Face silent living body detection method and device, storage medium and equipment

Abstract

The application discloses a face silent living body detection method and device, a storage medium and equipment, and belongs to the face recognition field. The method comprises the following steps: acquiring a face image to be detected comprising a face region and a background region and inputting the face image to a deep network model; the deep network model mines local features and global features of the face image to be detected through a first operation with a local receptive field and a second operation with a global receptive field, fuses the local features and the global features, and judges whether the face image to be detected is from a living body according to the fused features. The second operation with the global receptive field is used for the living body detection task of the deep network model, the second operation has the global receptive field, so that the deep network model captures global information of the face image to be detected under a larger receptive field. In combination with the first operation with the local receptive field, the local information and the global information obtained through the first operation and the second operation are fused, which is more conducive to the living body detection and improves the accuracy of the living body detection.

Description

Technical Field

[0001] This invention relates to the field of face recognition, and in particular to a method, apparatus, storage medium, and device for silent face liveness detection. Background Technology

[0002] Currently, facial recognition technology is widely used in fields such as finance and security. However, due to the advantages of convenient and contactless facial data acquisition, it is also easily exploited by others, such as by using photos or videos to breach facial recognition systems. Therefore, facial liveness detection technology, as the first hurdle in facial recognition technology, is particularly important.

[0003] Common face liveness detection technologies mainly fall into two categories: cooperative liveness detection and silent liveness detection. Cooperative liveness detection requires a high degree of user cooperation and involves only a single action, making it easily vulnerable to attack. Silent liveness detection, on the other hand, does not require user cooperation and determines whether a face is alive by distinguishing the differences between real and fake face images. Therefore, silent liveness detection is more widely accepted by users.

[0004] Deep learning is a commonly used method for silent face liveness detection. Based on a large amount of data, it automatically mines features to distinguish between real and fake faces through deep networks such as ResNet and DenseNet, and finally uses SoftMax to classify real and fake faces.

[0005] Deep learning uses multiple layers of convolutional kernels to gradually extract data features from shallow to deep. However, due to the limitations of convolutional kernels, they have only a small receptive field for the data and can only focus on the local differences between real and fake faces. They cannot focus on the global geometric relationships and differences of faces, making it difficult to improve the face liveness detection capabilities of deep networks. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method, apparatus, storage medium, and device for silent face liveness detection, thereby improving the accuracy of liveness detection.

[0007] The technical solution provided by this invention is as follows:

[0008] In a first aspect, the present invention provides a method for silent face liveness detection, the method comprising:

[0009] The face image to be detected is acquired and input into a trained deep network model. The face image to be detected includes a face region and a background region.

[0010] The deep network model mines local features of the face image to be detected through a first operation with a local receptive field, mines global features of the face image to be detected through a second operation with a global receptive field, and fuses the mined local features and global features to determine whether the face image to be detected is from a live subject based on the fused features.

[0011] Furthermore, the deep network model includes at least one operation block, pooling operation, fully connected operation, and Softmax operation, and each operation block includes at least one feature extraction module;

[0012] The deep network model mines local features from its input through the first operation performed in the feature extraction module, mines global features from its input through the second operation performed in the feature extraction module, and outputs the mined local features and global features by fusing them together. The output of each operation block is fused with the output of the previous operation block by the feature fusion module, so that the outputs of all operation blocks are fused to obtain the multi-scale fused features.

[0013] The fused features are sequentially processed through the pooling operation, the fully connected operation, and the Softmax operation to obtain the liveness score and spoofing score of the face image to be detected; the liveness score and spoofing score are used to determine whether the face image to be detected is from a live person.

[0014] Furthermore, the feature extraction module includes a local branch and a global branch. The local branch includes the first operation, and the global branch includes the second operation. A portion of the input channels of the feature extraction module are processed by the local branch to obtain the local features, and another portion of the input channels of the feature extraction module are processed by the global branch to obtain the global features.

[0015] Furthermore, the first operation includes a convolution operation, and the second operation includes a Fourier transform operation.

[0016] Furthermore, the local branch includes a first convolution module and a second convolution module, and the global branch includes a third convolution module and a Fourier spectrum transform module;

[0017] The results of processing a portion of the channels by the first convolution module are fused with the results of processing another portion of the channels by the third convolution module to obtain the local features; the results of processing a portion of the channels by the second convolution module are fused with the results of processing another portion of the channels by the Fourier spectrum transform module to obtain the global features.

