Face anonymization method and device
By obtaining the features and representation vectors of the face image, and combining the random vector and noise coefficient for perturbation processing, the random perturbation face image is generated, and the aesthetics and automation problems of face anonymization processing in the prior art are solved, and the effect of effectively hiding personal face features is achieved.
Patent Information
- Application Number
- CN202411792800.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-07
- Publication Date
- 2025-04-22
AI Technical Summary
The existing face anonymization technology has the risk of demosaic or defuzzy image super-resolution methods that may lead to the risk of blurred face restoration, and simple mosaic or blurred processing affects the aesthetics of the picture and lacks automated processing capabilities.
By filtering the face image, obtaining the feature vectors of key points and high-dimensional embedded representation vectors, constructing random vectors of the same dimensions and combining noise coefficients, calculating the perturbation representation vectors, and then stitching the feature vectors to generate the face image after random perturbation, replacing the original face image.
Effectively hide personal face feature information without changing the basic state of the face, protecting personal information, and improving the aesthetics of face images, avoiding the risk of face restoration.
Smart Images
Figure CN119863846A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a face processing technology, and in particular to a face anonymization method. The present application also relates to a face anonymization device. Background Art
[0002] The development of science and technology, especially artificial intelligence technology, has provided citizens with more convenient living and production tools, but at the same time its negative effects are gradually affecting citizens, among which the issue of personal information protection is becoming more and more urgent.
[0003] In today's society, personal facial data has become an important part of personal information protection. Personal facial data includes photos or videos containing personal facial data. Personal facial data is not only a personal portrait, but also the biometric information of the individual's face. Once leaked or illegally used by others, it may cause problems such as personal face impersonation, face recognition detection terminals being misled, face payment being misused, and personal bank face authentication information being stolen. Therefore, facial information has become an important part of sensitive personal information.
[0004] In the prior art, face anonymization technology that hides or masks face images is a major way to avoid face information leakage. Face anonymization is mainly achieved through mosaic processing or image covering technology, but there are still many problems to be solved in these technologies.
[0005] Mosaic processing is to manually select a specific area of a picture or video, and then use an image mosaic algorithm or a blurring algorithm to mosaic or blur the face area. This method is the simplest and most commonly used face anonymization process. However, there are currently a small number of image super-resolution methods for de-mosaicing or de-blurring, which can restore the blurred face to a certain extent. In addition, simple mosaic processing or blurring processing also affects the beauty of the picture or video. In addition, this method requires manual selection of the face area, and lacks efficient automatic processing capabilities for a large number of pictures and videos.
[0006] The avatar image overlay also requires manual selection of specific areas of the image or video, and then overlaying fixed avatar cartoons, non-personalized avatars, etc. on the selected area. This method also has problems such as affecting the overall appearance of the image, being inconsistent with the overall style of the image, and lacking automated processing capabilities. Summary of the invention
[0007] To solve the above technical problems, the present application provides a face anonymization method, which can avoid leakage of personal face information while displaying face images, and improve the aesthetics of videos or pictures. The present application also provides a face anonymization device.
[0008] The present application provides a face anonymization method, comprising:
[0009] Filtering face images detected from view data;
[0010] Obtaining a feature vector of key points and a high-dimensional embedded representation vector from the face image;
[0011] constructing a random vector of the same dimension as the characterization vector, combining the characterization vector with the random vector and a noise coefficient set for the random vector, and calculating a disturbance characterization vector;
[0012] The disturbance representation vector and the feature vector are concatenated to obtain a randomly disturbed face image, and the face image is replaced with the randomly disturbed face image.
[0013] Optionally, the feature vector of the key point is obtained:
[0014] The detection and output include: feature vector composed of position information of eyes, nose and mouth.
[0015] Optionally, the view data includes: video data and image data;
[0016] The video data is preprocessed into a single frame image.
[0017] Optionally, before acquiring the feature vector and the characterization vector, a mask is generated in the facial image, and when replacing the facial image, the facial image within the mask range is replaced.
[0018] Optionally, replacing the facial image within the mask range includes: using a blurring image method at the mask edge.
[0019] Optionally, the computational disturbance characterization vector is obtained using the following formula:
[0020] F'=F+aR
[0021] The F' represents a disturbance characterization vector, the F represents a characterization vector, the R represents a random vector, and the a represents a noise coefficient.
