AI generated video detection method and system fusing global and inter-frame semantics

Abstract

The invention relates to an AI generated video detection method and system fusing global and inter-frame semantics, and belongs to the technical field of artificial intelligence, computer vision and multimedia security. The method comprises the following steps: extracting advanced semantic features of a video frame by using a pre-trained CLIP model; modeling a video overall semantic distribution and narrative structure by adopting a Transform structure through a global semantic branch; capturing inter-frame local semantic mutation in parallel through an inter-frame semantic consistency branch by adopting BiLSTM and an attention mechanism; and finally, carrying out adaptive weighted fusion on discrimination results of the two branches through a dynamic fusion module, and outputting a final detection probability. According to the method, videos generated by multiple generation technologies such as GAN, a diffusion model and Sora can be effectively recognized from the high-level semantic dimension, and the method has high robustness to interference such as video compression and fuzziness and cross-model generalization ability and is suitable for actual scenes such as network content auditing and media evidence obtaining.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of artificial intelligence, computer vision and multimedia security technology, and particularly relates to an AI generated video detection method and system fusing global and inter-frame semantics. BACKGROUND

[0002] In recent years, generative artificial intelligence (AIGC) technologies centered on generative adversarial networks (GAN) and diffusion models have made breakthrough progress. In particular, the advent of video generation systems such as Sora, Pika, and Kling has enabled AI to generate content that is highly similar to real videos in terms of visual effects, motion coherence, and semantic expression, greatly promoting the development of industries such as content creation and special effects.

[0003] However, the misuse of AIGC technology has also brought serious security challenges. Highly realistic AI generated videos (AIGCVideo) can be used for malicious activities such as creating fake news, manipulating public opinion, and committing identity fraud, posing a serious threat to social security and public trust. Therefore, developing effective and reliable AI generated video detection technology has become an urgent need in the field of artificial intelligence security.

[0004] Existing video forgery detection methods can be mainly divided into several categories:

[0005] Detection based on low-level visual artifacts: Early detection methods mainly focus on low-level artifacts in images or video frames, such as color statistical anomalies, frequency domain noise, and compression traces. However, as the capabilities of generative models improve, these low-level artifacts become increasingly concealed, significantly reducing the effectiveness of such methods.

[0006] Detection based on spatio-temporal features: To address the temporal dimension unique to videos, some research has shifted to modeling spatio-temporal features. For example, some methods (such as AIGVDet) use optical flow estimation to capture inconsistencies in motion between video frames, such as motion irregularities and target outline blurring. Other methods (such as DuB3D and DeCoF) use VideoTransformer or 3D convolutional networks to directly learn spatio-temporal features from frame sequences to identify violations of physical laws and pixel abnormal drifts. Although these methods have improved detection capabilities to some extent, they still have the following problems: High computational cost: Optical flow calculation is usually computationally intensive, making it difficult to meet the needs of real-time or large-scale detection. Dependence on low-level features: Their core is still to capture pixel-level or motion-level abnormalities, and when high-quality generated videos (such as Sora generation) are visually and motionally indistinguishable from real videos, these methods will fail.

[0007] Detection based on multimodal and large models: Recently, researchers have begun to utilize large multimodal models (LMMs) or large language models (LLMs) for detection (such as LAVID and BusterX). These methods leverage the powerful visual understanding and reasoning capabilities of large models, enabling them not only to determine authenticity but also to provide explainable justifications. However, these methods typically involve large model sizes, high inference costs, and difficult deployment, limiting their application in many real-world scenarios.

[0008] In summary, existing technologies face two core challenges:

[0009] Insufficient modeling of high-level semantic anomalies: Most methods focus on low-level visual features or physical motion patterns, while having limited ability to capture anomalies at the high-level semantic level—such as logical sequences of events that contradict common sense (e.g., seeing the result before the cause), unexplained attribute mutations of the same object between different frames in a video (e.g., a sudden change in the color of a person's clothing), or drastic shifts in the overall narrative style of a video scene without reasonable transitions. These deep semantic deficiencies are becoming key clues in distinguishing high-quality AI-generated videos.

[0010] Insufficient generalization ability and robustness: Existing methods are often sensitive to specific types of generative models or artifacts. When faced with unseen generative models (cross-model generalization) or when the video has undergone common processing such as compression and blurring (robustness), their performance will significantly decrease.