[0018] Furthermore, the Fourier spectrum transform module sequentially includes a first combined layer, a Fourier transform layer, a second combined layer, an inverse Fourier transform layer, and a convolutional layer. Both the first combined layer and the second combined layer include convolution operations, BN operations, and activation operations.

[0019] Furthermore, the feature fusion module sequentially includes a first convolutional unit, a second convolutional unit, a BN unit, an activation unit, and a third convolutional unit.

[0020] Furthermore, all feature extraction modules of each operation block form a residual structure through shortcut links, the Fourier transform layer, the second combined layer, and the inverse Fourier transform layer of the Fourier spectrum transform module form a residual structure through shortcut links, and the second convolutional unit, BN unit, activation unit, and third convolutional unit of the feature fusion module form a residual structure through shortcut links.

[0021] Secondly, the present invention provides a silent face liveness detection device, the device comprising:

[0022] The data acquisition module is used to acquire the face image to be detected and input it into the trained deep network model. The face image to be detected includes a face region and a background region.

[0023] The inference module involves the deep network model mining local features of the face image to be detected through a first operation with a local receptive field, mining global features of the face image to be detected through a second operation with a global receptive field, fusing the mined local features and global features, and determining whether the face image to be detected is from a live subject based on the fused features.

[0024] Furthermore, the deep network model includes at least one operation block, pooling operation, fully connected operation, and Softmax operation, and each operation block includes at least one feature extraction module;

[0025] The deep network model mines local features from its input through the first operation performed in the feature extraction module, mines global features from its input through the second operation performed in the feature extraction module, and outputs the mined local features and global features by fusing them together. The output of each operation block is fused with the output of the previous operation block by the feature fusion module, so that the outputs of all operation blocks are fused to obtain the multi-scale fused features.

[0026] The fused features are sequentially processed through the pooling operation, the fully connected operation, and the Softmax operation to obtain the liveness score and spoofing score of the face image to be detected; the liveness score and spoofing score are used to determine whether the face image to be detected is from a live person.

[0027] Furthermore, the feature extraction module includes a local branch and a global branch. The local branch includes the first operation, and the global branch includes the second operation. A portion of the input channels of the feature extraction module are processed by the local branch to obtain the local features, and another portion of the input channels of the feature extraction module are processed by the global branch to obtain the global features.

[0028] Furthermore, the first operation includes a convolution operation, and the second operation includes a Fourier transform operation.

[0029] Furthermore, the local branch includes a first convolution module and a second convolution module, and the global branch includes a third convolution module and a Fourier spectrum transform module;

[0030] The results of processing a portion of the channels by the first convolution module are fused with the results of processing another portion of the channels by the third convolution module to obtain the local features; the results of processing a portion of the channels by the second convolution module are fused with the results of processing another portion of the channels by the Fourier spectrum transform module to obtain the global features.

[0031] Furthermore, the Fourier spectrum transform module sequentially includes a first combined layer, a Fourier transform layer, a second combined layer, an inverse Fourier transform layer, and a convolutional layer. Both the first combined layer and the second combined layer include convolution operations, BN operations, and activation operations.

[0032] Furthermore, the feature fusion module sequentially includes a first convolutional unit, a second convolutional unit, a BN unit, an activation unit, and a third convolutional unit.

[0033] Furthermore, all feature extraction modules of each operation block form a residual structure through shortcut links, the Fourier transform layer, the second combined layer, and the inverse Fourier transform layer of the Fourier spectrum transform module form a residual structure through shortcut links, and the second convolutional unit, BN unit, activation unit, and third convolutional unit of the feature fusion module form a residual structure through shortcut links.

[0034] Thirdly, the present invention provides a computer-readable storage medium for silent face liveness detection, including a memory for storing processor-executable instructions, which, when executed by the processor, implement the steps of the silent face liveness detection method described in the first aspect.

[0035] Fourthly, the present invention provides an apparatus for silent face liveness detection, comprising at least one processor and a memory storing computer-executable instructions, wherein the processor executes the instructions to implement the steps of the silent face liveness detection method described in the first aspect.

