A face forgery detection method and system based on dual-branch convolutional suppression texture network

Abstract

The application discloses a face forgery detection method and system based on a double-branch convolution texture suppression network, which inputs a to-be-predicted face image into the double-branch convolution texture suppression network to perform face forgery detection and obtain a face forgery detection result. The face forgery detection comprises the following steps: adding a first texture suppression feature map and a second texture suppression feature map to obtain a first feature map of a first branch; performing element multiplication on a texture suppression attention feature map and a feature map of a second branch to obtain a first feature map of the second branch; inputting the features of the first branch and the features of the second branch into a cross-flow fusion module to perform feature fusion of different dimensions, and obtaining fused features of the first branch and fused features of the second branch; and inputting the fused features in each dimension into a full connection layer after flattening and splicing operations to obtain the face forgery detection result. The application can improve the generalization performance of the model and improve the accuracy of face forgery detection.

Description

Technical Field

[0001] This invention relates to the field of face forgery detection technology, and in particular to a face forgery detection method and system based on a bi-branch convolutional suppression texture network. Background Technology

[0002] Deep learning-based face forgery detection has only emerged in recent years. While its detection results within a single dataset are already quite good, its performance across different datasets is less than satisfactory. This is partly because deep learning is still rapidly evolving, and older models have struggled with cross-dataset face forgery detection. Furthermore, face forgery datasets are constantly being updated, and forgery effects are becoming increasingly sophisticated. It is now difficult to rely on visually perceptible differences (e.g., significant color differences between the left and right eyes in a forged image, missing details and light reflections in the eye and teeth areas) for detection. Additionally, the different data distributions across datasets mean that features learned on the training dataset may not perform well on new datasets. Summary of the Invention

[0003] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention proposes a face forgery detection method and system based on a bi-branch convolutional suppression texture network, which can improve the generalization performance of the model and improve the accuracy of face forgery detection.

[0004] In a first aspect, embodiments of the present invention provide a face forgery detection method based on a dual-branch convolutional suppression texture network, the face forgery detection method based on a dual-branch convolutional suppression texture network comprising:

[0005] Obtain the image of the face to be predicted;

[0006] Construct a two-branch convolutional texture suppression network that includes a texture suppression module convolution, a cross-stream interaction module, and a cross-stream fusion module;

[0007] The image of the face to be predicted is input into the dual-branch convolutional suppression texture network for face forgery detection to obtain the face forgery detection result; wherein the dual-branch convolutional suppression texture network for face forgery detection includes:

[0008] The face image to be predicted is input into the texture suppression module convolution for feature extraction to obtain the first texture suppression feature map;

[0009] The face image to be predicted is passed through a convolutional layer to obtain the feature map of the second branch. The feature map of the second branch is then input into the texture suppression module in the cross-flow interaction module for feature extraction to obtain the second texture suppression feature map. The first texture suppression feature map and the second texture suppression feature map are then added to obtain the first feature map of the first branch of the dual-branch convolutional texture suppression network.

[0010] The first feature map of the first branch is input into the texture suppression feature attention in the cross-flow interaction module for feature extraction to obtain the texture suppression attention feature map. The texture suppression attention feature map and the feature map of the second branch are then multiplied element-wise to obtain the first feature map of the second branch of the dual-branch convolutional texture suppression network.

[0011] Features are extracted from the first feature map of the first branch to obtain the features of the first branch; features are extracted from the first feature map of the second branch to obtain the features of the second branch.

[0012] The features of the first branch and the features of the second branch are input into the cross-stream fusion module for feature fusion of different dimensions to obtain the fused features of the first branch and the fused features of the second branch.

[0013] The fused features in each dimension are flattened and spliced ​​to obtain a one-dimensional vector, which is then input into a fully connected layer for face forgery detection to obtain the face forgery detection result.

[0014] Compared with the prior art, the first aspect of the present invention has the following beneficial effects:

[0015] This method constructs a dual-branch convolutional texture suppression network (DCRMR) comprising a texture suppression module, a cross-stream interaction module, and a cross-stream fusion module. The method inputs the face image to be predicted into the DCRMR for face forgery detection, obtaining the face forgery detection result. Using the DCRMR for face forgery detection improves the accuracy of face forgery detection. The DCRMR for face forgery detection includes: inputting the face image to be predicted into the texture suppression module's convolution for feature extraction to obtain a first texture suppression feature map. The texture suppression module's convolution improves feature analysis capabilities and has better image texture suppression capabilities; passing the face image to be predicted through a convolutional layer to obtain a feature map for the second branch; inputting the second branch's feature map into the texture suppression module in the cross-stream interaction module for feature extraction to obtain a second texture suppression feature map; adding the first and second texture suppression feature maps to obtain the first feature map of the first branch of the DCRMR; and inputting the first feature map of the first branch into the texture suppression feature map in the cross-stream interaction module for feature extraction to obtain the texture suppression feature map. The textural suppression attention feature map is processed, and the textural suppression attention feature map and the feature map of the second branch are multiplied element-wise to obtain the first feature map of the second branch of the dual-branch convolutional suppression texture network. The cross-stream interaction module can guide the dual-branch convolutional suppression texture network to better extract more general features in face forgery. Features are extracted from the first feature map of the first branch to obtain the features of the first branch. Features are extracted from the first feature map of the second branch to obtain the features of the second branch. The features of the first branch and the features of the second branch are input into the cross-stream fusion module for feature fusion of different dimensions to obtain the fused features of the first branch and the fused features of the second branch. The cross-stream fusion module has image information fusion capability. Therefore, the cross-stream fusion module can improve the generalization performance and detection performance of the dual-branch convolutional suppression texture network. The fused features in each dimension are flattened and spliced ​​to obtain a one-dimensional vector. The one-dimensional vector is input into the fully connected layer for face forgery detection to obtain the face forgery detection result. Face forgery detection by fusing features can improve the accuracy of face forgery detection.

