Image detection method and device, storage medium and electronic equipment

Abstract

The invention discloses an image detection method and device, a storage medium and electronic equipment, and relates to the technical field of image processing, and the method comprises the steps: obtaining a to-be-detected face image from video data corresponding to a target video conference; analyzing an adjacent frame change condition corresponding to the to-be-detected face image to obtain an abnormal face region with an abnormality in the to-be-detected face image; performing texture feature recognition on the abnormal face region to obtain a target texture feature sequence corresponding to the abnormal face region, the target texture feature sequence being a texture feature sequence reflecting physiological and physical rules; and generating a detection result of the video data according to the target texture feature sequence, wherein the detection result is used for determining whether a face synthesis image exists in the to-be-detected face image. Compared with the prior art, the method has the advantages that pixel-level texture change can be considered, an accurate detection result is generated, and the detection accuracy of the AI synthesized face image is improved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to an image detection method, apparatus, storage medium and electronic device. Background Technology

[0002] AI-generated human face images are created using artificial intelligence techniques (especially deep learning) to produce realistic but non-existent human face images. The core of this technology is the generative model, which includes generative adversarial networks (GANs) and diffusion models. The generated faces are highly realistic.

[0003] Currently, image detection of AI-synthesized faces mainly involves examining areas such as noise, artifacts, and color inconsistencies in the image to identify whether the face in the image is AI-synthesized.

[0004] However, in video conferencing scenarios, images need to be compressed during the acquisition process. Using this method will cause image loss during image transmission and compression, resulting in lower accuracy of the obtained image detection results. Summary of the Invention

[0005] In view of this, this application provides an image detection method, apparatus, storage medium, and electronic device, the main purpose of which is to improve the technical problem of low accuracy of the obtained image detection results in video conferencing scenarios, where the compression of images during image acquisition causes image loss during image transmission and compression.

[0006] In a first aspect, this application provides an image detection method, including: Obtain the face image to be detected from the video data corresponding to the target video conference; The changes in adjacent frames corresponding to the face image to be detected are analyzed to obtain abnormal face regions in the face image to be detected. Texture feature recognition is performed on the abnormal face region to obtain the target texture feature sequence corresponding to the abnormal face region. The target texture feature sequence is a texture feature sequence that reflects physiological and physical laws. The detection result of the video data is generated based on the target texture feature sequence, and the detection result is used to determine whether there is a synthetic face image in the face image to be detected.

[0007] Secondly, this application provides an image detection apparatus, comprising: The acquisition module is configured to acquire the face image to be detected from the video data corresponding to the target video conference; The analysis module is configured to analyze the changes in adjacent frames corresponding to the face image to be detected, and to obtain abnormal face regions in the face image to be detected that contain abnormalities. The recognition module is configured to perform texture feature recognition on the abnormal face region to obtain a target texture feature sequence corresponding to the abnormal face region, wherein the target texture feature sequence is a texture feature sequence that reflects physiological and physical laws; The generation module is configured to generate detection results of the video data based on the target texture feature sequence, and the detection results are used to determine whether there is a synthetic face image in the face image to be detected.

[0008] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the image detection method described in the first aspect.

[0009] Fourthly, this application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the computer program to implement the image detection method described in the first aspect.

[0010] Fifthly, this application provides a computer program product, which includes a computer program that, when executed by a processor, implements the image detection method described in the first aspect.

[0011] Using the above technical solution, this application provides an image detection method, apparatus, storage medium, and electronic device, comprising: acquiring a face image to be detected from video data corresponding to a target video conference; analyzing the changes in adjacent frames corresponding to the face image to be detected to obtain an abnormal face region in the face image to be detected; performing texture feature recognition on the abnormal face region to obtain a target texture feature sequence corresponding to the abnormal face region, wherein the target texture feature sequence is a texture feature sequence reflecting physiological and physical laws; generating a detection result of the video data based on the target texture feature sequence, wherein the detection result is used to determine whether a synthetic face image exists in the face image to be detected. Compared with existing technologies, this application analyzes the changes in adjacent frames of the face image to be detected in the video data of a video conference to identify abnormal face regions. This allows the application to identify microscopic dynamic changes in face images, thereby identifying abnormal face regions. Furthermore, this application can perform texture feature recognition on abnormal face regions to obtain texture feature sequences that reflect physiological and physical laws. This allows the feature recognition of this application to take into account pixel-level texture changes, thereby generating accurate detection results. This avoids the problem of inaccurate recognition caused by image loss during image transmission and compression, and improves the detection accuracy of AI-synthesized face images. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A schematic flowchart of an image detection method provided in an embodiment of this application is shown; Figure 2 A schematic flowchart of an image detection method provided in an embodiment of this application is shown; Figure 3 The diagram shows a structural schematic of an example provided in an embodiment of this application; Figure 4 This paper shows a schematic diagram of the structure of an image detection device provided in an embodiment of this application; Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0015] The embodiments of this application will now be described in more detail with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0016] To address the issue of low accuracy in image detection results in video conferencing scenarios due to image compression during acquisition and subsequent image loss during transmission and compression, this embodiment provides an image detection method, such as... Figure 1 As shown, the method includes: Step 101: Obtain the face image to be detected from the video data corresponding to the target video conference.

[0017] In this application embodiment, the video data corresponding to the target video conference can be a continuous sequence of image frames acquired or received in real time in the video conference scene of the Metaverse. The Metaverse is an immersive digital world that integrates persistent, real-time interactive, and user-co-created virtual reality (VR).

[0018] In some examples, the face image to be detected can be an image region containing a face extracted from video data. It should be noted that the face image to be detected in the embodiments of this application is the direct object for face synthesis image detection.

[0019] For example, if it is necessary to detect synthetic face images in the video conference 1 of the metaverse, the corresponding video data 1 of the video conference 1 of the metaverse can be obtained, and then the face image to be detected 1 to be detected can be obtained from the video data 1, and the face image to be detected 1 can be determined as the detection object for whether there is a synthetic face in the video conference 1.