[0036] The present invention has the following beneficial effects:

[0037] This invention applies a second operation with a global receptive field to the liveness detection task of a deep network model. Compared to traditional convolution operations, the second operation has a global receptive field, enabling the deep network model to capture global information of the face image to be detected within a larger receptive field. Combining this with a first operation having a local receptive field, the local and global information obtained from the first and second operations are fused, which is more conducive to liveness detection and improves the accuracy of liveness detection. Attached Figure Description

[0038] Figure 1 This is a flowchart of the silent face liveness detection method of the present invention;

[0039] Figure 2 This is a schematic diagram illustrating an example of the deep network model of the present invention;

[0040] Figure 3 This is a schematic diagram of the feature extraction module;

[0041] Figure 4 This is a schematic diagram of the Fourier spectrum transform module;

[0042] Figure 5 This is a schematic diagram of the feature fusion module;

[0043] Figure 6 This is a schematic diagram of the silent liveness detection device for human faces according to the present invention. Detailed Implementation

[0044] To make the technical problems, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. The components of the embodiments of this invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0045] Example 1:

[0046] This invention provides a method for silent face liveness detection, such as... Figure 1 As shown, the method includes:

[0047] S100: Obtain the face image to be detected and input it into the trained deep network model, wherein the face image to be detected includes the face region and the background region.

[0048] Silent face liveness detection focuses not only on the local differences between real and fake faces, but also on the global geometric relationships and imaging differences of the face. For example, when photographing a face on a mobile phone screen, the face is often surrounded by the phone's border information; similarly, when photographing a piece of paper, the face is often surrounded by the paper's border.

[0049] To improve the ability of deep networks to focus on global facial geometric relationships and imaging differences, this invention introduces Fourier transform operations into the deep network model, enabling the deep network model to process the global background, which is very helpful for liveness detection, such as whether there are phone frames around the face, whether there are photo edges, and whether the background information matches objective reality; the Fourier transform operation is sensitive to periodic information, such as moiré patterns and paper textures, which is beneficial for liveness detection tasks.

[0050] Therefore, the face image to be detected input into the deep network model must include both the face region and a certain background region. If the face is artificial but the background is real, there will be differences between the face region and the background region in the imaging, such as the face region containing moiré patterns while the background region does not, which can be identified through Fourier transform operations.

[0051] This step involves acquiring images containing human faces using an acquisition device and performing face preprocessing on the acquired images to obtain the face image to be detected. Face preprocessing includes steps such as face detection, facial landmark localization, head pose estimation, and face normalization.

[0052] For example, this invention uses the MTCNN detection algorithm to achieve face detection and facial landmark localization. There are five facial landmarks: left eye, right eye, nose, left corner of mouth, and right corner of mouth. Furthermore, the 3D pose information of the face can be estimated using the 2D coordinate information of these facial landmarks.

[0053] Then, the face is normalized using the binocular coordinates to obtain the face image to be detected. Normalization specifically includes face alignment and scaling. Considering the importance of global background information, the face image to be detected can include more background regions and contain more background information. For example, the eyes are aligned to (94,108) and (129,108), the tip of the nose to (112,128), and the corners of the mouth to (98,148) and (126,148), and then scaled to a size of (224,224).

[0054] After obtaining the face image to be detected using the above method, it can be input into the trained deep network model.

[0055] Deep network models need to be trained before use, and the training process includes:

[0056] The real face image and the fake face image are preprocessed according to the aforementioned face preprocessing method to obtain real face samples and fake face samples.

[0057] Labels are assigned to real face samples and fake face samples, where the label for real face samples can be 1 and the label for fake face samples can be 0.

[0058] The deep network model was trained using real and fake face samples. During training, the parameters of the deep network model were optimized using stochastic gradient descent (SGD), and the loss function used during training was cross-entropy loss.

[0059] During training, data augmentation operations can be performed on the data online, such as random cropping, affine transformation, and color enhancement (ColorJitter).

[0060] S200: The deep network model mines the local features of the face image to be detected through a first operation with a local receptive field, mines the global features of the face image to be detected through a second operation with a global receptive field, and fuses the mined local features and global features. Based on the fused features, it is determined whether the face image to be detected is from a live subject.

[0061] The first operation can include convolution operations, which have a small local receptive field; the second operation can include Fourier transform operations, which have a global receptive field.

