image to obtain the spliced gradient supervision target G f and the unstitched gradient supervision target G b ; Using the splicing gradient supervision target as the supervision target of the splicing branch in the control tampering feature extraction network, and using the non-splicing gradient supervision target as the supervision target of the non-splicing branch in the control tampering feature extraction network; G f and G b The expression is: Where x represents the input image, y represents the tampered region annotation image corresponding to the input image, expand(·) represents the channel dimension to expand the truth matrix to 3 dimensions to ensure consistency with the input image, Sobel(·) represents the Sobel operator, and Y rev It represents the result after reversing Y from 0 to 1; The mean absolute error loss function MAE is determined as the optimization target, and the expression is: Among them, Y i represents the Ith spliced and unspliced gradient supervision target, represents the splicing and non-splicing targets calculated by the splicing and non-splicing branches, and n represents the total number of medical images used for training; The training is performed with minimizing the mean absolute error loss function as the optimization goal, and finally a trained control tampering feature extraction network is obtained.
6. The image tampering detection method based on cross feedback optimization network according to claim 5 is characterized in that: The step S3 specifically includes: Res2Net is used to perform multi-scale convolution on the standardized image. The input image first passes through the Res2Net initial convolution block, and then passes through the Res2Net downsampling convolution block repeated four times to obtain residual features of the first, second, third and fourth scales, respectively, as the multi-scale residual features.
7. The image tampering detection method based on cross feedback optimization network according to claim 6 is characterized in that: The step S4 specifically includes: Step S4-1, respectively connecting the spliced texture features of the first, second and third scales and the non-spliced texture features of the first, second and third scales to obtain the comparison tampering features of the first, second and third scales; Step S4-2, inputting the control tampering features of the first, second and third scales and the residual features of the first, second and third scales into three control interaction networks for fusion processing, to obtain the fusion features of the first, second and third scales; Step S4-3, processing the residual features of the fourth scale through an upsampling convolution block, connecting the processing result with the fusion features of the third scale, processing through an upsampling convolution block, connecting the processing result with the fusion features of the second scale, processing through an upsampling convolution block, connecting the processing result with the fusion features of the first scale, and processing through an upsampling convolution block to finally obtain a tampered area segmentation result; Step S4-4: perform global maximum pooling processing on the tampered region segmentation result, and perform classification prediction to finally obtain the tampered category.
8. The image tampering detection method based on cross feedback optimization network according to claim 7 is characterized in that: In step S4-2, the processing process of comparing the interactive network specifically includes: For the input residual features, firstly, the image with the tampered area annotation is intercepted to obtain the residual features of the spliced part and the residual features of the non-splicing part; For the residual features of the spliced part, they are input into the first shuffle fusion module together with the spliced texture features in the comparison tampering features; for the residual features of the non-spliced part, they are input into the second shuffle fusion module together with the non-spliced texture features in the comparison tampering features; The output features of the first and second shuffle fusion modules are processed by the comparison attention module to obtain a fused feature map containing the tampered area and a fused feature map containing the untampered area; The fused feature map containing the tampered area and the fused feature map containing the untampered area are respectively intercepted and merged using the tampered area annotation image to finally obtain the fused feature.