WO2023061122A1 - 人脸活体检测方法、装置、计算机可读存储介质及设备 - Google Patents

人脸活体检测方法、装置、计算机可读存储介质及设备 Download PDF

Info

Publication number
WO2023061122A1
WO2023061122A1 PCT/CN2022/118036 CN2022118036W WO2023061122A1 WO 2023061122 A1 WO2023061122 A1 WO 2023061122A1 CN 2022118036 W CN2022118036 W CN 2022118036W WO 2023061122 A1 WO2023061122 A1 WO 2023061122A1
Authority
WO
WIPO (PCT)
Prior art keywords
face
feature
image
light
ambient light
Prior art date
Application number
PCT/CN2022/118036
Other languages
English (en)
French (fr)
Inventor
王洪
Original Assignee
北京眼神科技有限公司
北京眼神智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京眼神科技有限公司, 北京眼神智能科技有限公司 filed Critical 北京眼神科技有限公司
Publication of WO2023061122A1 publication Critical patent/WO2023061122A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/10—Image acquisition
    • G06V10/12—Details of acquisition arrangements; Constructional details thereof
    • G06V10/14—Optical characteristics of the device performing the acquisition or on the illumination arrangements
    • G06V10/141—Control of illumination
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/40—Spoof detection, e.g. liveness detection
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02B—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO BUILDINGS, e.g. HOUSING, HOUSE APPLIANCES OR RELATED END-USER APPLICATIONS
    • Y02B20/00—Energy efficient lighting technologies, e.g. halogen lamps or gas discharge lamps
    • Y02B20/40—Control techniques providing energy savings, e.g. smart controller or presence detection