[0016] According to some embodiments of the present invention, the texture suppression module convolution includes a texture suppression module, multiple bi-branch convolutions, and multiple attention modules. The step of inputting the feature map of the second branch into the texture suppression module convolution for feature extraction to obtain a first texture suppression feature map includes:

[0017] The feature map of the second branch is input into the texture suppression module in the convolution of the texture suppression module to obtain a texture suppression feature map;

[0018] The texture suppression feature map is input into a two-branch convolution to obtain the convolutional texture suppression feature map;

[0019] The texture suppression feature map is concatenated with the convolved texture suppression feature map to obtain a concatenated texture suppression feature map;

[0020] The stitched texture suppression feature map is input into the attention module to obtain the attention feature map;

[0021] The attention feature map is processed through multiple bi-branch convolutions, concatenation, and attention modules to obtain the first texture suppression feature map.

[0022] According to some embodiments of the present invention, the step of inputting the first feature map of the first branch into the texture suppression feature attention in the cross-stream interaction module for feature extraction to obtain a texture suppression attention feature map includes:

[0023] The first feature map of the first branch is input into two parallel max pooling and average pooling layers in the texture suppression feature attention to obtain the pooled feature map.

[0024] The pooled feature map is then converted into a single-channel feature map using a 7*7 convolution.

[0025] The single-channel feature map is activated by an activation function to obtain a texture suppression attention feature map.

[0026] According to some embodiments of the present invention, features are extracted from the first feature map of the first branch to obtain the features of the first branch; features are extracted from the first feature map of the second branch to obtain the features of the second branch, including:

[0027] Multiple separable convolutional layers are used to extract intermediate features of the first branch from the first feature map of the first branch;

[0028] High-level features of the first branch are extracted from the intermediate features of the first branch using convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers.

[0029] Multiple separable convolutional layers are used to extract intermediate features of the second branch from the first feature map of the second branch;

[0030] High-level features of the second branch are extracted from the intermediate features of the second branch using convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers.

[0031] According to some embodiments of the present invention, the step of inputting the features of the first branch and the features of the second branch into the cross-stream fusion module for feature fusion of different dimensions to obtain the fused features of the first branch and the fused features of the second branch includes:

[0032] The intermediate features of the first branch and the intermediate features of the second branch are adjusted into multiple intermediate feature maps in the width dimension and multiple intermediate feature maps in the height dimension through the convolutional layer in the cross-stream fusion module;

[0033] Multiple intermediate feature maps along the width dimension are stitched together to obtain an intermediate width dimension stitched feature map;

[0034] Multiple intermediate feature maps in the height dimension are stitched together to obtain an intermediate height dimension stitched feature map;

[0035] Based on the intermediate features of the first branch, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map, the intermediate fusion features of the first branch are calculated.

[0036] Based on the intermediate features of the second branch, the intermediate width dimension spliced ​​feature map, and the intermediate height dimension spliced ​​feature map, the intermediate fusion features of the second branch are calculated.

[0037] The high-level features of the first branch and the second branch are adjusted into multiple high-level feature maps in the width dimension and multiple high-level feature maps in the height dimension through the convolutional layer in the cross-stream fusion module.

[0038] Multiple high-level feature maps along the width dimension are concatenated to obtain a high-level width dimension concatenated feature map;

[0039] Multiple high-level feature maps in the height dimension are stitched together to obtain a high-level height dimension stitched feature map;

[0040] Based on the high-level features of the first branch, the high-level width dimension stitched feature map, and the high-level height dimension stitched feature map, the high-level fusion features of the first branch are calculated.

[0041] The advanced fusion features of the second branch are calculated based on the advanced features of the second branch, the advanced width dimension stitched feature map, and the advanced height dimension stitched feature map.

[0042] According to some embodiments of the present invention, calculating the intermediate fusion feature of the first branch based on the intermediate features of the first branch, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map includes:

[0043] Based on the intermediate features of the first branch, the intermediate width dimension concatenated feature map, and the intermediate height dimension concatenated feature map, the intermediate attention map of the first branch is calculated as follows:

[0044]

[0045] Based on the intermediate features of the first branch and the intermediate attention map of the first branch, the intermediate fusion features of the first branch are calculated as follows:

[0046] T' = T + A T

[0047] Among them, A T Let T represent the intermediate attention map of the first branch, and let T represent the intermediate feature of the first branch. V represents element-wise multiplication, K represents the intermediate width dimension spliced ​​feature map, and T' represents the intermediate height dimension spliced ​​feature map.

[0048] According to some embodiments of the present invention, calculating the intermediate fusion feature of the second branch based on the intermediate features of the second branch, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map includes:

[0049] Based on the intermediate features of the second branch, the intermediate width dimension concatenated feature map, and the intermediate height dimension concatenated feature map, the intermediate attention map of the second branch is calculated as follows:

[0050]

[0051] Based on the intermediate features of the second branch and the intermediate attention map of the second branch, the intermediate fusion features of the second branch are calculated as follows:

[0052] T R ′=T R +A R

[0053] Among them, A R T represents the intermediate attention map of the second branch. R This represents the intermediate feature of the second branch. This represents element-wise multiplication, where V represents the mid-level width dimension concatenated feature map, K represents the mid-level height dimension concatenated feature map, and T represents the element-wise multiplication. R 'Indicates the intermediate fusion feature of the second branch.

[0054] Secondly, embodiments of the present invention also provide a face forgery detection system based on a dual-branch convolutional suppression texture network, the face forgery detection system based on a dual-branch convolutional suppression texture network comprising:

[0055] The image acquisition module is used to acquire the face image to be predicted;

[0056] The network building module is used to construct a two-branch convolutional texture suppression network that includes a texture suppression module convolution, a cross-stream interaction module, and a cross-stream fusion module;

[0057] The image detection module is used to input the face image to be predicted into the dual-branch convolutional suppression texture network for face forgery detection, and obtain the face forgery detection result; the image detection module includes a first feature map acquisition submodule, a second feature map acquisition submodule, a third feature map acquisition submodule, a feature extraction submodule, a feature fusion submodule, and a face detection submodule, wherein:

[0058] The first feature map acquisition submodule is used to input the face image to be predicted into the texture suppression module convolution for feature extraction to obtain the first texture suppression feature map;

[0059] The second feature map acquisition submodule is used to pass the face image to be predicted through a convolutional layer to obtain the feature map of the second branch, and input the feature map of the second branch into the texture suppression module in the cross-flow interaction module for feature extraction to obtain the second texture suppression feature map. The first texture suppression feature map and the second texture suppression feature map are added to obtain the first feature map of the first branch of the dual-branch convolutional texture suppression network.