[0020] As an alternative approach, if it is necessary to detect synthetic face images in the video conference 2 of the metaverse, the corresponding video data 2 of the video conference 2 of the metaverse can be obtained, and then the face image to be detected can be obtained from the video data 2. The face image to be detected 2 is determined as the detection object for whether there is a synthetic face in the video conference 2.

[0021] Step 102: Analyze the changes in adjacent frames corresponding to the face image to be detected to obtain abnormal face regions in the face image to be detected.

[0022] In this embodiment, the changes between adjacent frames can be the dynamic changes between the face images to be detected in adjacent frames at the pixel level or region level. For example, the face images to be detected include adjacent face images to be detected 1, face image 2, face image 3, and face image 4. Then, the changes between adjacent frames can include the dynamic changes between face image 1 and face image 2 at the pixel level or region level, the dynamic changes between face image 2 and face image 3 at the pixel level or region level, the dynamic changes between face image 3 and face image 4 at the pixel level or region level, and so on. These will not be listed one by one here.

[0023] In this embodiment of the application, the abnormal face region can be a local face region whose change pattern does not conform to physiological or physical laws, as identified by analyzing the changes in adjacent frames. For example, if the face image to be detected includes adjacent face images to be detected 1, 2, 3, and 4, then based on the changes between adjacent face images to be detected 1, 2, 3, and 4, abnormal face region 1, abnormal face region 2, etc., can be identified, and so on. No further examples will be given here.

[0024] Step 103: Perform texture feature recognition on the abnormal face region to obtain the target texture feature sequence corresponding to the abnormal face region.

[0025] Among them, the target texture feature sequence is a texture feature sequence that reflects physiological and physical laws.

[0026] In this embodiment, the target texture feature sequence can be texture features that reflect physiological (temporal consistency) and physical laws extracted from abnormal face regions. For example, if the face image to be detected includes adjacent face images to be detected 1, 2, 3, and 4, then based on the changes between adjacent face images to be detected 1, 2, 3, and 4, abnormal face regions 1 and 2 can be identified. Then, texture feature 1 that reflects physiological (temporal consistency) and physical laws can be extracted from abnormal face region 1, and texture feature 2 that reflects physiological (temporal consistency) and physical laws can be extracted from abnormal face region 2.

[0027] For example, the target texture feature sequence in this application embodiment can include features in three dimensions: 1. Pixel-level phase shift tracking: capturing the weak but physiologically rhythmic phase fluctuations in skin pixel brightness caused by heartbeats, micro-expressions, etc. AI-generated random noise or irregular jitter will destroy this phase correlation; 2. Multi-scale asymmetric gray-level co-occurrence matrix (MSA-GLCM): establishing the transfer relationship of pixel gray values ​​between consecutive frames, used to detect whether the AI-synthesized texture exhibits abrupt changes or pattern interruptions that do not conform to the laws of physical lighting and material changes when evolving across frames; 3. High-frequency noise spectral fingerprint (HFN-SF): analyzing the temporal spectral characteristics of high-frequency components of the image (including sensor noise and minute details). The noise of a real camera has specific random statistical characteristics, while the noise generated by AI may exhibit unnatural periodicity or pattern.

[0028] Step 104: Generate the detection results of video data based on the target texture feature sequence.

[0029] The detection results are used to determine whether a composite face image exists in the face image to be detected.

[0030] In this embodiment, face synthesis detection results for video data corresponding to the target meeting scene can be generated based on the identified target texture feature sequence. It should be noted that this embodiment can use the physiological-physical association microtexture descriptor (PMC-MTD) to fundamentally explore the micro-dynamics inherent in real-world faces that are difficult for AI models to perfectly simulate. It does not rely on macroscopic visual artifacts, but delves into the temporal changes at the pixel level.

[0031] Compared with existing technologies, this embodiment analyzes the changes in adjacent frames of the face image to be detected in the video data of a video conference to identify abnormal face regions. This allows the embodiment to identify microscopic dynamic changes in face images, thereby identifying abnormal face regions. Furthermore, this embodiment can perform texture feature recognition on abnormal face regions to obtain texture feature sequences reflecting physiological and physical laws. This allows the feature recognition in this embodiment to consider pixel-level texture changes, thereby generating accurate detection results. This avoids the problem of inaccurate recognition caused by image loss during image transmission and compression, improving the detection accuracy of AI-synthesized face images.

[0032] As a refinement and extension of the above embodiments, this embodiment provides an image detection method, such as... Figure 2 As shown, the method includes: Step 201: Obtain the face image to be detected from the video data corresponding to the target video conference.

[0033] For example, in this application embodiment, video stream data (i.e. video data in this application embodiment) of a video conference (i.e. the target video conference in this application embodiment) can be received, and then the input video frames can be processed to detect and locate face regions (i.e., the face image to be detected in this application embodiment).

[0034] Step 202: Analyze the changes in adjacent frames corresponding to the face image to be detected to obtain abnormal face regions in the face image to be detected.

[0035] Optionally, when performing the task of "analyzing the changes in adjacent frames corresponding to the face image to be detected to obtain abnormal face regions in the face image to be detected", the following methods can be used, but are not limited to: analyzing the changes in adjacent frames corresponding to the face image to be detected to obtain directional data and amplitude consistency data of pixel grayscale changes between adjacent frames corresponding to the face image to be detected; and determining abnormal face regions in the face image to be detected based on the directional data and amplitude consistency data.

[0036] As an alternative approach, after processing the input video frames and detecting and locating the face region (i.e., the face image to be detected in this embodiment), this embodiment can also preprocess the input video. It should be noted that related technologies mainly use fixed-size or salient region division to directly operate on the frames. In this embodiment, adaptive micro-region division needs to be introduced.