Definitions

  • the present application relates to the field of face recognition, in particular to a face detection method, device, computer-readable storage medium and equipment.
  • face recognition technology has been widely used in the fields of finance and security. Because the face has the advantages of easy acquisition and non-contact, but it is also very easy to be used by others to break through the face recognition system by taking photos or remakes of videos. Therefore, face liveness detection technology is particularly important as the first threshold of face recognition technology.
  • the first way is face movement liveness detection.
  • the liveness detection system issues random head and face movement instructions, and the user completes the corresponding actions according to the instructions to determine that the person is alive;
  • the first way is to extract the features of RGB (Red Green Blue, color system) images used to distinguish real faces and prosthetic faces, and do two classifications of real faces and prosthetic faces;
  • the third way is the first two
  • RGB Red Green Blue, color system
  • Human face motion liveness detection requires a high degree of cooperation from users, and the random head and face motion commands issued are generally a single motion command, such as nodding, turning the head, blinking, and opening the mouth. There is a high probability that it can be deceived through the liveness detection system, or videos that take multiple actions of the user can also easily pass through the liveness detection system.
  • the way of extracting the features of the RGB image for judging the real face and the prosthetic face to perform liveness detection is silent liveness detection. This kind of liveness detection does not require the cooperation of the user and is more easily accepted by the user.
  • deep learning methods are generally used. Driven by a large amount of live and non-living face data, automatic learning can effectively distinguish the features of real faces and prosthetic faces. The difference between real and fake face imaging.
  • RGB images are prone to misjudgment due to imaging differences such as lighting, resolution, and different mobile phone lenses.
  • a face detection method, device, readable storage medium and equipment are provided, which improves the accuracy of face detection, and does not require active cooperation of the user, so that the user experience is good.
  • the present application provides a method for face detection, the method comprising:
  • Preprocessing the light-filling face image to obtain an overall image of a light-filling face and a partial image of a face-filling face;
  • feature extraction is performed on the overall image of the human face in ambient light and the overall image of the human face in supplementary light, to obtain the first feature and the second feature; Performing feature extraction on the partial image of the human face and the partial image of the human face with supplementary light to obtain the third feature and the fourth feature;
  • the user is classified according to the fusion feature, and it is judged whether the ambient light human face image and the supplementary light human face image are from a living body.
  • the first convolutional neural network and the second convolutional neural network are trained by the following method:
  • the first convolutional neural network is trained according to the overall image sample of the ambient light human face and the overall image sample of the supplementary light human face, and according to the local image sample of the ambient light human face and the supplementary light human face Face partial image samples, training the second convolutional neural network to obtain the overall loss in the model training process of the first convolutional neural network and the second convolutional neural network;
  • the convolutional neural network Updating the parameters of the first convolutional neural network and the second convolutional neural network through backpropagation until the overall loss meets a preset loss condition, the first convolutional neural network and the second convolutional neural network The convolutional neural network is trained.
  • the method also includes:
  • the overall image samples of the ambient light face, the partial image samples of the ambient light face, the overall image sample of the supplementary light face and the partial image samples of the supplementary light face belonging to the same living body or prosthesis are respectively preset with labels; It is used to identify whether the images in the multiple samples belong to the same living body or the same prosthesis.
  • the first convolutional neural network is trained according to the overall image sample of the ambient light face and the overall image sample of the supplementary light face
  • the human face is trained according to the ambient light image sample. Partial image samples and the partial image samples of the supplementary light human face
  • the second convolutional neural network is trained to obtain the model training process of the first convolutional neural network and the second convolutional neural network Overall loss, including:
  • the first convolutional neural network feature extraction is performed on the overall image sample of the ambient light human face and the overall image sample of the supplementary light human face to obtain the fifth feature and the sixth feature; through the second convolutional neural network The network performs feature extraction on the partial image sample of the human face in the ambient light and the partial image sample of the human face in the supplementary light to obtain the seventh feature and the eighth feature;
  • the first comparison loss L 1 , the second comparison loss L 2 and the cross-entropy loss L 3 are weighted and summed to obtain an overall loss L All .
  • the fusion of the fifth feature, the sixth feature, the seventh feature and the eighth feature to obtain the fusion sample feature includes:
  • the fifth feature, the seventh feature, the sixth feature and the eighth feature are sequentially connected to obtain the fused sample feature.
  • the fusion of the first feature, the second feature, the third feature and the fourth feature to obtain the fusion feature includes:
  • the first feature, the third feature, the second feature and the fourth feature are sequentially connected to obtain the fusion feature; the fusion feature is used to distinguish a living human face from a prosthetic human face, so
  • the above fusion features include the overall 3D information of the face and the light spot information of the eyes.
  • the partial face image in ambient light and the partial face image in supplementary light are eye images; the third feature and the fourth feature include eye spot information.
  • the preprocessing of the ambient light face image to obtain the overall image of the face under ambient light and the partial image of the face under ambient light includes:
  • an eye center and radius are determined using a radially symmetric transformation
  • the partial image of the human face under ambient light is intercepted.
  • the preprocessing of the light-filling human face image to obtain the light-filling overall image of the human face and the partial image of the light-filling human face includes:
  • an eye center and radius are determined using a radially symmetric transformation
  • the partial image of the human face with supplementary light is intercepted.
  • the acquisition of the ambient light face image and fill light face image within the same user preset time period includes:
  • the supplementary light is supplemented by a supplementary light, and an RGB image is collected under supplementary light conditions to obtain the supplementary light face image.
  • the method also includes:
  • the present application provides a human face biopsy detection device, the device comprising:
  • An image acquisition module configured to acquire ambient light face images and supplementary light face images within the same user preset time period
  • the first preprocessing module is used to preprocess the face image under ambient light to obtain the overall image of the face under ambient light and the partial image of the face under ambient light;
  • the second preprocessing module is used to preprocess the light-filling face image to obtain the overall image of the light-filling face and the partial image of the light-filling face;
  • the image feature extraction module is used to perform feature extraction on the overall image of the human face in the ambient light and the overall image of the human face in the supplementary light through the first convolutional neural network to obtain the first feature and the second feature; through the second convolution The neural network performs feature extraction on the partial image of the human face with ambient light and the partial image of the human face with supplementary light to obtain the third feature and the fourth feature;
  • An image feature fusion module configured to fuse the first feature, the second feature, the third feature, and the fourth feature to obtain a fusion feature
  • the living body detection module is used for classifying through the fusion feature, and judging whether the ambient light face image and the supplementary light face image are from living bodies.
  • the first convolutional neural network and the second convolutional neural network are obtained through the following module training:
  • the sample construction module is used to construct an overall image sample of a human face with ambient light, a partial image sample of a human face with ambient light, an overall image sample of a human face with fill light, and a partial image sample of a human face with fill light, and set labels respectively;
  • the sample feature extraction module is used to perform feature extraction on the overall image sample of the ambient light face and the overall image sample of the supplementary light face through the first convolutional neural network to obtain the fifth feature and the sixth feature; through the second The convolutional neural network performs feature extraction on the partial image samples of the ambient light face and the partial image samples of the fill light face to obtain the seventh feature and the eighth feature;
  • a comparison loss calculation module configured to calculate a comparison loss L 1 between the fifth feature and the sixth feature and a comparison loss L 2 between the seventh feature and the eighth feature;
  • a cross-entropy loss calculation module configured to fuse the fifth feature, the sixth feature, the seventh feature, and the eighth feature to obtain a fusion sample feature, and calculate the cross-entropy of the fusion sample feature loss L 3 ;
  • a backpropagation module configured to weight and sum the comparison loss L 1 , the comparison loss L 2 and the cross-entropy loss L 3 to obtain an overall loss L All , and update the first volume through back propagation convolutional neural network and the parameters of the second convolutional neural network.
  • the fifth feature, the seventh feature, the sixth feature and the eighth feature are sequentially connected to obtain the fusion sample feature
  • the first feature, the third feature, the second feature and the fourth feature are sequentially connected to obtain a fusion feature.
  • the partial face image in ambient light and the partial face image in supplementary light are eye images.
  • the device also includes:
  • the image processing module is used to perform face detection, face key point positioning, head pose estimation and face normalization on the ambient light face image and the supplementary light face image to obtain the overall image of the ambient light face and Fill in the overall image of the face;
  • the image area determination module is used for centering on the eye coordinates obtained by positioning the key points of the human face, and expanding some pixels around to obtain a region of interest including the eyes, wherein the number of pixels expanded around is determined according to the distance between the left and right eyes;
  • An image information determination module configured to determine the eye center and radius using radial symmetric transformation within the region of interest
  • the image interception module is used for intercepting and obtaining the partial image of the human face in ambient light and the partial image of the human face in supplementary light according to the eye center and radius.
  • the image acquisition module is used to collect RGB images under ambient light through a face lens to obtain the ambient light face image
  • the supplementary light is supplemented by a supplementary light, and an RGB image is collected under supplementary light conditions to obtain the supplementary light face image.
  • the present application provides a computer-readable storage medium for human face liveness detection, including a memory for storing processor-executable instructions, and when the instructions are executed by the processor, the implementation includes the first aspect. The steps of the human face living body detection method described above.
  • the present application provides a device for live face detection, including at least one processor and a memory storing computer-executable instructions.
  • the processor executes the instructions, the human face described in the first aspect is realized. Steps of the liveness detection method.
  • This application obtains the face image under ambient light and supplementary light, obtains the overall area and local area of the face image under ambient light and supplementary light through preprocessing, and extracts features through a convolutional neural network, and fuses the extracted features Perform live face and prosthetic face classification.
  • This application uses the characteristics of local changes in the living face when it encounters a point light source to analyze the local changes of the face before and after the light supplement, and uses the overall change characteristics of the imaging of the living face before and after the light supplement to analyze the overall face image before and after the light supplement. Changes, liveness detection, high accuracy, simple and convenient, does not require active cooperation from users, and has a good user experience.
  • Fig. 1 is a schematic flow chart of a method for detecting human face liveness provided in an exemplary embodiment
  • Fig. 2 is a schematic diagram of a convolutional neural network training provided in an exemplary embodiment
  • Fig. 3 is a schematic flow chart of a face image acquisition step provided in an exemplary embodiment
  • Fig. 4 is a schematic flow chart of a human face image screening step provided in an exemplary embodiment
  • Fig. 5 is a schematic diagram of a preprocessing process provided in an exemplary embodiment
  • Fig. 6 is a schematic flowchart of a training method for a first convolutional neural network and a second convolutional neural network provided in an exemplary embodiment
  • Fig. 7 is a schematic diagram of a convolutional neural network training provided in an exemplary embodiment
  • Fig. 8 is a schematic flowchart of a method for obtaining a training set provided in an exemplary embodiment
  • Fig. 9 is a schematic flowchart of a method for preprocessing an ambient light face image and a supplementary light face image provided in an exemplary embodiment
  • Fig. 10 is a schematic diagram of a preprocessing process provided in an exemplary embodiment
  • Fig. 11 is a schematic diagram of a preprocessing process provided in an exemplary embodiment
  • Fig. 12 is a schematic structural diagram of a human face liveness detection device provided in an exemplary embodiment.
  • the embodiment of the present application provides a face detection method, which can be used for face recognition devices, especially for mobile face recognition devices, the mobile face recognition device has a screen or supplementary light Lights for supplementary light during face detection.
  • the mobile face recognition devices include but are not limited to smartphones, notebook computers, tablet computers, and handheld computers.
  • Fig. 1 is the schematic flow chart of the human face living body detection method of the present application, and this method comprises:
  • S110 Acquiring ambient light face images and supplementary light face images within the same user-preset time period.
  • the face recognition device on the mobile terminal acquires the ambient light face image and supplementary light image within the same user-preset time period.
  • the changes in the partial image of the face with ambient light and the partial image of the face with supplementary light such as the change of the light spot in the pupil
  • the changes in the overall image of the face with ambient light and the overall image of the face with supplementary light it is possible to distinguish between live faces and Prosthetic face.
  • S120 Perform preprocessing on the face image under ambient light to obtain an overall image of the face under ambient light and a partial image of the face under ambient light.
  • the face recognition device at the mobile terminal preprocesses the face image under ambient light to obtain the overall image of the face under ambient light and the partial image of the face under ambient light.
  • the process of preprocessing face images in ambient light mainly includes face detection, face key point positioning, head pose estimation, and face normalization, etc.
  • the face images in ambient light are processed , the overall image of the face in ambient light and the partial image of the face in ambient light corresponding to the face image in ambient light can be obtained.
  • the preprocessing process of the face image under ambient light can be adaptively adjusted according to the actual application process, and the embodiment of the present application does not limit the specific preprocessing process.
  • S130 Perform preprocessing on the fill-in face image to obtain a fill-in face overall image and a fill-in face partial image.
  • the face recognition device at the mobile terminal preprocesses the fill-in face image to obtain the fill-in face overall image and fill-in face partial image.
  • the process of preprocessing the face image with supplementary light may also include face detection, face key point location, head pose estimation and face normalization, etc. The processing process is similar, and will not be repeated in this embodiment of the present application.
  • S140 Use the first convolutional neural network to perform feature extraction on the overall image of the face in ambient light and the overall image of the face in supplementary light to obtain the first feature and the second feature; use the second convolutional neural network to extract the partial image of the face in ambient light
  • the third feature and the fourth feature are obtained by performing feature extraction with the partial image of the face with supplementary light.
  • two convolutional neural networks are used, namely the first convolutional neural network and the second convolutional neural network, the first convolutional neural network And the second convolutional neural network can choose a lightweight or small network, such as the Base network of the Siamese Network used.
  • the structure of the first convolutional neural network and the second convolutional neural network can be the same, and based on the requirement of algorithm efficiency, the quantity processing capacity of the second convolutional neural network can be relatively small and lighter, so as to improve the algorithm efficiency.
  • the face recognition device on the mobile terminal analyzes the changes of the partial image of the face in ambient light and the partial image of the face in supplementary light through the convolutional neural network, and analyzes the changes in the overall image of the face in ambient light and the overall image of the face in supplementary light.
  • the face recognition device at the mobile terminal performs feature extraction on the overall image of the face in ambient light and the overall image of the face in supplementary light through the first convolutional neural network to obtain the first feature and the second feature.
  • feature extraction is carried out on the partial image of the face in ambient light and the partial image of the face in supplementary light to obtain the third feature and the fourth feature.
  • FIG. 2 a structural embodiment of the first convolutional neural network and the second convolutional neural network is shown in Figure 2, wherein Net1 is the first convolutional neural network, which is used to extract features from the overall area of the face, respectively Extract the features of the overall image of the face in ambient light and the overall image of the face in supplementary light.
  • face_b and face_f are the overall image of the face in ambient light and the overall image of the face in supplementary light, respectively
  • the extracted features are the first feature f1 and the second feature f2, respectively, and the lengths of f1 and f2 are 256.
  • Net2 is the second convolutional neural network, which is used for feature extraction of local areas, and extracts the features of ambient light partial face images and supplementary light face partial images respectively.
  • eye_b and eye_f are the partial face image of the ambient light and the partial face image of the supplementary light respectively
  • the extracted features are the third feature f3 and the fourth feature f4 respectively
  • the lengths of f3 and f4 are 256 respectively.
  • the face recognition device on the mobile terminal can fuse the features of the overall image of the face with ambient light, the overall image of the face with fill light, the partial image of the face with ambient light and the partial image of the face with fill light, and fully tap the 3D image of the overall face.
  • Information such as eye spots and other information, are used to distinguish prosthetic faces from live faces.
  • the face recognition device at the mobile terminal fuses the first feature, the second feature, the third feature, and the fourth feature according to a preset feature fusion algorithm to obtain a fusion feature.
  • S160 Classify by fusing features, and judge whether the ambient light face image and the supplementary light face image are from a living body.
  • the face recognition device on the mobile terminal performs classification through fusion features, and judges whether the ambient light face image and the supplementary light face image are from a living body. Specifically, by analyzing the fusion features, the mobile face recognition device can obtain the degree of change between the partial image of the face in ambient light and the partial image of the face in supplementary light, as well as the difference between the overall image of the face in ambient light and the overall image of the face in supplementary light. The degree of variation to distinguish live and prosthetic faces. The smaller the degree of change, the more similar the ambient light face image and fill light face image are, and the more it indicates that the ambient light face image and fill light face image come from a prosthetic face. The greater the degree of change, the less similar the ambient light face image and the supplementary light face image are, and the more it indicates that the ambient light face image and the supplementary light face image are from living human faces.
  • the face image under ambient light and supplementary light is acquired through the mobile terminal face recognition device, the overall area and local area of the face image under ambient light and supplementary light are obtained by preprocessing, and the convolution neural
  • the network extracts features and fuses the extracted features to classify live faces and prosthetic faces.
  • using this method using the characteristics of local changes in the live face when it encounters a point light source, analyze the local changes of the face before and after light supplement, and use the overall imaging change characteristics of the live face before and after light supplement to analyze the face image before and after light supplement Changes in liveness detection, high accuracy, simple and convenient, no active cooperation from users, and good user experience.
  • the feature fusion method adopted can be feature stacking, and the four feature connections (Concat) can be directly fused, or the four features can also be fused using a convolution method.
  • the embodiment of the present application is for feature fusion The method is not limited.
  • the fusion feature includes the overall information of the face image under ambient light and local information such as eyes, and the overall information of the face image under supplementary light and local information such as eyes, so as to distinguish the face image in the environment. Changes that occur under light and fill light (light source changes).
  • step 110 includes:
  • S301 Collect an RGB image under ambient light through a face lens to obtain a face image under ambient light.
  • the face recognition device can fill in the light through the fill light. If the face recognition device is a mobile face recognition device such as a mobile phone, it can also use its built-in screen fill light without adding additional hardware devices, and Very convenient to implement.
  • the screen fill light can maximize the brightness of the screen, and except for the face area, the pixels are set to be all white or one of red, yellow, and blue, or flash alternately.
