System and method for detecting face activity in an image and electronic device

Abstract

The application discloses an image detection system and method and an electronic device. The technology relates to the field of facial attribute recognition. The technology can identify a face in an image as real or deceptive. The system and method provided by the application include recognizing facial attributes and using a multi-task learning network. The neural network includes segmentation functions and classification functions. The final output result is used to obtain pixel-level semantic information and high-level feature information.

Description

Technical Field

[0001] This application relates to systems and methods for face liveness detection, which aim to facilitate the extraction of facial attributes to determine whether a presented face is real or deceptive. More specifically, this application aims to provide a novel multi-task learning network for face anti-spoofing, combining segmentation and classification tasks for liveness detection of face images. Background Technology

[0002] Image-based face despoofing refers to face despoofing techniques that use only the three primary colors (Red, Green, Blue, RGB) as input without requiring additional information such as depth or heat. In past work, such as Blacksesame's "RGB-Near Infrared Ray (NIR) Dual-Camera Face Despoofing System Based on Deep Neural Networks," multi-sensor methods have achieved excellent performance in face despoofing. Existing image-based face despoofing methods can be divided into two categories: traditional methods and deep learning-based methods.

[0003] Recently, deep learning methods, especially Convolutional Neural Networks (CNNs), have been utilized in face anti-spoofing research. As Yaojie Liu mentioned in "Computer Vision and Pattern Recognition," some publicly available CNN results generally outperform traditional methods. Furthermore, the "Cross-Database Face Anti-Spoofing with Robust Feature Representations" method proposed by Lei Li and Keyurkumar Patel utilizes pre-trained Convolutional Architecture for Fast Feature Embedding Networks (CaffeNet) or Visual Geometry Group (VGG) face models as feature extractors to distinguish between live and spoofed faces.

[0004] Furthermore, Sergey Tulyakov utilized multiple spatial scales to classify liveness and spoofing images in his work "Adaptive Matrix Completion for Heart Rate Estimation from Facial Videos under Realistic Conditions." Additional information, such as remote photovolumetric methods and spoofing noise, has also been developed.

[0005] J. Yang, in "Learning Convolutional Neural Networks for Face Anti-Spoofing," proposed using image stacks of different proportions cropped from the original frames as input to a CNN model. Furthermore, Z. Xu, in "Learning Transient Features for Face Anti-Spoofing Using a Long Short-Term Memory (LSTM) CNN Architecture," proposed utilizing transient features between frames through an LSTM CNN network. Y. Atoum, in "Face Anti-Spoofing Using Patch-Based and Depth-Based Convolutional Neural Networks," proposed using a fully convolutional neural network to estimate the depth of the input face image, then feeding the generated depth map into a support vector machine (SVM) to distinguish between live and spoofed faces.

[0006] Another prior art, U.S. Patent 8,856,541, discloses a technique for detecting liveness in images using a graphical user interface (GUI). The system in this patent utilizes GUI elements to provide challenging patterns for detecting human eye movements based on first and second images of a face. While this prior art provides comprehensive analysis and gives accurate results, it still cannot detect liveness when only a single image is provided.

[0007] Although the aforementioned techniques represent an improvement over traditional methods, their detection accuracy and robustness to external interference variables remain unsatisfactory for practical applications. Furthermore, while all the references address the activity detection problem, none disclose segmentation-guided anti-spoofing classification networks, thus limiting their effectiveness.

[0008] Therefore, to overcome the limitations of the aforementioned references, this application proposes a novel multi-task learning network for face anti-spoofing that combines segmentation and classification tasks. Furthermore, the system of this application includes the following innovations over the prior art.

[0009] • A novel multi-task learning network for face anti-spoofing.

[0010] • A multi-scale attention input solution for face anti-spoofing tasks.

[0011] • A decision-making fusion scheme for primary and secondary tasks.

[0012] This application aims to provide an improvement in recognizing facial activity in images. This application provides a multi-task learning network, named the Segmentation-Guided Face Anti-Spoofing Classification Network.

[0013] Clearly, various methods and systems suitable for a wide range of purposes have been developed in the prior art. However, while these applications may be applicable to their specific purposes, they are not applicable to the purposes of this application as described above. Therefore, there is a need for an advanced image detection system to detect the liveness of an image. Summary of the Invention

[0014] This application provides a system and method for liveness detection of facial images, utilizing facial attributes to determine whether the presented face is real or fake. This application also provides a novel multi-task learning network for anti-spoofing faces. Furthermore, the learning network combines segmentation and classification tasks to perform liveness detection of facial images. The system of this application includes several basic elements as described below to achieve the goal of detecting whether a face in an input image is real or fake.