[0062] In liveness detection tasks, prosthetic faces are typically printed photos, cropped faces, or captured videos. For printed photos and captured videos, there is usually a border for the carrier on which the prosthetic face is made, such as the edge of the photo or the border of a mobile phone.

[0063] This invention's deep network model uses a first operation to mine local features of the face image to be detected, and a second operation to mine global features. The local and global features are then fused. While focusing on the essential characteristics of real and fake faces, it integrates information from the background and face regions, making it easier to distinguish between them. Taking a convolution operation as the first operation and a Fourier transform operation as the second operation as an example, due to the significant differences in imaging between objects of different materials, real faces have relatively closed imaging characteristics. This also means that background information and real faces also have relatively closed imaging characteristics. Fake faces made of different materials have significant imaging differences between themselves and real faces. These differences are easily captured by the Fourier transform, which helps distinguish between real and fake faces. Furthermore, the Fourier transform operation is sensitive to periodic information, such as the rough surface of paper or the moiré pattern on a photographed screen, which is easily captured by the Fourier transform operation, also aiding in the distinction between real and fake faces.

[0064] This invention applies a second operation with a global receptive field to the liveness detection task of a deep network model. Compared to traditional convolution operations, the second operation has a global receptive field, enabling the deep network model to capture global information of the face image to be detected within a larger receptive field. Combining this with a first operation (such as convolution) that has a local receptive field, the local and global information obtained from the first and second operations are fused, which is more beneficial for liveness detection and improves its accuracy.

[0065] This invention does not limit the structure of deep network models; in one example, such as Figure 2 As shown, the deep network model includes at least one operation block (Block1 to Block4), pooling operations (for example, average pooling), fully connected operations (FC), and softmax operations, etc.

[0066] Each operation block includes at least one feature extraction module, where n1, n2, n3, and n4 represent the number of feature extraction modules within each operation block; for example, n1 = n2 = n3 = n4 = 2. Furthermore, all feature extraction modules within each operation block can be linked together to form a residual structure.

[0067] The deep network model mines local features from its input through a first operation performed in the feature extraction module, mines global features from its input through a second operation performed in the feature extraction module, and then outputs the mined local features and global features by fusing them together.

[0068] Several operation blocks are executed sequentially. The output of the previous operation block is downsampled and used as the input of the next operation block. Furthermore, the output of each operation block is fused with the output of the previous operation block through the feature fusion module (CRB), so that the outputs of all operation blocks are fused to obtain the multi-scale fused features.

[0069] The fused features are sequentially processed through pooling, fully connected, and Softmax operations to obtain the liveness score and spoofing score of the face image to be detected; and the liveness score and spoofing score are used to determine whether the face image to be detected is from a live person.

[0070] For example, a face image to be detected with a size of W*H*3 (i.e., the width of the face image to be detected is W, the height is H, and the number of channels is 3) is input into a deep network model. After Block 1, Block 2, and Block 3, it is downsampled by 1 / 4, 1 / 8, and 1 / 16 respectively, resulting in the following values: The feature map. Downsampling is performed using a 1*1 convolution with a stride of 2 (Conv 1*1,2).

[0071] After each operation block, the feature information from the previous operation block is fused using CRB, thus achieving multi-scale feature fusion. After feature fusion using 1 / 16 sampling, the feature map size is... This refers to the multi-scale fusion feature.

[0072] Then, after average pooling, a 1*1*1024 feature value is obtained. After a fully connected operation, a 2D feature value is obtained. Softmax is used to obtain the liveness score S1 when the face is identified as real and the fakeness score S2 when it is identified as fake, where S1 + S2 = 1.0. If S1 > S2, the face is identified as live; otherwise, it is identified as fake. Alternatively, using S1 or S2 as the final score, a threshold T is set to determine whether the face is real or fake. If S1 > T (or S2 <= T), the face is considered live; otherwise, it is considered fake.

[0073] This invention fuses information at multiple scales, combining local and global information, as well as image spectral information. Through local information mining, global information mining, and multi-scale information integration, it enhances the classification capabilities of deep network models.

[0074] The feature extraction module of this invention includes local branches and global branches. The local branches include convolution operations to mine local information using a smaller receptive field, while the global branches include Fourier transform operations to mine global information. The two branches are then fused together using a concat operation, achieving the mining and fusion of local and global information. Figure 3 As shown.