  • the ambient light face image After the ambient light face image is successfully collected, turn on the fill light or the screen for fill light, collect multiple fill light face images within the time period t2, and perform face detection at the same time, the face will be detected and the image quality meets the requirements The ambient light face image is retained. If no face is detected within the time period t2 or the image quality does not meet the requirements, it will return a failure, reminding the user to adjust the acquisition position or acquisition posture and re-acquire the fill-in face image.
  • the method further includes:
  • the face recognition device at the mobile terminal selects the qualified fill-light face image and ambient light face image from the ambient light face image and fill-in face image with the head posture as a frontal face according to the image quality, and more Good increases algorithm robustness.
  • the face image with supplementary light and the face image with ambient light that meet the conditions can be selected through traditional methods such as image signal-to-noise ratio, or through a face image quality judgment algorithm, which is not limited in this application.
  • the fill-in light face image and the ambient light face image with a quality score higher than a set quality score threshold may be selected.
  • the ambient light face image and the fill light face image are preprocessed to obtain the ambient light face overall image, the ambient light partial face image, the fill light overall face image and the fill light face partial image.
  • Prosthetic faces generally include printed face photos, face photos or videos played on the screen after duplication, etc. In this application, it is necessary to distinguish prosthetic faces from live faces to achieve liveness detection.
  • the eyes in the prosthetic face are not sensitive to point light sources, so the eyes in the prosthetic face have no difference before and after the light supplement, and there are light spots in the pupils before and after the light supplement, or there are no light spots at the same time, such as Figure 5 shows. That is to say, there is a light spot change in the pupil before and after the real face is filled with light, but there is no light spot change in the pupil before and after the prosthetic face is filled with light. Because the prosthetic face is flat, its cheeks, nose, forehead, mouth and other parts have no shadows under ambient light and fill light conditions. A real human face and a prosthetic human face can be distinguished through the above characteristics.
  • the present application obtains the partial image of the face in ambient light and the partial image of the face in supplementary light through preprocessing, and uses the characteristics of changes in the living face under the conditions of ambient light and supplementary light to distinguish the live face from the prosthetic face.
  • the smaller the change between the ambient light face partial image and the fill light face partial image that is, the more similar the ambient light face partial image and the fill light face partial image
  • the characteristics of pupils forming spots when they meet point light sources it is possible to analyze whether there are changes in pupil spots in partial face images under ambient light and supplementary light, so as to distinguish live faces from prosthetic faces.
  • the surface of the printed face photo and the screen surface on which the face photo or video is played are basically flat. From the light reflection principle, the reflection angles of each position on the plane are basically the same. Therefore, under the supplementary light, the flat prosthetic face will not show the characteristics of real face imaging (that is, the parts of the real face such as the tip of the nose and forehead that are closer to the lens are brighter, and the parts that are farther away from the lens, such as eye sockets and cheeks, are brighter. darker), resulting in a smaller change in the imaging of the prosthetic face before and after the supplementary light.
  • the 3D information of the face constructed by the distorted photo is relatively simple, which is far from the 3D information of the real face, and the distance between the tip of the nose and the forehead of the real face cannot be achieved.
  • the imaging of the real face has a large change before and after filling light, while the imaging change of the prosthetic face is small before and after filling light.
  • real faces and prosthetic faces can be distinguished.
  • the present application obtains the overall image of the face with ambient light and the overall image of the face with supplementary light through preprocessing, and analyzes the changes in the overall image of the face with ambient light and the overall image of the face with supplementary light to distinguish living human faces from prosthetic human faces.
  • the smaller the change between the overall image of the ambient light face and the overall image of the fill light face that is, the more similar the overall image of the ambient light face and the overall image of the fill light face), the more it indicates that the ambient light face image and the fill light face image are from Prosthetic face.
  • the aforementioned first convolutional neural network and the second convolutional neural network need to be trained before use, as shown in Figure 6, the training method is as follows:
  • S10 Construct an overall face image sample with ambient light, a partial face image sample with ambient light, an overall face image sample with fill light, and a partial face image sample with fill light, and set labels respectively.
  • training sample pair includes a prosthetic face image pair composed of an ambient light face image and a supplementary light face image of a prosthetic face and an ambient light human face image composed of a live face.
  • a live face image pair composed of a face image and a fill-in face image.
  • the label of a pair of prosthetic face images can be set to 1, indicating that they are similar and are prosthetics, and the label of a pair of live face images can be set to 0, indicating that they are not similar and are living.
  • the training sample pairs are preprocessed to obtain an overall face image sample under ambient light, a partial face image sample under ambient light, an overall face image sample with supplementary light, and a partial image sample with supplemental light.
  • S20 Use the first convolutional neural network to perform feature extraction on the overall image sample of the face with ambient light and the overall image sample of the face with supplementary light to obtain the fifth feature and the sixth feature; Feature extraction is performed on the partial image samples and the partial image samples of the fill-in face to obtain the seventh feature and the eighth feature.
  • the feature extraction method in this step is the same as the feature extraction method in S120, which can be understood by referring to the content described in step S120, and will not be repeated here.
  • This application uses the same convolutional neural network (i.e., the first convolutional neural network) to train and extract features for the overall image sample of the face in ambient light and the overall image sample in supplementary light, and the weights are shared.
  • the neural network is a Siamese Network, which uses the Siamese network to analyze and extract features in pairs of samples, and more specifically mines the imaging difference of the face before and after filling the light, and uses the difference for classification. Ambient light face partial image samples and fill light partial face image samples are the same.
  • Contrastive Loss that is, the contrast loss
  • its expression is as follows:
  • L is the contrast loss
  • d
  • Softmax can be used to classify the fusion feature fc, where the classification result 1 represents the prosthesis and 0 represents the living body.
  • this contrastive loss is used to measure whether two samples of the same person's overall face image and face partial image match under ambient light and fill light.
  • the mismatch represents a pair of real samples, and the match represents a pair of fake samples.
  • f1 and f2, f3 and f4 compute the contrastive losses L 1 and L 2 in pairs, respectively.
  • S40 Fusing the fifth feature, the sixth feature, the seventh feature, and the eighth feature to obtain a fusion sample feature, and calculating a cross-entropy loss L 3 of the fusion sample feature.
  • step S130 The method of performing feature fusion in this step is the same as that in S130, which can be understood by referring to the content described in step S130, and will not be repeated here.
  • the fifth feature, the seventh feature, the sixth feature and the eighth feature are sequentially connected to obtain the fused sample feature.
  • This application uses the cross entropy loss function (Cross Entropy Loss) for the classification of fusion sample features, and calculates the cross entropy loss L 3 , as shown in FIG. 7 .
  • Cross entropy loss function Cross Entropy Loss
  • the training set consisting of training sample pairs can be obtained by the following method, as shown in Figure 8:
  • the live face is in front of the face camera.
  • face images turn on the face lens, collect face images of live faces under ambient light, detect faces through the face detection algorithm and judge that the image quality meets the requirements, continue to collect, and continuously obtain n_b faces image to obtain the ambient light face image of the live face.
  • S12 Acquiring several fill-in light face images of living human faces collected by the face lens under the fill light or the fill light of the screen.
  • the live face is in front of the face camera.
  • Control the fill light or screen highlight through the upper application App and collect the face image of the live face in the fill light environment.
  • n_f face images After reaching the number of collections, turn off the fill light or return to the normal state of the screen, and get the fill light of the live face face image.
  • each pair of live face images includes ambient light face images and fill light face images of live faces collected at the same time period, from n_b ambient light face images and n_f
  • Each of the supplementary light face images randomly selects one image multiple times to form a live face image pair.
  • the prosthetic face is in front of the face camera, and the acquisition method is the same as that of S11, which can be understood by referring to the content described in step S11, and will not be repeated here.
  • S15 Acquire several fill-in face images of prosthetic faces captured by the face lens under fill-in light or screen fill-in light.
  • the prosthetic face is in front of the face camera, and the acquisition method is the same as that of S12, which can be understood by referring to the content described in step S12, and will not be repeated here.
  • step S13 The implementation method of this step is the same as that of S13, which can be understood with reference to the content described in step S13, and will not be repeated here.
  • the partial human face of the present application may be partial areas of the human face such as eyes, cheeks, nose, forehead, mouth, etc., preferably the eye area.
  • the ambient light face image and the supplementary light face image can be preprocessed by the method shown in Figure 9:
  • S910 Perform face detection, face key point positioning, head pose estimation and face normalization on the ambient light face image and fill light face image, and obtain the ambient light overall face image and fill light overall face image .
  • the MTCNN detection algorithm is used to realize face detection and face feature point positioning .
  • the 3D pose information is estimated by the 2D coordinate information of the key points. Normalize the face through the coordinates of both eyes, including face alignment and scaling, align the eyes to (38,52) and (74,52), and the scaling size is (112,112), as shown in Figure 10.
  • S920 Taking the coordinates of the eyes obtained by locating the key points of the face as the center, expanding several pixels around to obtain a region of interest including the eyes, wherein the number of pixels expanded around is determined according to the distance between the left and right eyes.
  • Eye ROI region of interest
  • , x right is the abscissa of the right eye, and x left is the abscissa of the left eye.
  • the number of pixels expanded to the surrounding area is d iris /3.
  • S940 Intercept and obtain a partial face image in ambient light and a partial face image in supplementary light according to the center and radius of the eyes.
  • the eye circle center (taking the left eye as an example, the circle center is x l_center , y l_center ) as the center, intercepting an image with an image width of 4.4R and a height of 2.2R;
  • the left-eye image and right-eye image are scaled to 112*56, and the scaled left-eye image and right-eye image are stitched and stacked together to obtain a partial face image with ambient light and a partial face image with fill light, as shown in Figure 11 shown.
  • the processing method is the same as the fish hammer method for ambient light face images and fill light face images.
  • the embodiment of the present application provides a face detection device 1200, as shown in Figure 12, the device 1200 includes:
  • the image acquisition module 1201 is configured to acquire the ambient light face image and fill light face image within the same user-preset time period.
  • the first preprocessing module 1202 is configured to preprocess the face image under ambient light to obtain an overall image of the face under ambient light and a partial image of the face under ambient light.
  • the second preprocessing module 1203 is configured to perform preprocessing on the fill-in face image to obtain the fill-in full face image and the fill-in face partial image.
  • the image feature extraction module 1204 is used to perform feature extraction on the overall image of the face with ambient light and the overall image of the face with supplementary light through the first convolutional neural network to obtain the first feature and the second feature; Feature extraction is performed on the partial face image under ambient light and the partial face image under supplementary light to obtain the third feature and the fourth feature.
  • the image feature fusion module 1205 is configured to fuse the first feature, the second feature, the third feature and the fourth feature to obtain a fusion feature.
  • the living body detection module 1206 is used for performing classification by fusion features, and judging whether the ambient light face image and the supplementary light face image are from living bodies.
  • This application obtains the face image under ambient light and supplementary light, obtains the overall area and local area of the face image under ambient light and supplementary light through preprocessing, and extracts features through a convolutional neural network, and fuses the extracted features Perform live face and prosthetic face classification.
  • This application uses the characteristics of local changes in the living face when it encounters a point light source to analyze the local changes of the face before and after the light supplement, and uses the overall change characteristics of the imaging of the living face before and after the light supplement to analyze the overall face image before and after the light supplement. Changes, liveness detection, high accuracy, simple and convenient, does not require active cooperation from users, and has a good user experience.
  • the face detection device 1200 also includes:
  • the sample construction module is used to construct the overall image sample of the face with ambient light, the partial image sample of the face with ambient light, the overall image sample of the face with fill light and the partial image sample of the face with fill light belonging to the same living body or prosthesis.
  • the loss calculation module is used to train the first convolutional neural network according to the overall image sample of the face with ambient light and the overall image sample of the face with fill light, and to train the first convolutional neural network according to the partial image sample of the face with ambient light and the partial image sample of the face with fill light , train the second convolutional neural network, and obtain the overall loss during the model training process of the first convolutional neural network and the second convolutional neural network.
  • the backpropagation module is used to update the parameters of the first convolutional neural network and the second convolutional neural network through backpropagation until the overall loss meets the preset loss condition, the first convolutional neural network and the second convolutional neural network Network training is complete.
  • the aforementioned first convolutional neural network and the second convolutional neural network need to be trained before use, which are obtained through the following module training:
  • the sample construction module is used to construct the overall image sample of the face with ambient light, the partial image sample of the face with ambient light, the overall image sample of the face with fill light and the partial image sample of the face with fill light, and set labels respectively.
  • the sample feature extraction module is used to perform feature extraction on the overall image sample of the ambient light face and the overall image sample of the face with supplementary light through the first convolutional neural network to obtain the fifth feature and the sixth feature; through the second convolutional neural network Feature extraction is performed on the partial face image sample under ambient light and the partial face image sample under supplementary light to obtain the seventh feature and the eighth feature.
  • a comparison loss calculation module configured to calculate a comparison loss L 1 between the fifth feature and the sixth feature and a comparison loss L 2 between the seventh feature and the eighth feature.
  • the cross-entropy loss calculation module is used to fuse the fifth feature, the sixth feature, the seventh feature and the eighth feature to obtain the fusion sample features, and calculate the cross-entropy loss L 3 of the fusion sample features.
  • the backpropagation module is used to weight and sum L 1 , L 2 and L 3 to obtain the overall loss L All , and update the parameters of the first convolutional neural network and the second convolutional neural network through backpropagation.
  • the aforementioned sample construction modules include:
  • the first acquiring unit is used to acquire ambient light face images of several living human faces captured by the face lens under ambient light.
  • the second acquiring unit is used to acquire the fill-in light face images of several living human faces captured by the face lens under the fill-in light or the fill-in light of the screen.
  • the living body face image pair acquisition unit is used to randomly extract an image from the ambient light face image and the supplementary light face image of the living body face respectively to form a living body face image pair.
  • the third acquisition unit is used to acquire the ambient light face images of several prosthetic faces collected by the face lens under ambient light.
  • the fourth acquisition unit is used to acquire the fill-in face images of several prosthetic faces captured by the face lens under the fill-in light or the fill-in light of the screen.
  • the prosthetic face image pair acquisition unit is used to randomly extract an image from the ambient light face image and the supplementary light face image of the prosthetic face to form a prosthetic face image pair.
  • the fusion method of the first feature, the third feature, the second feature and the fourth feature there is no limit to the fusion method of the first feature, the third feature, the second feature and the fourth feature.
  • the fifth feature, the seventh feature, the sixth feature and the eighth feature are sequentially connected, Get the fusion sample features.
  • the first feature, the third feature, the second feature and the fourth feature are sequentially connected to obtain a fusion feature.
  • the ambient light partial face image and supplementary light partial face image of this application are eye images.
  • the preprocessing of the ambient light face image and supplementary light face image includes:
  • the face processing unit is used to perform face detection, face key point positioning, head pose estimation and face normalization on the ambient light face image and the supplementary light face image, and obtain the overall image of the ambient light face and the supplementary face image. Overall image of the light face.
  • the eye region of interest acquisition unit is used to expand a number of pixels around the center of the eye coordinates obtained by positioning the key points of the face to obtain the region of interest including the eyes, wherein the number of pixels expanded to the surroundings is determined according to the distance between the left and right eyes.
  • a pupil location unit for determining the center and radius of the eye using a radially symmetric transformation within the region of interest.
  • the image intercepting unit is used to intercept and obtain the partial image of the human face in ambient light and the partial image of the human face in supplementary light according to the center and radius of the eyes.
  • the method described in the above-mentioned embodiment 1 provided by this application can realize the business logic through a computer program and record it on a storage medium, and the storage medium can be read and executed by a computer to realize the embodiment of this specification 1 Effect of the described regimen. Therefore, the present application also provides a computer-readable storage medium for live face detection, including a memory for storing processor-executable instructions. When the instructions are executed by the processor, the steps of the face live detection method in Embodiment 1 are realized. .
  • This application obtains the face image under ambient light and supplementary light, obtains the overall area and local area of the face image under ambient light and supplementary light through preprocessing, and extracts features through a convolutional neural network, and fuses the extracted features Perform live face and prosthetic face classification.
  • This application uses the characteristics of local changes in the living face when it encounters a point light source to analyze the local changes of the face before and after the light supplement, and uses the overall change characteristics of the imaging of the living face before and after the light supplement to analyze the overall face image before and after the light supplement. Changes, liveness detection, high accuracy, simple and convenient, does not require active cooperation from users, and has a good user experience.
  • the storage medium may include a physical device for storing information, and information is usually digitized and then stored using an electrical, magnetic, or optical medium. Described storage medium can include: the device that utilizes electric energy mode to store information such as, various memory, as RAM, ROM etc.; USB stick; a device that stores information optically, such as a CD or DVD. Of course, there are other readable storage media, such as quantum memory, graphene memory and so on.
  • the above-mentioned storage medium may also include other implementations according to the description of the method embodiment 1.
  • the implementation principle and technical effect of this embodiment are the same as those of the aforementioned method embodiment 1.
  • please refer to the description of the related method embodiment 1. will not be repeated here.
  • the present application also provides a device for live face detection.
  • the device may be a separate computer, or may include one or more of the methods or one or more of the methods described in this specification.
  • the device for live face detection may include at least one processor and a memory storing computer-executable instructions. When the processor executes the instructions, any one or more steps of the method for live face detection in Embodiment 1 above are implemented. .
  • This application obtains the face image under ambient light and supplementary light, obtains the overall area and local area of the face image under ambient light and supplementary light through preprocessing, and extracts features through a convolutional neural network, and fuses the extracted features Perform live face and prosthetic face classification.
  • This application uses the characteristics of local changes in the living face when it encounters a point light source to analyze the local changes of the face before and after the light supplement, and uses the overall change characteristics of the imaging of the living face before and after the light supplement to analyze the overall face image before and after the light supplement. Changes, liveness detection, high accuracy, simple and convenient, does not require active cooperation from users, and has a good user experience.
  • the above-mentioned device may also include other implementations 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.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Human Computer Interaction (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Image Processing (AREA)

