Image Forgery Localization Method, Medium and System Based on Multi-Prior Fusion Strategy

Abstract

The present invention claims protection for an image forgery localization method, medium and system based on a multi-prior fusion strategy, which relates to technical fields such as digital image processing, computer vision, and machine learning. The specific steps are as follows: 1) To produce a forged dataset, collect a dataset of real images and corresponding region masks in the existing field; 2) Use the collected real images and masks to produce a forged dataset through an algorithm. In order to further improve the generalization ability of the deep learning model, three types of forged datasets are produced according to the different qualities of the original images; 3) Use the three produced forged datasets to train the deep learning model with an encoder-decoder structure; 4) At the same time, in order to further enhance the generalization forgery localization ability of deep learning, use a multi-prior fusion strategy for training. 5) Use the trained deep learning model to localize the forged areas of the tested tampered images. The model can detect and localize the tampered images in reality.

Description

Technical Field

[0001] The present invention relates to a method for digital image forgery detection and localization, and belongs to the technical fields of digital image processing, computer vision, machine learning, etc. Background Art

[0002] With the increasing development of information technology, digital images have become an important carrier for information storage and dissemination. Nowadays, digital cameras and mobile phones have become very common electronic products in people's lives, and people can easily take and store photos. At the same time, with the popularization of social networks and high-speed networks, people can easily share and spread their pictures on the Internet, and digital images have become an indispensable important part of people's daily lives. The wide application of digital images is not only in personal daily life, but also widely used in industries, medicine, science and other fields. Therefore, the development of digital image technology has increasingly become a research hotspot. There is an old Chinese saying: "What you hear is false, what you see is true." However, with the development of some image editing software, such as Photoshop, GIMP, etc. These image editing software often have the characteristics of easy-to-operate user interfaces, and people can easily use the image editing software to forge images in various ways. Forged images usually distort the content of the original image, and it is often difficult for the human eye to distinguish them, which conveys false information to people and will cause various negative social impacts. According to whether the semantic content of the forged image changes, image forgery can be divided into two types: global image forgery and local image forgery. Among them, global image forgery is to operate on the global content of the image through some image processing methods (such as image contrast enhancement, image histogram equalization) to change the visual effect of the image, and this process will not change the semantic content information of the image. Local image forgery is to change the local content of an image in some way. This kind of locally forged image has semantic content that did not exist originally, and can convey false information to people, so there is a greater potential security risk and may cause greater negative impacts. For local image forgery, it can be divided into three categories according to different forgery methods: image splicing forgery (Splicing), image copy-move forgery (Copy-Move), and image removal forgery (Removal). Different forgery methods have their own characteristics and will cause different negative impacts. With the development of deep learning technology, more and more researchers use deep learning technology for image forgery detection. In fact, deep learning technology has become the most effective way for current image forensics research. With the development of deep learning technology, researchers have gradually paid attention to the generalization detection and localization ability of deep learning models, that is, deep learning can detect forged samples that have not participated in training, which is more meaningful in reality. In addition, locating the forged content of the forged image and giving a clear forged area can enable people to distinguish better.

[0003] With the increasing development of artificial intelligence-related technologies, image forgery detection through deep learning has become the most important approach currently and in the future. At the same time, deep learning technology has also been proven to have strong performance. For methods of using deep learning technology for image forgery detection, they can be divided into two categories according to different training categories and test categories: methods for specific forgery types and methods for multiple forgery types. Currently, the methods often cannot detect datasets that are not involved in training, that is, the model lacks generalization detection and localization capabilities.

[0004] CN114677332A, a method for detecting image tampering based on deep feature fusion of JPEG fingerprints, relates to technical fields such as digital image processing, computer vision, and deep learning. The specific steps are as follows: 1) Collect and organize publicly available uncompressed image samples; 2) Tamper with the image samples in two ways and label them to obtain tampered image samples and labels to complete the construction of the image tampering dataset; 3) Use the made image tampering dataset to train the deep feature fusion network; 4) Use the trained model to test the two types of tampered images to obtain the final effect. The model trained by using the feature fusion convolutional neural network in this method can detect tampered images in reality, has practical significance, and achieves good detection accuracy.