[0022] Optionally, the representation vector is obtained by an encoder-decoder neural network, and the encoder-decoder neural network is obtained by pre-training, including:
[0023] Randomly select a face image and input it into the encoder network and the key point detection network to obtain the high-dimensional embedded representation vector and the feature vector of the key point respectively;
[0024] splicing the feature vector and the representation vector and inputting them into a decoder network, wherein the decoder network outputs a generated face image;
[0025] According to the deviation between the face image and the generated face image, adjusting parameters by back propagation method;
[0026] Repeat the above steps until the deviation meets the preset requirements to obtain the encoder-decoder neural network.
[0027] The present application also provides a face anonymization device, comprising:
[0028] A screening module, used for screening face images detected from the view data;
[0029] An extraction module, used for obtaining a feature vector of key points and a high-dimensional embedded representation vector from the face image;
[0030] A calculation module, used for constructing a random vector of the same dimension as the characterization vector, combining the characterization vector with the random vector and a noise coefficient set for the random vector, and calculating a disturbance characterization vector;
[0031] A splicing module is used to splice the disturbance representation vector and the feature vector to obtain a randomly disturbed face image, and replace the face image with the randomly disturbed face image.
[0032] Optionally, also include:
[0033] The key point detection module is used to detect and output a feature vector consisting of the position information of the eyes, nose and mouth.
[0034] Optionally, also include:
[0035] Video data processing module: used for preprocessing the video data into a single frame image.
[0036] Optionally, also include:
[0037] The mask module is used to generate a mask in the face image before acquiring the feature vector and the characterization vector, and replace the face image within the mask range when replacing the face image.
[0038] Optionally, replacing the facial image within the mask range includes: using a blurring image method at the mask edge.
[0039] Optionally, the computational disturbance characterization vector is obtained using the following formula:
[0040] F'=F+aR
[0041] The F' represents a disturbance characterization vector, the F represents a characterization vector, the R represents a random vector, and the a represents a noise coefficient.
[0042] Optionally, it further includes: a model training module, the model training module including:
[0043] An acquisition module, used for randomly selecting a face image and inputting it into an encoder network and a key point detection network, and obtaining the representation vector of the high-dimensional embedding and the feature vector of the key point respectively;
[0044] An output module, used for concatenating the feature vector and the representation vector and inputting them into a decoder network, wherein the decoder network outputs a generated face image;
[0045] An adjustment module, used for adjusting parameters by a back propagation method according to the deviation between the face image and the generated face image;
[0046] A loop module is used to repeat the above steps until the deviation meets the preset requirements to obtain the encoder-decoder neural network.
[0047] The advantages of this application over the prior art are as follows:
[0048] The present application provides a method for anonymizing a face, including: screening a face image detected from view data; obtaining a feature vector of key points and a high-dimensional embedded representation vector from the face image; constructing a random vector of the same dimension as the representation vector, combining the representation vector with the random vector and a noise coefficient set for the random vector, and calculating a perturbed representation vector; concatenating the perturbed representation vector and the feature vector to obtain a randomly perturbed face image, and replacing the face image with the randomly perturbed face image. By modifying the face attribute features, the attribute feature information of a specific individual can be hidden without changing the basic state of the face, and the visual effect will not be affected. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 This is a flowchart of the face anonymization processing in this application.
[0050] Figure 2 This is a flow chart of face anonymization data in this application.
[0051] Figure 3 This is a schematic diagram of face anonymization processing in this application.
[0052] Figure 4 This is a schematic diagram of the key point detection network training in this application.
[0053] Figure 5 This is a schematic diagram of the face masking network training in this application.
[0054] Figure 6 This is a schematic diagram of the encoder-decoder network training in this application.
[0055] Figure 7 It is a schematic diagram of the face anonymization processing device in this application.
[0056] Figure 8 It is a schematic diagram of the training module in the anonymization device in this application. DETAILED DESCRIPTION
[0057] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.