[0075] When data is input into the feature extraction module, some channels are processed through the local branch to obtain the local features, while other channels are processed through the global branch to obtain the global features.

[0076] Assuming the number of input channels of the feature extraction module is NC, and it is split into segments with a split ratio of RC∈[0,1], then the number of local branch channels (i.e., the partial channels) is Round(NC*RC), and the number of global branch channels (i.e., the other partial channels) is Round(NC*(1-RC)). In this invention, RC can be set to 0.5.

[0077] After a series of operations, the local and global branches maintain their respective number of channels. Finally, the local and global features are fused together and output through the Concat operation. The output channels of the feature extraction module are consistent with the input channels, which is equal to NC.

[0078] The aforementioned local branches include the first convolutional module (Conv 3*3) and the second convolutional module (Conv 3*3), while the global branches include the third convolutional module (Conv 3*3) and the Fourier spectrum transform module (SpectralTransform).

[0079] The results of processing a portion of the channels by the first convolution module are fused with the results of processing another portion of the channels by the third convolution module, and can also be processed by BN and ReLU operations to obtain local features; the results of processing a portion of the channels by the second convolution module are fused with the results of processing another portion of the channels by the Fourier spectrum transform module, and can also be processed by BN and ReLU operations to obtain global features.

[0080] like Figure 4 As shown, the Fourier spectrum transform module sequentially includes a first combination layer (Conv-BN-ReLU, including convolution operation, BN operation and activation operation), a Fourier transform layer (Real FFT2d), a second combination layer (Conv-BN-ReLU), an inverse Fourier transform layer (Inv Real FFT2d) and a convolution layer (Conv 1*1), and the Fourier transform layer, the second combination layer and the inverse Fourier transform layer can also be linked together to form a residual structure.

[0081] FFT stands for Fast Fourier Transform, FFT2d refers to Two-Dimensional Fast Fourier Transform, and Real represents real numbers. Because the Fast Fourier Transform algorithm is for complex signals, while the time domain of this invention is a real signal, Real FFT refers to the Fast Fourier Transform for real numbers. Real FFT only uses half of the spectrum of FFT.

[0082] like Figure 5 As shown, the aforementioned feature fusion module sequentially includes a first convolutional unit (Conv 1*1), a second convolutional unit (Conv 3*3), a batch normalization (BN) unit, an activation unit (ReLU), and a third convolutional unit (Conv 3*3). The number of channels in each convolutional unit in the feature fusion module remains consistent with its input. The second convolutional unit, BN unit, activation unit, and third convolutional unit can also form a residual structure through shortcut links.

[0083] Example 2:

[0084] This invention provides a silent face liveness detection device, such as... Figure 6 As shown, the device includes:

[0085] Data acquisition module 1 is used to acquire the face image to be detected and input it into the trained deep network model. The face image to be detected includes the face region and the background region.

[0086] Inference module 2, the deep network model mines local features of the face image to be detected through a first operation with a local receptive field, mines global features of the face image to be detected through a second operation with a global receptive field, and fuses the mined local features and global features to determine whether the face image to be detected is from a live subject based on the fused features.

[0087] This invention applies a second operation with a global receptive field to the liveness detection task of a deep network model. Compared to traditional convolution operations, the second operation has a global receptive field, enabling the deep network model to capture global information of the face image to be detected within a larger receptive field. Combining this with a first operation having a local receptive field, the local and global information obtained from the first and second operations are fused, which is more conducive to liveness detection and improves the accuracy of liveness detection.

[0088] This invention does not limit the structure of deep network models; in one example, such as Figure 2 As shown, the deep network model includes at least one operation block, pooling operation, fully connected operation and Softmax operation, and each operation block includes at least one feature extraction module;

[0089] The deep network model mines local features from its input through the first operation performed in the feature extraction module, mines global features from its input through the second operation performed in the feature extraction module, and outputs the mined local features and global features by fusing them together. The output of each operation block is fused with the output of the previous operation block by the feature fusion module, so that the outputs of all operation blocks are fused to obtain the multi-scale fused features.

[0090] The fused features are sequentially processed through the pooling operation, the fully connected operation, and the Softmax operation to obtain the liveness score and spoofing score of the face image to be detected; the liveness score and spoofing score are used to determine whether the face image to be detected is from a live person.