[0037] In this embodiment, motion vector discrepancy (MVD) analysis can be used to dynamically divide micro-regions (i.e., abnormal face regions in this embodiment). Instead of directly extracting motion vectors, the directionality of pixel grayscale changes between adjacent frames (i.e., directional data in this embodiment) and amplitude consistency entropy (i.e., amplitude consistency data in this embodiment) are calculated. Regions with low and persistent entropy values ​​may indicate potential micro-anomalies, which can then be identified as micro-regions (i.e., abnormal face regions in this embodiment). For example, in face regions, physiological activities such as blinking and speaking cause the direction and amplitude of pixel grayscale changes to exhibit certain regularity. However, in AI-synthesized videos, even in local areas, this regularity may be broken, leading to abnormal entropy values.

[0038] Step 203: Determine the pixel grayscale change sequence between adjacent frames corresponding to the abnormal face region, and determine the phase offset feature vector corresponding to the abnormal face region based on the pixel grayscale change sequence.

[0039] For example, based on the example in step 202, after determining the micro-regions (i.e., the abnormal face regions in the embodiments of this application), features that can reflect temporal consistency (i.e., the target texture feature sequence in the embodiments of this application) can be extracted from these regions.

[0040] It should be noted that related texture feature extraction (such as LBP, Gabor) may not be able to capture sufficiently fine temporal changes. However, in the embodiments of this application, a Physiological-Physical Correlated Micro-Texture Descriptor (PMC-MTD) can be introduced, which enables the embodiments of this application to extract texture features that reflect physiological and physical laws in a defined micro-region.

[0041] In some examples, for each identified micro-region Rk (i.e., the abnormal face region in the embodiments of this application), such as an eye region, pixel-level phase shift tracking can be performed to obtain the phase shift feature vector in the embodiments of this application.

[0042] As an optional approach, for each pixel (x,y) within the micro-region Rk (i.e., the abnormal face region in this embodiment), this embodiment can track its grayscale value change sequence Gt(x,y) in consecutive frames (i.e., the pixel grayscale change sequence in this embodiment). Then, within a small temporal window (e.g., 5-10 frames), a local Fourier transform can be performed on the Gt(x,y) sequence (i.e., the pixel grayscale change sequence in this embodiment) using Formula 1 to extract its dominant frequency and phase information. Formula 1 is specifically shown below: F(x,y,f)=F(Gt(x,y) fort∈[t0,t0+Δt])(Formula 1) For example, in the embodiments of this application, a relative phase offset matrix M for the micro-region Rk (i.e., the pixel grayscale change sequence in the embodiments of this application) can be constructed using Formula 2. ,k(t0) (i.e., the phase offset feature vector in the embodiments of this application), where each element (p,q) in the matrix represents the relative phase difference between pixels (xp,yp) and (xq,yq) at time t0, and Formula 2 is specifically shown below: Δ pq(t0) = (xp,yp,t0) (xq,yq,t0) (Formula 2) It should be noted that related technologies directly calculate optical flow or displacement, while the embodiments of this application focus on the fluctuation phase of pixel values ​​on the time axis. Specifically, the micro-movements of real human facial skin (such as skin micro-movements caused by pulse and respiration) will produce highly correlated pixel-level brightness fluctuations, and their phase shifts in local areas should conform to certain physiological models. AI synthesis may find it difficult to accurately simulate this micro-phase coupling relationship between pixels, resulting in a relative phase difference Δ. pq(t0) exhibits unusual randomness or inconsistency.

[0043] Step 204: Identify the grayscale transition information corresponding to the abnormal face region based on the pixel grayscale change sequence, and determine the grayscale change feature vector between different pixel scales in the abnormal face region based on the grayscale transition information.

[0044] For example, based on the example in step 203, for the micro-region Rk (i.e., the abnormal face region in this application embodiment), the co-occurrence matrix (i.e., the gray-scale change feature vector in this application embodiment) can be constructed in two frames Gt and Gt+1 at times t and t+1.

[0045] As an alternative approach, the multi-scale asymmetric gray-level co-occurrence matrix (MSA-GLCM) (i.e., the gray-level change feature vector in the embodiments of this application) can be used to measure the frequency of occurrence of the gray-level value i of pixel (x,y) in Gt and the gray-level value j of pixel (x+dx,y+dy) located at a specific relative position (dx,dy) in Gt+1 using Formula 3, as shown below: Ct,t+1(i,j,dx,dy)=∑(x,y)∈RkI(Gt(x,y)=iandGt+1(x+dx,y+dy)=j) (Formula 3) In Formula 3, I represents the indicator function.

[0046] For example, the embodiments of this application focus on the grayscale transfer from Gt to Gt+1, so the matrix is ​​asymmetric, and multiple (dx,dy) offsets can be selected, including not only adjacent pixels but also slightly distant pixels, in order to capture the temporal dependence of different spatial scales and extract texture features from MSA-GLCM, such as contrast, homogeneity, energy, correlation, etc., but these features are based on cross-frame temporal relationships.

[0047] It should be noted that the GLCM in the related technologies is static and mostly symmetrical. In the embodiments of this application, temporal and asymmetric characteristics can be introduced. Specifically, the changes in the texture of a real human face between consecutive frames follow physical laws (e.g., slight deformation, subtle changes in lighting). The texture synthesized by AI may have unnatural jumps or pattern interruptions when transitioning between frames. This will manifest as an abnormal frequency distribution of a specific i,j pair in MSA-GLCM, or an abnormal feature value under a specific (dx,dy).

[0048] Step 205: Identify the noise spectrum information corresponding to the abnormal face region, and determine the noise spectrum feature vector corresponding to the abnormal face region based on the noise spectrum information.