Abstract

本申请公开了一种人脸活体检测方法、装置、可读存储介质及设备，属于人脸识别领域。本申请获取环境光和补光下的人脸图像，通过预处理得到环境光和补光下的人脸图像的整体区域和局部区域，并通过卷积神经网络提取特征，将提取的特征融合后进行活体人脸和假体人脸分类。

Description

人脸活体检测方法、装置、计算机可读存储介质及设备
本申请要求于2021年10月15日申请的，申请号为202111202019.0、名称为“人脸活体检测方法、装置、计算机可读存储介质及设备”的中国专利申请的优先权，在此将其全文引入作为参考。
技术领域
本申请涉及人脸识别领域，特别是指一种人脸活体检测方法、装置、计算机可读存储介质及设备。
背景技术
目前，人脸识别技术已被广泛应用于金融和安全领域。由于人脸具有获取方便，非接触等优点，但也极易被他人利用，以照片或翻拍视频的方式，攻破人脸识别系统。因此，人脸活体检测技术作为人脸识别技术第一道门槛，显的尤为重要。
目前，移动端人脸活体检测主要有三种方式，第一种方式是人脸动作活体检测，活体检测系统下发随机头脸部动作指令，用户按指令完成相应的动作则判断为活体；第二种方式是对RGB(Red Green Blue，颜色系统)图像提取用于判别真实人脸和假体人脸的特征，做真实人脸和假体人脸的二分类；第三种方式是前两种方式的整合，在利用第一种方式进行人脸动作活体检测的过程中，为第二种方式的人脸活体检测提供多张多表情的RGB图像用于判别是否为真实人脸。
人脸动作活体检测需要用户高度配合，且下发的随机头脸部动作指令一般为单一的动作指令，比如点头、转头、眨眼和张嘴等，不法分子通过抖动、扭曲或旋转照片等动作有很大概率能欺骗通过活体检测系统，或者拍摄用户的多个动作的视频同样能较轻松通过活体检测系统。
第二种方式中对RGB图像提取用于判断真实人脸和假体人脸的特征进行活体检测的方式为静默活体检测，这种活体检测方式无需用户配合，更容易被用户所接受。提取用于判断真实人脸和假体人脸的特征时一般使用深度学习方法，通过大量的活体和非活体人脸数据驱动，自动学习能够有效判别真实人脸和假体人脸的特征，区分真假人脸成像差异。但是RGB图像由于光照、分辨率、不同手机镜头等成像差异，容易导致误判。
发明内容
根据本申请的各种实施例，提供一种人脸活体检测方法、装置、可读存储介质及设备，提高 了人脸活体检测的准确率，并且不需要用户主动配合，用户体验好。
本申请提供技术方案如下：
第一方面，本申请提供一种人脸活体检测方法，所述方法包括：
获取同一用户预设时间段内的环境光人脸图像和补光人脸图像；
对所述环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像；
对所述补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像；
通过第一卷积神经网络对所述环境光人脸整体图像和所述补光人脸整体图像进行特征提取，得到第一特征和第二特征；通过第二卷积神经网络对所述环境光人脸局部图像和所述补光人脸局部图像进行特征提取，得到第三特征和第四特征；
将所述第一特征、所述第二特征、所述第三特征和所述第四特征进行融合，得到融合特征；
通过所述融合特征对所述用户进行分类，判断所述环境光人脸图像和所述补光人脸图像是否来自活体。
在其中一个实施例中，所述第一卷积神经网络和所述第二卷积神经网络通过如下方法训练得到：
构造隶属于同一活体或假体的环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本；
根据所述环境光人脸整体图像样本和所述补光人脸整体图像样本，对所述第一卷积神经网络进行训练，以及根据所述环境光人脸局部图像样本和所述补光人脸局部图像样本，对所述第二卷积神经网络进行训练，得到所述第一卷积神经网络和所述第二卷积神经网络的模型训练过程中的总体损失；
通过反向传播更新所述第一卷积神经网络和所述第二卷积神经网络的参数，直至所述总体损失满足预设的损失条件，所述第一卷积神经网络和所述第二卷积神经网络训练完成。
在其中一个实施例中，所述方法还包括：
将隶属于同一活体或假体的环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本分别预设置有标签；所述标签用于标识所述多个样本中的图像隶属于同一活体还是同一假体。
在其中一个实施例中，所述根据所述环境光人脸整体图像样本和所述补光人脸整体图像样本，对所述第一卷积神经网络进行训练，以及根据所述环境光人脸局部图像样本和所述补光人脸局部 图像样本，对所述第二卷积神经网络进行训练，得到所述第一卷积神经网络和所述第二卷积神经网络的模型训练过程中的总体损失，包括：
通过所述第一卷积神经网络对所述环境光人脸整体图像样本和所述补光人脸整体图像样本进行特征提取，得到第五特征和第六特征；通过所述第二卷积神经网络对所述环境光人脸局部图像样本和所述补光人脸局部图像样本进行特征提取，得到第七特征和第八特征；
计算所述第五特征与所述第六特征的第一对比损失L 1以及所述第七特征与所述第八特征的第二对比损失L 2；
将所述第五特征、所述第六特征、所述第七特征和所述第八特征进行融合，得到融合样本特征，并根据所述融合样本特征，以及所述标签计算所述融合样本特征的交叉熵损失L 3；
将所述第一对比损失L 1、所述第二对比损失L 2和所述交叉熵损失L 3加权求和，得到总体损失L All。
在其中一个实施例中，所述将所述第五特征、所述第六特征、所述第七特征和所述第八特征进行融合，得到融合样本特征，包括：
所述第五特征、所述第七特征、所述第六特征和所述第八特征顺序连接，得到所述融合样本特征。
在其中一个实施例中，所述将所述第一特征、所述第二特征、所述第三特征和所述第四特征进行融合，得到融合特征，包括：
将所述第一特征、所述第三特征、所述第二特征和所述第四特征顺序连接，得到所述融合特征；所述融合特征用于区分活体人脸和假体人脸，所述融合特征包含人脸整体3D信息和眼部的光斑信息。
在其中一个实施例中，所述环境光人脸局部图像和所述补光人脸局部图像为眼睛图像；所述第三特征和所述第四特征包括眼部的光斑信息。
在其中一个实施例中，所述对所述环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像，包括：
对所述环境光人脸图像进行人脸检测、人脸关键点定位、头部姿态估计和人脸归一化，得到环境光人脸整体图像；
以人脸关键点定位得到的眼睛坐标为中心，向四周扩充若干像素，得到包含眼睛的感兴趣区域，其中，向四周扩充的像素数根据左右眼间距确定；
在所述感兴趣区域内，使用径向对称变换确定眼睛中心和半径；
根据所述眼睛中心和所述半径截取得到环境光人脸局部图像。
在其中一个实施例中，所述对所述补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像，包括：
对所述补光人脸图像进行人脸检测、人脸关键点定位、头部姿态估计和人脸归一化，得到补光人脸整体图像；
以人脸关键点定位得到的眼睛坐标为中心，向四周扩充若干像素，得到包含眼睛的感兴趣区域，其中，向四周扩充的像素数根据左右眼间距确定；
在所述感兴趣区域内，使用径向对称变换确定眼睛中心和半径；
根据所述眼睛中心和所述半径截取得到补光人脸局部图像。
在其中一个实施例中，所述获取同一用户预设时间段内的环境光人脸图像和补光人脸图像，包括：
通过人脸镜头在环境光下采集RGB图像，得到所述环境光人脸图像；
通过补光灯进行补光，在补光条件下采集RGB图像，得到所述补光人脸图像。
在其中一个实施例中，所述方法还包括：
通过关键点回归得到多幅所述环境光人脸图像和所述补光人脸图像的头部姿态；
选择所述头部姿态为正脸的所述环境光人脸图像和所述补光人脸图像。
第二方面，本申请提供一种人脸活体检测装置，所述装置包括：
图像获取模块，用于获取同一用户预设时间段内的环境光人脸图像和补光人脸图像；
第一预处理模块，用于对所述环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像；
第二预处理模块，用于对所述补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像；
图像特征提取模块，用于通过第一卷积神经网络对所述环境光人脸整体图像和所述补光人脸整体图像进行特征提取，得到第一特征和第二特征；通过第二卷积神经网络对所述环境光人脸局部图像和所述补光人脸局部图像进行特征提取，得到第三特征和第四特征；
图像特征融合模块，用于将所述第一特征、所述第二特征、所述第三特征和所述第四特征进行融合，得到融合特征；
活体检测模块，用于通过所述融合特征进行分类，判断所述环境光人脸图像和所述补光人脸图像是否来自活体。
在其中一个实施例中，所述第一卷积神经网络和第二卷积神经网络通过如下模块训练得到：
样本构造模块，用于构造环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本，并分别设置标签；
样本特征提取模块，用于通过第一卷积神经网络对所述环境光人脸整体图像样本和所述补光人脸整体图像样本进行特征提取，得到第五特征和第六特征；通过第二卷积神经网络对所述环境光人脸局部图像样本和所述补光人脸局部图像样本进行特征提取，得到第七特征和第八特征；
对比损失计算模块，用于计算所述第五特征与所述第六特征的对比损失L 1以及所述第七特征与所述第八特征的对比损失L 2；
交叉熵损失计算模块，用于将所述第五特征、所述第六特征、所述第七特征和所述第八特征进行融合，得到融合样本特征，并计算所述融合样本特征的交叉熵损失L 3；
反向传播模块，用于将所述对比损失L 1、所述对比损失L 2和所述交叉熵损失L 3加权求和，得到总体损失L All，并通过反向传播更新所述第一卷积神经网络和所述第二卷积神经网络的参数。
在其中一个实施例中，所述第五特征、所述第七特征、所述第六特征和所述第八特征顺序连接，得到所述融合样本特征；
所述第一特征、所述第三特征、所述第二特征和所述第四特征顺序连接，得到融合特征。
在其中一个实施例中，所述环境光人脸局部图像和所述补光人脸局部图像为眼睛图像。
在其中一个实施例中，所述装置还包括：
图像处理模块，用于对所述环境光人脸图像和补光人脸图像进行人脸检测、人脸关键点定位、头部姿态估计和人脸归一化，得到环境光人脸整体图像和补光人脸整体图像；
图像区域确定模块，用于以人脸关键点定位得到的眼睛坐标为中心，向四周扩充若干像素，得到包含眼睛的感兴趣区域，其中，向四周扩充的像素数根据左右眼间距确定；
图像信息确定模块，用于在所述感兴趣区域内，使用径向对称变换确定眼睛中心和半径；
图像截取模块，用于根据眼睛中心和半径截取得到环境光人脸局部图像和补光人脸局部图像。
在其中一个实施例中，所述图像获取模块用于通过人脸镜头在环境光下采集RGB图像，得到所述环境光人脸图像；
通过补光灯进行补光，在补光条件下采集RGB图像，得到所述补光人脸图像。
第三方面，本申请提供一种用于人脸活体检测的计算机可读存储介质，包括用于存储处理器可执行指令的存储器，所述指令被所述处理器执行时实现包括第一方面所述的人脸活体检测方法的步骤。
第四方面，本申请提供一种用于人脸活体检测的设备，包括至少一个处理器以及存储计算机可执行指令的存储器，所述处理器执行所述指令时实现第一方面所述的人脸活体检测方法的步骤。
本申请具有以下有益效果：
本申请获取环境光和补光下的人脸图像，通过预处理得到环境光和补光下的人脸图像的整体区域和局部区域，并通过卷积神经网络提取特征，将提取的特征融合后进行活体人脸和假体人脸分类。本申请利用活体人脸遇点光源后局部发生变化的特性，分析补光前后人脸局部的变化，并利用活体人脸在补光前后的成像整体变化特性，分析补光前后人脸图像的整体变化，进行活体检测，准确率高，简单方便，不需要用户主动配合、用户体验好。
附图说明
为了更好地描述和说明这里公开的那些发明的实施例，可以参考一幅或多幅附图。用于描述附图的附加细节或示例不应当被认为是对所公开的发明、目前描述的实施例以及目前理解的这些发明的最佳模式中的任何一者的范围的限制。
图1为一示例性实施例中提供的一种人脸活体检测方法的流程示意图；
图2为一示例性实施例中提供的一种卷积神经网络训练的示意图；
图3为一示例性实施例中提供的一种人脸图像采集步骤的流程示意图；
图4为一以示例性实施例中提供的一种人脸图像筛选步骤的流程示意图；
图5为一示例实施例性实施例中提供的一种预处理过程示意图；
图6为一示例性实施例中提供的一种第一卷积神经网络和第二卷积神经网络的训练方法流程示意图；
图7为一示例性实施例中提供的一种卷积神经网络训练的示意图；
图8为一示例性实施例中提供的一种训练集的获取方法流程示意图；
图9为一示例性实施例中提供的一种对环境光人脸图像和补光人脸图像进行预处理的方法流程示意图；
图10为一示例性实施例中提供的一种预处理过程示意图；
图11为一示例性实施例中提供的一种预处理过程示意图；
图12为一实施例性实施例中提供的一种人脸活体检测装置的结构示意图。
具体实施方式
为使本申请要解决的技术问题、技术方案和优点更加清楚，下面将结合附图及具体实施例对本申请的技术方案进行清楚、完整地描述。显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。通常在此处附图中描述和示出的本申请实施例的组件可以以各种不同的配置来布置和设计。因此，以下对在附图中提供的本申请的实施例的详细描述并非旨在限制要求保护的本申请的范围，而是仅仅表示本申请的选定实施例。基于本申请的实施例，本领域技术人员在没有做出创造性劳动的前提下所获得的所有其他实施例，都属于本申请保护的范围。