[0015] The face detection module processes and scales the input image to generate feature maps. Input image scaling is typically used to extract a small number or multiple facial attributes as needed.

[0016] A segmentation-guided anti-spoofing classification network is used to perform image segmentation and classification.

[0017] The fusion module is used to combine the results of the segmentation module and the classification module to identify the liveness of face images.

[0018] The main objective of this application is to provide a system for detecting liveness in face images by utilizing a multi-task learning network with segmentation and classification modules.

[0019] The face detection module receives the input image, processes it, and identifies the Region of Interest (ROI). It recognizes faces and highlights them using bounding boxes. The recognized faces are scaled up to rescale the image and identify detailed facial attributes. Furthermore, the face detection module also acquires information beyond the facial regions and parts of the face.

[0020] The segmentation module extracts pixel-level semantic information. Accurate semantic information for each pixel is obtained by labeling the image. Four labels are used to annotate live and spoofed images at the pixel level; these labels are color representations of the pixel-level image used to identify certain attributes of the image at the pixel level. After segmenting each pixel, the segmentation module counts the number of spoofed and live pixels.

[0021] The classification module extracts high-level feature information from the input image. A probability determination operation is performed to detect whether the features belong to a spoofing or a live entity. Furthermore, probability estimation is implemented using the softmax function.

[0022] Another objective of this application is to provide a segmentation module that serves as an auxiliary module for collecting the final results.

[0023] Another objective of this application is to provide a classification module and a fusion module. The classification module determines a probability value, and the fusion module merges the results of the segmentation module with the results of the classification module. The classification module and the fusion module are only used when the segmentation module fails to detect the activity of the image.

[0024] Another objective of this application is to provide a labeling unit for labeling individual colors onto face, foreground, and background areas according to a defined color palette.

[0025] Another objective of this application is to provide annotations based on real faces, fake faces, real foreground, fake foreground, real background, and fake background. Furthermore, the annotation unit generates feature maps of the image.

[0026] However, another objective of this application is to provide a decoding head to obtain feature representations carrying local and global contextual information.

[0027] Other objects and aspects of this application will be set forth in the following detailed description taken in conjunction with the accompanying drawings, which illustrate features according to embodiments of this application.

[0028] In order to achieve the above and related objectives, this application may be implemented in the form shown in the accompanying drawings. However, it should be noted that the drawings are merely illustrative and may be modified in terms of the specific structure illustrated and described within the scope of the appended claims.

[0029] Although this application has been described above with reference to various exemplary embodiments and implementations, it should be understood that the various features, aspects, and functions described in one or more individual embodiments are not limited in applicability to the specific embodiments described, and may be applied individually or in various combinations to one or more other embodiments of this application, whether or not such embodiments are described and whether such features are presented as part of the described embodiments. Therefore, the breadth and scope of this application should not be limited by any of the exemplary embodiments described above.

[0030] In some examples, terms and phrases such as “one or more,” “at least,” “but not limited to,” or other similar expansion words and phrases should not be interpreted as referring to or requiring a narrower case in the absence of such expansion words. Attached Figure Description

[0031] The objectives and features of this application will be fully revealed from the following description and appended claims in conjunction with the accompanying drawings. It should be understood that these drawings depict only exemplary embodiments of the application and are therefore not intended to limit its scope. This application will be described and explained with reference to the following drawings and additional features and details.

[0032] Figure 1A An image detection system according to the first case of this application is shown.

[0033] Figure 1B An image detection system according to the second case of this application is shown.

[0034] Figure 2 A face detection module of an image detection system according to this application is shown.

[0035] Figure 3 A segmentation module of an image detection system according to this application is shown.

[0036] Figure 4 An overview of the segmentation-guided face anti-spoofing classification network according to this application is shown.

[0037] Figure 5 The image illustrates the pixel-level labeling of live and spoofed images using four different labels, according to this application.

[0038] Figure 6 The data preprocessing for multi-scale generation according to this application is shown.

[0039] Figure 7 A method for detecting activity in an image according to this application is shown.