[0005] Defect 1: Overly dependent on JPEG format images and fails when the forged images are in other formats.

[0006] Solution 1: Made 3 types of datasets, provided more prior information, and was more in line with real-world problems.

[0007] Defect 2: Sensitive to the JEPG compression factor QF

[0008] Solution 2: Does not depend on a specific factor and is simpler and more effective.

[0009] Defect 3: Need to collect uncompressed images to make the forged dataset

[0010] Solution 3: Has no excessive restrictions on real images.

[0011] Defect 4: When training the model with two types of tampered images mixed, the features may interfere with each other and it is impossible to fit the model with the optimal effect.

[0012] Solution 4: Use a multi-prior fusion strategy for training to avoid the problems brought by blind mixed training. Summary of the Invention

[0013] The present invention aims to solve the problems existing in the existing image forgery detection methods. By creating forgery datasets with different original image qualities, various image quality situations in the real world are simulated. To avoid mutual interference between features caused by direct mixed training, a multi-prior fusion strategy is used for training, so that the model can have better generalization detection ability. The technical solution of the present invention is as follows:

[0014] An image forgery localization method based on a multi-prior fusion strategy, which comprises the following steps:

[0015] (1) Obtain publicly available real original images and region-corresponding mask samples;

[0016] (2) Use an image local forgery algorithm written in Python to create forgery datasets for the obtained real original images and region-corresponding mask samples; According to different original image qualities, three image forgery datasets are created, namely Set_ori, Set_jpeg, and Set_blur;

[0017] (3) Use the three forgery datasets created in step (2) to train a deep learning model with an encoder-decoder structure;

[0018] (4) Use a multi-prior fusion strategy to train the deep learning model in step (3);

[0019] (5) Use the model trained in step (4) to test the input image, and obtain the final forgery localization result, which is shown as a binary mask, where the white area represents the forgery area and the black area represents the real area.

[0020] Further, the step (1) of obtaining publicly available real original images and region-corresponding mask samples specifically includes:

[0021] Forgery datasets are created using the COCO dataset and the corresponding binary masks; First, parse the validation set of COCO2017, and generate binary masks corresponding to a certain object in the image, and remove those images with too large object areas (>50%) and too small areas (<5%); Then, from the remaining images, randomly select 4000 original images and object corresponding masks.

[0022] Further, the step (2) of creating three image forgery datasets according to different original image qualities specifically includes:

[0023] Three types of forged datasets were created. The original images of each dataset were taken from the same 4,000 images, with the difference being the operations before the forgery process; the dataset obtained by directly forging the original images is called Set_ori; the dataset obtained by first performing random Gaussian blur on the original images and then forging is called Set_blur; the dataset obtained by first performing random JPEG compression on the original images and then forging is called Set_jpeg.

[0024] Furthermore, the specific process of image forgery in step (2) is implemented through an algorithm. First, a mask of the forged object is generated based on the mask of the real object, and then a forged image is generated, as shown in the following formula:

[0025]

[0026] where M fake represents the binary mask corresponding to the forged image, I fake refers to the finally forged image; I real is the MSCOCO real original image, mask is the binary mask corresponding to the real object of I real , mask 2 represents an image with all pixel values being 0 and the same size as the original image; Box(·) represents calculating the maximum contour bounding box of the cropped object area; this operation is to crop out the object while avoiding too many redundant pixels; T(·) represents random size transformation; represents randomly pasting the former into the latter; in addition, T(·) and have exactly the same operations in the formulas for making the forged mask and the forged image, thus ensuring the precise correspondence between the forged image and the binary mask of the forged area.

[0027] Furthermore, in step (3), the three forged datasets made in step (2) are used to train the deep learning model with an encoder-decoder structure, specifically including:

[0028] The designed model uses the three forged datasets made as the training set, and each dataset provides a prior information for the deep learning model; in order to avoid direct mixed training, a multi-prior fusion strategy is used to train the model; the basic network architecture of the deep learning model adopts an encoder-decoder structure. The encoder extracts features from the input image, and the encoder part uses existing network structures for combination, including ResNet-50, VGG-16, and Transformer; the decoder part adopts a combination of multi-layer upsampling and convolution, uses the features extracted by the encoder to distinguish between the forged area and the real area, and finally outputs a binary mask of the same size as the input image to identify the forged area.