[0060] The third feature map acquisition submodule is used to input the first feature map of the first branch into the texture suppression feature attention in the cross-flow interaction module for feature extraction, to obtain the texture suppression attention feature map, and to perform element-wise multiplication of the texture suppression attention feature map and the feature map of the second branch to obtain the first feature map of the second branch of the dual-branch convolutional texture suppression network.

[0061] The feature extraction submodule is used to extract features from the first feature map of the first branch to obtain the features of the first branch; and to extract features from the first feature map of the second branch to obtain the features of the second branch.

[0062] The feature fusion submodule is used to input the features of the first branch and the features of the second branch into the cross-stream fusion module to perform feature fusion of different dimensions, so as to obtain the fused features of the first branch and the fused features of the second branch.

[0063] The face detection submodule is used to perform flattening and splicing operations on the fused features in each dimension to obtain a one-dimensional vector, and input the one-dimensional vector into a fully connected layer to perform face forgery detection to obtain the face forgery detection result.

[0064] Thirdly, embodiments of the present invention also provide a face forgery detection device based on a dual-branch convolutional suppression texture network, including at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, the instructions being executed by the at least one control processor to enable the at least one control processor to perform a face forgery detection method based on a dual-branch convolutional suppression texture network as described above.

[0065] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing computer-executable instructions for causing a computer to execute a face forgery detection method based on a dual-branch convolutional suppression texture network as described above.

[0066] It is understood that the beneficial effects of the second to fourth aspects compared with the related technologies are the same as the beneficial effects of the first aspect compared with the related technologies. Please refer to the relevant description in the first aspect above, which will not be repeated here. Attached Figure Description

[0067] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0068] Figure 1 This is a flowchart of a face forgery detection method based on a dual-branch convolutional suppression texture network according to an embodiment of the present invention;

[0069] Figure 2 yes Figure 1 The flowchart in step S300;

[0070] Figure 3 This is a schematic diagram of the structure of a dual-branch convolutional suppression texture network according to an embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram of the convolutional structure of a texture suppression module according to an embodiment of the present invention;

[0072] Figure 5 This is a schematic diagram of the cross-stream interaction module according to an embodiment of the present invention;

[0073] Figure 6 This is a schematic diagram of the cross-stream fusion module according to an embodiment of the present invention;

[0074] Figure 7 This is a structural diagram of a face forgery detection system based on a dual-branch convolutional suppression texture network according to an embodiment of the present invention;

[0075] Figure 8 yes Figure 7 The structural diagram of the sub-modules in the image detection module. Detailed Implementation

[0076] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0077] In the description of this invention, the use of terms such as "first," "second," etc., is for the purpose of distinguishing technical features only and should not be construed as indicating or implying relative importance, or implicitly indicating the number of technical features indicated, or implicitly indicating the order of the technical features indicated.

[0078] In the description of this invention, it should be understood that the orientation descriptions, such as up, down, etc., are based on the orientation or positional relationship shown in the drawings and are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0079] In the description of this invention, it should be noted that, unless otherwise explicitly defined, terms such as "setting," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0080] Deep learning-based face forgery detection has only emerged in recent years. While its detection results within a single dataset are already quite good, its performance across different datasets is less than satisfactory. This is partly because deep learning is still rapidly evolving, and older models have struggled with cross-dataset face forgery detection. Furthermore, face forgery datasets are constantly being updated, and forgery effects are becoming increasingly sophisticated. It is now difficult to rely on visually perceptible differences (e.g., significant color differences between the left and right eyes in a forged image, missing details and light reflections in the eye and teeth areas) for detection. Additionally, the different data distributions across datasets mean that features learned on the training dataset may not perform well on new datasets.

[0081] This invention constructs a dual-branch convolutional texture suppression network (DCRMR) comprising a texture suppression module, a cross-stream interaction module, and a cross-stream fusion module. The network inputs a face image to be predicted into the DCRMR for face forgery detection, obtaining the detection result. Using the DCRMR for face forgery detection improves the accuracy of face forgery detection. The DCRMR for face forgery detection includes: inputting the face image to be predicted into the texture suppression module's convolution for feature extraction to obtain a first texture suppression feature map. The texture suppression module's convolution improves feature analysis capabilities and has better texture suppression capabilities; passing the face image to be predicted through a convolutional layer to obtain a feature map for the second branch; inputting the second branch's feature map into the texture suppression module in the cross-stream interaction module for feature extraction to obtain a second texture suppression feature map; adding the first and second texture suppression feature maps to obtain a first feature map for the first branch of the DCRMR; and inputting the first feature map of the first branch into the texture suppression feature map in the cross-stream interaction module for feature extraction to obtain a texture suppression feature map. The textural suppression attention feature map is processed, and the textural suppression attention feature map and the feature map of the second branch are multiplied element-wise to obtain the first feature map of the second branch of the dual-branch convolutional suppression texture network. The cross-stream interaction module can guide the dual-branch convolutional suppression texture network to better extract more general features in face forgery. Features are extracted from the first feature map of the first branch to obtain the features of the first branch. Features are extracted from the first feature map of the second branch to obtain the features of the second branch. The features of the first branch and the features of the second branch are input into the cross-stream fusion module for feature fusion of different dimensions to obtain the fused features of the first branch and the fused features of the second branch. The cross-stream fusion module has image information fusion capability. Therefore, the cross-stream fusion module can improve the generalization performance and detection performance of the dual-branch convolutional suppression texture network. The fused features in each dimension are flattened and spliced ​​to obtain a one-dimensional vector. The one-dimensional vector is input into the fully connected layer for face forgery detection to obtain the face forgery detection result. Face forgery detection by fusing features can improve the accuracy of face forgery detection.