[0049] For example, based on the example in step 204, in each frame Gt of this application embodiment, the high-frequency component Nt(x,y) of the micro-region Rk (i.e., the abnormal face region in this application embodiment) can be extracted by high-pass filtering (e.g., using the Laplacian operator or the difference of Gaussian filter). This represents the minute details and noise in the image. Then, a Fourier transform can be performed on Nt(x,y) on the time axis. For each pixel (x,y), the spectrum of its high-frequency component in time (i.e., the noise spectrum feature vector in this application embodiment) can be analyzed by Formula 4. Formula 4 is as follows: S(x,y,f)=F(Nt(x,y) fort∈[t0,t0+Δt])(Formula 4) For example, embodiments of this application may further extract spectral statistical features of S(x,y,f), such as the energy distribution of the spectrum, peak frequency, bandwidth, and the smoothness or randomness measure of the spectrum (e.g., by calculating the entropy of the spectrum or the autocorrelation of the power spectral density function).

[0050] It should be noted that the embodiments of this application not only focus on image content, but also on noise features that are difficult to detect. Specifically, noise captured by real cameras (such as thermal noise and shot noise) has a specific statistical distribution and temporal correlation. Its spectrum usually presents a broad spectrum and random pattern without a specific period. The noise synthesized by AI is often learned by the generative model during training, or introduced to cover up defects. It may exhibit unnatural periodicity or pattern, or have abnormally concentrated energy at a specific frequency, lacking the characteristics of real random noise. These anomalies can be captured in the spectral fingerprint.

[0051] Step 206: Perform feature fusion based on the phase shift feature vector, grayscale change feature vector, and noise spectrum feature vector to obtain the target texture feature sequence corresponding to the abnormal face region.

[0052] Optionally, when performing the "feature fusion based on phase shift feature vector, grayscale change feature vector, and noise spectrum feature vector to obtain the target texture feature sequence corresponding to the abnormal face region", the following methods can be used, but are not limited to: extracting appearance feature vectors from the target image blocks in the abnormal face region; using a gating network to identify the appearance feature vectors and generating gating weights corresponding to the phase shift feature vector, grayscale change feature vector, and noise spectrum feature vector respectively; and fusing the phase shift feature vector, grayscale change feature vector, and noise spectrum feature vector based on the gating weights to obtain the target texture feature sequence corresponding to the abnormal face region.

[0053] It should be noted that the feature fusion methods of related technologies (such as direct stitching or simple weighted summation) require the assumption that all features are equally important for any region of the face. However, when AI simulates different regions of the face, the types of flaws it exposes are different.

[0054] As an alternative approach, for the eye region, forgery traces are mainly manifested in physiological activities, such as unnatural blinking or eye movements. Therefore, pixel-level phase shift tracking features (i.e., the phase shift feature vector in the embodiments of this application) should have a higher weight in this region.

[0055] As an alternative approach, for the skin areas of the cheeks or forehead, since there is a lack of vigorous movement, but forgery traces are often manifested in unnatural skin texture and light reflection, the multi-scale asymmetric gray-level co-occurrence matrix (i.e., the gray-level change feature vector in the embodiments of this application) and high-frequency noise spectral fingerprint features (i.e., the noise spectral feature vector in the embodiments of this application) should have higher weights.

[0056] As an alternative approach, for the mouth region, since it involves both physiological movement (speaking) and complex deformation texture, all three features may be required (i.e., the phase shift feature vector, grayscale change feature vector, and noise spectrum feature vector in the embodiments of this application), and the weights can be dynamically adjusted according to its speaking state.

[0057] For example, embodiments of this application can replace simple feature combinations with a Region-Aware Adaptive Gating Network (RAGN) to achieve personalized forgery feature analysis for different micro-regions.

[0058] In some examples, the embodiments of this application can specifically extract features of each dimension. For each micro-region Rk to be analyzed (i.e., the abnormal face region in the embodiments of this application), three independent feature vectors are calculated, including: pixel-level phase offset tracking feature vector Fphase(Rk,t) (i.e., phase offset feature vector in the embodiments of this application), multi-scale asymmetric gray-level co-occurrence matrix feature vector Fglcm(Rk,t) (i.e., gray-level change feature vector in the embodiments of this application), and high-frequency noise spectrum fingerprint feature vector Fnoise(Rk,t) (i.e., noise spectrum feature vector in the embodiments of this application).

[0059] In some examples, embodiments of this application can further determine the semantic representation of micro-regions Rk (i.e., the abnormal face region in embodiments of this application), which is the key to achieving region awareness. Specifically, an appearance feature vector Eregion(Rk) can be extracted from the original image patch of the micro-region Rk using a pre-trained small CNN (such as the shallow part of MobileNet). This vector can capture the local visual pattern of the region.

[0060] As an optional approach, embodiments of this application can further perform gating weight generation. Specifically, the region representation vector Eregion(Rk) can be input into a small gating network, which can be a simple multilayer perceptron (MLP). The input layer dimension is equal to the dimension of Eregion(Rk), the hidden layer can be set to 1-2 layers, and the output layer has 3 neurons, corresponding to the weights of the three features respectively. The network outputs a 3-dimensional weight vector W(Rk)=[wphase, wglcm, wnoise]. Then, the Softmax function is applied to the output weight vector to ensure that the sum of all weights is 1, i.e., α=Softmax(W(Rk)), to obtain the final gating weight α(Rk)=[αphase,αglcm,αnoise].

[0061] As an optional approach, this embodiment of the application can further employ weighted feature fusion. Specifically, the gating weights generated in the previous step can be used to perform weighted fusion on the three original feature vectors, generating the final, comprehensive PMC-MTD feature vector FPMC(Rk,t) (i.e., the target texture feature sequence in this embodiment of the application) through Formula 5. Formula 5 is shown below: FPMC(Rk,t)=αphase Fphase(Rk,t)⊕αglcm Fglcm(Rk,t)⊕αnoise Fnoise(Rk,t) (Formula 5) In Formula 5, ⊕ represents vector concatenation. This method preserves the independent information of each feature while adjusting the importance through weights.

[0062] As an alternative approach, embodiments of this application may further perform temporal consistency modeling and anomaly detection on the final generated FPMC(Rk,t) feature vector sequence (i.e., the target texture feature sequence in embodiments of this application) through adaptive sparse temporal topology network (ASTT-Net) mechanical energy processing or other classifiers.