[0029] Furthermore, the specific process of training the deep learning model using the multi-prior fusion strategy includes:

[0030] In each iteration, a data training batch is constructed using samples from a single type of forged dataset, and the corresponding loss is calculated through forward propagation. Then, instead of backpropagating to update the parameters, another training batch is sampled for another type of forged dataset to obtain the corresponding loss, but again, the parameters are not updated through backpropagation. When each type of forged dataset has undergone forward propagation, the losses calculated for each batch are accumulated, and the entire gradient is backpropagated. Then, this multi-prior fusion training strategy is repeated alternately. Each update of the model weights through backpropagation includes the contributions of each type of forged data. The cross-entropy loss function is used to calculate the loss during the model training process.

[0031] Furthermore, the loss calculation formula during the model training process is as follows:

[0032]

[0033] where N represents the number of different types of forged training sets, M represents the number of pixels in an image, represents the predicted probability of the j-th pixel of the sample on the i-th forged dataset, represents the corresponding label of the j-th pixel of the sample on the i-th forged dataset; λ represents the weight when calculating the loss on the i-th forged dataset, used to adjust the importance of different forged datasets. Here, λ i = 1, considering that each prior information is important.

[0034] A medium stores a computer program internally. The computer program, when read by a processor, executes the image forgery localization method based on the multi-prior fusion strategy described in any one of the above items.

[0035] An image forgery localization system based on the method described in any one of the above items includes:

[0036] An acquisition module: used to acquire publicly available real original images and region-corresponding mask samples;

[0037] A forged dataset production module: used to produce a forged dataset for the acquired real original images and region-corresponding mask samples using an image local forgery algorithm written in Python; Three image forgery datasets are produced according to different original image qualities, namely Set_ori, Set_jpeg, and Set_blur;

[0038] Training module: used to train a deep learning model with an encoder-decoder structure by using three fabricated datasets; for the deep learning model, a multi-prior fusion strategy is used for training;

[0039] Testing module: used to test the input image by using the trained model, and obtain the final forgery localization result, which is shown as a binary mask. The white area represents the forged area, and the black area represents the real area.

[0040] The advantages and beneficial effects of the present invention are as follows:

[0041] The present invention utilizes technologies such as digital image processing, computer vision, and machine learning to achieve the task of digital image forgery localization. The present invention is a deep learning-based method that improves the generalization ability of the deep learning model. The present invention has the following advantages:

[0042] (1) It is convenient and effective to build a network for training by using Pytorch;

[0043] (2) Collect real images and generate forged images through algorithms, saving a large amount of human resources;

[0044] (3) The present invention has effects on three types: image splicing, image copy-pasting, and image deletion.

[0045] (4) The present invention locates the forged area, which is more conducive to people's discrimination.

[0046] (5) It has good model generalization ability. For datasets not involved in training, it also has good forgery localization effects, which has practical significance.

[0047] Innovation point 1: Three forged datasets are made according to the quality of the original image, which are respectively called Set_ori, Set_jpeg, and Set_blur. Each type of dataset provides a kind of prior information, and no one has considered such prior information before. It is beneficial to the generalization forgery localization ability of the model (Advantage 3, Advantage 5).

[0048] Innovation Point 2: When training with multiple datasets, directly mixing training may cause interference between features, and the model may only fit a certain type of feature. The present invention uses a multi-prior fusion strategy to avoid the above problems and further improves the forgery localization performance. Description of the Drawings

[0049] Figure 1 It is the system flowchart of the preferred embodiment provided by the present invention;

[0050] Figure 2 (a) is the original image of the forged image;

[0051] Figure 2 (b) is a forged image. The first row is forged by image splicing, and the second row is forged by image copy-pasting.

[0052] Figure 2 (c) Labels of the forged image, where the white area corresponds to the forged area.