[0058] The present application mainly provides a face anonymization method and a device corresponding to the anonymization method. The anonymization method in the present application includes: screening face images detected from view data; obtaining feature vectors of key points and high-dimensional embedded representation vectors from the face images; constructing a random vector of the same dimension as the representation vector, combining the representation vector with the random vector and the noise coefficient set for the random vector, and calculating the perturbation representation vector; splicing the perturbation representation vector and the feature vector to obtain a face image after random perturbation, and replacing the face image with the face image after random perturbation. Through the processing of the above steps, the face does not need to change the basic features of the face, and the feature information about the individual can be hidden, thereby achieving the purpose of protecting citizens' personal information, while also increasing the beauty of the face image.
[0059] Figure 1 This is the flow chart of face anonymization processing in this application. Figure 2 This is the flow diagram of the face anonymization data in this application. Figure 3 This is a schematic diagram of face anonymization processing in this application.
[0060] Please refer to Figure 1 , Figure 2 as well as Figure 3 As shown, S101 screens the face image A detected from the view data 001;
[0061] In this application, the view data 001 includes video data and image data, but in actual operation applications, all data that can depict or represent a face can be included. In the case where the view data 001 may leak personal face information, it can be reasonably expanded for use under the connotation of this application.
[0062] The sources of the view data 001 of the present application are various, such as images collected by video cameras, cameras, etc. Therefore, in the view data 001, it is not guaranteed that there is only a face image A, and there is only one face image A. Therefore, before processing the face image A, the view data 001 needs to be pre-processed.
[0063] After obtaining the view data 001, face detection 002 needs to be performed in advance to detect the face image A. The other parts of the view data 001 are not within the scope of discussion of this application. Therefore, the face image A mentioned below refers to the detected face image part of the person, which will not be described in detail below.
[0064] If the view data 001 is video data, it is necessary to process the video data into a single frame image in advance. If the view data 001 is an image, it can directly enter the step of face detection 002. Preferably, the face detection can be performed by a face detection module.
[0065] After the face image A is detected, it is necessary to select the face image A according to the specific situation, that is, face screening 003. For all the face images A in a view data 001, some face images A need to be anonymized, while some do not need to be anonymized, so it is necessary to select the face images A that need to be anonymized in advance for further processing. For the sake of simplicity of description and understanding, in the following, face image A refers to the face image A that needs to be anonymized, and the face image A that does not need to be anonymized is not within the scope of discussion of the core content of this application and will not be repeated in the following.
[0066] Step S102: acquiring a feature vector V of key points and a high-dimensional embedded representation vector F from the face image A;
[0067] Through step S101, the face image A has been acquired, and this step will start to anonymize the face image A.
[0068] like Figure 3 As shown, after face screening 003, the face image A will be sent to different modules for different processing, mainly including the key point detection network L and the encoder network E.
[0069] The key point detection network L in the present application can be a designed multi-layer convolutional neural network, but is not limited thereto, and those skilled in the art can make other choices according to actual conditions. The key point detection network L is pre-trained.
[0070] Figure 4 It is a schematic diagram of the training of the key point detection network L in this application.
[0071] Please refer to Figure 4 As shown, a facial key point dataset 012 is obtained; in this step, the facial key point dataset 012 is a pre-prepared training image material.
[0072] Input the facial key point data set 012 into the key point detection network L;
[0073] Detect the key point position information, that is, the feature vector V.
[0074] After the above steps, it is determined whether the key point information is correct, and the above steps are repeated until the key point detection network is trained.
[0075] The key point information in the face image will be detected and output by the key point detection network, including: a feature vector V composed of the position information of the eyes, nose and mouth.
[0076] In the present application, the encoder network E is one of the core modules in the present application: the encoder network E part of the encoder-decoder neural network E / D. After the face image A is input into the encoder network E, the encoder network E will detect the face image A and output the high-dimensional embedding representation vector F of the face image. For simple description, the high-dimensional embedding representation vector F in this article is referred to as the representation vector F.
[0077] The key point detection network L and the face encoder network E respectively detect the feature vector V and the representation vector F about the face image A. The present application also performs other operations on the face image A, including: masking.
[0078] The mask is to mask the face image through the face mask analysis network M, that is, to prevent the face image A from being observed or detected normally, so that personal information can be protected when the face image A is not anonymized, and at the same time, provide conditions for subsequent anonymization processing, including: when replacing the face image A, the face image A within the mask range is replaced, and during the replacement process, a fuzzy image method can be used at the mask edge to enhance the naturalness of the edge transition. The face mask analysis network M is obtained through pre-training.