[0063] It should be noted that the Region-Aware Adaptive Gated Fusion Network (RAGN) in this application embodiment can intelligently combine the aforementioned PMC-MTD features. Specifically, it can semantically or visually represent the micro-regions to be analyzed (such as the corners of the eyes, the wings of the nose, and the cheeks) through facial key points or local appearance, in order to understand which specific part of the face is being analyzed. Then, the representation information of the region can be input into a lightweight gating network, which will dynamically generate a set of exclusive fusion weights based on the characteristics of the current region. Then, this set of weights is used to perform weighted fusion of multiple PMC-MTD features extracted from the region, so that this application embodiment can automatically decide which feature should be trusted more according to different regions. For example, in the eye region, the phase shift feature (i.e., the phase shift feature vector in this application embodiment) can be automatically given higher weights; while in the forehead region, texture and noise features (i.e., the noise spectrum feature vector in this application embodiment) may be given higher weights.

[0064] In this embodiment, the static and unchanging feature combination method is abandoned. The fusion process is dynamic, and the weights are entirely determined by the inherent attributes (semantics and appearance) of the current micro-region, making the detection more targeted. During the decision analysis, the gating weights of each region can be traced back. For example, if the system determines that a video is fake and finds that the anomaly in its eye region mainly comes from phase features (because this feature has the highest weight), this provides a more specific explanation for the reason for the fakeness. By allowing the appropriate features to play their maximum role in the appropriate regions, the noise interference caused by irrelevant features can be effectively suppressed, and the signal strength of key fake traces can be amplified, thereby improving the overall detection sensitivity and accuracy. The gating network itself has a very lightweight structure, with minimal added computational overhead, making it easy to integrate into existing processes without significantly affecting real-time performance.

[0065] Step 207: Generate the detection results of video data based on the target texture feature sequence.

[0066] The detection results are used to determine whether a composite face image exists in the face image to be detected.

[0067] Optionally, when performing the "generating detection results of video data based on the target texture feature sequence", the following methods may be used, but are not limited to: determining the similarity data between different target texture features in the target texture feature sequence; generating a feature network corresponding to the target texture feature sequence based on the similarity data; determining the propagation coefficient of each target texture feature in the target texture feature sequence in the feature network, and generating the first confidence data corresponding to the feature network based on the propagation coefficient; generating the detection results of video data based on the first confidence data and the feature network.

[0068] For example, based on the example in step 205, the embodiments of this application can construct a dynamic topology graph (i.e., a feature network in this application) from the extracted PMC-MTD feature sequence (i.e., the target texture feature sequence in this application embodiment). In this dynamic topology graph (i.e., the feature network in this application embodiment), nodes represent micro-feature vectors at each time step, and edges are connected based on the similarity between feature vectors (e.g., based on a customized distance metric, such as physiological-physical distance) and temporal proximity. Physiological-physical distance will combine the directionality, amplitude, and expected physiological patterns of pixel value changes (e.g., skin micro-movements caused by pulse or respiration).

[0069] It should be noted that the dynamic topology graph (i.e. the feature network in this embodiment) does not establish full connections, but rather establishes sparse connections based on preset physiological / physical laws and feature similarity. For example, it can be assumed that the microscopic changes in real videos will form relatively smooth or predictable paths in the topology graph. If the microscopic features synthesized by AI deviate from these laws, it will lead to the formation of abnormal sparse connections or breaks in the topology graph.

[0070] In some examples, embodiments of this application can also determine the Anomaly Propagation Index (API) (i.e., the propagation coefficient in embodiments of this application). The Anomaly Propagation Index (API) (i.e., the propagation coefficient in embodiments of this application) measures the degree of influence of an anomalous feature point on its neighboring feature points in the ASTT-Net (i.e., the feature network in embodiments of this application). If an anomalous feature point can quickly and extensively affect multiple normal connections around it, it indicates that it is a strong anomaly. Embodiments of this application can measure the degree of disorder of this topology by means of API, without relying on general anomaly detection algorithms.

[0071] It should be noted that the embodiments of this application introduce an Adaptive Sparse Temporal Topology Network (ASTT-Net) (i.e., the feature network in the embodiments of this application) to establish a temporal model of micro-features and detect inconsistencies therein.

[0072] Optionally, when performing the "detection result of video data generated based on the first confidence data and the feature network", the following methods can be used, but are not limited to these: determining the target deviation data corresponding to the face image to be detected based on the first confidence data and the feature network, wherein the target deviation data is comprehensive deviation data reflecting physiological and physical laws; mapping the target deviation data to the face image to be detected as the second confidence data of the face synthesis image, and generating the detection result of video data based on the second confidence data.

[0073] For example, in the above example, the embodiments of this application can comprehensively consider different types of anomalies detected in the ASTT-Net (i.e., the feature network in the embodiments of this application) (e.g., phase shift anomalies, high-frequency noise anomalies, topological connection breaks, etc.) to calculate a comprehensive physiological-physical deviation score (i.e., the comprehensive deviation data in the embodiments of this application). The comprehensive physiological-physical deviation score (i.e., the comprehensive deviation data in the embodiments of this application) will be more interpretable than the detection methods in related technologies because it is directly related to the shortcomings of AI synthesis technology in simulating real physiological and physical laws.

[0074] For example, embodiments of this application can further map the physiological-physical deviation score to a detection confidence score, wherein a higher deviation score indicates a higher confidence that the video is AI-synthesized, thus obtaining a final output: a one-dimensional feature vector composed of one or more of the above indicators. This vector accurately quantifies the stability of the face edge background region, effectively capturing contradictions and inconsistencies generated between frames due to AI synthesis, and is then fused with other features and fed into a classifier for final judgment.

[0075] It should be noted that the embodiments of this application can use multi-dimensional physiological-physical deviation measurement (i.e., the comprehensive deviation data in the embodiments of this application) to judge the video and give a confidence level based on the detected degree of abnormality.