[0040] Figure 8 A schematic diagram of the cascading result fusion process according to this application is shown.

[0041] Figure 9 A schematic diagram of the structure of an electronic device provided according to an exemplary embodiment of this application is shown. Detailed Implementation

[0042] Facial liveness detection is a crucial task in computer vision, aiming to enable facial interaction systems to verify whether a presented face is real or deceptive. With the successful application of mobile phone unlocking, access control, and e-wallet payments, facial interaction systems have become an indispensable part of the real world. However, these systems pose significant threats. Imagine an attacker possessing your photo or video, capable of unlocking your phone and even making payments using your e-wallet. Therefore, facial liveness detection has become a vital technology for protecting our privacy and property from unauthorized use.

[0043] Typically, sources of malicious attacks include printed photos, screen images or videos, highly realistic masks, or 3D models of authorized clients. Among these types of attacks, the most sophisticated are printed photos or screen images obtained from the internet.

[0044] Figure 1A The illustration shows an image detection system for detecting liveness in an image according to a first embodiment of this application. The image detection system 100 includes a face detection module 200 and a segmentation module 300. The face detection module 200 includes a processing unit, a scaling unit, and a labeling unit.

[0045] The processing unit is configured to process the image to identify the region of interest. The scaling unit scales the region of interest in the image to identify at least one facial feature and generate a first image.

[0046] The annotation unit is configured to annotate a single color from a plurality of defined colors. A first color is annotated to faces in the first image, a second color is annotated to the foreground region of the first image, and a third color is annotated to the background region of the first image to generate a second image. The annotation unit assigns colors based on the activity of the foreground and background regions.

[0047] A segmentation module 300 is used to extract pixel-level semantic information from a first image and generate a second image. The segmentation module is configured to count the number of live pixels and spoof pixels in the first image. If the segmentation module fails to detect liveness in the video, a classification module is used.

[0048] In addition, the face detection module 200 performs face detection and generates bounding boxes. The segmentation module 300 extracts feature representations from the image. The feature representations carry local and global contextual information. Furthermore, the prediction for each pixel is obtained from the last convolutional layer in the segmentation head.

[0049] The segmentation-guided face anti-spoofing classification network's backbone consists of a cascade of convolutional layers, pooling layers, and activation layers to extract features for each block. A decoding head (also called a segmentation head) is then used to obtain the final feature representation, which carries both local and global contextual information. Finally, the last convolutional layer in the decoding head yields the final prediction for each pixel. The feature maps are also fed into a classification head, which consists of a convolutional layer and a fully connected layer, to obtain the final feature representation. A softmax function is then used to obtain the final liveness / spoofing probability.

[0050] Figure 1B The illustration shows an image detection system for detecting liveness in an image according to a second embodiment of the present invention. The segmentation-guided classification image detection system 100B includes a face detection module 200, a segmentation module 300, a classification module 400, and a fusion module 500. The face detection module 200 performs face detection and generates bounding boxes.

[0051] The segmentation module 300 extracts feature representations from the image carrying local and global contextual information. Furthermore, the prediction result for each pixel is obtained from the last convolutional layer in the segmentation head. The classification module 400 is a hierarchical architecture with one convolutional layer and a fully connected layer. The feature map is input into the classification module 400 to obtain the final feature representation, and a softmax function is used to obtain the final live / spoof probability. The fusion module 500 verifies the live / spoof probability, which is also used for a second verification.

[0052] A classification module is used to extract high-level feature information from the second image and generate a third image. The classification module is configured to estimate the probability that the high-level features belong to a live or spoofed individual. A fusion module is used to fuse the second and third images to detect the liveness of the images.

[0053] In recent decades, research on attribute representations based on objects, faces, and scenes has attracted widespread attention as a supplement to classification representations. However, few have attempted to utilize semantic information in face deception. In fact, for face deception, additional semantic information can be used to characterize target images through attributes, rather than classifying them into a single category, i.e., live or deceiving. In this paper, we design a multi-task learning network, named the Segmentation-Guided Face Deception Classification Network, as shown in Figure 1.

[0054] The backbone is a cascade of multiple convolutional layers, pooling layers, and activation layers, with a segmentation head, also known as a decoding head, that decodes facial feature representations block by block. Finally, the final feature containing predictions for each pixel is obtained from the last convolutional layer in the segmentation or decoding head.