[0079] Figure 5 This is a training diagram of the face mask training network M in this application.
[0080] Please refer to Figure 5 As shown, a face mask dataset 014 is obtained; the face mask dataset 014 is a pre-prepared training image. The face mask analysis network M can be a designed multi-layer convolutional neural network, but is not limited thereto. Those skilled in the art can make other choices according to actual conditions.
[0081] Inputting the face mask data set into the face mask analysis network M;
[0082] Generate face mask 017 image.
[0083] After the above steps, it is determined whether the face mask 017 image is correct, and the above steps are repeated until the face mask analysis network M is trained.
[0084] Step S103 constructs a random vector R of the same dimension as the characterization vector F, combines the characterization vector F with the random vector R and a noise coefficient a set for the random vector, and calculates a disturbance characterization vector F';
[0085] The feature vector V and the representation vector F of the face image have been obtained above. In order to anonymize the face, the representation vector F needs to be corrected to achieve the purpose of removing personal information.
[0086] First, a random vector R is constructed, wherein the random vector R and the characterization vector F are of the same latitude, and the characterization vector F is modified by the random vector R.
[0087] Before modifying the characterization vector F, a noise parameter a needs to be preset. The noise parameter a is used to indicate the influence of the random vector R on the characterization vector F. Specifically, the characterization vector F is modified by formula 1:
[0088] F'=F+aR
[0089] The F' is a modified representation vector, which is referred to as a perturbed representation vector F' in this application. Preferably, in the same view data 001, using the same random vector R and noise parameter a will help the stability and consistency of face anonymization.
[0090] Step S104 concatenates the disturbance representation vector F' and the feature vector V to obtain a randomly disturbed face image A', and replaces the face image with the randomly disturbed face image A'.
[0091] The disturbance characterization vector F' is obtained above. In this step, the disturbance characterization vector F' and the feature vector V are concatenated to form a concatenated vector 007, and then a disturbed face image A' is formed.
[0092] In step S102, the facial image A has been masked, and the perturbed facial image A' and the facial mask 017 are input into the decoder network D in the encoder-decoder neural network E / D, and the perturbed facial image A' is replaced to the masked position of the facial image A in the decoder network D, thus completing the facial image anonymization processing and obtaining the facial data anonymization result 011.
[0093] In the present application, face anonymization processing is mainly performed through an encoder-decoder neural network, and the encoder-decoder neural network obtains the representation vector, and the encoder-decoder neural network is obtained through pre-training.
[0094] Figure 6 It is a schematic diagram of training the encoder-decoder neural network in this application.
[0095] Please refer to Figure 6 As shown, step S401 randomly selects a face data and inputs it into the encoder network E and the key point detection network L to obtain the high-dimensional embedded representation vector F and the feature vector V of the key point respectively;
[0096] The encoder-decoder neural network E / D is established with a network including a multi-layer convolutional neural network, a residual network, etc. as the main structure. In order to train the encoder-decoder neural network E / D, it is first necessary to collect a large number of public face data sets 016, and then divide the face data into different batches, and select one face data each time to input into the encoder-decoder neural network E / D.
[0097] After receiving the face data, the encoder-decoder neural network E / D sends it to the trained face mask analysis network M, key point detection network L and face encoder network E respectively.
[0098] Step S402: concatenate the feature vector V and the representation vector F and input them into a decoder network D, and the decoder network D outputs a generated face image A”;
[0099] The face key point detection network L and the encoder network E respectively obtain the key feature vector V and the characterization vector F of the face, and then splice the characterization vector F and the feature vector V of the face and input them into the decoder network D. At the same time, the face mask 017 image generated by the face mask analysis network M is also sent to the decoder network D. The training splicing vector F" after the feature vector V and the characterization vector F are spliced replaces the face mask 017 in the mask image, and finally the output generates a face image A".
[0100] Step S403: adjusting parameters by back propagation method according to the deviation between the face data and the generated face image A″;
[0101] The generated face image A" is compared with the face data, and the deviation between the generated face image A" and the face data is calculated. The deviation in this application is the error function. Parameters are adjusted according to the error function through the back propagation method.
