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

Abstract

Embodiments of the present application disclose a face silent living body detection method and device, electronic equipment and storage medium, belonging to the technical field of machine learning. The face silent living body detection method comprises: acquiring a face image to be detected; normalizing the acquired face image; using a pre-trained convolutional neural network to extract features of the normalized face image to generate a feature map, and calculating a living body judgment score according to the generated feature map, wherein the noise map of the normalized face image is utilized and boundary enhancement processing is performed when the feature map is generated; and predicting the probability that the face image to be detected comes from a living body according to the living body judgment score. The embodiments of the present application have good generalization and good detection effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to a face silent live body detection method and device, electronic equipment and storage medium. BACKGROUND

[0002] At present, face recognition technology has been widely used in the fields of finance and security. Due to the advantages of convenient face acquisition, non-contact, etc., it is also easy for others to use it in the form of photos or video to break through the face recognition system. Therefore, face live body detection technology as the first threshold of face recognition technology is particularly important.

[0003] Common face live body detection technologies mainly include cooperative and silent types. The cooperative type requires high user cooperation and single action, which is easy to be broken. The silent type does not require user cooperation and distinguishes the imaging differences between real and fake faces to determine whether it is a live body. Therefore, silent live detection is more acceptable to users.

[0004] Face silent live detection can be divided into traditional machine learning method and deep learning method from the technical means.

[0005] The traditional machine learning method starts from the intuitive impression and manually designs the features that can distinguish real and fake faces. Such intuitive impressions are generally the imaging characteristics of fake face, such as moire of video, rough surface of paper photo containing paper lines, etc. Common feature extraction methods include LBP (Local Binary Pattern) feature, Fourier spectrum, Difference of Gaussians (DOG), etc. Single or multiple features are combined, and SVM (support vector machines) and other machine learning methods are used to classify real and fake faces.

[0006] The deep learning method generally uses a large amount of data as the basis, automatically mines network parameters for distinguishing real and fake faces through deep networks such as ResNet and DenseNet, and finally classifies real and fake faces through SoftMax. Advanced attention mechanisms (Attention) or additional supervision such as Fourier spectrum are usually added. Some also use data preprocessing or online data enhancement (Crop, Color, Affine, etc.) to increase the diversity of data. Generally, real faces contain rich 3D (three-dimensional) information, and fake faces do not. By collecting face three-dimensional information through professional equipment or restoring face three-dimensional information through deep learning means, the network robustness can be improved by supervising the face three-dimensional information in network training.

[0007] The above methods have relatively good performance under the attack mode of the relative closed set, but often have low generalization ability and a sharp increase in misrecognition rate under unknown attack modes, SUMMARY

[0008] The technical problem to be solved by the embodiments of the present application is to provide a face silent live body detection method and device with good generalization and detection effect, an electronic device, and a storage medium.

[0009] To solve the above technical problems, the technical solutions of the embodiments of the present application are as follows:

[0010] In one aspect, a face silent live body detection method is provided, comprising:

[0011] obtaining a face image to be detected;

[0012] normalizing the obtained face image;

[0013] using a pre-trained convolutional neural network to extract features from the normalized face image to generate a feature map, and calculating a live body judgment score according to the generated feature map, wherein the normalized face image noise map is used when generating the feature map and boundary enhancement processing is performed;

[0014] predicting a probability that the face image to be detected comes from a live body according to the live body judgment score.

[0015] In some embodiments of the present application, the normalized face image contains background information.

[0016] And / or, the normalization of the obtained face image comprises:

[0017] performing face detection on the obtained face image to determine the position of the face;

[0018] performing key point positioning on the face;

[0019] performing face normalization according to the results of face detection and key point positioning.

[0020] In some embodiments of the present application, the use of a pre-trained convolutional neural network to extract features from the normalized face image to generate a feature map, and the calculation of a live body judgment score according to the generated feature map, comprises:

[0021] converting the normalized face image into a grayscale image;

[0022] obtaining a noise map by using BayarConv convolution on the grayscale image;

[0023] connecting the normalized face image and the noise map to obtain a combined face image;

[0024] Multi-stage convolution operations are performed on the combined face images to obtain the final feature maps, wherein boundary enhancement processing is performed on the feature maps obtained in at least one stage of convolution operations.

[0025] Perform a fully connected operation on the final feature map to obtain the fully connected operation result;

[0026] The result of the fully connected operation is fed into the classification function to obtain the liveness detection score.