[0076] As an optional approach, this embodiment of the application can further fuse feature vectors from local texture feature extraction branches at different scales to form a comprehensive texture feature vector. The fusion method can include simple concatenation of feature vectors, weighted averaging, or learning fusion through an additional neural network layer, etc. The fused multi-scale local texture feature vector can then be input into a pre-trained classifier (e.g., Support Vector Machine (SVM), Logistic Regression, or a small fully connected neural network). The classifier outputs a discrimination result indicating whether the current face image is a real image or an AI-synthesized image, and may provide a confidence score (i.e., the second confidence data in this embodiment of the application). Based on the confidence score (i.e., the second confidence data in this embodiment of the application), a detection result is output, such as "real" or "fake". The result can be sent to a video conferencing system for further processing (such as alarm or blocking).

[0077] As an alternative approach, this application also provides an example that may include, but is not limited to, the following steps: Step S1: Obtain the video stream from the video conference.

[0078] Step S2: Extract the current video frame from the video stream.

[0079] Step S3: Perform face detection on the current video frame and obtain the bounding box coordinates of the face region. If no face is detected, skip the current frame or perform other processing. Step S4: Crop the face region image based on the bounding box coordinates and perform preprocessing, including but not limited to: converting the image to grayscale, standardizing the image size, and performing local contrast enhancement to eliminate the influence of factors such as lighting and size. Simultaneously, extract the background in the vicinity of the face, potentially extracting edge regions of ten to tens of pixels outwards.

[0080] Step S5: Extract local texture features from the preprocessed face region image at multiple preset scales. For example: Scale 1: Use a smaller receptive field or texture description operator to capture fine textures.

[0081] Scale 2: Use medium receptive field or texture description operator to capture medium textures.

[0082] Scales 3 and 4: Use larger receptive fields or texture descriptors to capture macroscopic texture patterns.

[0083] LBP and its variants, Gabor filter banks, or feature maps output by pre-trained convolutional neural networks at different layers can be used as texture representations at different scales.

[0084] Scale 5: Wavelet texture feature extraction.

[0085] Extracting texture features after wavelet transform. Wavelet transform can decompose an image into different frequency sub-bands, thereby analyzing texture at different resolutions. It is suitable for processing noise and edge features, realizing the decomposition and reconstruction of facial edge artifacts. Based on the reconstructed artifact signal, a threshold is constructed for the background of the facial edges to form artifact features.

[0086] Scale 6: Extraction of temporal micro-consistency features.

[0087] AI Synthesis Trace Detection Based on Temporal Microscopic Consistency Analysis: While AI synthesis technology can maintain good consistency on a macroscopic level when generating consecutive frames, temporal inconsistencies often exist at the microscopic level (pixel-level subtle changes). The microscopic changes in real human faces must conform to physiological and physical laws, while the microscopic changes in AI synthesis may exhibit random or periodic abnormal characteristics.

[0088] Step S6: Fuse the local texture features extracted from different scales. For example, feature vectors from different scales can be directly concatenated into a longer feature vector. Alternatively, an attention mechanism or learning network can be used to dynamically weight and fuse features from different scales.

[0089] Step S7: Input the fused multi-scale local texture feature vector into a pre-trained classifier for discrimination. The classifier outputs a discrimination result based on the input feature vector, such as binary classification (real / fake) or giving the probability of belonging to the real or fake category.

[0090] Step S8: Based on the classifier's output, determine whether the face image in the current video frame is AI-generated. The detection result can be compared with a set threshold. Step S9: Repeat steps S2-S8 to process subsequent frames of the video stream, thereby achieving real-time or near real-time detection of the entire video.

[0091] As an example, this application also provides an example, such as Figure 3The diagram illustrates the system results of this example. The system can receive video stream data (i.e., video data) from a video conference (i.e., the target video conference in this embodiment) via a video stream input module. The input video frames can then be processed by a face detection module to detect and locate face regions (i.e., the face image to be detected in this embodiment). The extracted face regions are then standardized using a preprocessing module, such as cropping, scaling, grayscale conversion, and contrast enhancement. Finally, detection is performed using multiple parallel processing branches within a multi-scale local texture feature extraction module. Each branch is responsible for extracting local texture features of the face region at a specific scale. For example, filters of different sizes (such as Gabor filter banks) or Local Binary Pattern (LBP) operators with different sampling rates can be used, or feature maps output from different convolutional layers of a deep learning model can be used to represent texture information at different scales. Each branch outputs a local texture feature vector corresponding to its scale.

[0092] Among related technologies, AI-synthesized face detection technology mainly includes methods based on image quality analysis, physiological signal detection (such as heart rate and eye movement), spatiotemporal feature analysis, and deep learning. Among these, image quality analysis-based methods typically check for noise, artifacts, and color inconsistencies in images; physiological signal-based methods need to capture and analyze physiological signals that are difficult to forge; spatiotemporal feature analysis-based methods examine the coherence and consistency between video frames; and deep learning-based methods automatically learn features that distinguish real and forged images by training models such as convolutional neural networks. However, these related technologies have limitations in video conferencing scenarios: image quality-based methods are susceptible to video compression and transmission losses; physiological signal detection requires high-quality video sources and complex analysis techniques; simple spatiotemporal features are insufficient to capture complex temporal artifacts; while some general-purpose deep learning models, although powerful, may not be optimized for the characteristics of AI synthesis in video conferencing scenarios (such as real-time requirements, limited computing resources, and specific synthesis artifacts), and may ignore key texture information in local facial regions; as AI synthesis technology continues to advance, synthesized images become increasingly realistic, leading to a reduction in global feature differences; images in video conferencing are usually compressed and transmitted, introducing noise and artifacts, which interfere with image quality-based detection; complex detection algorithms may cause delays, affecting user experience; deep learning methods focus more on global or mid-level features and are insufficient in capturing the temporal consistency of AI synthesis at the micro-texture level.