[0082] Reference Figures 1 to 2 This invention provides a face forgery detection method based on a dual-branch convolutional suppression texture network. This method includes, but is not limited to, steps S100 to S300 and steps S310 to S360, wherein:

[0083] Step S100: Obtain the face image to be predicted;

[0084] Step S200: Construct a dual-branch convolutional texture suppression network that includes a texture suppression module convolution, a cross-stream interaction module, and a cross-stream fusion module;

[0085] Step S300: Input the face image to be predicted into a dual-branch convolutional suppression texture network for face forgery detection, and obtain the face forgery detection result; wherein the dual-branch convolutional suppression texture network for face forgery detection includes:

[0086] Step S310: Input the face image to be predicted into the texture suppression module convolution for feature extraction to obtain the first texture suppression feature map;

[0087] Step S320: Pass the face image to be predicted through a convolutional layer to obtain the feature map of the second branch. Input the feature map of the second branch into the texture suppression module in the cross-flow interaction module for feature extraction to obtain the second texture suppression feature map. Add the first texture suppression feature map and the second texture suppression feature map to obtain the first feature map of the first branch of the dual-branch convolutional texture suppression network.

[0088] Step S330: Input the first feature map of the first branch into the texture suppression feature attention in the cross-flow interaction module for feature extraction to obtain the texture suppression attention feature map, and multiply the texture suppression attention feature map and the feature map of the second branch element by element to obtain the first feature map of the second branch of the dual-branch convolutional texture suppression network.

[0089] Step S340: Extract features from the first feature map of the first branch to obtain the features of the first branch; extract features from the first feature map of the second branch to obtain the features of the second branch;

[0090] Step S350: Input the features of the first branch and the features of the second branch into the cross-stream fusion module to perform feature fusion of different dimensions, and obtain the fused features of the first branch and the fused features of the second branch.

[0091] Step S360: Flatten and concatenate the fused features in each dimension to obtain a one-dimensional vector, and input the one-dimensional vector into the fully connected layer for face forgery detection to obtain the face forgery detection result.

[0092] In steps S100 to S300 and steps S310 to S360 of some embodiments, to improve the accuracy of face forgery detection, this embodiment acquires the face image to be predicted, constructs a dual-branch convolutional texture suppression network including a texture suppression module convolution, a cross-stream interaction module, and a cross-stream fusion module, and inputs the face image to be predicted into the dual-branch convolutional texture suppression network for face forgery detection to obtain the face forgery detection result; to improve feature analysis capability and have better image texture suppression capability, this embodiment inputs the face image to be predicted into the texture suppression module convolution for feature extraction to obtain a first texture suppression feature map; to extract more general features in face forgery, this embodiment passes the face image to be predicted through a convolutional layer to obtain a second branch feature map, inputs the second branch feature map into the texture suppression module in the cross-stream interaction module for feature extraction to obtain a second texture suppression feature map, and adds the first texture suppression feature map and the second texture suppression feature map to obtain a dual-branch convolutional texture suppression feature map. The first feature map of the first branch of the dual-branch convolutional suppression texture network is input into the texture suppression feature attention module in the cross-stream interaction module for feature extraction, resulting in a texture suppression attention feature map. The texture suppression attention feature map and the feature map of the second branch are then element-wise multiplied to obtain the first feature map of the second branch of the dual-branch convolutional suppression texture network. To improve generalization performance and detection capability, this embodiment extracts features from the first feature map of the first branch to obtain the features of the first branch. Features are also extracted from the first feature map of the second branch to obtain the features of the second branch. The features of the first and second branches are then input into the cross-stream fusion module for feature fusion across different dimensions to obtain the fused features of the first and second branches. To improve the accuracy of face forgery detection, this embodiment performs flattening and concatenation operations on the fused features in each dimension to obtain a one-dimensional vector. This one-dimensional vector is then input into a fully connected layer for face forgery detection to obtain the face forgery detection result.

[0093] In some embodiments, the texture suppression module convolution includes a texture suppression module, multiple bi-branch convolutions, and multiple attention modules. The feature map of the second branch is input into the texture suppression module convolution for feature extraction to obtain a first texture suppression feature map, including:

[0094] The feature map of the second branch is input into the texture suppression module in the texture suppression module convolution to obtain the texture suppression feature map;

[0095] The texture suppression feature map is input into a two-branch convolution to obtain the convolutional texture suppression feature map;

[0096] The texture suppression feature map is concatenated with the convolved texture suppression feature map to obtain the concatenated texture suppression feature map;

[0097] The stitched texture suppression feature map is input into the attention module to obtain the attention feature map;

[0098] The attention feature map is processed through multiple bi-branch convolutions, concatenation, and attention modules to obtain the first texture suppression feature map.

[0099] In this embodiment, the dual-branch convolutional structure can significantly improve the feature analysis capability and representation capability of the dual-branch convolutional texture suppression network, enabling it to have better image texture suppression capabilities. By using an attention module, the dual-branch convolutional texture suppression network can focus more on forged regions.

[0100] In some embodiments, the first feature map of the first branch is input into the texture suppression feature attention in the cross-stream interaction module for feature extraction to obtain a texture suppression attention feature map, including:

[0101] The first feature map of the first branch is input into two parallel max pooling and average pooling layers in the texture suppression feature attention to obtain the pooled feature map.

[0102] The pooled feature map is adjusted into a single-channel feature map through a 7*7 convolution;

[0103] The single-channel feature map is activated by an activation function to obtain a texture suppression attention feature map.

[0104] In this embodiment, a texture suppression attention feature map is obtained through texture suppression feature attention, which can be used to guide the extraction of forgery clues.

[0105] In some embodiments, features are extracted from the first feature map of the first branch to obtain the features of the first branch; features are extracted from the first feature map of the second branch to obtain the features of the second branch, including:

[0106] Multiple separable convolutional layers are used to extract intermediate features of the first branch from the first feature map of the first branch;

[0107] High-level features of the first branch are extracted from the intermediate features of the first branch using convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers.

[0108] Multiple separable convolutional layers are used to extract intermediate features of the second branch from the first feature map of the second branch;

[0109] High-level features of the second branch are extracted from the intermediate features of the second branch using convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers.