[0027] In some embodiments of the present invention, the step of using a pre-trained convolutional neural network to extract features from a normalized face image to generate a feature map, and calculating a liveness detection score based on the generated feature map, includes:

[0028] A multi-stage convolution operation is performed on the normalized face image to obtain the final first feature map, wherein the feature map obtained in at least one stage of the convolution operation is subjected to boundary enhancement processing.

[0029] Convert the normalized face image into a grayscale image;

[0030] The grayscale image is convolved with BayarConv to obtain a noise image;

[0031] Perform multi-stage convolution operations on the noise map to obtain the final second feature map;

[0032] Connect the final first feature map and the final second feature map to obtain the combined feature map;

[0033] Perform a fully connected operation on the combined feature map to obtain the fully connected operation result;

[0034] The result of the fully connected operation is fed into the classification function to obtain the liveness detection score.

[0035] In some embodiments of the present invention, performing a multi-stage convolution operation on the noise map to obtain the final second feature map includes:

[0036] Boundary enhancement processing is performed on the feature maps obtained from at least one stage of convolution operation.

[0037] In some embodiments of the present invention, the boundary enhancement processing of the feature map obtained in at least one stage of convolution operation includes:

[0038] The feature maps obtained from at least one stage of convolution operation are sequentially subjected to boundary feature extraction, batch normalization, norm normalization, and activation function activation to obtain the boundary map.

[0039] The boundary map is multiplied by the feature map obtained in the convolution operation of this stage.

[0040] In some embodiments of the present application, the classification function is a softmax function.

[0041] And / or, the norm normalization is L2Norm normalization.

[0042] And / or, the activation function activation is Sigmod function activation.

[0043] In another aspect, a face silent living body detection device is provided, comprising:

[0044] An acquisition module is configured to acquire a face image to be detected.

[0045] A normalization module is configured to normalize the acquired face image.

[0046] A calculation module is configured to perform feature extraction on the normalized face image using a pre-trained convolutional neural network to generate a feature map, and calculate a living body judgment score according to the generated feature map, wherein the normalized face image noise map is utilized and boundary enhancement processing is performed when the feature map is generated.

[0047] A prediction module is configured to predict a probability that the face image to be detected comes from a living body according to the living body judgment score.

[0048] In still another aspect, an electronic device is provided, comprising: a housing, a processor, a memory, a circuit board and a power supply circuit, wherein the circuit board is arranged inside a space enclosed by the housing, and the processor and the memory are arranged on the circuit board; the power supply circuit is configured to supply power to each circuit or device of the electronic device; the memory is configured to store executable program codes; and the processor is configured to run programs corresponding to the executable program codes by reading the executable program codes stored in the memory, and execute any of the above-mentioned methods.

[0049] In yet another aspect, a computer readable storage medium is provided, which stores one or more programs executable by one or more processors to implement any of the above-mentioned methods.

[0050] The embodiments of the present application have the following beneficial effects:

[0051] The face silent live body detection method, device, electronic equipment and storage medium provided by the embodiment of the present application first acquire a face image to be detected, then normalize the acquired face image, then use a pre-trained convolutional neural network to extract features of the normalized face image to generate a feature map, and calculate a live body judgment score according to the generated feature map, wherein the noise map of the normalized face image is used when the feature map is generated and boundary enhancement processing is performed, and finally the probability that the face image to be detected comes from a live body is predicted according to the live body judgment score. In the embodiment of the present application, the boundary feature extraction and noise estimation are integrated into the face end-to-end live body detection classification model, and the prosthesis in the live body detection task is abstracted as a model of real face + noise, the prior knowledge of the prosthesis face is effectively used, that is, the prosthesis face is greatly different from the real skin due to the material, and there is a difference in lens optical imaging, and this difference is easier to capture in the noise image; the prosthesis carrier such as a mobile phone display screen and a printed photo has an intuitive carrier boundary around the face, in addition, the cut face and the hole-cut face have obvious boundaries at the edge of the face and the edge of the hole-cut facial features, such information is captured by the boundary extraction operator in the network and integrated into the original features, which is beneficial to improve the classification ability of the model. Therefore, the embodiment of the present application has good generalization and good detection effect. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the drawings shown.