[0055] As required, the obtained feature maps can be further analyzed by a classification head, which adopts a hierarchical architecture, including a convolutional layer and a fully connected layer to obtain more detailed image probability estimates. The soft-max function is used to obtain the final probabilities of spoofed and live pixels.

[0056] To obtain accurate semantic information for each pixel, the segmentation network uses four labels at the pixel level to annotate live and spoofed images. Yellow represents fake faces, green represents fake foreground, and blue represents fake background. Additionally, red represents real faces, purple represents real foreground, and black represents real background.

[0057] To focus more intently on the face region, the input image is further processed using a standard face detection model to obtain face bounding boxes, which are then rescaled to capture more features. Edge information is contained within the rescaled bounding boxes.

[0058] The learning network outputs two results during inference: one is the segmentation task output, which divides each pixel of the input image into 1) real face / foreground pixels, 2) fake face / foreground pixels, 3) real background pixels, and 4) fake background pixels. The number of real face pixels and fake face pixels is counted and used to calculate the probability of a real face. The other result is the liveness / deception probability, which is the output of the classification task. The final prediction result is obtained by applying formula (1) as follows.

[0059] P l =α∑P sl / (∑P sl +∑P ss )+βP cl

[0060] In formula (1): P l ∑P represents the probability that the final result belongs to the "living" type. sl and ∑P ss P represents the total number of pixels representing real faces and fake faces in the output segmentation image, respectively. cl This represents the probability of a "live" type from the classification head. α and β represent the weights of the segmentation and classification results, respectively. (Default α = β = 0.5)

[0061] First, the liveness probability or spoofing probability from the segmentation network is verified. If the condition is met, the final result is output; otherwise, a second verification is performed using the liveness / spoofing probability. The fusion module combines the outputs from the segmentation and classification modules to detect the liveness of the image.

[0062] Figure 2 A face detection module of the image detection system according to the present invention is shown. A rescaled bounding box is obtained from a conventional face detection module 200. Furthermore, information beyond the face region and face portions is obtained.

[0063] The face detection module includes a processing unit 202, a scaling unit 204, and an annotation unit 206. The processing unit 202 processes the input image to generate a bounding box. After the bounding box is generated, the scaling unit performs attention scaling 204 to rescale the image to a larger size to obtain more features.

[0064] To obtain accurate semantic information for each pixel, the system uses annotation units 206. These annotation units use different colors to label the face, foreground region, and background region.

[0065] A first color is applied to the face in the image, a second color is applied to the foreground region, and a third color is applied to the background region of the first image to generate the second image. The annotation units assign colors based on the activity of the foreground and background regions.

[0066] To obtain accurate semantic information for each pixel, four labels were used at the pixel level to annotate both liveness and spoofing images. Yellow represents fake faces, green represents fake foreground, and blue represents fake background. Additionally, red represents real faces, purple represents real foreground, and black represents real background.

[0067] Figure 3 The internal components of the segmentation module 300 are shown. The segmentation module segments each pixel of the input image into 1) real face / foreground pixels, 2) fake face / foreground pixels, 3) real background pixels and 4) fake background pixels.

[0068] The segmentation module includes an extractor unit 302 and a counter unit 304. The extractor unit 302 extracts features from each block of the feature map. Pixel-level semantic information is extracted by the segmentation module 300. The counter unit 304 counts the number of spoofed pixels and live pixels respectively to predict the recognition result and detect the liveness of the image.

[0069] Figure 4 An overview of the segmentation-guided face anti-spoofing classification network according to the present invention is shown. An input image 402 is fed into the network. The network includes a backbone 404, which is a cascade of convolutional layers, pooling layers 412, and activation layers. The backbone (with the cascade of convolutional layers, pooling layers, and activation layers) is used to extract features for each block. A decoding head (also called a segmentation head) is then used to obtain the final feature representation, which carries local and global contextual information.

[0070] Features from each block are applied to obtain the final feature representation. The prediction for each pixel is obtained from the last convolutional layer in the segmentation head or decoding head. The feature map thus obtained is fed into a classification head 406 consisting of a convolutional layer and a fully connected layer to obtain the final feature representation, and the probability of deception / liveness is obtained using the softmax function.

[0071] Figure 5 This invention illustrates the pixel-level labeling of live and spoofed images using four types of labels. The segmentation module, in order to obtain accurate semantic information for each pixel, labels the live and spoofed images using four types of labels at the pixel level.