[0110] In some embodiments, the features of the first branch and the features of the second branch are input into the cross-stream fusion module for feature fusion of different dimensions to obtain the fused features of the first branch and the fused features of the second branch, including:

[0111] The intermediate features of the first branch and the intermediate features of the second branch are adjusted into multiple intermediate feature maps in the width dimension and multiple intermediate feature maps in the height dimension through the convolutional layer in the cross-stream fusion module;

[0112] Multiple intermediate feature maps in the width dimension are concatenated to obtain an intermediate width dimension concatenated feature map;

[0113] Multiple intermediate feature maps in the height dimension are stitched together to obtain an intermediate height dimension stitched feature map;

[0114] Based on the intermediate features, the intermediate width dimension spliced ​​feature map, and the intermediate height dimension spliced ​​feature map of the first branch, the intermediate fusion features of the first branch are calculated.

[0115] Based on the intermediate features, the intermediate width dimension spliced ​​feature map, and the intermediate height dimension spliced ​​feature map of the second branch, the intermediate fusion features of the second branch are calculated.

[0116] The high-level features of the first and second branches are adjusted into multiple high-level feature maps in the width dimension and multiple high-level feature maps in the height dimension through the convolutional layer in the cross-stream fusion module.

[0117] Multiple high-level feature maps in the width dimension are concatenated to obtain a high-level width-dimensional concatenated feature map;

[0118] Multiple high-level feature maps in the height dimension are concatenated to obtain a high-level height dimension concatenated feature map;

[0119] Based on the high-level features, high-level width dimension stitched feature map, and high-level height dimension stitched feature map of the first branch, the high-level fusion features of the first branch are calculated.

[0120] Based on the high-level features, high-level width dimension stitched feature map, and high-level height dimension stitched feature map of the second branch, the high-level fusion features of the second branch are calculated.

[0121] In this embodiment, feature fusion is performed by extracting intermediate and high-level features, and detection is performed by fusing the features, which can improve the accuracy of face detection.

[0122] In some embodiments, the intermediate fusion features of the first branch are calculated based on the intermediate features, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map of the first branch, including:

[0123] Based on the intermediate features, the concatenated feature map of the intermediate width dimension, and the concatenated feature map of the intermediate height dimension of the first branch, the intermediate attention map of the first branch is calculated as follows:

[0124]

[0125] Based on the intermediate features of the first branch and the intermediate attention map of the first branch, the intermediate fusion features of the first branch are calculated as follows:

[0126] T' = T + A T

[0127] Among them, A T Let T represent the intermediate attention map of the first branch, where T represents the intermediate features of the first branch. denoted as element-wise multiplication, V represents the mid-level width dimension concatenated feature map, K represents the mid-level height dimension concatenated feature map, and T' represents the mid-level fusion feature of the first branch.

[0128] In some embodiments, the intermediate fusion features of the second branch are calculated based on the intermediate features, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map of the second branch, including:

[0129] Based on the intermediate features, the concatenated feature map of the intermediate width dimension, and the concatenated feature map of the intermediate height dimension of the second branch, the intermediate attention map of the second branch is calculated as follows:

[0130]

[0131] Based on the intermediate features of the second branch and the intermediate attention map of the second branch, the intermediate fusion features of the second branch are calculated as follows:

[0132] T R ′=T R +A R

[0133] Among them, A R T represents the intermediate attention map of the second branch. R This represents the intermediate features of the second branch. This represents element-wise multiplication, where V represents the concatenated feature map in the intermediate width dimension, K represents the concatenated feature map in the intermediate height dimension, and T represents the element-wise multiplication. R 'Indicates the intermediate fusion characteristics of the second branch.

[0134] To facilitate understanding by those skilled in the art, a set of preferred embodiments is provided below:

[0135] Reference Figure 3This embodiment utilizes the learning and representation capabilities of CNNs to adaptively suppress image texture and extract traces of forgery. Then, the obtained image with suppressed texture and the original image are input into the backbone network respectively to establish a dual-branch convolutional texture suppression network. To fully utilize the image with suppressed texture, this embodiment creates two modules: a cross-stream interaction module and a cross-stream fusion block. The specific process includes:

[0136] Reference Figure 4 To enhance the representational and analytical capabilities of the two-branch convolutional texture suppression network, this embodiment employs a two-branch structure. The texture suppression module (TSM) first convolves the input image with both 3x3 and 1x1 convolutional kernels to obtain two feature maps. Then, it subtracts the original image element-wise from the obtained feature maps to obtain two texture suppression feature maps. Finally, these two texture suppression feature maps are summed to obtain the texture suppression feature map F (to enhance the texture information representation capability of the two-branch convolutional texture suppression network). The TbCon (Two-branch Convolution) structure is a two-branch structure. After convolving the input texture suppression feature map with both 3x3 and 1x1 convolutional kernels to obtain two feature maps, these two feature maps are element-wise summed and activated using the ReLU activation function to obtain the convolved texture suppression feature map.

[0137] like Figure 4 As shown, in this embodiment, TbCon is applied to the texture suppression feature map F obtained by TSM to obtain the convolved texture suppression feature map F1. Then, F and F1 are concatenated to obtain the concatenated texture suppression feature map. This indicates a concatenation operation. The resulting concatenated texture suppression feature map is then input into the attention module (CBAM) to obtain an attention feature map. The attention module makes the bi-branch convolutional texture suppression network focus more on fake regions. Finally, the obtained attention feature map is repeated twice, involving TbCon, concatenation, and the attention module, to obtain a texture suppression feature map that suppresses texture information. The final result is as follows: Figure 3 As shown in the TSM stream (i.e., the first branch), the resulting texture suppression feature map, which suppresses texture information, is then convolved again to obtain the first texture suppression feature map F of the TSM stream. T .

[0138] like Figure 5 As shown, the cross-stream interaction module designed in this embodiment is mainly implemented through a texture suppression module (TSM) and texture suppression feature attention (TSFA). Specifically:

[0139] In the cross-stream interaction module, TSM is used to enrich the information of the feature maps with inhibitory textures. Given an RGB face image to be predicted, a convolutional layer is used to obtain the RGB stream feature map (i.e., the feature map of the second branch) F. R The feature map F of the RGB stream is obtained using TSM. R Convert to a feature map with suppressed texture (i.e., a second texture suppression feature map) F RT Then, F RT The first texture suppression feature map F of the TSM stream T Element-wise addition yields the first feature map F of the first branch of the dual-branch convolutional suppression texture network. E Due to F RT and F T Coming from different sources and different operations, they include different information, therefore, enhanced F can be obtained. T The first feature map F of the feature information E .