[0011] Therefore, there is an urgent need to design a new detection method that can effectively capture the essential and ubiquitous semantic defects in AI-generated videos from a high-level semantic space, and that is efficient, has strong generalization ability and robustness. Summary of the Invention

[0012] The purpose of this invention is to overcome the aforementioned deficiencies of existing technologies and provide an AI-generated video detection method and system that integrates global and inter-frame semantics. This invention aims to address the issue from a high-level semantic dimension of video content, rather than relying on easily mutated and hidden low-level visual artifacts. By modeling and analyzing the global narrative structure and local temporal semantic logic of the video in parallel, this method can effectively identify high-quality, highly deceptive AI-generated videos and exhibits superior generalization ability and robustness against different generation models, video compression, and other common interferences.

[0013] To achieve the above objectives, this invention provides an AI-generated video detection method that integrates global and inter-frame semantics, comprising the following core steps:

[0014] Step S1: High-level semantic feature extraction of the video frame sequence. The video to be detected is subjected to frame sequence extraction to obtain a video frame sequence. Each video frame in the sequence is input into an image encoder in a multimodal contrastive language-image (CLIP) model pre-trained on massive image-text pairs to extract the high-level semantic feature vector of that frame. The feature vector captured by the CLIP model not only contains the visual attributes of objects, but also contains rich semantic information aligned with natural language concepts, such as object category, action logic, scene context, etc., thereby constructing a robust semantic perception feature space that is crucial for subsequent semantic analysis.

[0015] Step S2: Global Semantic Structure Modeling (Global Semantic Branch, GST) The frame-level semantic feature vector sequence extracted in Step S1 is input into a global semantic modeling branch. This branch employs a Transformer sequence model architecture to perform deep modeling of the semantic feature sequence of the entire video. Its purpose is to capture the overall semantic distribution, narrative rhythm, and thematic coherence of the video over a global time span. This branch can effectively detect macroscopic semantic anomalies in AI-generated videos caused by single text prompts or defects in the generation mechanism, such as unnatural repetition of scene themes, chaotic overall narrative logic, or abrupt shifts in video style without reasonable transitions. By pooling the output sequence of the Transformer model (e.g., average pooling), a global semantic representation vector for the video is formed and input into a classifier to obtain preliminary global semantic discrimination results.

[0016] Step S3: Inter-frame Semantic Temporal Consistency Modeling (Inter-frame Consistency Branch, IFSC) In parallel, the frame-level semantic feature sequence extracted in Step S1 is input into an inter-frame semantic consistency modeling branch. This branch aims to capture, with fine granularity, the local semantic dependencies and temporal logical jumps between video frame sequences. It utilizes a Bidirectional Long Short-Term Memory (BiLSTM) network combined with a frame-level attention mechanism to model the evolution of inter-frame semantic features. This branch focuses on identifying common local semantic inconsistencies in AI-generated videos, such as: the color of clothing of the same person changing arbitrarily in adjacent frames, the trajectory of an object not conforming to physical laws, and the interruption of semantic coherence of specific actions. Through an attention mechanism, it dynamically focuses on the frame with the most significant semantic abrupt change, ultimately forming the inter-frame semantic consistency evaluation result of the video, which is then input into a classifier to obtain preliminary inter-frame consistency discrimination results.

[0017] Step S4: Dynamic Fusion and Final Judgment of Dual-Branch Discrimination Results. To comprehensively utilize semantic information from both global and local dimensions, this invention designs a dynamic fusion module. This module receives preliminary discrimination results (such as classification probabilities) from the global semantic branch (step S2) and the inter-frame consistency branch (step S3). A lightweight neural network (e.g., a multilayer perceptron MLP) is trained to predict an optimal set of fusion weights for the current input video sample based on the output characteristics of the two branches. This dynamic weighting mechanism adaptively adjusts the contribution of the two information streams; for example, when the video has significant global narrative problems, the weight of the global semantic branch increases. Finally, the weighted fusion result is fed into the final classifier, outputting a final discrimination score indicating that the video to be detected is AI-generated content.

[0018] Step S5: Compare the final discrimination score with a preset threshold, and output the final judgment result of whether the video to be detected is AI-generated content.

[0019] Preferably, in step S1, the frame sequence extraction adopts a uniform sampling method.

[0020] Preferably, in step S1, the multimodal contrastive language-image model is the CLIP model, and its image encoder is a visual Transformer architecture.

[0021] Preferably, in step S2, the Transformer sequence model is a multi-layer Transformer encoder; when the frame-level semantic feature sequence is input, it is first combined with position encoding to introduce temporal information.