[0072] like Figure 5 In the left real image 502 and left feature image 504, yellow represents a fake face, green represents a fake foreground, and blue represents a fake background. In the right real image 506 and right feature image 508, red represents a real face, purple represents a real foreground, and black represents a real background.

[0073] Figure 6The illustration depicts data preprocessing for multi-scale generation according to the present invention. To focus more on the face region, the input image is processed using a conventional face detection model to obtain face bounding boxes. Edge information is contained within the rescaled bounding boxes.

[0074] The input image 602 is presented to the face detection module 600 to obtain a face bounding box. Attention scaling is performed 604 to rescale it to a larger or smaller image 608 to obtain more features.

[0075] Figure 7 A method for detecting the liveness of an image according to the present invention is illustrated. In step 702, the input image is presented to a face detection module, which identifies regions of interest by creating bounding boxes around the image, and a scaling unit rescales the image to obtain larger and more features.

[0076] In step 704, the image detected by the face detection module is annotated to generate a feature map. The feature map is generated based on the annotations: yellow represents fake faces, green represents fake foreground, blue represents fake background, red represents real faces, purple represents real foreground, and black represents real background.

[0077] Next, in step 706, the segmentation module extracts pixel-level semantic information from the feature map. The extractor unit extracts pixels from the feature map, and the counter unit calculates the number of spoofed and live pixels to generate a first result. In this case, the classification module is not used.

[0078] If high-level feature information needs to be extracted, a classification module will be used. Additionally, in step 708, the probability that a feature belongs to a deception or a living entity is calculated to generate a second result.

[0079] Finally, in step 710, the fusion module combines the first result with the second result to detect the activity of the image.

[0080] Figure 8 A flowchart of the cascaded result fusion according to the present invention is shown. Input image 802 is encoded by encoder 804. It is then presented to segmentation decoder 806 and classification decoder 816. The liveness / spoofing probability (1) 808 is calculated, and if the result is satisfactory, a first-stage verification 810 is performed. Later, this output becomes the final result; otherwise, classification decoder 816 calculates the liveness / spoofing probability (2) 812. The result obtained from the segmentation task is combined with the classification task to obtain the final result 814.

[0081] In some embodiments, only segmentation level information is needed, and in some embodiments, the results of the segmentation module are used as an aid in collecting the final results.

[0082] Below, for reference Figure 9 This describes an electronic device according to embodiments of the present application. Figure 9 The diagram shown is a structural schematic of an electronic device provided in an exemplary embodiment of this application.

[0083] like Figure 9 As shown, the electronic device 90 includes one or more processors 901 and memory 902.

[0084] The processor 901 may be a central processing unit (CPU) or other form of processing unit with data processing and / or instruction execution capabilities, and may control other components in the electronic device 90 to perform desired functions.

[0085] The memory 902 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 901 may execute the program instructions to implement the method for detecting face liveness on an input image using multi-task learning, as described in the embodiments of this application above, and / or other desired functions. Various contents, such as a first image, a second image, etc., may also be stored in the computer-readable storage medium.

[0086] In one example, the electronic device 90 may also include an input device 903 and an output device 904, which are interconnected via a bus system and / or other forms of connection mechanism (not shown).

[0087] The input device 903 may include, for example, a keyboard, a mouse, etc.

[0088] The output device 904 can output various information to the outside, including facial liveness detected based on the input image. The output device 904 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0089] Of course, for the sake of simplicity, Figure 9 Only some of the components of the electronic device 90 relevant to this application are shown in this illustration; components such as buses, input / output interfaces, etc., are omitted. In addition, the electronic device 90 may include any other suitable components depending on the specific application.

[0090] In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps of the methods for detecting face liveness on an input image using multi-task learning according to various embodiments of this application as described above.

[0091] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0092] Furthermore, embodiments of this application may also be computer-readable storage media storing computer program instructions that, when executed by a processor, cause the processor to perform the steps of the method for detecting face liveness on an input image using multi-task learning according to various embodiments of this application described above.

[0093] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0094] While various embodiments of the invention have been described above, it should be understood that they are presented by way of example only and not in a limiting manner. Similarly, the figures may depict exemplary architectures or other configurations of the invention, done to aid in understanding the features and functions that may be included in the invention. The invention is not limited to the illustrated exemplary architectures or configurations, but various alternative architectures and configurations can be used to achieve the desired features.