[0093] It's important to note that when fusing multiple forgery features (such as texture, noise, and temporal variations), related technologies typically employ simple splicing or fixed-weighting, which essentially assumes all features are equally important for any region of the face. This static, non-differentiated fusion approach fails to accurately capture the different types of flaws exposed by AI when forging different parts of the face. Due to this lack of specificity, irrelevant features in certain facial regions can actually become noise, interfering with and masking the key signals that truly reveal forgery. For example, in the eye region, subtle anomalies in texture features may be far less important than anomalies in physiological motion (phase) features, but traditional methods cannot reflect this difference, resulting in limited detection sensitivity and accuracy.

[0094] In this embodiment, the application effectively amplifies key forgery signals and suppresses noise interference from irrelevant features through an adaptive mechanism that "maximizes the effect of appropriate features in appropriate regions," thereby enabling more accurate and sensitive capture of subtle flaws in highly realistic AI-synthesized images. Furthermore, it eliminates the "black box" of forgery detection; when the system determines a video to be forged, the judgment basis can be clearly explained by tracing back the gating weights of each region. For example, it can be concluded that "the video is determined to be forged primarily because the physiological phase features (highest weight) of its eye region exhibit severe abnormalities," providing clear clues for subsequent analysis and evidence collection. Since this method is based on the inherent physiological and physical partitioning characteristics of the human face and is not targeted at specific types of generative models, it possesses good robustness and versatility against various evolving AI synthesis technologies. The embodiment also makes the core gating network structure lightweight, with minimal added computational overhead, allowing it to be seamlessly integrated into existing detection processes as an efficient "plugin," improving performance without affecting the system's real-time capabilities.

[0095] Compared with existing technologies, this embodiment analyzes the changes in adjacent frames of the face image to be detected in the video data of a video conference to identify abnormal face regions. This allows the embodiment to identify microscopic dynamic changes in face images, thereby identifying abnormal face regions. Furthermore, this embodiment can perform texture feature recognition on abnormal face regions to obtain texture feature sequences reflecting physiological and physical laws. This allows the feature recognition in this embodiment to consider pixel-level texture changes, thereby generating accurate detection results. This avoids the problem of inaccurate recognition caused by image loss during image transmission and compression, improving the detection accuracy of AI-synthesized face images.

[0096] Furthermore, as Figure 1 and Figure 2 The specific implementation of the method shown in this embodiment provides an image detection device, such as... Figure 4As shown, the device includes: an acquisition module 31, an analysis module 32, an identification module 33, and a generation module 34.

[0097] The acquisition module 31 is configured to acquire the face image to be detected from the video data corresponding to the target video conference; Analysis module 32 is configured to analyze the changes in adjacent frames corresponding to the face image to be detected, and to obtain abnormal face regions in the face image to be detected that are abnormal. The recognition module 33 is configured to perform texture feature recognition on the abnormal face region to obtain a target texture feature sequence corresponding to the abnormal face region, wherein the target texture feature sequence is a texture feature sequence that reflects physiological and physical laws; The generation module 34 is configured to generate a detection result of the video data based on the target texture feature sequence, and the detection result is used to determine whether there is a synthetic face image in the face image to be detected.

[0098] In some examples of this embodiment, the recognition module 33 is specifically configured to: determine the pixel grayscale change sequence between adjacent frames corresponding to the abnormal face region; determine the phase shift feature vector corresponding to the abnormal face region based on the pixel grayscale change sequence; identify the grayscale transfer information corresponding to the abnormal face region based on the pixel grayscale change sequence; determine the grayscale change feature vector between different pixel scales in the abnormal face region based on the grayscale transfer information; identify the noise spectrum information corresponding to the abnormal face region; determine the noise spectrum feature vector corresponding to the abnormal face region based on the noise spectrum information; and perform feature fusion based on the phase shift feature vector, the grayscale change feature vector, and the noise spectrum feature vector to obtain the target texture feature sequence corresponding to the abnormal face region.

[0099] In some examples of this embodiment, the recognition module 33 is further configured to extract appearance feature vectors from target image blocks in the abnormal face region; use a gating network to recognize the appearance feature vectors to generate gating weights corresponding to the phase shift feature vector, the grayscale change feature vector, and the noise spectrum feature vector, respectively; and perform feature fusion on the phase shift feature vector, the grayscale change feature vector, and the noise spectrum feature vector according to the gating weights to obtain the target texture feature sequence corresponding to the abnormal face region.

[0100] In some examples of this embodiment, the generation module 34 is specifically configured to: determine similarity data between different target texture features in the target texture feature sequence; generate a feature network corresponding to the target texture feature sequence based on the similarity data; determine the propagation coefficient of each target texture feature in the target texture feature sequence in the feature network, and generate first confidence data corresponding to the feature network based on the propagation coefficient; and generate the detection result of the video data based on the first confidence data and the feature network.

[0101] In some examples of this embodiment, the generation module 34 is further configured to determine the target deviation data corresponding to the face image to be detected based on the first confidence data and the feature network, wherein the target deviation data is comprehensive deviation data reflecting physiological and physical laws; map the target deviation data to the face image to be detected as the second confidence data of the face composite image, and generate the detection result of the video data based on the second confidence data.

[0102] In some examples of this embodiment, the analysis module 32 is specifically configured to analyze the changes in adjacent frames corresponding to the face image to be detected, and obtain directional data and amplitude consistency data of pixel grayscale changes between adjacent frames corresponding to the face image to be detected; based on the directional data and the amplitude consistency data, determine abnormal face regions with abnormalities in the face image to be detected.

[0103] It should be noted that other corresponding descriptions of the functional units involved in the image detection device provided in this embodiment can be found in [reference needed]. Figure 1 and Figure 2 The corresponding descriptions in [the document] will not be repeated here.

[0104] Based on the above, Figure 1 and Figure 2 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 and Figure 2 The method shown.

[0105] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0106] like Figure 5 The diagram shown is a hardware structure schematic of an electronic device according to the present invention, comprising: At least one processor 401; and, A memory 402 is communicatively connected to at least one of the processors 401; wherein, The memory 402 stores instructions that can be executed by at least one of the processors to enable the at least one of the processors to perform the image detection method as described above.