[0102] Step S404 repeats the above steps until the deviation meets the preset requirements, and obtains the encoder-decoder neural network E / D.
[0103] After adjusting the parameters, the face data is input again and steps S401 to S404 are repeated until the deviation meets the preset requirements, thereby obtaining the trained encoder-decoder neural network E / D.
[0104] The above is the face anonymization method in this application. The face image obtained by the face anonymization method no longer carries personal information, which protects personal information while not affecting the aesthetics. In addition to the above face anonymization method, this application also provides a face anonymization device.
[0105] The face anonymization device described in the present application includes: a screening module 101, used to screen face images detected from view data; an extraction module 102, used to obtain feature vectors of key points and high-dimensional embedded representation vectors from the face images; a calculation module 103, used to construct a random vector of the same dimension as the representation vector, combine the representation vector with the random vector and a noise coefficient set for the random vector, and calculate a perturbation representation vector; a splicing module 104, used to splice the perturbation representation vector and the feature vector to obtain a randomly perturbed face image, and replace the face image with the randomly perturbed face image.
[0106] Figure 7 It is a schematic diagram of the face anonymization device described in this application.
[0107] Please refer to Figure 7 As shown, the screening module 101 is used to screen the face images detected from the view data;
[0108] In the present application, the view data 001 shown includes video data and image data, but in actual operational applications, all data that can depict or represent a human face may be included.
[0109] The sources of the view data 001 of the present application are various, such as images collected by video cameras, cameras, etc. Therefore, in the view data 001, it is not guaranteed that there is only a face image A, and there is only one face image A. Before processing the face image A, the view data 001 needs to be pre-processed.
[0110] After the view data 001 is acquired, face detection 002 needs to be performed in advance to detect the face image A. The other parts of the view data 001 are not within the scope of discussion of this application.
[0111] The anonymization device further includes: a video data processing module, which is used to pre-process the video data into a single frame image. If the view data 001 is video data, it is necessary to pre-process the video data into a single frame image. If the view data 001 is an image, it can directly enter the step of face detection 002. Preferably, the face detection can be performed by a face detection module.
[0112] After the face image A is detected, it is necessary to select the face image A according to the specific situation, that is, face screening 003. For all the face images A in a view data 001, some face images A need to be anonymized, while some do not need to be anonymized, so it is necessary to select the face images A that need to be anonymized in advance for further processing.
[0113] An extraction module 102, used to obtain a feature vector of key points and a high-dimensional embedded representation vector from the face image;
[0114] The face image A has been acquired through the screening module 101 , and then the face image A will be anonymized.
[0115] After face screening 003, the face image A will be sent to different modules for different processing, mainly including the key point detection network L and the encoder network E.
[0116] The key point detection network L in the present application can be a designed multi-layer convolutional neural network, but is not limited thereto, and those skilled in the art can make other choices according to actual conditions. The key point detection network L is pre-trained.
[0117] Obtain a facial key point data set 012; in this step, the facial key point data set 012 is a pre-prepared training image material.
[0118] Input the facial key point data set 012 into the key point detection network L;
[0119] Detect the key point position information, that is, the feature vector V.
[0120] After the above steps, it is determined whether the key point information is correct, and the above steps are repeated until the key point detection network is trained.
[0121] The key point information in the face image will be detected and output by the key point detection network, including: a feature vector V composed of the position information of the eyes, nose and mouth.
[0122] In the present application, the encoder network E is one of the core modules in the present application: the encoder network E part of the encoder-decoder neural network E / D. After the face image A is input into the encoder network E, the encoder network E will detect the face image A and output the high-dimensional embedding representation vector F of the face image. For simple description, the high-dimensional embedding representation vector F in this article is referred to as the representation vector F.
[0123] The key point detection network L and the face encoder network E respectively detect the feature vector V and the characterization vector F about the face image A. The anonymization device described in the present application also includes: a mask module, which is used to generate a face mask before obtaining the feature vector and the characterization vector, and replace the face image within the mask range when replacing the face image.
[0124] The mask is to mask the face image through the face mask analysis network M, that is, to prevent the face image A from being observed or detected normally, so that personal information can be protected when the face image A is not anonymized, and at the same time, provide conditions for subsequent anonymization processing, including: when replacing the face image A, the face image A within the mask range is replaced, and during the replacement process, a fuzzy image method can be used at the mask edge to enhance the naturalness of the edge transition. The face mask analysis network M is obtained through pre-training.