[0053] Figure 2 (d) Detection effect of this paper. Detailed implementation manners

[0054] Next, the technical solutions in the embodiments of the present invention will be clearly and detailedly described in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only a part of the embodiments of the present invention.

[0055] The technical solution for the present invention to solve the above technical problems is:

[0056] The system flowchart is as Figure 1 shown. A contrast enhancement detection method based on histogram features includes the following steps:

[0057] The first step: Collect real images and corresponding region mask datasets; in this paper, a forged dataset is made using the COCO dataset and the corresponding binary masks. First, this paper parsed the validation set of COCO2017 and generated binary masks corresponding to a certain object in the image, removing those images with too large object areas (>50%) and too small areas (<5%). Then, from the remaining images, 4000 original images and the corresponding object masks were randomly selected.

[0058] The second step: Use an algorithm to make a forged dataset for the collected image samples. According to the different qualities of the original images, three image forged datasets were made; this paper made 3 types of forged datasets, and the original images of each dataset came from the same 4000 images. The difference lies in the operations before the forgery process. The dataset directly forged from the original images is called Set_ori; the dataset first randomly Gaussian blurred and then forged is called Set_blur; the dataset first randomly JPEG compressed and then forged is called Set_jpeg. The specific process of image forgery is realized through an algorithm. First, a mask of the forged object is generated according to the mask of the real object, and then a forged image is generated, as shown in the following formula.

[0059]

[0060]

[0061] Where M fakeRepresents the binary mask corresponding to the forged image. fake Refers to the final forged image. real is the real original image of MSCOCO, and the mask is I real The real object corresponds to a binary mask, mask 2 Represents an image with the same size as the original image and all pixel values ​​are 0. Box(·) represents the maximum contour bounding box calculated for the cropped object area. This operation is to crop the object while avoiding too many redundant pixels. T(·) represents a random size transformation. Indicates pasting the former into the latter at a random position. In addition, T(·) and The operations in Formula 4-1 and Formula 4-2 are exactly the same, thereby ensuring that the forged image and the forged region binary mask correspond exactly.

[0062] Step 3: Use the three fake datasets to train the deep learning model of the encoder-decoder structure and train it through a multi-prior fusion strategy. The specific steps are as follows:

[0063] Specifically, the model designed in this paper uses three fake data sets as training sets, and each data set provides a priori information for the deep learning model. In order to avoid direct mixed training, a one-to-many prior fusion strategy is used to train the model. The basic network architecture of the deep learning model adopts an encoder-decoder structure. The encoder extracts features from the input image. The encoder part can be combined with existing network structures, such as ResNet-50, VGG-16, and Transformer. The decoder part uses a combination of multi-layer upsampling and convolution, and uses the features extracted by the encoder to distinguish between fake and real areas, and finally outputs a binary mask of the same size as the input image to identify the fake area.

[0064] In order to avoid the problems caused by direct mixed training as much as possible, this paper trains deep learning models through a multi-prior fusion strategy. Specifically, in each iteration, a data training batch is constructed using samples of a single type of forged data set, and the corresponding loss is calculated through forward propagation. Then, there is no rush to backpropagate and update the parameters, but continue to sample training batches for another type of forged data set to obtain the corresponding loss, but there is no rush to backpropagate and update the parameters. When each type of forged data set has undergone forward propagation, the losses calculated for each batch will be accumulated, and the entire gradient will be backpropagated. Then this paper repeats this multi-prior fusion training strategy. Each backpropagation parameter update of the model weights includes the contribution of each type of forged data. This paper uses the cross entropy loss function to calculate the loss in the process of training the model. The loss calculation formula for the model training process is as follows:

[0065]

[0066] Where N represents the number of different types of forged training sets, and M represents the number of pixels in an image. represents the predicted probability of the j-th pixel of the sample on the i-th forged data set. represents the corresponding label of the j-th pixel of the sample on the i-th forged data set. λ represents the weight when calculating the loss on the i-th forged data set, which is used to adjust the importance of different forged data sets. Here, this paper sets λ i = 1, and this paper believes that each prior information is very important.