[0053] Figure 1 The flowchart of an embodiment of the face silent live body detection method of the present application;

[0054] Figure 2 For Figure 1 The structure diagram of a convolutional neural network used in the method embodiment shown in the figure;

[0055] Figure 3 For Figure 1 The structure diagram of another convolutional neural network used in the method embodiment shown in the figure;

[0056] Figure 4 For Figure 2 And Figure 3 The structure diagram of the boundary extraction operator in the convolutional neural network shown in the figure;

[0057] Figure 5 The structure diagram of an embodiment of the face silent live body detection device of the present application;

[0058] Figure 6 Structure diagram of an embodiment of the electronic device of the present application. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work under the premise that the embodiments in the present application are within the scope of protection of the present application.

[0060] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative positional relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly.

[0061] In addition, the description of “first”, “second” and the like in the present application is only for the purpose of description, and cannot be understood as indicating or implying the relative importance of the indicated technical features or implicitly indicating the number of the indicated technical features. Therefore, the features defined with “first”, “second” can explicitly or implicitly include at least one of the features. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the fact that a person of ordinary skill in the art can realize it, and when the combination of technical solutions contradicts each other or cannot be realized, it should be considered that the combination of technical solutions does not exist, nor is it within the scope of protection required by the present application.

[0062] In the prior art, based on deep learning face silent detection, under the condition of a large amount of data, the normalized RGB face image is generally used as input, and the data features are gradually excavated from shallow to deep through multiple layers of convolution kernels, and then the features are classified into true and false face through SoftMax. This processing method generally has a deep network, and has relatively good performance under the attack mode of a relatively closed set, but often has low generalization ability and a sharp increase in false recognition rate for unknown attack modes, that is, overfitting.

[0063] The above situation is analyzed in two aspects. On the one hand, because the face is silent and the attention is concentrated on the face area, the existing normalization scheme usually takes the form of cutting a rectangular frame outside the face, and the relationship between the face and the background is ignored, which often plays a crucial role in determining whether it is a fake. For example, the face around the screen of the mobile phone often contains the mobile phone frame; the face around the paper photo often contains the paper frame; the face around the cut face has a relatively obvious cutting boundary, and similar boundaries also exist in the five feature positions of the cut-out face; the screen reflection, moire, also contains very obvious boundary information, and the moire reaction period wave boundary. Setting a component in the network structure to capture this kind of boundary information will greatly help improve the classification ability of the model. On the other hand, the RGB data is used without brain to throw the network black box learning, and the prior knowledge of the fake face is ignored. For example, moire, coarse-grained particles on the surface of the photo, and imaging differences caused by different materials. This prior knowledge can be abstracted as noise in the field of computer vision, that is, fake face = real face + noise, and the noise is evaluated from the original image, and the noise is used to assist in real and fake face classification, which is also very helpful to improve the classification ability of the model.

[0064] Based on the above analysis, the embodiment of the present application adds a boundary detection operator in the network to extract boundary information in the feature map at different stages, and then fuses the extracted boundary information into the feature map to enrich the boundary information in the feature map for distinguishing fake faces. At the same time, the BayarConv module is introduced to generate a noise image of the original RGB face, and then through network learning, the difference between real and fake faces in noise is captured to help distinguish real and fake faces.

[0065] On the one hand, the embodiment of the present application provides a face silent live body detection method, as shown in Figure 1 The method of the embodiment comprises the following steps:

[0066] Step 101: obtaining a face image to be detected;

[0067] In this step, the face image to be detected is usually an RGB face image.

[0068] Step 102: normalizing the obtained face image;

[0069] In order to facilitate image detection, the image usually needs to be normalized, and various normalization methods in the art can be used here. In the embodiment of the present application, preferably, the normalized face image contains background information, so as to retain the relationship between the face and the background and help distinguish real and fake faces.

[0070] As an optional embodiment, the normalizing the obtained face image (step 102) can comprise:

[0071] Step 1021: face detection is performed on the acquired face image to determine the position of the face;

[0072] Step 1022: key point positioning is performed on the face;

[0073] In steps 1021-1022, the MTCNN (Multi-task convolutional neural network) detection algorithm can be used to realize face detection and face key point positioning. There are five key points in the face, i.e., left eye, right eye, nose, left corner of the mouth, and right corner of the mouth. It should be noted that face detection can also use SSD (Single Shot MultiBox Detector), CRAFT, ADABOOST, etc. Similarly, face key point positioning can also use TDCNN, SDM (Supvised Descent Method), etc. Step 1022 only needs to obtain the coordinates of the five key points of the face image.