[0125] Obtain a face mask dataset 014; the face mask dataset 014 is a pre-prepared training image. The face mask analysis network M can be a designed multi-layer convolutional neural network, but is not limited thereto. Those skilled in the art can make other choices according to actual conditions.
[0126] Inputting the face mask data set into the face mask analysis network M;
[0127] Generate face mask 017 image.
[0128] After the above steps, it is determined whether the face mask 017 image is correct, and the above steps are repeated until the face mask analysis network M is trained.
[0129] A calculation module 103, configured to construct a random vector of the same dimension as the characterization vector, combine the characterization vector with the random vector and a noise coefficient set for the random vector, and calculate a disturbance characterization vector;
[0130] The feature vector V and the representation vector F of the face image have been obtained above. In order to anonymize the face, the representation vector F needs to be corrected to achieve the purpose of removing personal information.
[0131] First, a random vector R is constructed, wherein the random vector R and the characterization vector F are of the same latitude, and the characterization vector F is modified by the random vector R.
[0132] Before modifying the characterization vector F, a noise parameter a needs to be preset. The noise parameter a is used to indicate the influence of the random vector R on the characterization vector F. Specifically, the characterization vector F is modified by formula 1:
[0133] F'=F+aR
[0134] The F' is a modified representation vector, which is referred to as a perturbed representation vector F' in this application. Preferably, in the same view data 001, using the same random vector R and noise parameter a will help the stability and consistency of face anonymization.
[0135] The splicing module 104 is used to splice the disturbance representation vector and the feature vector to obtain a randomly disturbed face image, and replace the face image with the randomly disturbed face image.
[0136] The disturbance representation vector F' is obtained above. In this step, the disturbance representation vector F' and the feature vector V are concatenated to form a disturbed face image A'.
[0137] In the extraction module 102, the face image A has been masked, and the perturbed face image A' and the face mask 017 are input into the decoder network D in the encoder-decoder neural network E / D. In the decoder network D, the perturbed face image A' is replaced to the masked position of the face image A, thereby completing the anonymization of the face image.
[0138] In the present application, face anonymization processing is mainly performed through an encoder-decoder neural network, and the encoder-decoder neural network obtains the representation vector, and the encoder-decoder neural network is obtained through pre-training.
[0139] Figure 8 It is a schematic diagram of the training module in the anonymization device in this application.
[0140] Please refer to Figure 8 As shown, the anonymization device described in the present application also includes: a training module.
[0141] The training module includes:
[0142] An acquisition unit 201 is used to randomly select a face data and input it into an encoder network and a key point detection network to respectively acquire the high-dimensional embedded representation vector and the feature vector of the key point;
[0143] The encoder-decoder neural network E / D is established with a network including a multi-layer convolutional neural network, a residual network, etc. as the main structure. In order to train the encoder-decoder neural network E / D, it is first necessary to collect a large number of public face data sets 016, and then divide the face data into different batches, and select one face data each time to input into the encoder-decoder neural network E / D.
[0144] After receiving the face data, the encoder-decoder neural network E / D sends it to the trained face mask analysis network M, key point detection network L and face encoder network E respectively.
[0145] An output unit 202, configured to concatenate the feature vector and the representation vector and input the concatenated vector into a decoder network, wherein the decoder network outputs a generated face image;
[0146] The face key point detection network L and the encoder network E respectively obtain the key feature vector V and the characterization vector F of the face, and then splice the characterization vector F and the feature vector V of the face and input them into the decoder network D. At the same time, the face mask 017 image generated by the face mask analysis network M is also sent to the decoder network D. The splicing vector F" after the feature vector V and the characterization vector F are spliced replaces the face mask 017 in the mask image, and the final output generates a face image A".
[0147] An adjustment unit 203, configured to adjust parameters by a back propagation method according to the deviation between the face data and the generated face image;
[0148] The generated face image A" is compared with the face data, and the deviation between the generated face image A" and the face data is calculated. The deviation in this application is the error function. Parameters are adjusted according to the error function through the back propagation method.