[0091] The feature extraction module includes a local branch and a global branch. The local branch includes the first operation, and the global branch includes the second operation. A portion of the input channels of the feature extraction module are processed by the local branch to obtain the local features, and another portion of the input channels of the feature extraction module are processed by the global branch to obtain the global features.

[0092] The first operation includes a convolution operation, and the second operation includes a Fourier transform operation.

[0093] The local branch includes a first convolution module and a second convolution module, and the global branch includes a third convolution module and a Fourier spectrum transform module;

[0094] The results of processing a portion of the channels by the first convolution module are fused with the results of processing another portion of the channels by the third convolution module to obtain the local features; the results of processing a portion of the channels by the second convolution module are fused with the results of processing another portion of the channels by the Fourier spectrum transform module to obtain the global features.

[0095] The Fourier spectrum transform module sequentially includes a first combined layer, a Fourier transform layer, a second combined layer, an inverse Fourier transform layer, and a convolutional layer. Both the first and second combined layers include convolution operations, BN operations, and activation operations.

[0096] The feature fusion module includes, in sequence, a first convolutional unit, a second convolutional unit, a BN unit, an activation unit, and a third convolutional unit.

[0097] All feature extraction modules of each operation block form a residual structure through shortcut links. The Fourier transform layer, the second combined layer, and the inverse Fourier transform layer of the Fourier spectrum transform module form a residual structure through shortcut links. The second convolutional unit, the BN unit, the activation unit, and the third convolutional unit of the feature fusion module form a residual structure through shortcut links.

[0098] The device provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment 1. For the sake of brevity, any parts not mentioned in this device embodiment can be referred to the corresponding content in the aforementioned method embodiment 1. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the aforementioned device and unit can all be referred to the corresponding processes in the aforementioned method embodiment 1, and will not be repeated here.

[0099] Example 3:

[0100] The method described in Embodiment 1 of this invention can implement business logic through a computer program and record it on a storage medium. This storage medium can be read and executed by a computer, achieving the effect of the solution described in Embodiment 1 of this specification. Therefore, this invention also provides a computer-readable storage medium for silent face liveness detection, including a memory for storing processor-executable instructions. When executed by a processor, the instructions implement the steps of the silent face liveness detection method of Embodiment 1.

[0101] This invention applies a second operation with a global receptive field to the liveness detection task of a deep network model. Compared to traditional convolution operations, the second operation has a global receptive field, enabling the deep network model to capture global information of the face image to be detected within a larger receptive field. Combining this with a first operation having a local receptive field, the local and global information obtained from the first and second operations are fused, which is more conducive to liveness detection and improves the accuracy of liveness detection.

[0102] The storage medium may include a physical device for storing information, typically digitizing the information and then storing it using electrical, magnetic, or optical methods. The storage medium may include: devices that store information using electrical energy, such as various types of memory, like RAM and ROM; devices that store information using magnetic energy, such as hard disks, floppy disks, magnetic tapes, magnetic core memory, bubble memory, and USB flash drives; and devices that store information using optical methods, such as CDs or DVDs. Of course, there are other readable storage media, such as quantum memories and graphene memories.

[0103] The storage medium described above may also include other implementation methods according to the description of method embodiment 1. The implementation principle and technical effects of this embodiment are the same as those of the aforementioned method embodiment 1. For details, please refer to the description of the relevant method embodiment 1, which will not be repeated here.

[0104] Example 4:

[0105] The present invention also provides a device for silent face liveness detection. The device may be a standalone computer, or it may include an actual operating device that uses one or more of the methods or embodiments described in this specification. The device for silent face liveness detection may include at least one processor and a memory storing computer-executable instructions. When the processor executes the instructions, it implements the steps of any one or more of the silent face liveness detection methods described in Embodiment 1.

[0106] This invention applies a second operation with a global receptive field to the liveness detection task of a deep network model. Compared to traditional convolution operations, the second operation has a global receptive field, enabling the deep network model to capture global information of the face image to be detected within a larger receptive field. Combining this with a first operation having a local receptive field, the local and global information obtained from the first and second operations are fused, which is more conducive to liveness detection and improves the accuracy of liveness detection.

[0107] The device described above may include other implementation methods according to the description of method embodiment 1. The implementation principle and technical effects of this embodiment are the same as those of the aforementioned method embodiment 1. For details, please refer to the description of the relevant method embodiment 1, which will not be repeated here.