[0074] Step 1023: face normalization is performed according to the results of face detection and key point positioning.

[0075] In this step, the face image is normalized, specifically including face alignment and scaling. Considering the importance of global background information, more background information is needed, and the present embodiment can align the eyes to (94, 108) and (129, 108), align the nose tip to (112, 128), align the corners of the mouth to (98, 148) and (126, 148), and scale the size to (224, 224). In this way, the normalized face image contains a larger background area in addition to the face area, thereby preserving the background information.

[0076] It can be understood that other key points can also be selected for face key points, and the number of key points can be more or less than 5, and other methods of aligning and scaling images can also be used for normalization, which does not affect the implementation of the technical solution of the present application. Steps 1021-1023 described above constitute an image preprocessing step.

[0077] Step 103: a pre-trained convolutional neural network (CNN) is used to extract features from the normalized face image to generate a feature map, and a live body judgment score is calculated according to the generated feature map, wherein the noise map of the normalized face image is used when generating the feature map and boundary enhancement processing is performed;

[0078] In this step, when the normalized face image is extracted by the convolutional neural network to generate a feature map, the boundary enhancement processing is performed on the feature map (specifically, a boundary detection operator can be added in the network to extract the boundary information in the feature map, and then the extracted boundary information is fused into the feature map) to enrich the boundary information in the feature map for distinguishing the false face. At the same time, the noise map of the normalized face image is used to generate a feature map (specifically, a BayarConv module can be introduced to generate a noise image of the original RGB face image, and then the network is learned to capture the difference between the true and false face in the noise), to help distinguish the true and false face.

[0079] To achieve the above purpose, the convolutional neural network in the embodiment of the present application can have various structural forms, respectively as shown in Figure 2 and Figure 3 The following are described as follows:

[0080] Corresponding to Figure 2 , the feature extraction of the normalized face image by using the pre-trained convolutional neural network to generate a feature map, and calculating the live judgment score according to the generated feature map (step 103), can include:

[0081] Step 1031: converting the normalized face image into a gray image;

[0082] In this step, various conventional techniques in the art can be used to convert the face image into a gray image, which will not be described in detail here.

[0083] In Figure 2 , the convolutional neural network takes ResNet18 as the basic network, and the sampling rate of each stage is 4, 8, 16 and 32 respectively. The input image (i.e. the normalized face image) is WxHx3, where W and H are the width and height of the normalized face image (i.e. both are 224 as described above), and 3 represents the three channels of the RGB face image. After conversion to a gray image (Gray), it becomes a WxHx1 image, where 1 represents a single channel.

[0084] Step 1032: obtaining a noise map by BayarConv convolution on the gray image;

[0085] In Figure 2 , the BayarConv convolution kernel can be 5x5x3 in size, so as to obtain a WxHx3 noise map. Specifically, the BayarConv convolution kernel can be 3 stacked 5*5 convolution kernels, each 5*5 convolution kernel center is -1, and the 24 neighborhood sum is 1, wherein the 24 neighborhood parameters can be updated, but each time needs to be normalized.

[0086] Step 1033: connecting the normalized face image and the noise map to obtain a combined face image;

[0087] In Figure 2 , the input image and the noise map are concatenated to obtain a combined face image with a size of WxHx6.

[0088] Step 1034: performing multi-stage convolution operation on the combined face image to obtain a final feature map, wherein the feature map obtained in at least one stage of the convolution operation is subjected to boundary enhancement processing;

[0089] In Figure 2 , the convolutional neural network includes four times (stages) of convolution operation of Block1, Block2, Block3 and Block4, wherein the feature map of at least one stage or each stage can be subjected to boundary enhancement processing. The boundary enhancement processing is to add a boundary detection operator (Sobel operator or other operators) in the network to extract boundary information in the feature map of different stages, and then fuse the extracted boundary information into the feature map of the stage to enrich the boundary information in the feature map for distinguishing the false face.

[0090] Step 1035: performing full connection operation on the final feature map to obtain a full connection operation result;

[0091] In Figure 2 , the full connection operation corresponds to FC (fully connected layers) with a size of 1x1x1024.

[0092] Step 1036: sending the full connection operation result into a classification function to obtain a live body judgment score.