在其中一个实施例中，本申请实施例提供一种人脸活体检测方法，可以用于人脸识别设备，尤其适用于移动端人脸识别设备，该移动端人脸识别设备具有屏幕或补光灯，以便在人脸活体检测时进行补光，该移动端人脸识别设备包括但不限于智能手机、笔记本电脑、平板电脑、掌上电脑等。
图1为本申请的人脸活体检测方法的流程示意图，该方法包括：
S110：获取同一用户预设时间段内的环境光人脸图像和补光人脸图像。
在实施中，移动端人脸识别设备获取同一用户预设时间段内的环境光人脸图像和补光图像。通过分析环境光人脸局部图像和补光人脸局部图像的变化(例如瞳孔内光斑的变化)，以及分析环境光人脸整体图像和补光人脸整体图像的变化，以区分活体人脸和假体人脸。
S120：对环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像。
在实施中，移动端人脸识别设备对环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像。具体地，针对环境光人脸图像进行预处理的过程主要包括人脸检测、人脸关键点定位、头部姿态估计和人脸归一化等，经过预处理过程对环境光人脸图像进行处理，可以得到环境光人脸图像对应的环境光人脸整体图像和环境光人脸局部图像。其中，针对环境光人脸图像的预处理过程，可以根据实际应用过程进行适应性调整，本申请实施例对于具体的预处理过程不做限定。
S130：对补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像。
在实施中，移动端人脸识别设备对补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像。其中，对补光人脸图像进行预处理的过程也可以包括人脸检测、人脸关键点定位、头部姿态估计和人脸归一化等，与步骤S120中对环境光人脸图像进行预处理的过程相似，本申请实施例不再赘述。
S140：通过第一卷积神经网络对环境光人脸整体图像和补光人脸整体图像进行特征提取，得到第一特征和第二特征；通过第二卷积神经网络对环境光人脸局部图像和补光人脸局部图像进行特征提取，得到第三特征和第四特征。
其中，由于本申请实施例需要使用人脸整体区域和局部区域信息，因此用到两个卷积神经网络网络，即第一卷积神经网络和第二卷积神经网络，第一卷积神经网络和第二卷积神经网络可以选用轻量或者小型网络，例如采用的Siamese Network的Base网络。第一卷积神经网络和第二卷积神经网络的结构可以相同，并且基于算法效率的需求，第二卷积神经网络的数量处理能力可相对较小，更加轻量化，以此提升算法效率。
在实施中，移动端人脸识别设备通过卷积神经网络分析环境光人脸局部图像和补光人脸局部图像的变化，以及分析环境光人脸整体图像和补光人脸整体图像的变化。具体地，移动端人脸识别设备通过第一卷积神经网络对环境光人脸整体图像和补光人脸整体图像进行特征提取，得到第一特征和第二特征。然后，又通过第二卷积神经网络对环境光人脸局部图像和补光人脸局部图像 进行特征提取，得到第三特征和第四特征。
可选的，第一卷积神经网络和第二卷积神经网络的一种结构实施例如图2所示，其中Net1为第一卷积神经网络，用于对人脸整体区域进行特征提取，分别提取环境光人脸整体图像和补光人脸整体图像的特征。其中face_b和face_f分别为环境光人脸整体图像和补光人脸整体图像，提取的特征分别为第一特征f1和第二特征f2，f1和f2的长度各为256。Net2为第二卷积神经网络，用于对局部区域进行特征提取，分别提取环境光人脸局部图像和补光人脸局部图像的特征。其中eye_b和eye_f分别为环境光人脸局部图像和补光人脸局部图像，提取的特征分别为第三特征f3和第四特征f4，f3和f4的长度各为256。
S150：将第一特征、第二特征、第三特征和第四特征进行融合，得到融合特征。
在实施中，移动端人脸识别设备可以将环境光人脸整体图像、补光人脸整体图像、环境光人脸局部图像和补光人脸局部图像的特征融合，充分挖掘人脸整体的3D信息以及眼部的光斑等信息，用于区分假体人脸和活体人脸。具体地，移动端人脸识别设备根据预设的特征融合算法，将第一特征、第二特征、第三特征和第四特征进行融合，得到融合特征。
其中，特征融合算法的具体内容，在下述实施例中进行详细列举，在此不再赘述。
S160：通过融合特征进行分类，判断环境光人脸图像和补光人脸图像是否来自活体。
在实施中，移动端人脸识别设备通过融合特征进行分类，判断环境光人脸图像和补光人脸图像是否来自活体。具体地，移动端人脸识别设备通过对融合特征进行分析，即可得到环境光人脸局部图像和补光人脸局部图像的变化程度以及环境光人脸整体图像和补光人脸整体图像的变化程度，以区分活体人脸和假体人脸。其变化程度越小，说明环境光人脸图像和补光人脸图像越相似，越表明环境光人脸图像和补光人脸图像来自假体人脸。其变化程度越大，说明环境光人脸图像和补光人脸图像越不相似，越表明环境光人脸图像和补光人脸图像来自活体人脸。
本申请实施例中，通过移动端人脸识别设备获取环境光和补光下的人脸图像，预处理得到环境光和补光下的人脸图像的整体区域和局部区域，并通过卷积神经网络提取特征，将提取的特征融合后进行活体人脸和假体人脸分类。采用本方法，利用活体人脸遇点光源后局部发生变化的特性，分析补光前后人脸局部的变化，并利用活体人脸在补光前后的整体成像变化特性，分析补光前后人脸图像的变化，进行活体检测，准确率高，简单方便，不需要用户主动配合、用户体验好。
在其中一个实施例中，采用的特征融合方法可以为特征堆叠，直接将四个特征连接(Concat)进行融合，或者也可使用卷积方式将四个特征进行融合，本申请实施例对于特征融合方法不做限定。
在其中一个实施例中，融合特征包括了环境光下的人脸图像整体信息和眼睛等局部信息，以及补光下的人脸图像整体信息和眼睛等局部信息，以便于区分人脸图像在环境光和补光(光源变化)下发生的变化。
本实施例中，如图3所示，步骤110的具体处理过程包括：
S301，通过人脸镜头在环境光下采集RGB图像，得到环境光人脸图像。
S302，通过补光灯进行补光，在补光条件下采集RGB图像，得到补光人脸图像。
在实施中，人脸识别设备可以通过补光灯进行补光，若人脸识别设备为手机等移动端人脸识别设备，也可以使用其自带的屏幕补光，无需增加额外硬件设备，并十分方便实现。屏幕补光可以使屏幕亮度拉到最大，并且除人脸区域外，像素设置全白或者红、黄、蓝中一种，或交替爆闪。
在t1时间段内采集多幅环境光人脸图像，并同时进行人脸检测，将检测到人脸并且图像质量符合要求的环境光人脸图像保留，如果t1时间段内未检测到人脸或图像质量不符合要求，则返回失败，提醒用户调整人脸位置后重新采集环境光人脸图像。
环境光人脸图像采集成功后，开启补光灯或屏幕进行补光，在时间段t2内采集多幅补光人脸图像，并同时进行人脸检测，将检测到人脸并且图像质量符合要求的环境光人脸图像保留，如果t2时间段内未检测到人脸或图像质量不符合要求，则返回失败，提醒用户调整采集位置或者采集姿势后重新采集补光人脸图像。
在其中一个实施例中，如图4所示，在步骤302之后，该方法还包括：
S401，通过关键点回归得到多幅环境光人脸图像和补光人脸图像的头部姿态。
S402，选择头部姿态为正脸的环境光人脸图像和补光人脸图像。
在实施中，移动端人脸识别设备根据图像质量从头部姿态为正脸的环境光人脸图像和补光人脸图像中选择符合条件的补光人脸图像和环境光人脸图像，更好的增加了算法鲁棒性。
选择符合条件的补光人脸图像和环境光人脸图像，可以通过图像信噪比等传统方法，也可通过人脸图像质量判定算法进行选择，本申请对此不作限定。一般的，可以选择质量分数高于设定的质量分数阈值的补光人脸图像和环境光人脸图像。
然后，对环境光人脸图像和补光人脸图像进行预处理，得到环境光人脸整体图像、环境光人脸局部图像、补光人脸整体图像和补光人脸局部图像。
假体人脸一般包括打印的人脸照片、翻拍后通过屏幕播放的人脸照片或视频等，本申请需要将假体人脸与活体人脸区分开，实现活体检测。
研究表明，人的眼睛、脸颊、鼻子、额头、嘴巴等人脸局部区域具有对点光源光照敏感的特性，在补光前后发生变化。以眼睛为例，在环境光下瞳孔内基本无光斑，在补光条件下，黑瞳孔形成光斑且瞳孔收缩变小，瞳孔内含有较亮光斑。因此真实的人眼(即活体)在补光前后存在明显差异，补光前瞳孔内无光斑，补光后瞳孔内含有较亮的光斑，如图5所示。脸颊、鼻子、额头、嘴巴等部位在补光条件下具有因凹凸不平形成的阴影等。
而假体人脸中的眼睛不具备对点光源光照敏感的特性，因此假体人脸中的眼睛在补光前后不存差异，补光前后瞳孔内均同时有光斑或均同时无光斑，如图5所示。即真实人脸补光前后在瞳孔内有光斑变化，而假体人脸补光前后在瞳孔内无光斑变化。假体人脸中因为是平面，其脸颊、鼻子、额头、嘴巴等部位在环境光和补光条件下均无阴影。通过上述特性可以区分出真实人脸和假体人脸。
因此，本申请通过预处理得到环境光人脸局部图像和补光人脸局部图像，利用活体人脸在环境光和补光条件下变化的特性，以区分活体人脸和假体人脸。环境光人脸局部图像和补光人脸局部图像的变化越小(即环境光人脸局部图像和补光人脸局部图像越相似)，越表明环境光人脸图像和补光人脸图像来自假体人脸。
例如利用瞳孔遇点光源形成光斑的特性，分析环境光人脸局部图像和补光人脸局部图像内是否有瞳孔光斑的变化，以区分活体人脸和假体人脸。环境光人脸局部图像和补光人脸局部图像内的瞳孔光斑变化越小，越表明环境光人脸图像和补光人脸图像来自假体人脸。
但是，目前配戴眼镜的用户较多，眼睛的镜片会形成反光光斑，反光光斑十分容易覆盖掉瞳孔内的光斑，导致在用户佩戴眼镜的情况下利用瞳孔的光斑区分活体人脸和假体人脸存在较高的误判。同时，使用相纸等表面光滑容易反光的纸张打印假体人脸，通过扭曲假体人脸照片，也能小概率在瞳孔内人为造成光斑，或者造成阴影，从而影响活体检测的准确性。
为解决这一问题，本申请从人脸整体着手。相较假体人脸，真实(活体)人脸包含丰富的3D信息。在镜头同方向的补光下，真实人脸中鼻尖和额头等离镜头较近的部位较亮，眼窝和脸颊等离镜头较远的部位较暗，与补光前在环境光下的成像有较大差距。在环境光下，真实人脸中各个部位的亮度差别不大，离镜头较近的部位与离镜头较远的部位的亮暗对比不明显。即由于真实人脸丰富的3D信息，使得真实人脸在补光前后的成像有较大变化。
而打印的人脸照片表面和播放人脸照片或视频的屏幕表面基本是平面，从光反射原理解释，平面上各个位置的反射角基本一致。因此在补光下，平面的假体人脸不会出现真实人脸成像的特性(即真实人脸中鼻尖和额头等离镜头较近的部位较亮，眼窝和脸颊等离镜头较远的部位较暗)，导致在补光前后，假体人脸的成像变化较小。
即使通过扭曲照片的方式使得人脸照片表面不是平面，但是因为扭曲的照片构造的人脸3D信息较单一，与真实人脸3D信息差距较大，也不能达到真实人脸的鼻尖和额头等离镜头较近的部位较亮且眼窝和脸颊等离镜头较远的部位较暗的成像效果。
可见，真实人脸在补光前后的成像有较大变化，而假体人脸在补光前后的成像变化较小。通过这一特性可以区分出真实人脸和假体人脸。
因此，本申请通过预处理得到环境光人脸整体图像和补光人脸整体图像，分析环境光人脸整体图像和补光人脸整体图像的变化，以区分活体人脸和假体人脸。环境光人脸整体图像和补光人脸整体图像的变化越小(即环境光人脸整体图像和补光人脸整体图像越相似)，越表明环境光人脸图像和补光人脸图像来自假体人脸。
前述的第一卷积神经网络和第二卷积神经网络在使用前需要训练，如图6所示，其训练方法如下：
S10：构造环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本，并分别设置标签。
其具体实现方式包括：
获取由训练样本对组成的训练集，其中，训练样本对包括由假体人脸的环境光人脸图像和补光人脸图像组成的假体人脸图像对和由活体人脸的环境光人脸图像和补光人脸图像组成的活体人脸图像对。
分别为假体人脸图像对和活体人脸图像对设置不同的标签。
示例性的，可以将假体人脸图像对的标签设置为1，表示相似且为假体，活体人脸图像对的标签设置为0，表示不相似且为活体。
对训练样本对进行预处理，得到环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本。
S20：通过第一卷积神经网络对环境光人脸整体图像样本和补光人脸整体图像样本进行特征提取，得到第五特征和第六特征；通过第二卷积神经网络对环境光人脸局部图像样本和补光人脸局部图像样本进行特征提取，得到第七特征和第八特征。
本步骤的特征提取方法与S120的特征提取方法相同，可参照步骤S120中阐述的内容进行理解，此处不再赘述。
S30：计算第五特征与第六特征的对比损失L 1以及第七特征与第八特征的对比损失L 2。
本申请对环境光人脸整体图像样本和补光人脸整体图像样本使用一个相同的卷积神经网络(即第一卷积神经网络)训练和提取特征，并且权值共享，该第一卷积神经网络为孪生网络(Siamese Network)，使用孪生网络对样本成对分析提取特征，更具针对性挖掘人脸在补光前后的成像差异，利用该差异进行分类。环境光人脸局部图像样本和补光人脸局部图像样本同理。
对提取的特征，使用Contrastive Loss进行监督。
Contrastive Loss，即对比损失，其表达式如下：
Figure PCTCN2022118036-appb-000001
其中，L为对比损失，d＝||a n-b n|| 2，为两个样本特征a n、b n的欧式距离，n＝1,2,…,N，N为样本特征a n、b n的数量，y为两个样本特征是否匹配的标签，y＝1代表两个样本特征相似或者匹配，y＝0代表不匹配，margin为设置的阈值。分类时，可以使用Softmax对融合特征fc进行分类，其中分类结果1代表假体，0代表活体。