[0067] The model trained by this multi-prior fusion strategy can learn richer effective information in the feature space. Generally speaking, when using one prior information to train a deep learning network, in an ideal situation, the deep learning model will learn relevant feature information, and its performance in the feature space is that it has good effects on data with feature distances close to those of the training data. However, when the feature distance of the test data is very different from that of the training data, it has no effect on such data. Therefore, when only training one forged data set, the detection data of the model is limited to a certain type of data. However, the data in the real world is diverse, which is why when most forged detection and localization methods detect samples in some data sets that are not involved in training, the detection and localization effects become very poor or even ineffective. When using the three types of forged data sets made in this paper, more prior information is provided during the training process, and at the same time, the multi-prior fusion strategy is used to train the deep learning model. The model learns richer forged clue features, thereby improving the generalization detection and localization ability of the model.

[0068] Step 4: Use the trained model to test the input image to obtain the final forged localization result, which is shown as a binary mask. The white area represents the forged area, and the black area represents the real area.

[0069] Experimental method:

[0070] During this experiment, we parsed the validation set of COCO2017 in this paper and generated binary masks corresponding to a certain object in the image, removing those images with object areas that are too large (>50%) and too small (<5%). Then, from the remaining images, 4000 original images and the corresponding object masks were randomly selected. Then, three forged data sets were made using these original images, and these data were used as the training set. To verify the generalization detection and localization ability of the model, some publicly available data sets were used for testing, including the CASIA and Coverage data sets.

[0071] Step 1: Process the original images according to different image processing methods to simulate different image qualities. Forge the original images of different qualities through an algorithm to obtain three forged datasets as the training sets for the deep learning model.

[0072] Step 2: Build a deep learning model using the Pytorch framework. Input the training set images and their corresponding labels into the deep learning model. After training and calculating the loss using the multi-prior fusion strategy, perform backpropagation iteration to update the parameters. After optimizing the trained parameters, obtain the finally trained model.

[0073] Step 3: Use the trained model to test the images in the public dataset, output a forged localization binary mask, where black represents the real area and white represents the forged area, and calculate the accuracy of the forged localization according to the corresponding labels.

[0074] Experiments prove that the method proposed in the present invention can effectively perform forged localization on the public dataset after training. The model has the ability of generalized forged localization and has practical significance.

[0075] The systems, devices, modules or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0076] Computer-readable media includes both permanent and non-permanent, removable and non-removable media and can be implemented by any method or technology for storing information. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette tapes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission media that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0077] It should also be noted that the term "comprising", "including" or any other variation thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such process, method, commodity or device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the process, method, commodity or device comprising said element.

[0078] The above embodiments should be understood as being only for illustrative purposes of the present invention and not for limiting the scope of protection of the present invention. After reading the content described in the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. An image forgery localization method based on a multi-prior fusion strategy, characterized in that, it includes the following steps: (1) Obtain publicly available real original images and region-corresponding mask samples; (2) Use an image local forgery algorithm written in Python on the obtained real original images and region-corresponding mask samples to produce forgery datasets; According to different original image qualities, three image forgery datasets are produced, namely Set_ori, Set_jpeg, and Set_blur; (3) Use the three forgery datasets produced in step (2) to train a deep learning model with an encoder-decoder structure; (4) Use a multi-prior fusion strategy to train the deep learning model in step (3); (5) Use the model trained in step (4) to test the input image, and obtain the final forgery localization result, which is shown as a binary mask, where the white area represents the forgery area and the black area represents the real area; The specific process of image forgery in step (2) is implemented through an algorithm. First, a mask of the forgery object is generated based on the mask of the real object, and then a forgery image is generated, as shown in the following formula: Among them, M fake represents the binary mask corresponding to the forged image, and I fake refers to the finally forged image; I real is the MSCOCO real original image, and mask is the binary mask corresponding to the real object of I real ; mask 2 represents an image with all pixel values being 0 and the same size as the original image; Box(·) represents calculating the maximum contour bounding box of the cropped object area; this operation is to crop out the object while avoiding too many redundant pixels; T(·) represents a random size transformation; represents pasting the former randomly into the latter; in addition, T(·) and have exactly the same operations in the formulas for making the forged mask and the forged image, so as to ensure the exact correspondence between the forged image and the binary mask of the forged area; The step (3) uses the three forgery datasets produced in step (2) to train a deep learning model with an encoder-decoder structure, specifically including: The designed model uses the three forgery datasets produced as the training set, and each dataset provides a prior information for the deep learning model; In order to avoid direct mixed training, a multi-prior fusion strategy is used to train the model; The basic network architecture of the deep learning model adopts an encoder-decoder structure. The encoder extracts features from the input image, and the encoder part uses existing network structures for combination, including ResNet-50, VGG-16, and Transformer; The decoder part adopts a combination of multi-layer upsampling and convolution, uses the features extracted by the encoder to distinguish the forgery area and the real area, and finally outputs a binary mask with the same size as the input image to identify the forgery area.