[0093] In this step, the classification function can be a binary classification function, and in Figure 2 , the classification function is a softmax function, thereby obtaining the live body judgment score.

[0094] Corresponding to Figure 3 , the feature extraction on the normalized face image using the pre-trained convolutional neural network to generate a feature map, and the calculation of the live body judgment score according to the generated feature map (step 103) can include:

[0095] Step 1031': performing multi-stage convolution operation on the normalized face image to obtain a final first feature map, wherein the feature map obtained in at least one stage of the convolution operation is subjected to boundary enhancement processing;

[0096] Figure 3The illustrated convolutional neural network still takes ResNet18 as the base network, and the sampling rates of each stage are 4, 8, 16 and 32 respectively. The network architecture includes a main network located at the top of the figure and a branch network located at the bottom, and both the main network and the branch network include four times (stages) of convolution operations of Block1, Block2, Block3 and Block4, and the weights can not be shared. The input image (i.e. the normalized face image) is input into the main network for multi-stage convolution operation. Figure 3 The same parts as Figure 2 above can refer to the above description of Figure 2 , which will not be repeated here.

[0097] Step 1032': converting the normalized face image into a grayscale image;

[0098] Step 1033': obtaining a noise map from the grayscale image using BayarConv convolution;

[0099] In Figure 3 , the size of the BayarConv convolution kernel is also 5x5x3.

[0100] Step 1034': performing multi-stage convolution operation on the noise map to obtain a final second feature map;

[0101] In Figure 3 , the noise map is input into the branch network for multi-stage convolution operation.

[0102] It can be understood that the branch network can not include boundary enhancement processing, however, preferably, the step 1034' includes: performing boundary enhancement processing on the feature map obtained in at least one stage of convolution operation.

[0103] Step 1035': connecting the final first feature map and the final second feature map to obtain a combined feature map;

[0104] In Figure 3 , that is, the final first feature map and the final second feature map are concatenated (connected / fused) to obtain a combined feature map.

[0105] Step 1036': performing full connection operation on the combined feature map to obtain a full connection operation result;

[0106] In Figure 3 , the size of the full connection layer FC is also 1x1x1024.

[0107] Step 1037': inputting the full connection operation result into a classification function to obtain a live body judgment score.

[0108] In Figure 3In this context, the classification function is also the softmax function.

[0109] As an optional embodiment, the boundary enhancement processing of the feature map obtained in at least one stage of convolution operation (in step 1034, step 1031', and step 1034') is as follows: Figure 4 As shown, it may include:

[0110] S11: Perform boundary feature extraction, batch normalization, norm normalization, and activation function activation on the feature maps obtained in at least one stage of convolution operation to obtain the boundary map;

[0111] S12: Multiply the boundary map dotwise by the feature map obtained in the convolution operation of this stage.

[0112] exist Figure 4 In this process, boundary feature extraction uses the SobelConv convolution kernel, which can be obtained by combining Sobel_x and Sobel_y, where the parameters do not need to be updated.

[0113] Sobel_x can specifically be:

[0114] 1 0 -1 2 0 -2 1 0 -1

[0115] Sobel_y can specifically be:

[0116] 1 2 1 0 0 0 -1 -2 -1

[0117] The Sobel operator adapts to the number of channels in the feature map at each stage. For example, in stage 1 (corresponding to Block 1), the output feature map has 64 channels. SobelConv can then use a 64*3*3*1 kernel in the X direction and a 64*3*3*1 kernel in the Y direction. Then, batch normalization (BN) + norm normalization (such as L2 Norm) + activation function (such as Sigmoid) are performed in the X and Y directions to generate a W / 4*H / 4*1 boundary map. This boundary map is then multiplied by the input W / 4*H / 4*64 feature map to strengthen the boundary, which is reflected in the feature map as having higher feature values ​​at the boundary.

[0118] In step 103 above, the training of the convolutional neural network can employ conventional techniques in this field. For example, real and fake face images can be preprocessed using the same methods described above to obtain normalized real and fake face images, with real faces labeled as 1 and fake faces labeled as 0. The network is then trained using the SGD (stochastic gradient descent) optimization method and cross-entropy loss. During training, data augmentation operations can also be performed online, such as random cropping, affine, and color jitter.

[0119] Step 104: Predict the probability that the face image to be detected is from a living person based on the liveness detection score.