在本申请中，使用该对比损失来度量同一个人的人脸整体图像和人脸局部图像在环境光和补光下的两个样本是否匹配，不匹配代表真人样本对，匹配代表假体样本对。f1和f2，f3和f4分别 成对计算对比损失L 1和L 2。
S40：将第五特征、第六特征、第七特征和第八特征进行融合，得到融合样本特征，并计算融合样本特征的交叉熵损失L 3。
本步骤进行特征融合的方法与S130相同，可参照步骤S130中阐述的内容进行理解，此处不再赘述。在其中一个实施例中，第五特征、第七特征、第六特征和第八特征顺序连接，得到融合样本特征。
本申请使用交叉熵损失函数(Cross Entropy Loss)用于融合样本特征的分类，计算交叉熵损失L 3，如图7所示。
S50：将对比损失L 1、对比损失L 2和交叉熵损失L 3加权求和，得到总体损失L All，并通过反向传播更新第一卷积神经网络和第二卷积神经网络的参数。
例如在其中一个实施例中，L 1、L 2和L 3的权重可分别为0.25、0.25、0.5，L All＝0.25L 1+0.25L 2+0.5L 3。
前述的S10中，可以通过如下方法获取由训练样本对组成的训练集，如图8所示：
S11：获取由人脸镜头在环境光下采集的若干幅活体人脸的环境光人脸图像。
本步骤中，人脸镜头前为活体人脸。采集人脸图像时，打开人脸镜头，在环境光下采集活体人脸的人脸图像，通过人脸检测算法检测到人脸并判断图像质量满足要求后，继续采集，持续获取n_b幅人脸图像，得到活体人脸的环境光人脸图像。
S12：获取由人脸镜头在补光灯或屏幕的补光下采集的若干幅活体人脸的补光人脸图像。
本步骤中，人脸镜头前为活体人脸。通过上层应用App控制补光灯或者屏幕高亮，在补光环境下采集活体人脸的人脸图像。当在人脸图像中检测到人脸并且图像质量满足要求时，继续采集，持续获取n_f幅人脸图像，达到采集数量后关闭补光灯或回复屏幕正常态，得到活体人脸的补光人脸图像。
S13：分别从活体人脸的环境光人脸图像和补光人脸图像中各自随机抽取一幅图像，组成活体人脸图像对。
本步骤用于制作活体人脸图像对，每一活体人脸图像对包含同时间段采集的活体人脸的环境光人脸图像和补光人脸图像，从n_b幅环境光人脸图像和n_f幅补光人脸图像中各自随机多次抽取一幅图像，组成活体人脸图像对。
S14：获取由人脸镜头在环境光下采集的若干幅假体人脸的环境光人脸图像。
本步骤中，人脸镜头前为假体人脸，采集方法与S11相同，可参照步骤S11中阐述的内容进行理解，此处不再赘述。
S15：获取由人脸镜头在补光灯或屏幕的补光下采集的若干幅假体人脸的补光人脸图像。
本步骤中，人脸镜头前为假体人脸，采集方法与S12相同，可参照步骤S12中阐述的内容进行理解，此处不再赘述。
S16：从假体人脸的环境光人脸图像和补光人脸图像中各自随机抽取一幅图像，组成假体人脸图像对。
本步骤的实现方法与S13相同，可参照步骤S13中阐述的内容进行理解，此处不再赘述。
本申请的人脸局部可以为眼睛、脸颊、鼻子、额头、嘴巴等人脸局部区域，优选为眼睛区域。
本申请中，以人脸局部为眼睛区域为例，可以通过如图9所示的方法对环境光人脸图像和补光人脸图像进行预处理：
S910：对环境光人脸图像和补光人脸图像进行人脸检测、人脸关键点定位、头部姿态估计和人脸归一化，得到环境光人脸整体图像和补光人脸整体图像。
本申请不限制人脸检测、人脸关键点定位、头部姿态估计和人脸归一化的具体方法，在其中一个实施例中，使用MTCNN检测算法，实现人脸检测和人脸特征点定位。人脸关键点共5个，左眼、右眼、鼻子、左嘴角、右嘴角。通过关键点的2D坐标信息，估计3D姿态信息。通过双眼坐标，对人脸进行归一化，具体包含人脸对齐和缩放，将双眼对齐到(38,52)和(74,52)，缩放大小 为(112,112)，如图10所示。
S920：以人脸关键点定位得到的眼睛坐标为中心，向四周扩充若干像素，得到包含眼睛的感兴趣区域，其中，向四周扩充的像素数根据左右眼间距确定。
为了充分挖掘补光前后瞳孔内变化，做如下精细的归一化操作：
在人脸关键点定位得到的眼睛坐标为中心，向外扩充若干像素，获取包含眼睛的感兴趣区域(眼睛ROI)。
其中，向四周扩充的像素数根据左右眼间距确定，左右眼间距为d iris＝|x right-x left|，x right为右眼横坐标，x left为左眼横坐标，在其中一个实施例中，向四周扩充的像素数为d iris/3。
S930：在感兴趣区域内，使用径向对称变换(Radial Symmetry Transform，RST)确定眼睛中心和半径R。
S940：根据眼睛中心和半径截取得到环境光人脸局部图像和补光人脸局部图像。
其中，对左眼和右眼的每个眼睛，以眼睛圆心(以左眼为例，圆心为x l_center，y l_center)为中心，截取图像宽度为4.4R，高度为2.2R的图像；将截取的左眼图像和右眼图像缩放到112*56大小，并将缩放后的左眼图像和右眼图像拼接堆叠在一起，得到环境光人脸局部图像和补光人脸局部图像，如图11所示。
前述的S10中构造环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本时，也需要对训练样本对进行预处理，该预处理方法与环境光人脸图像和补光人脸图像的鱼锤方法相同。
在其中一个实施例中，本申请实施例提供一种人脸活体检测装置1200，如图12所示，该装置1200包括：
图像获取模块1201，用于获取同一用户预设时间段内的环境光人脸图像和补光人脸图像。
第一预处理模块1202，用于对环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像。
第二预处理模块1203，用于对补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像。
图像特征提取模块1204，用于通过第一卷积神经网络对环境光人脸整体图像和补光人脸整体图像进行特征提取，得到第一特征和第二特征；通过第二卷积神经网络对环境光人脸局部图像和补光人脸局部图像进行特征提取，得到第三特征和第四特征。
图像特征融合模块1205，用于将第一特征、第二特征、第三特征和第四特征进行融合，得到融合特征。
活体检测模块1206，用于通过融合特征进行分类，判断环境光人脸图像和补光人脸图像是否来自活体。
本申请获取环境光和补光下的人脸图像，通过预处理得到环境光和补光下的人脸图像的整体区域和局部区域，并通过卷积神经网络提取特征，将提取的特征融合后进行活体人脸和假体人脸分类。本申请利用活体人脸遇点光源后局部发生变化的特性，分析补光前后人脸局部的变化，并利用活体人脸在补光前后的成像整体变化特性，分析补光前后人脸图像的整体变化，进行活体检测，准确率高，简单方便，不需要用户主动配合、用户体验好。
在其中一个实施例中，该人脸活体检测装置1200还包括：
样本构造模块，用于构造隶属于同一活体或假体的环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本。
损失计算模块，用于根据环境光人脸整体图像样本和补光人脸整体图像样本，对第一卷积神经网络进行训练，以及根据环境光人脸局部图像样本和补光人脸局部图像样本，对第二卷积神经网络进行训练，得到第一卷积神经网络和第二卷积神经网络的模型训练过程中的总体损失。
反向传播模块，用于通过反向传播更新第一卷积神经网络和第二卷积神经网络的参数，直至总体损失满足预设的损失条件，第一卷积神经网络和第二卷积神经网络训练完成。
前述的第一卷积神经网络和第二卷积神经网络在使用前需要训练，其通过如下模块训练得到：
样本构造模块，用于构造环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本，并分别设置标签。
样本特征提取模块，用于通过第一卷积神经网络对环境光人脸整体图像样本和补光人脸整体图像样本进行特征提取，得到第五特征和第六特征；通过第二卷积神经网络对环境光人脸局部图像样本和补光人脸局部图像样本进行特征提取，得到第七特征和第八特征。
对比损失计算模块，用于计算第五特征与第六特征的对比损失L 1以及第七特征与第八特征的对比损失L 2。
交叉熵损失计算模块，用于将第五特征、第六特征、第七特征和第八特征进行融合，得到融合样本特征，并计算融合样本特征的交叉熵损失L 3。
反向传播模块，用于将L 1、L 2和L 3加权求和，得到总体损失L All，并通过反向传播更新第一卷积神经网络和第二卷积神经网络的参数。
在其中一个实施例中，前述的样本构造模块包括：
第一获取单元，用于获取由人脸镜头在环境光下采集的若干幅活体人脸的环境光人脸图像。
第二获取单元，用于获取由人脸镜头在补光灯或屏幕的补光下采集的若干幅活体人脸的补光人脸图像。
活体人脸图像对获取单元，用于分别从活体人脸的环境光人脸图像和补光人脸图像中各自随机抽取一幅图像，组成活体人脸图像对。
第三获取单元，用于获取由人脸镜头在环境光下采集的若干幅假体人脸的环境光人脸图像。
第四获取单元，用于获取由人脸镜头在补光灯或屏幕的补光下采集的若干幅假体人脸的补光人脸图像。
假体人脸图像对获取单元，用于从假体人脸的环境光人脸图像和补光人脸图像中各自随机抽取一幅图像，组成假体人脸图像对。
本申请中，不限制第一特征、第三特征、第二特征和第四特征的融合方式，在其中一个实施例中，第五特征、第七特征、第六特征和第八特征顺序连接，得到融合样本特征。
相应的，第一特征、第三特征、第二特征和第四特征顺序连接，得到融合特征。
本申请的环境光人脸局部图像和补光人脸局部图像为眼睛图像，相应的，对环境光人脸图像和补光人脸图像的预处理包括：
人脸处理单元，用于对环境光人脸图像和补光人脸图像进行人脸检测、人脸关键点定位、头部姿态估计和人脸归一化，得到环境光人脸整体图像和补光人脸整体图像。
眼睛感兴趣区域获取单元，用于以人脸关键点定位得到的眼睛坐标为中心，向四周扩充若干像素，得到包含眼睛的感兴趣区域，其中，向四周扩充的像素数根据左右眼间距确定。
瞳孔定位单元，用于在感兴趣区域内，使用径向对称变换确定眼睛中心和半径。
图像截取单元，用于根据眼睛中心和半径截取得到环境光人脸局部图像和补光人脸局部图像。
本申请实施例所提供的装置，其实现原理及产生的技术效果和前述方法实施例1相同，为简要描述，该装置实施例部分未提及之处，可参考前述方法实施例1中相应内容。所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，前述描述的装置和单元的具体工作过程，均可以参考上述方法实施例1中的对应过程，在此不再赘述。
在其中一个实施例中，本申请提供的上述实施例1所述的方法可以通过计算机程序实现业务逻辑并记录在存储介质上，所述的存储介质可以计算机读取并执行，实现本说明书实施例1所描述方案的效果。因此，本申请还提供用于人脸活体检测的计算机可读存储介质，包括用于存储处理器可执行指令的存储器，指令被处理器执行时实现包括实施例1的人脸活体检测方法的步骤。
本申请获取环境光和补光下的人脸图像，通过预处理得到环境光和补光下的人脸图像的整体 区域和局部区域，并通过卷积神经网络提取特征，将提取的特征融合后进行活体人脸和假体人脸分类。本申请利用活体人脸遇点光源后局部发生变化的特性，分析补光前后人脸局部的变化，并利用活体人脸在补光前后的成像整体变化特性，分析补光前后人脸图像的整体变化，进行活体检测，准确率高，简单方便，不需要用户主动配合、用户体验好。
所述存储介质可以包括用于存储信息的物理装置，通常是将信息数字化后再以利用电、磁或者光学等方式的媒体加以存储。所述存储介质有可以包括：利用电能方式存储信息的装置如，各式存储器，如RAM、ROM等；利用磁能方式存储信息的装置如，硬盘、软盘、磁带、磁芯存储器、磁泡存储器、U盘；利用光学方式存储信息的装置如，CD或DVD。当然，还有其他方式的可读存储介质，例如量子存储器、石墨烯存储器等等。
上述所述的存储介质根据方法实施例1的描述还可以包括其他的实施方式，本实施例的实现原理及产生的技术效果和前述方法实施例1相同，具体可以参照相关方法实施例1的描述，在此不作一一赘述。
在其中一个实施例中，本申请还提供一种用于人脸活体检测的设备，所述的设备可以为单独的计算机，也可以包括使用了本说明书的一个或多个所述方法或一个或多个实施例装置的实际操作装置等。所述人脸活体检测的设备可以包括至少一个处理器以及存储计算机可执行指令的存储器，处理器执行所述指令时实现上述任意一个或者多个实施例1中所述人脸活体检测方法的步骤。
本申请获取环境光和补光下的人脸图像，通过预处理得到环境光和补光下的人脸图像的整体区域和局部区域，并通过卷积神经网络提取特征，将提取的特征融合后进行活体人脸和假体人脸分类。本申请利用活体人脸遇点光源后局部发生变化的特性，分析补光前后人脸局部的变化，并利用活体人脸在补光前后的成像整体变化特性，分析补光前后人脸图像的整体变化，进行活体检测，准确率高，简单方便，不需要用户主动配合、用户体验好。
上述所述的设备根据方法实施例1的描述还可以包括其他的实施方式，本实施例的实现原理及产生的技术效果和前述方法实施例1相同，具体可以参照相关方法实施例1的描述，在此不作一一赘述。
最后应说明的是：以上所述实施例，仅为本申请的具体实施方式，用以说明本申请的技术方案，而非对其限制，本申请的保护范围并不局限于此，尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，其依然可以对前述实施例所记载的技术方案进行修改或可轻易想到变化，或者对其中部分技术特征进行等同替换；而这些修改、变化或者替换，并不使相应技术方案的本质脱离本申请实施例技术方案的精神和范围。都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (15)