[0108] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. All should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for silent face liveness detection, characterized in that, The method includes: The face image to be detected is acquired and input into a trained deep network model. The face image to be detected includes a face region and a background region. The deep network model mines local features of the face image to be detected through a first operation with a local receptive field, mines global features of the face image to be detected through a second operation with a global receptive field, and fuses the mined local features and global features to determine whether the face image to be detected is from a live subject based on the fused features. The first operation includes a convolution operation, and the second operation includes a Fourier transform operation; The deep network model includes at least one operation block, pooling operation, fully connected operation and softmax operation, and each operation block includes at least one feature extraction module; The feature extraction module includes a local branch and a global branch. The global branch includes the second operation and a third convolution module and a Fourier spectrum transform module. The Fourier spectrum transform module sequentially includes a first combined layer, a Fourier transform layer, a second combined layer, an inverse Fourier transform layer, and a convolutional layer. Both the first and second combined layers include convolution operations, BN operations, and activation operations.

2. The silent face liveness detection method according to claim 1, characterized in that, The deep network model mines local features from its input through the first operation performed in the feature extraction module, mines global features from its input through the second operation performed in the feature extraction module, and outputs the mined local features and global features by fusing them together. The output of each operation block is fused with the output of the previous operation block by the feature fusion module, so that the outputs of all operation blocks are fused to obtain the multi-scale fused features. The fused features are sequentially processed through the pooling operation, the fully connected operation, and the Softmax operation to obtain the liveness score and spoofing score of the face image to be detected; the liveness score and spoofing score are used to determine whether the face image to be detected is from a live person.

3. The silent face liveness detection method according to claim 2, characterized in that, The local branch includes the first operation, wherein a portion of the input channels of the feature extraction module are processed by the local branch to obtain the local features, and another portion of the input channels of the feature extraction module are processed by the global branch to obtain the global features.

4. The silent face liveness detection method according to claim 3, characterized in that, The local branch includes a first convolutional module and a second convolutional module; The results of processing a portion of the channels by the first convolution module are fused with the results of processing another portion of the channels by the third convolution module to obtain the local features; the results of processing a portion of the channels by the second convolution module are fused with the results of processing another portion of the channels by the Fourier spectrum transform module to obtain the global features.

5. The silent face liveness detection method according to claim 4, characterized in that, The feature fusion module includes, in sequence, a first convolutional unit, a second convolutional unit, a BN unit, an activation unit, and a third convolutional unit.

6. The silent face liveness detection method according to claim 5, characterized in that, All feature extraction modules of each operation block form a residual structure through shortcut links. The Fourier transform layer, the second combined layer, and the inverse Fourier transform layer of the Fourier spectrum transform module form a residual structure through shortcut links. The second convolutional unit, the BN unit, the activation unit, and the third convolutional unit of the feature fusion module form a residual structure through shortcut links.

7. A silent face liveness detection device, characterized in that, The device includes: The data acquisition module is used to acquire the face image to be detected and input it into the trained deep network model. The face image to be detected includes a face region and a background region. The inference module, wherein the deep network model mines local features of the face image to be detected through a first operation with a local receptive field, mines global features of the face image to be detected through a second operation with a global receptive field, and fuses the mined local features and global features, and determines whether the face image to be detected is from a live subject based on the fused features. The first operation includes a convolution operation, and the second operation includes a Fourier transform operation; The deep network model includes at least one operation block, pooling operation, fully connected operation and softmax operation, and each operation block includes at least one feature extraction module; The feature extraction module includes a local branch and a global branch. The global branch includes the second operation and a third convolution module and a Fourier spectrum transform module. The Fourier spectrum transform module sequentially includes a first combined layer, a Fourier transform layer, a second combined layer, an inverse Fourier transform layer, and a convolutional layer. Both the first and second combined layers include convolution operations, BN operations, and activation operations.

8. A computer-readable storage medium for silent face liveness detection, characterized in that, It includes a memory for storing processor-executable instructions, which, when executed by the processor, implement the steps of the silent face liveness detection method according to any one of claims 1-6.

9. A device for silent face liveness detection, characterized in that, It includes at least one processor and a memory storing computer-executable instructions, wherein the processor executes the instructions to implement the steps of the silent face liveness detection method according to any one of claims 1-6.

Images