[0120] In this step, the probability of a live face can be predicted based on the liveness detection score. For example, the score is S1 when the face is identified as real and S2 when it is identified as fake. S1 + S2 = 1. If S1 > S2, the face is identified as live; otherwise, it is identified as fake. Alternatively, S1 or S2 can be used as the final score alone. By setting a threshold T, the real or fake face can be determined. If S1 > T (or S2 <= T), the face is considered live; otherwise, it is considered fake.

[0121] In summary, the silent liveness detection method for faces provided in this embodiment of the invention first acquires a face image to be detected, then normalizes the acquired face image, then uses a pre-trained convolutional neural network to extract features from the normalized face image to generate a feature map, and calculates a liveness detection score based on the generated feature map. The feature map generation process utilizes the noise map of the normalized face image and performs boundary enhancement processing. Finally, the probability that the face image to be detected is from a live person is predicted based on the liveness detection score. In this embodiment of the invention, boundary feature extraction and noise estimation are integrated into an end-to-end face liveness detection classification model. The prosthetic face in the liveness detection task is abstracted as a model of a real face plus noise. This effectively utilizes prior knowledge about the prosthetic face; that is, because the material of the prosthetic face differs significantly from real skin, there are differences in lens optical imaging, and these differences are more easily captured in noisy images. The prosthetic face carrier, such as a mobile phone screen or a printed photo, has a clear carrier boundary around the face. Furthermore, cropped and cutout faces have obvious boundaries at the edges of the face and the edges of the cutout features. This information is captured by the boundary extraction operator in the network and integrated into the original features, which helps improve the model's classification ability. Therefore, this embodiment of the invention has good generalization and excellent detection performance.

[0122] After testing, the invention was found to be effective. Figure 2 The method of the illustrated embodiment achieves an accuracy of 99.56% on the constructed test set; the method employing this invention... Figure 3The accuracy of the method of the embodiment on the constructed test set is 99.87%, and the accuracy of a conventional ResNet18 network on the constructed test set is 94.31%. Therefore, the method of the embodiment can greatly improve the face liveness detection accuracy.

[0123] In another aspect, the embodiment of the present application provides a face silent liveness detection device, which comprises Figure 5 as shown, comprising:

[0124] The acquisition module 11 is configured to acquire a face image to be detected.

[0125] The normalization module 12 is configured to normalize the acquired face image.

[0126] The calculation module 13 is configured to perform feature extraction on the normalized face image using a pre-trained convolutional neural network to generate a feature map, and calculate a liveness judgment score according to the generated feature map, wherein the noise map of the normalized face image is utilized and boundary enhancement processing is performed when the feature map is generated.

[0127] The prediction module 14 is configured to predict the probability that the face image to be detected comes from a live body according to the liveness judgment score.

[0128] The device of the embodiment can be used to execute Figure 1 The technical scheme of the method embodiment as shown has similar implementation principles and technical effects, and will not be described here.

[0129] Preferably, the normalized face image contains background information.

[0130] Preferably, the normalization module 12 comprises:

[0131] The detection unit is configured to perform face detection on the acquired face image to determine the position of the face.

[0132] The positioning unit is configured to perform key point positioning on the face.

[0133] The normalization unit is configured to perform face normalization according to the results of face detection and key point positioning.

[0134] Preferably, the calculation module 13 comprises:

[0135] The conversion unit is configured to convert the normalized face image into a grayscale image.

[0136] The first convolution unit is configured to obtain a noise map by utilizing BayarConv convolution on the grayscale image.

[0137] a connection unit, configured to connect the normalized face image and the noise map to obtain a combined face image;

[0138] a second convolution unit, configured to perform multi-stage convolution operation on the combined face image to obtain a final feature map, wherein the feature map obtained in at least one stage of the convolution operation is subjected to boundary enhancement processing;

[0139] a full connection unit, configured to perform full connection operation on the final feature map to obtain a full connection operation result;

[0140] a classification unit, configured to input the full connection operation result into a classification function to obtain a live body judgment score.