[0149] The loop unit 204 is used to repeat the above steps until the deviation meets the preset requirements to obtain the encoder-decoder neural network.
[0150] After adjusting the parameters, the face data is input again, and the processing of units 201 to 204 is repeated until the deviation meets the preset requirements, and the trained encoder-decoder neural network E / D is obtained.
Claims
1. A face anonymization method, characterized in that: include: Filtering face images detected from view data; Obtaining a feature vector of key points and a high-dimensional embedded representation vector from the face image; constructing a random vector of the same dimension as the characterization vector, combining the characterization vector with the random vector and a noise coefficient set for the random vector, and calculating a disturbance characterization vector; The disturbance representation vector and the feature vector are concatenated to obtain a randomly disturbed face image, and the face image is replaced with the randomly disturbed face image.
2. The face anonymization method according to claim 1, characterized in that: The method of obtaining the feature vector of the key point includes: The detection and output include: feature vector composed of position information of eyes, nose and mouth.
3. The face anonymization method according to claim 1, characterized in that: The view data includes: video data and image data; The video data is preprocessed into a single frame image.
4. The face anonymization method according to claim 1, characterized in that: Before acquiring the feature vector and the characterization vector, a face mask is generated, and when replacing the face image, the face image within the mask range is replaced.
5. The face anonymization method according to claim 4, characterized in that: The replacing of the face image within the mask range includes: using a blurring image method at the mask edge.
6. The face anonymization method according to claim 1, characterized in that: The calculation disturbance characterization vector is obtained using the following formula: F'=F+aR The F' represents a disturbance characterization vector, the F represents a characterization vector, the R represents a random vector, and the a represents a noise coefficient.
7. The face anonymization method according to claim 1, characterized in that: The representation vector is obtained by an encoder-decoder neural network, and the encoder-decoder neural network is obtained by pre-training, including: Randomly select a face data and input it into the encoder network and the key point detection network to obtain the high-dimensional embedded representation vector and the feature vector of the key point respectively; splicing the feature vector and the representation vector and inputting them into a decoder network, wherein the decoder network outputs a generated face image; According to the deviation between the face data and the generated face image, adjusting parameters by back propagation method; Repeat the above steps until the deviation meets the preset requirements to obtain the encoder-decoder neural network.
8. A face anonymization device, characterized in that: include: A screening module, used for screening face images detected from the view data; An extraction module, used for obtaining a feature vector of key points and a high-dimensional embedded representation vector from the face image; A calculation module, used for constructing a random vector of the same dimension as the characterization vector, combining the characterization vector with the random vector and a noise coefficient set for the random vector, and calculating a disturbance characterization vector; A splicing module is used to splice the disturbance representation vector and the feature vector to obtain a randomly disturbed face image, and replace the face image with the randomly disturbed face image.
9. The face anonymization method according to claim 8, characterized in that: Also includes: The key point detection network is used to detect and output a feature vector consisting of the position information of the eyes, nose and mouth.
10. The face anonymization method according to claim 8, characterized in that: Also includes: Video data processing module: used for preprocessing the video data into a single frame image.
11. The face anonymization method according to claim 8, characterized in that: Also includes: The mask module is used to generate a face mask before acquiring the feature vector and the characterization vector, and replace the face image within the mask range when replacing the face image.
12. The face anonymization method according to claim 11, characterized in that: The replacing of the face image within the mask range comprises: using a blurring image method at the mask edge.
13. The face anonymization method according to claim 8, characterized in that: The calculation disturbance characterization vector is obtained using the following formula: F'=F+aR The F' represents a disturbance characterization vector, the F represents a characterization vector, the R represents a random vector, and the a represents a noise coefficient.
14. The face anonymization method according to claim 8, characterized in that: Also includes: A model training module, wherein the model training module comprises: An acquisition unit, used for randomly selecting a face data and inputting it into the encoder network and the key point detection network, and respectively acquiring the representation vector of the high-dimensional embedding and the feature vector of the key point; An output unit, used for concatenating the feature vector and the representation vector and inputting them into a decoder network, wherein the decoder network outputs a generated face image; An adjustment unit, used for adjusting parameters by a back propagation method according to the deviation between the face data and the generated face image; The cyclic unit is used to repeat the above steps until the deviation meets the preset requirements to obtain the encoder-decoder neural network.