[0095] Although the invention has been described above with reference to various exemplary embodiments and implementations, it should be understood that the various features, aspects, and functions described in one or more individual embodiments are not limited to their applicability to particular embodiments. The embodiments described therein may be applied individually or in various combinations to one or more other embodiments of the invention, whether or not such embodiments are described and whether or not such features are presented as part of the described embodiments. Therefore, the breadth and scope of the invention should not be limited by any of the exemplary embodiments described above.

[0096] In some cases, the presence of broadening words and phrases such as “one or more,” “at least,” “but not limited to,” or other similar phrases should not be interpreted as implying or requiring a narrower situation to exist in situations where such broadening phrases may not exist.

Claims

1. An image detection system for detecting the liveness of an image, characterized in that, include: The face detection module includes: A processing unit is used to process the image to identify regions of interest; A scaling unit is used to scale the region of interest to identify one or more facial features to generate a first image; A labeling unit is configured to label a first color from a plurality of defined colors onto the face region of the first image, a second color from the plurality of defined colors onto the foreground region of the first image, and a third color from the plurality of defined colors onto the background region of the first image, based on a liveness image and a spoofing image, to generate a second image, wherein different defined colors are used to characterize spoofing pixels and liveness pixels in different regions; and The segmentation module includes: Extractor unit, configured to extract the deception pixels and the live pixels within the second image based on the plurality of defined colors to generate a third image; and A counter unit is used to count the live pixels and the spoof pixels in the third image to detect the liveness of the image.

2. The image detection system according to claim 1, characterized in that, The annotations are based on real faces, fake faces, real foreground, fake foreground, real background, and fake background.

3. The image detection system according to claim 1, characterized in that, The annotation unit assigns colors based on the activity of the foreground region and the background region.

4. The image detection system according to claim 1, characterized in that, The annotation unit generates a feature map of the image.

5. An image detection system for detecting the liveness of an image, characterized in that, include: The face detection module includes: A processing unit is used to process the image to identify regions of interest; A scaling unit is used to scale the region of interest to identify one or more facial features to generate a first image; A labeling unit is configured to label a first color from a plurality of defined colors onto the foreground region of the first image, and a second color from the plurality of defined colors onto the background region of the first image, based on a liveness image and a spoofing image, to generate a second image, wherein different defined colors are used to characterize spoofing pixels and liveness pixels in different regions; and The segmentation module includes: Extractor unit, configured to extract the deception pixels and the live pixels within the second image based on the plurality of defined colors to generate a third image; and A counter unit is used to count the live pixels and the spoof pixels in the third image to detect the activity in the image and generate a first result; A classification module is used to extract high-level feature information from the second image and generate a second result, wherein the classification module is configured to estimate the probability that the high-level feature information belongs to a live face; and A fusion module is used to fuse the first result and the second result to generate a final image for detecting the activity of the image.

6. The image detection system according to claim 5, characterized in that, The classification module is part of the backbone network and adopts a block architecture.

7. The image detection system according to claim 6, characterized in that, The backbone network is a cascaded structure of convolutional layers, pooling layers, and activation layers.

8. The image detection system according to claim 6, characterized in that, The backbone network is a block architecture.

9. The image detection system according to claim 7, characterized in that, The convolutional layer and the activation layer identify the features of each block.

10. The image detection system according to claim 5, characterized in that, The segmentation module includes a decoding head to obtain feature representations carrying local and global context information.

11. A method for detecting face liveness on an input image using multi-task learning, characterized in that, include: The input image is obtained through a face detection module, the region of interest in the image is identified, and the image is scaled to perform appropriate feature extraction. The face detection module marks the spoof pixels and live pixels in the image with different colors and generates a feature map based on the markings. The different colors are used to represent spoof pixels and live pixels in different regions. Pixel-level semantic information is extracted from the feature map and the number of deceitful pixels and live pixels is counted to generate a first result; The probability of the features obtained after feature extraction belonging to a deception or a living entity is calculated to extract high-level feature information to generate a second result; The first result and the second result are combined to generate the activity of the image.

12. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the method for detecting face liveness on an input image using multi-task learning as described in claim 11.

13. A computer-readable storage medium, characterized in that, The storage medium stores a computer program for executing the method for detecting face liveness on an input image using multi-task learning as described in claim 11.

Images