[0141] Preferably, the computing module 13 comprises:

[0142] a first convolution unit, configured to perform multi-stage convolution operation on the normalized face image to obtain a final first feature map, wherein the feature map obtained in at least one stage of the convolution operation is subjected to boundary enhancement processing;

[0143] a conversion unit, configured to convert the normalized face image into a grayscale image;

[0144] a second convolution unit, configured to obtain a noise map by performing BayarConv convolution on the grayscale image;

[0145] a third convolution unit, configured to perform multi-stage convolution operation on the noise map to obtain a final second feature map;

[0146] a connection unit, configured to connect the final first feature map and the final second feature map to obtain a combined feature map;

[0147] a full connection unit, configured to perform full connection operation on the combined feature map to obtain a full connection operation result;

[0148] a classification unit, configured to input the full connection operation result into a classification function to obtain a live body judgment score.

[0149] Preferably, the third convolution unit comprises:

[0150] a boundary enhancement sub-unit, configured to perform boundary enhancement processing on the feature map obtained in at least one stage of the convolution operation.

[0151] Preferably, the boundary enhancement processing on the feature map obtained in at least one stage of the convolution operation comprises:

[0152] performing, in sequence, boundary feature extraction, batch normalization, norm normalization and activation function activation on the feature map obtained in at least one stage of the convolution operation to obtain a boundary map;

[0153] multiplying the boundary map points to the feature map obtained in the stage convolution operation.

[0154] Preferably, the classification function is a softmax function.

[0155] Preferably, the norm normalization is L2Norm normalization.

[0156] Preferably, the activation function is a Sigmod function activation.

[0157] The embodiment of the present application also provides an electronic device, Figure 6 The structure diagram of an embodiment of the electronic device of the present application can realize the present application Figure 1 The flow of the embodiment shown in the figure, such as Figure 6 The electronic device can include a housing 41, a processor 42, a memory 43, a circuit board 44 and a power supply circuit 45, wherein the circuit board 44 is arranged inside the space surrounded by the housing 41, and the processor 42 and the memory 43 are arranged on the circuit board 44; the power supply circuit 45 is used to supply power to each circuit or device of the electronic device; the memory 43 is used to store executable program codes; the processor 42 runs the program corresponding to the executable program codes by reading the executable program codes stored in the memory 43, and is used to execute the method of any one of the preceding method embodiments.

[0158] The specific execution process of the processor 42 for the above steps and the steps further executed by the processor 42 by running the executable program codes can be referred to the description of the embodiment shown in the figure, which will not be repeated here. Figure 1

[0159] The electronic device exists in various forms, including but not limited to:

[0160] (1) Mobile communication device: the feature of this kind of device is to have mobile communication function, and to provide voice and data communication as the main target. This kind of terminal includes: smart phone (such as iPhone), multimedia phone, functional phone, and low-end phone, etc.

[0161] (2) Ultra-mobile personal computer device: this kind of device belongs to the category of personal computer, has computing and processing function, and generally has the feature of mobile Internet. This kind of terminal includes: PDA, MID and UMPC device, such as iPad.

[0162] (3) Portable entertainment device: this kind of device can display and play multimedia content. This kind of device includes: audio and video player (such as iPod), palm game machine, electronic book, and smart toy and portable vehicle navigation device.

[0163] ​(4) Server: a device providing computing services, the configuration of the server includes a processor, a hard disk, a memory, a system bus, etc., the server is similar to the general computer architecture, but since it is required to provide high reliable services, therefore, in the processing capacity, stability, reliability, security, scalability, manageability, etc. higher requirements.

[0164] (5) Other electronic devices with data interaction function.

[0165] The embodiment of the present application also provides a computer readable storage medium, the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the method steps of any method embodiment.

[0166] The embodiment of the present application also provides an application program, which is executed to realize the method provided by any method embodiment of the present application.