[0107] Figure 5 Take a processor 401 as an example.

[0108] The electronic device may also include an input device 403 and a display device 404.

[0109] The processor 401, memory 402, input device 403, and display device 404 can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0110] Memory 402, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the image detection method in the embodiments of this application, for example, Figure 1 and Figure 2 The method flow is shown. The processor 401 executes various functional applications and data processing by running non-volatile software programs, instructions, and modules stored in the memory 402, thereby implementing the image detection method in the above embodiments.

[0111] The memory 402 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the image detection method, etc. Furthermore, the memory 402 may include high-speed random access memory and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 402 may optionally include memory remotely located relative to the processor 401, and these remote memories may be connected to the apparatus performing the image detection method via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0112] The input device 403 can receive user clicks and generate signal inputs related to user settings and function control of the image detection method. The display device 404 may include a display screen or other display device.

[0113] The one or more modules are stored in the memory 402, and when run by the one or more processors 401, they execute the image detection method in any of the above method embodiments.

[0114] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0115] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0116] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0117] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms, or it can be implemented by hardware. By applying the solution of this embodiment, compared with the existing technology, this embodiment analyzes the changes of adjacent frames of the face image to be detected in the video data of the video conference, thereby determining the abnormal face region. This embodiment can identify the micro-dynamic changes in the face image, thereby determining the abnormal face region. This embodiment can also perform texture feature recognition on the abnormal face region to obtain a texture feature sequence reflecting physiological and physical laws. This allows the feature recognition of this embodiment to take into account pixel-level texture changes, thereby generating accurate detection results and avoiding the problem of inaccurate recognition caused by image loss during image transmission and compression, thus improving the detection accuracy of AI-synthesized face images.

[0118] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0119] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. An image detection method, characterized in that, include: Obtain the face image to be detected from the video data corresponding to the target video conference; The changes in adjacent frames corresponding to the face image to be detected are analyzed to obtain abnormal face regions in the face image to be detected. Texture feature recognition is performed on the abnormal face region to obtain the target texture feature sequence corresponding to the abnormal face region. The target texture feature sequence is a texture feature sequence that reflects physiological and physical laws. The detection result of the video data is generated based on the target texture feature sequence, and the detection result is used to determine whether there is a synthetic face image in the face image to be detected.

2. The method according to claim 1, characterized in that, The step of performing texture feature recognition on the abnormal face region to obtain the target texture feature sequence corresponding to the abnormal face region includes: Determine the pixel grayscale change sequence between adjacent frames corresponding to the abnormal face region, and determine the phase offset feature vector corresponding to the abnormal face region based on the pixel grayscale change sequence; The grayscale transition information corresponding to the abnormal face region is identified based on the pixel grayscale change sequence, and the grayscale change feature vector between different pixel scales in the abnormal face region is determined based on the grayscale transition information. Identify the noise spectrum information corresponding to the abnormal face region, and determine the noise spectrum feature vector corresponding to the abnormal face region based on the noise spectrum information; Based on the phase offset feature vector, the grayscale change feature vector, and the noise spectrum feature vector, feature fusion is performed to obtain the target texture feature sequence corresponding to the abnormal face region.

3. The method according to claim 2, characterized in that, The step of fusing features based on the phase shift feature vector, the grayscale change feature vector, and the noise spectrum feature vector to obtain the target texture feature sequence corresponding to the abnormal face region includes: Extract appearance feature vectors from target image blocks in the abnormal face region; A gating network is used to identify the appearance feature vector, and gating weights corresponding to the phase shift feature vector, the grayscale change feature vector, and the noise spectrum feature vector are generated respectively. Based on the gating weights, the phase offset feature vector, the grayscale change feature vector, and the noise spectrum feature vector are fused to obtain the target texture feature sequence corresponding to the abnormal face region.

4. The method according to claim 1, characterized in that, The step of generating the detection result of the video data based on the target texture feature sequence includes: Determine the similarity data between different target texture features in the target texture feature sequence; Based on the similarity data, a feature network corresponding to the target texture feature sequence is generated; Determine the propagation coefficient of each target texture feature in the target texture feature sequence in the feature network, and generate the first confidence data corresponding to the feature network based on the propagation coefficient; The detection results of the video data are generated based on the first confidence data and the feature network.

5. The method according to claim 4, characterized in that, The step of generating the detection result of the video data based on the first confidence data and the feature network includes: The target deviation data corresponding to the face image to be detected is determined based on the first confidence data and the feature network. The target deviation data is a comprehensive deviation data that reflects physiological and physical laws. The target deviation data is mapped to the second confidence data of the face image to be detected as a synthesized face image, and the detection result of the video data is generated based on the second confidence data.

6. The method according to any one of claims 1 to 5, characterized in that, The analysis of changes in adjacent frames corresponding to the face image to be detected identifies abnormal face regions within the face image to be detected, including... The changes in adjacent frames corresponding to the face image to be detected are analyzed to obtain the directional data and amplitude consistency data of the pixel grayscale changes between adjacent frames corresponding to the face image to be detected. Based on the directional data and the amplitude consistency data, abnormal face regions with anomalies are identified from the face image to be detected.

7. An image detection device, characterized in that, include: The acquisition module is configured to acquire the face image to be detected from the video data corresponding to the target video conference; The analysis module is configured to analyze the changes in adjacent frames corresponding to the face image to be detected, and to obtain abnormal face regions in the face image to be detected that contain abnormalities. The recognition module is configured to perform texture feature recognition on the abnormal face region to obtain a target texture feature sequence corresponding to the abnormal face region, wherein the target texture feature sequence is a texture feature sequence that reflects physiological and physical laws; The generation module is configured to generate detection results of the video data based on the target texture feature sequence, and the detection results are used to determine whether there is a synthetic face image in the face image to be detected.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

10. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images