[0140] Utilizing F in Texture Suppression Feature Attention (TSFA) E To guide the RGB stream, first set F E The input is fed into two parallel max-pooling and average-pooling layers, then adjusted to a single-channel feature map by a 7x7 convolution, and finally the texture suppression attention feature F is obtained through an activation function. EA Finally, F EA Feature map F of RGB stream R Element-wise multiplication is performed to obtain the first feature map of the second branch of the bi-branch convolutional suppression texture network. By using F... EA Feature map F of RGB stream R Element-wise multiplication is used to guide the extraction of forgery clues from the RGB stream.

[0141] The cross-stream interaction module is mainly used to exchange and interact on the low-level features of two streams, thereby enriching the information of the two streams.

[0142] As mentioned above, this embodiment considers TSM streams and RGB streams, although different, to be complementary. Therefore, to further fuse the features of TSM and RGB streams, this embodiment proposes a cross-stream fusion module to fuse mid-to-high-level features of TSM and RGB streams. The specific process is as follows:

[0143] Reference Figure 3The method employs multiple separable convolutional layers to extract intermediate features of the TSM stream from the first feature map; it employs convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers to extract high-level features of the TSM stream from the intermediate features; it employs multiple separable convolutional layers to extract intermediate features of the RGB stream from the first feature map; and it employs convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers to extract high-level features of the RGB stream from the intermediate features.

[0144] Reference Figure 6 The intermediate features T of the input TSM stream and the intermediate features T of the RGB stream are combined. R The features are concatenated along the channel (C) dimension and reshaped by convolution into four feature maps V1, V2, K1, and K2. Then, V1 and V2 are concatenated along the width dimension, and K1 and K2 are concatenated along the height dimension, producing two components V and K. Component V represents the correlation between two different domains along the width dimension, while component K represents the correlation between two different domains along the height dimension. Then, the following formula is used: and Obtain the intermediate attention map A used to guide the TSM flow. T And intermediate attention map A guiding RGB flow R Finally, through T' = T + A T The intermediate fusion feature T' of the output TSM stream, through T R '=T R +A R Intermediate blending feature T of the output RGB stream R '. T' and T R It not only possesses the characteristics unique to its own flow, but also includes the characteristics of another flow, thus exhibiting richness and complementarity.

[0145] By repeating the above process of fusing intermediate features of the TSM stream and intermediate features of the RGB stream, the same operation is performed on the advanced features of the TSM stream and the advanced features of the RGB stream to obtain advanced fused features of the TSM stream and the advanced fused features of the RGB stream.

[0146] Then, the intermediate fusion features of the TSM and RGB streams and the advanced fusion features of the TSM and RGB streams are flattened into one-dimensional vectors (2048 values). The two one-dimensional vectors are then concatenated (CAT) to obtain a one-dimensional vector (4096 values). Finally, this one-dimensional vector is input into a fully connected layer for prediction to determine whether the input face image to be predicted is a fake image.

[0147] To better illustrate this, the following experiment was conducted in this embodiment:

[0148] 1. Experimental environment of the embodiment.

[0149] The dual-branch convolutional suppression texture network in this embodiment will be trained on three NVIDIA GeForce GTX TITAN X GPU platforms with 12GB of memory.

[0150] This embodiment uses three datasets in its experiments: Faceforensics++ (FF++), CelebDF, and DeeperForensics-1.0 (DF1.0). The FF++ dataset, proposed by Rossler et al. in "Faceforensics++: Learning to detect manipulated facial images," includes four forgery methods: Deepfakes (DF), Face2Face (F2F), FaceSwap (FS), and NeuralTextures (NT). CelebDF was proposed by Li et al. in "Celeb-df: A large-scale challenging dataset for deepfakeforensics," and DF1.0 was proposed by Jiang et al. in "Deeperforensics-1.0: A large-scale dataset for real-world face forgery detection." This embodiment uses the FF++ dataset as the training set for training and then tests on the CelebDF and DF1.0 datasets.

[0151] 2. Experimental results of the examples.

[0152] Table 1 shows a comparison of the generalization performance of the technical solution of this embodiment with other state-of-the-art methods on the Celeb-DF dataset, with Area Under Curve (AUC) as the benchmark. The best results are indicated in bold. These models were trained on the FF++ dataset and tested on CelebDF. The technical solution of this embodiment outperforms other methods on CelebDF and performs well on FF++. This demonstrates the effectiveness and superior generalization ability of the technical solution of this embodiment.

[0153] Table 1

[0154]

[0155]

[0156] Table 2 shows a comparison of the generalization performance of the proposed solution with other state-of-the-art methods on the DF1.0 dataset, using Area Under Curve (AUC) as the benchmark. The best results are indicated in bold. These models were trained on four different forgery methods on the FF++ dataset and tested on the DF1.0 dataset. The proposed solution performs best on the Deepfake and Face2Face subsets. More importantly, the proposed solution has the best average result, demonstrating its effectiveness and superior generalization ability.

[0157] Table 2

[0158]

[0159] Reference Figures 7 to 8 This invention also provides a face forgery detection system based on a dual-branch convolutional suppression texture network. This system includes an image acquisition module 100, a network construction module 200, and an image detection module 300, wherein:

[0160] Image acquisition module 100 is used to acquire the face image to be predicted;

[0161] Network building module 200 is used to build a two-branch convolutional texture suppression network that includes a texture suppression module convolution, a cross-stream interaction module, and a cross-stream fusion module;

[0162] Image detection module 300 is used to input the face image to be predicted into a dual-branch convolutional suppression texture network for face forgery detection and obtain the face forgery detection result. The image detection module includes a first feature map acquisition submodule 310, a second feature map acquisition submodule 320, a third feature map acquisition submodule 330, a feature extraction submodule 340, a feature fusion submodule 350, and a face detection submodule 360, wherein:

[0163] The first feature map acquisition submodule 310 is used to input the face image to be predicted into the texture suppression module convolution for feature extraction and to obtain the first texture suppression feature map.

[0164] The second feature map acquisition submodule 320 is used to pass the face image to be predicted through a convolutional layer to obtain the feature map of the second branch, input the feature map of the second branch into the texture suppression module in the cross-flow interaction module for feature extraction to obtain the second texture suppression feature map, and add the first texture suppression feature map and the second texture suppression feature map to obtain the first feature map of the first branch of the dual-branch convolutional texture suppression network.