[0167] The above is the preferred embodiment of the present application, it should be pointed out that, for those skilled in the art, without departing from the principles of the present application, can make a number of improvements and refinements, these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A face silent live body detection method, characterized in that, The method comprises: obtaining a face image to be detected; normalizing the obtained face image; extracting features of the normalized face image using a pre-trained convolutional neural network to generate a feature map, and calculating a live body judgment score according to the generated feature map, wherein the normalized face image is used to generate the feature map and boundary enhancement processing is performed; predicting a probability that the face image to be detected comes from a live body according to the live body judgment score; wherein the feature extraction of the normalized face image using the pre-trained convolutional neural network to generate the feature map and the calculation of the live body judgment score according to the generated feature map comprise: converting the normalized face image into a grayscale image; obtaining a noise map by using BayarConv convolution on the grayscale image; connecting the normalized face image and the noise map to obtain a combined face image; performing multi-stage convolution operation on the combined face image to obtain a final feature map, wherein boundary enhancement processing is performed on the feature map obtained in at least one stage of the convolution operation; performing full connection operation on the final feature map to obtain a full connection operation result; sending the full connection operation result into a classification function to obtain the live body judgment score. Alternatively, the feature extraction of the normalized face image using the pre-trained convolutional neural network to generate the feature map and the calculation of the live body judgment score according to the generated feature map comprise: performing multi-stage convolution operation on the normalized face image to obtain a final first feature map, wherein boundary enhancement processing is performed on the feature map obtained in at least one stage of the convolution operation; converting the normalized face image into a grayscale image; obtaining a noise map by using BayarConv convolution on the grayscale image; performing multi-stage convolution operation on the noise map to obtain a final second feature map; connecting the final first feature map and the final second feature map to obtain a combined feature map; performing full connection operation on the combined feature map to obtain a full connection operation result; sending the full connection operation result into a classification function to obtain the live body judgment score.

2. The method of claim 1, wherein, The normalized face image contains background information. And / or, the normalization of the obtained face image comprises: performing face detection on the obtained face image to determine the position of the face; performing key point positioning on the face; performing face normalization according to the results of the face detection and the key point positioning.

3. The method of claim 1, wherein, The multi-stage convolution operation on the noise map to obtain the final second feature map comprises: performing boundary enhancement processing on the feature map obtained in at least one stage of the convolution operation.

4. The method according to any one of claims 1 to 3, characterized in that, The boundary enhancement processing on the feature map obtained in at least one stage of the convolution operation comprises: performing boundary feature extraction, batch normalization, norm normalization and activation function activation on the feature map obtained in at least one stage of the convolution operation in sequence to obtain a boundary map; point multiplying the boundary map to the feature map obtained in the stage of the convolution operation.

5. The method of claim 4, wherein, The classification function is a softmax function. And / or, the norm normalization is L2Norm normalization. And / or, the activation function activation is Sigmod function activation.

6. A face silent live body detection device, characterized in that, The method comprises: An acquisition module is configured to acquire a face image to be detected; A normalization module is configured to normalize the acquired face image; A calculation module is configured to perform feature extraction on the normalized face image using a pre-trained convolutional neural network to generate a feature map, and calculate a live body judgment score according to the generated feature map, wherein the normalized face image is used to generate the feature map and boundary enhancement processing is performed; A prediction module is configured to predict a probability that the face image to be detected comes from a live body according to the live body judgment score; The calculation module includes: A conversion unit is configured to convert the normalized face image into a grayscale image; A first convolution unit is configured to obtain a noise map by performing BayarConv convolution on the grayscale image; A connection unit is configured to connect the normalized face image and the noise map to obtain a combined face image; A second convolution unit is configured to perform multi-stage convolution operation on the combined face image to obtain a final feature map, wherein boundary enhancement processing is performed on the feature map obtained in at least one stage of the convolution operation; A full connection unit is configured to perform full connection operation on the final feature map to obtain a full connection operation result; A classification unit is configured to input the full connection operation result into a classification function to obtain a live body judgment score. Alternatively, the calculation module includes: A first convolution unit is configured to perform multi-stage convolution operation on the normalized face image to obtain a final first feature map, wherein boundary enhancement processing is performed on the feature map obtained in at least one stage of the convolution operation; A conversion unit is configured to convert the normalized face image into a grayscale image; A second convolution unit is configured to obtain a noise map by performing BayarConv convolution on the grayscale image; A third convolution unit is configured to perform multi-stage convolution operation on the noise map to obtain a final second feature map; A connection unit is configured to connect the final first feature map and the final second feature map to obtain a combined feature map; A full connection unit is configured to perform full connection operation on the combined feature map to obtain a full connection operation result; A classification unit is configured to input the full connection operation result into a classification function to obtain a live body judgment score.

7. An electronic device, comprising: The electronic device includes a housing, a processor, a memory, a circuit board and a power circuit, wherein the circuit board is arranged inside a space enclosed by the housing, and the processor and the memory are arranged on the circuit board; the power circuit is configured to supply power to each circuit or device of the electronic device; the memory is configured to store executable program codes; the processor is configured to run programs corresponding to the executable program codes by reading the executable program codes stored in the memory, and execute the method of any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores one or more programs, which can be executed by one or more processors to implement the method of any one of claims 1-5.

Images