  1. 一种人脸活体检测方法，其特征在于，所述方法包括：
    获取同一用户预设时间段内的环境光人脸图像和补光人脸图像；
    对所述环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像；
    对所述补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像；
    通过第一卷积神经网络对所述环境光人脸整体图像和所述补光人脸整体图像进行特征提取，得到第一特征和第二特征；通过第二卷积神经网络对所述环境光人脸局部图像和所述补光人脸局部图像进行特征提取，得到第三特征和第四特征；
    将所述第一特征、所述第二特征、所述第三特征和所述第四特征进行融合，得到融合特征；
    通过所述融合特征对所述用户进行分类，判断所述环境光人脸图像和所述补光人脸图像是否来自活体。
  2. 根据权利要求1所述的人脸活体检测方法，其特征在于，所述第一卷积神经网络和所述第二卷积神经网络通过如下方法训练得到：
    构造隶属于同一活体或假体的环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本；
    根据所述环境光人脸整体图像样本和所述补光人脸整体图像样本，对所述第一卷积神经网络进行训练，以及根据所述环境光人脸局部图像样本和所述补光人脸局部图像样本，对所述第二卷积神经网络进行训练，得到所述第一卷积神经网络和所述第二卷积神经网络的模型训练过程中的总体损失；
    通过反向传播更新所述第一卷积神经网络和所述第二卷积神经网络的参数，直至所述总体损失满足预设的损失条件，所述第一卷积神经网络和所述第二卷积神经网络训练完成。
  3. 根据权利要求2所述的人脸活体检测方法，其特征在于，所述方法还包括：
    将隶属于同一活体或假体的环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本分别预设置有标签；所述标签用于标识多个样本中的图像隶属于同一活体还是同一假体。
  4. 根据权利要求3中所述的人脸活体检测方法，其特征在于，所述根据所述环境光人脸整体图像样本和所述补光人脸整体图像样本，对所述第一卷积神经网络进行训练，以及根据所述环境光人脸局部图像样本和所述补光人脸局部图像样本，对所述第二卷积神经网络进行训练，得到所述第一卷积神经网络和所述第二卷积神经网络的模型训练过程中的总体损失，包括：
    通过所述第一卷积神经网络对所述环境光人脸整体图像样本和所述补光人脸整体图像样本进行特征提取，得到第五特征和第六特征；通过所述第二卷积神经网络对所述环境光人脸局部图像样本和所述补光人脸局部图像样本进行特征提取，得到第七特征和第八特征；
    计算所述第五特征与所述第六特征的第一对比损失L 1以及所述第七特征与所述第八特征的第二对比损失L 2；
    将所述第五特征、所述第六特征、所述第七特征和所述第八特征进行融合，得到融合样本特征，并根据所述融合样本特征，以及所述标签计算所述融合样本特征的交叉熵损失L 3；
    将所述第一对比损失L 1、所述第二对比损失L 2和所述交叉熵损失L 3加权求和，得到总体损失L All。
  5. 根据权利要求4所述的人脸活体检测方法，其特征在于，所述将所述第五特征、所述第六特征、所述第七特征和所述第八特征进行融合，得到融合样本特征，包括：
    所述第五特征、所述第七特征、所述第六特征和所述第八特征顺序连接，得到所述融合样本特征。
  6. 根据权利要求1所述的人脸活体检测方法，其特征在于，所述将所述第一特征、所述第二特征、所述第三特征和所述第四特征进行融合，得到融合特征，包括：
    将所述第一特征、所述第三特征、所述第二特征和所述第四特征顺序连接，得到所述融合特征；所述融合特征用于区分活体人脸和假体人脸，所述融合特征包含人脸整体3D信息和眼部的光斑信息。
  7. 根据权利要求1-3中任一项所述的人脸活体检测方法，其特征在于，所述环境光人脸局部图像和所述补光人脸局部图像为眼睛图像；所述第三特征和所述第四特征包括眼部的光斑信息。
  8. 根据权利要求6所述的人脸活体检测方法，其特征在于，所述对所述环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像，包括：
    对所述环境光人脸图像进行人脸检测、人脸关键点定位、头部姿态估计和人脸归一化，得到环境光人脸整体图像；
    以人脸关键点定位得到的眼睛坐标为中心，向四周扩充若干像素，得到包含眼睛的感兴趣区域，其中，向四周扩充的像素数根据左右眼间距确定；
    在所述感兴趣区域内，使用径向对称变换确定眼睛中心和半径；
    根据所述眼睛中心和所述半径截取得到环境光人脸局部图像。
  9. 根据权利要求6所述的人脸活体检测方法，其特征在于，所述对所述补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像，包括：
    对所述补光人脸图像进行人脸检测、人脸关键点定位、头部姿态估计和人脸归一化，得到补光人脸整体图像；
    以人脸关键点定位得到的眼睛坐标为中心，向四周扩充若干像素，得到包含眼睛的感兴趣区域，其中，向四周扩充的像素数根据左右眼间距确定；
    在所述感兴趣区域内，使用径向对称变换确定眼睛中心和半径；
    根据所述眼睛中心和所述半径截取得到补光人脸局部图像。
  10. 根据权利要求1所述的人脸活体检测方法，其特征在于，所述获取同一用户预设时间段内的环境光人脸图像和补光人脸图像，包括：
    通过人脸镜头在环境光下采集RGB图像，得到所述环境光人脸图像；
    通过补光灯进行补光，在补光条件下采集RGB图像，得到所述补光人脸图像。
  11. 根据权利要求10所述的人脸活体检测方法，其特征在于，所述方法还包括：
    通过关键点回归得到多幅所述环境光人脸图像和所述补光人脸图像的头部姿态；
    选择所述头部姿态为正脸的所述环境光人脸图像和所述补光人脸图像。
  12. 一种人脸活体检测装置，其特征在于，所述装置包括：
    图像获取模块，用于获取同一用户预设时间段内的环境光人脸图像和补光人脸图像；
    第一预处理模块，用于对所述环境光人脸图像进行预处理，得到环境光人脸整体图像和环境光人脸局部图像；
    第二预处理模块，用于对所述补光人脸图像进行预处理，得到补光人脸整体图像和补光人脸局部图像；
    图像特征提取模块，用于通过第一卷积神经网络对所述环境光人脸整体图像和所述补光人脸整体图像进行特征提取，得到第一特征和第二特征；通过第二卷积神经网络对所述环境光人脸局部图像和所述补光人脸局部图像进行特征提取，得到第三特征和第四特征；
    图像特征融合模块，用于将所述第一特征、所述第二特征、所述第三特征和所述第四特征进行融合，得到融合特征；
    活体检测模块，用于通过所述融合特征进行分类，判断所述环境光人脸图像和所述补光人脸图像是否来自活体。
  13. 根据权利要求12所述的人脸活体检测装置，其特征在于，所述人脸活体检测装置还包括：
    样本构造模块，用于构造隶属于同一活体或假体的环境光人脸整体图像样本、环境光人脸局部图像样本、补光人脸整体图像样本和补光人脸局部图像样本；
    损失计算模块，用于根据所述环境光人脸整体图像样本和所述补光人脸整体图像样本，对所述第一卷积神经网络进行训练，以及根据所述环境光人脸局部图像样本和所述补光人脸局部图像样本，对所述第二卷积神经网络进行训练，得到所述第一卷积神经网络和所述第二卷积神经网络的模型训练过程中的总体损失；
    反向传播模块，用于通过反向传播更新所述第一卷积神经网络和所述第二卷积神经网络的参数，直至所述总体损失满足预设的损失条件，所述第一卷积神经网络和所述第二卷积神经网络训练完成。
  14. 一种用于人脸活体检测的计算机可读存储介质，其特征在于，包括用于存储处理器可执行指令的存储器，所述指令被所述处理器执行时实现包括权利要求1-11任一所述人脸活体检测方法的步骤。
  15. 一种用于人脸活体检测的设备，其特征在于，包括至少一个处理器以及存储计算机可执行指令的存储器，所述处理器执行所述指令时实现权利要求1-11中任意一项所述人脸活体检测方法的步骤。
PCT/CN2022/118036 2021-10-15 2022-09-09 人脸活体检测方法、装置、计算机可读存储介质及设备 WO2023061122A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111202019.0A CN115995103A (zh) 2021-10-15 2021-10-15 人脸活体检测方法、装置、计算机可读存储介质及设备
CN202111202019.0 2021-10-15