[0165] The third feature map acquisition submodule 330 is used to input the first feature map of the first branch into the texture suppression feature attention in the cross-flow interaction module for feature extraction, to obtain the texture suppression attention feature map, and to multiply the texture suppression attention feature map and the feature map of the second branch by element-wise multiplication to obtain the first feature map of the second branch of the dual-branch convolutional texture suppression network.

[0166] The feature extraction submodule 340 is used to extract features from the first feature map of the first branch to obtain the features of the first branch; and to extract features from the first feature map of the second branch to obtain the features of the second branch.

[0167] The feature fusion submodule 350 is used to input the features of the first branch and the features of the second branch into the cross-stream fusion module to perform feature fusion of different dimensions, so as to obtain the fused features of the first branch and the fused features of the second branch.

[0168] The face detection submodule 360 ​​is used to flatten and stitch the fused features in each dimension to obtain a one-dimensional vector, and then input the one-dimensional vector into the fully connected layer to perform face forgery detection and obtain the face forgery detection result.

[0169] It should be noted that since the face forgery detection system based on a dual-branch convolutional suppression texture network in this embodiment is based on the same inventive concept as the face forgery detection method based on a dual-branch convolutional suppression texture network described above, the corresponding content in the method embodiment is also applicable to this system embodiment, and will not be described in detail here.

[0170] This invention also provides a face forgery detection device based on a dual-branch convolutional suppression texture network, comprising: at least one control processor and a memory for communicatively connecting to the at least one control processor.

[0171] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0172] The non-transient software program and instructions required to implement the face forgery detection method based on a dual-branch convolutional suppression texture network in the above embodiments are stored in memory. When executed by a processor, the face forgery detection method based on a dual-branch convolutional suppression texture network in the above embodiments is executed. For example, the method described above is executed. Figure 1 Method steps S100 to S300 and Figure 2 Method steps S310 to S360.

[0173] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0174] This invention also provides a computer-readable storage medium storing computer-executable instructions. These instructions are executed by one or more control processors, causing the processors to perform a face forgery detection method based on a dual-branch convolutional suppression texture network as described in the above-described method embodiments. For example, the method described above... Figure 1 The functions of steps S100 to S300 in the method and Figure 2 The functions of steps S310 to S360 in the method.

[0175] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0176] The above is a detailed description of the preferred embodiments of this application. However, the embodiments of this application are not limited to the above-described implementation methods. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the embodiments of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of the embodiments of this application.

Claims

1. A face forgery detection method based on a dual-branch convolutional suppression texture network, characterized in that, The face forgery detection method based on a dual-branch convolutional suppression texture network includes: Obtain the image of the face to be predicted; Construct a two-branch convolutional texture suppression network that includes a texture suppression module convolution, a cross-stream interaction module, and a cross-stream fusion module; The image of the face to be predicted is input into the dual-branch convolutional suppression texture network for face forgery detection to obtain the face forgery detection result; wherein the dual-branch convolutional suppression texture network for face forgery detection includes: The face image to be predicted is input into the texture suppression module convolution for feature extraction to obtain the first texture suppression feature map; The face image to be predicted is passed through a convolutional layer to obtain the feature map of the second branch. The feature map of the second branch is then input into the texture suppression module in the cross-flow interaction module for feature extraction to obtain the second texture suppression feature map. The first texture suppression feature map and the second texture suppression feature map are then added to obtain the first feature map of the first branch of the dual-branch convolutional texture suppression network. The first feature map of the first branch is input into the texture suppression feature attention in the cross-flow interaction module for feature extraction to obtain the texture suppression attention feature map. The texture suppression attention feature map and the feature map of the second branch are then multiplied element-wise to obtain the first feature map of the second branch of the dual-branch convolutional texture suppression network. Features are extracted from the first feature map of the first branch to obtain the features of the first branch; features are extracted from the first feature map of the second branch to obtain the features of the second branch. The features of the first branch and the features of the second branch are input into the cross-stream fusion module for feature fusion of different dimensions to obtain the fused features of the first branch and the fused features of the second branch. The fused features in each dimension are flattened and spliced ​​to obtain a one-dimensional vector, which is then input into a fully connected layer for face forgery detection to obtain the face forgery detection result.

2. The face forgery detection method based on a dual-branch convolutional suppression texture network according to claim 1, characterized in that, The texture suppression module convolution includes a texture suppression module, multiple bi-branch convolutions, and multiple attention modules. The step of inputting the feature map of the second branch into the texture suppression module convolution for feature extraction to obtain a first texture suppression feature map includes: The feature map of the second branch is input into the texture suppression module in the convolution of the texture suppression module to obtain a texture suppression feature map; The texture suppression feature map is input into a two-branch convolution to obtain the convolutional texture suppression feature map; The texture suppression feature map is concatenated with the convolved texture suppression feature map to obtain a concatenated texture suppression feature map; The stitched texture suppression feature map is input into the attention module to obtain the attention feature map; The attention feature map is processed through multiple bi-branch convolutions, concatenation, and attention modules to obtain the first texture suppression feature map.

3. The face forgery detection method based on a dual-branch convolutional suppression texture network according to claim 1, characterized in that, The step of inputting the first feature map of the first branch into the texture suppression feature attention in the cross-stream interaction module for feature extraction to obtain a texture suppression attention feature map includes: The first feature map of the first branch is input into two parallel max pooling and average pooling layers in the texture suppression feature attention to obtain the pooled feature map. The pooled feature map is then converted into a single-channel feature map using a 7*7 convolution. The single-channel feature map is activated by an activation function to obtain a texture suppression attention feature map.

4. The face forgery detection method based on a dual-branch convolutional suppression texture network according to claim 1, characterized in that, Extract features from the first feature map of the first branch to obtain the features of the first branch; Features are extracted from the first feature map of the second branch to obtain the features of the second branch, including: Multiple separable convolutional layers are used to extract intermediate features of the first branch from the first feature map of the first branch; High-level features of the first branch are extracted from the intermediate features of the first branch using convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers. Multiple separable convolutional layers are used to extract intermediate features of the second branch from the first feature map of the second branch; High-level features of the second branch are extracted from the intermediate features of the second branch using convolutional layers, multiple separable convolutional layers, max pooling layers, and global average pooling layers.