2. The image forgery localization method based on a multi-prior fusion strategy according to claim 1, characterized in that, the step (1) of obtaining publicly available real original images and region-corresponding mask samples specifically includes: A forgery dataset is produced using the COCO dataset and the corresponding binary masks; First, the validation set of COCO2017 is parsed, and a binary mask corresponding to a certain object in the image is generated, and the images with object areas greater than 50% and areas less than 5% are removed; Then, from the remaining images, 4000 original images and object corresponding masks are randomly selected.

3. The image forgery localization method based on a multi-prior fusion strategy according to claim 1, characterized in that, the step (2) of producing three image forgery datasets according to different original image qualities specifically includes: Three types of forged datasets were created. The original images of each dataset were taken from the same 4,000 images, with the difference being the operations before the forgery process; the dataset obtained by directly forging the original images is called Set_ori; the dataset obtained by first performing random Gaussian blur on the original images and then forging is called Set_blur; the dataset obtained by first performing random JPEG compression on the original images and then forging is called Set_jpeg.

4. The image forgery localization method based on the multi-prior fusion strategy according to claim 1, characterized in that the training of the deep learning model using the multi-prior fusion strategy specifically includes: In each iteration, a data training batch was constructed using samples of a single type of forged dataset, and the corresponding loss was calculated through forward propagation; then, at this time, the parameters were not updated by backpropagation, but instead, a training batch was sampled for another type of forged dataset to obtain the corresponding loss, but again, the parameters were not updated by backpropagation; when each type of forged dataset has undergone forward propagation, the losses calculated for each batch will be accumulated, and the entire gradient will be backpropagated; then this multi-prior fusion training strategy is repeated alternately; each backpropagation parameter update of the model weights includes the contributions of each type of forged data; the cross-entropy loss function was used to calculate the loss during the training of the model.

5. The image forgery localization method based on the multi-prior fusion strategy according to claim 4, characterized in that the loss calculation formula during the model training process is as follows: where N represents the number of different types of forged training sets, and M represents the number of pixels in an image, represents the predicted probability of the j-th pixel of the sample on the i-th forged data set, represents the corresponding label of the j-th pixel of the sample on the i-th forged data set; λ represents the weight when calculating the loss on the i-th forged dataset, which is used to adjust the importance of different forged datasets. Here, λ i is set to 1, assuming that each prior information is equally important.

6. A medium that stores a computer program internally, characterized in that when the computer program is read by a processor, it executes the image forgery localization method based on the multi-prior fusion strategy according to any one of claims 1 to 5 above.

7. An image forgery localization system based on the method according to any one of claims 1-5, characterized in that it includes: Acquisition module: used to acquire publicly available real original images and region corresponding mask samples; Forged dataset production module: used to produce a forged dataset for the acquired real original images and region corresponding mask samples using an image local forgery algorithm written in python; according to the different qualities of the original images, three image forged datasets were produced, called Set_ori, Set_jpeg, and Set_blur respectively; Training module: used to train a deep learning model with an encoder-decoder structure using the three forged datasets produced; For the deep learning model, the multi-prior fusion strategy is used for training; Testing module: used to test the input images using the trained model to obtain the final forgery localization result, which is presented as a binary mask, where the white area represents the forged area and the black area represents the real area.

Images