Publications (1)

Publication Number Publication Date
WO2023061122A1 true WO2023061122A1 (zh) 2023-04-20

Family

ID=85987278

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/118036 WO2023061122A1 (zh) 2021-10-15 2022-09-09 人脸活体检测方法、装置、计算机可读存储介质及设备

Country Status (2)

Country Link
CN (1) CN115995103A (zh)
WO (1) WO2023061122A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117372604A (zh) * 2023-12-06 2024-01-09 国网电商科技有限公司 一种3d人脸模型生成方法、装置、设备及可读存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106203305A (zh) * 2016-06-30 2016-12-07 北京旷视科技有限公司 人脸活体检测方法和装置
CN107886070A (zh) * 2017-11-10 2018-04-06 北京小米移动软件有限公司 人脸图像的验证方法、装置及设备
CN110490041A (zh) * 2019-05-31 2019-11-22 杭州海康威视数字技术股份有限公司 人脸图像采集装置及方法
US20200320184A1 (en) * 2017-12-21 2020-10-08 Yoti Holding Limited Biometric User Authentication

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106203305A (zh) * 2016-06-30 2016-12-07 北京旷视科技有限公司 人脸活体检测方法和装置
CN107886070A (zh) * 2017-11-10 2018-04-06 北京小米移动软件有限公司 人脸图像的验证方法、装置及设备
US20200320184A1 (en) * 2017-12-21 2020-10-08 Yoti Holding Limited Biometric User Authentication
CN110490041A (zh) * 2019-05-31 2019-11-22 杭州海康威视数字技术股份有限公司 人脸图像采集装置及方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117372604A (zh) * 2023-12-06 2024-01-09 国网电商科技有限公司 一种3d人脸模型生成方法、装置、设备及可读存储介质
CN117372604B (zh) * 2023-12-06 2024-03-08 国网电商科技有限公司 一种3d人脸模型生成方法、装置、设备及可读存储介质

Also Published As

Publication number Publication date
CN115995103A (zh) 2023-04-21

Similar Documents

Publication Publication Date Title
WO2020151489A1 (zh) 基于面部识别的活体检测的方法、电子设备和存储介质
WO2022206319A1 (zh) 图像处理方法、装置、设备、存储介质计算机程序产品
CN110889312B (zh) 活体检测方法和装置、电子设备、计算机可读存储介质
JP5629803B2 (ja) 画像処理装置、撮像装置、画像処理方法
JP4743823B2 (ja) 画像処理装置、撮像装置、画像処理方法
CN109948566B (zh) 一种基于权重融合与特征选择的双流人脸反欺诈检测方法
WO2019076227A1 (zh) 人脸图像分类方法、装置及服务器
WO2015149534A1 (zh) 基于Gabor二值模式的人脸识别方法及装置
WO2019137178A1 (zh) 人脸活体检测
CN108596082A (zh) 基于图像扩散速度模型和色彩特征的人脸活体检测方法
CN114333078A (zh) 活体检测方法、装置、电子设备及存储介质
WO2005008593A1 (ja) 画像処理装置、撮像装置、画像処理方法
CN105956572A (zh) 一种基于卷积神经网络的活体人脸检测方法
WO2016010724A1 (en) Multispectral eye analysis for identity authentication
WO2016010721A1 (en) Multispectral eye analysis for identity authentication
WO2016010720A1 (en) Multispectral eye analysis for identity authentication
CN106203391A (zh) 基于智能眼镜的人脸识别方法
CN112818722A (zh) 模块化动态可配置的活体人脸识别系统
Liu et al. Light field-based face liveness detection with convolutional neural networks
CN112434647A (zh) 一种人脸活体检测方法
CN113536843A (zh) 基于多模态融合卷积神经网络的防伪人脸识别系统
WO2023061122A1 (zh) 人脸活体检测方法、装置、计算机可读存储介质及设备
CN111259757A (zh) 一种基于图像的活体识别方法、装置及设备
Li et al. A dual-modal face anti-spoofing method via light-weight networks
CN112949353A (zh) 虹膜静默活体检测方法、装置、可读存储介质及设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22880064

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 14.08.2024)