5. The face forgery detection method based on a dual-branch convolutional suppression texture network according to claim 4, characterized in that, The step of inputting the features of the first branch and the features of the second branch into the cross-stream fusion module for feature fusion of different dimensions to obtain the fused features of the first branch and the fused features of the second branch includes: The intermediate features of the first branch and the intermediate features of the second branch are adjusted into multiple intermediate feature maps in the width dimension and multiple intermediate feature maps in the height dimension through the convolutional layer in the cross-stream fusion module; Multiple intermediate feature maps along the width dimension are stitched together to obtain an intermediate width dimension stitched feature map; Multiple intermediate feature maps in the height dimension are stitched together to obtain an intermediate height dimension stitched feature map; Based on the intermediate features of the first branch, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map, the intermediate fusion features of the first branch are calculated. Based on the intermediate features of the second branch, the intermediate width dimension spliced ​​feature map, and the intermediate height dimension spliced ​​feature map, the intermediate fusion features of the second branch are calculated. The high-level features of the first branch and the second branch are adjusted into multiple high-level feature maps in the width dimension and multiple high-level feature maps in the height dimension through the convolutional layer in the cross-stream fusion module. Multiple high-level feature maps along the width dimension are concatenated to obtain a high-level width dimension concatenated feature map; Multiple high-level feature maps in the height dimension are stitched together to obtain a high-level height dimension stitched feature map; Based on the high-level features of the first branch, the high-level width dimension stitched feature map, and the high-level height dimension stitched feature map, the high-level fusion features of the first branch are calculated. The advanced fusion features of the second branch are calculated based on the advanced features of the second branch, the advanced width dimension stitched feature map, and the advanced height dimension stitched feature map.

6. The face forgery detection method based on a dual-branch convolutional suppression texture network according to claim 5, characterized in that, The step of calculating the intermediate fusion feature of the first branch based on the intermediate features of the first branch, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map includes: Based on the intermediate features of the first branch, the intermediate width dimension concatenated feature map, and the intermediate height dimension concatenated feature map, the intermediate attention map of the first branch is calculated as follows: Based on the intermediate features of the first branch and the intermediate attention map of the first branch, the intermediate fusion features of the first branch are calculated as follows: T'=T+A T Among them, A T Let T represent the intermediate attention map of the first branch, and let T represent the intermediate feature of the first branch. V represents element-wise multiplication, K represents the intermediate width dimension spliced ​​feature map, and T' represents the intermediate height dimension spliced ​​feature map.

7. The face forgery detection method based on a dual-branch convolutional suppression texture network according to claim 5, characterized in that, The step of calculating the intermediate fusion feature of the second branch based on the intermediate features of the second branch, the intermediate width dimension stitched feature map, and the intermediate height dimension stitched feature map includes: Based on the intermediate features of the second branch, the intermediate width dimension concatenated feature map, and the intermediate height dimension concatenated feature map, the intermediate attention map of the second branch is calculated as follows: Based on the intermediate features of the second branch and the intermediate attention map of the second branch, the intermediate fusion features of the second branch are calculated as follows: T R ′=T R +A R Among them, A R T represents the intermediate attention map of the second branch. R This represents the intermediate feature of the second branch. This represents element-wise multiplication, where V represents the mid-level width dimension concatenated feature map, K represents the mid-level height dimension concatenated feature map, and T represents the element-wise multiplication. R 'Indicates the intermediate fusion feature of the second branch.

8. A face forgery detection system based on a dual-branch convolutional suppression texture network, characterized in that, The face forgery detection system based on a dual-branch convolutional suppression texture network includes: The image acquisition module is used to acquire the face image to be predicted; The network building module is used to construct a two-branch convolutional texture suppression network that includes a texture suppression module convolution, a cross-stream interaction module, and a cross-stream fusion module; The image detection module is used to input the face image to be predicted into the dual-branch convolutional suppression texture network for face forgery detection, and obtain the face forgery detection result; the image detection module includes a first feature map acquisition submodule, a second feature map acquisition submodule, a third feature map acquisition submodule, a feature extraction submodule, a feature fusion submodule, and a face detection submodule, wherein: The first feature map acquisition submodule is used to input the face image to be predicted into the texture suppression module convolution for feature extraction to obtain the first texture suppression feature map; The second feature map acquisition submodule is used to pass the face image to be predicted through a convolutional layer to obtain the feature map of the second branch, input the feature map of the second branch to the texture suppression module in the cross-flow interaction module for feature extraction to obtain the second texture suppression feature map, and add the first texture suppression feature map and the second texture suppression feature map to obtain the first feature map of the first branch of the dual-branch convolutional texture suppression network. The third feature map acquisition submodule is used to input the first feature map of the first branch into the texture suppression feature attention in the cross-flow interaction module for feature extraction, to obtain the texture suppression attention feature map, and to perform element-wise multiplication of the texture suppression attention feature map and the feature map of the second branch to obtain the first feature map of the second branch of the dual-branch convolutional texture suppression network. The feature extraction submodule is used to extract features from the first feature map of the first branch to obtain the features of the first branch; and to extract features from the first feature map of the second branch to obtain the features of the second branch. The feature fusion submodule is used to input the features of the first branch and the features of the second branch into the cross-stream fusion module to perform feature fusion of different dimensions, so as to obtain the fused features of the first branch and the fused features of the second branch. The face detection submodule is used to perform flattening and splicing operations on the fused features in each dimension to obtain a one-dimensional vector, and input the one-dimensional vector into a fully connected layer to perform face forgery detection to obtain the face forgery detection result.

9. A face forgery detection device based on a dual-branch convolutional suppression texture network, characterized in that, The method includes at least one control processor and a memory for communicatively connecting to the at least one control processor; the memory stores instructions executable by the at least one control processor, which, when executed by the at least one control processor, enable the at least one control processor to perform the face forgery detection method based on a dual-branch convolutional suppression texture network as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for causing a computer to perform the face forgery detection method based on a bi-branch convolutional suppression texture network as described in any one of claims 1 to 7.

Images