[0022] Preferably, in step S2, the pooling is an average pooling operation performed over time.

[0023] Preferably, in step S3, the bidirectional long short-term memory network is a BiLSTM containing multiple hidden layers; the frame-level attention mechanism assigns attention weights to each time step in the frame sequence by calculating a weight vector.

[0024] Preferably, in step S4, the dynamic fusion module performs the following operations: concatenating the two preliminary discrimination results, inputting the concatenated vector into the lightweight neural network, which is then followed by a Softmax layer to output two dynamic fusion weights.

[0025] This invention also provides an AI-generated video detection system that integrates global and inter-frame semantics, for implementing the method, comprising:

[0026] The feature extraction module is used to extract frame sequences from the input video and obtain frame-level high-level semantic feature sequences through a pre-trained image encoder.

[0027] The global semantic modeling module performs global semantic structure modeling on the feature sequence based on the Transformer architecture and outputs global semantic discrimination results.

[0028] The inter-frame consistency modeling module performs inter-frame semantic consistency modeling on the feature sequence based on BiLSTM and attention mechanism, and outputs the inter-frame consistency discrimination result.

[0029] The dynamic fusion module receives two discrimination results, predicts fusion weights through a neural network, performs weighted fusion, and outputs the final detection probability.

[0030] The determination module is used to compare the final probability with the threshold and output the detection result.

[0031] Preferably, the image encoder used by the feature extraction module is a CLIP model visual encoder.

[0032] Preferably, the dynamic fusion module includes a multilayer perceptron and a Softmax layer for generating input-adaptive fusion weights.

[0033] Compared with the prior art, the present invention has the following significant advantages:

[0034] Detection from a high-level semantic dimension to perceive deeper forgery traces: By introducing the CLIP model to extract high-level semantic features, this invention gets rid of the dependence on pixel-level artifacts and can capture the deep semantic rules (such as narrative logic and causal relationships) inherent in real videos that are difficult for AI to perfectly imitate, thereby effectively detecting visually highly realistic forgeries.

[0035] The complementary dual-branch structure enhances detection accuracy and robustness: the global semantic branch focuses on the macroscopic narrative structure of the video, while the inter-frame consistency branch focuses on the microscopic temporal logic. This design, which performs complementary analysis at different granularities and levels, enables the method to demonstrate good detection capabilities against diverse forgery techniques (such as GANs, diffusion models, Sora, etc.).

[0036] Superior generalization ability and robustness: The core of this method lies in analyzing more fundamental semantic defects that do not depend on the implementation details of specific generative models. Temporal perturbation experiments have demonstrated that even when the order of video frames is shuffled, the performance of this method is almost unaffected, confirming that its detection is based on the semantic content itself rather than temporal artifacts. Furthermore, high-level semantic features have a natural resistance to common interferences such as video compression, blurring, and scaling, enabling this method to achieve leading performance and robustness on multiple public benchmark datasets (GVD and GVF).

[0037] Highly efficient, practical, and easy to deploy: Compared to detection schemes based on large language models, this invention achieves top-notch detection performance without relying on high computing resources, possesses good practicality and deployment value, and is suitable for various practical application scenarios such as network content review and media forensics. Attached Figure Description

[0038] Figure 1 The advanced semantic feature visualization analysis comparison chart shows the difference in the ability of CLIP features (left) and traditional ResNet features (right) used in this invention to locate semantic subjects in images. Figure 2 : The overall model structure diagram of the detection method proposed in this invention. Figure 3 The figure shows the performance results of the method of this invention compared with a number of existing methods on the GVD dataset. Figure 4 The figure shows the performance results of the method of this invention compared with the existing state-of-the-art methods on the GVF dataset. Figure 5 The figure shows the robustness test results of the detection performance of the method of the present invention under different video compression rates (H.264). Figure 6 The figure shows the results of a robustness test of the detection performance of the method of the present invention on video content after different degrees of Gaussian blurring. Figure 7 The image shows the robustness test results of the detection performance of the method of the present invention on video content after being scaled at different ratios.

[0039] Figure 8 The method of this invention is for video frame order perturbation.

[0040] Figure 9 Ablation experiments of the method of the present invention. Detailed Implementation

[0041] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0042] Example 1

[0043] This invention proposes an AI-generated video detection method that integrates global and inter-frame semantics. The overall model structure of its preferred embodiment is as follows: Figure 2 As shown. The specific implementation steps of this method are as follows:

[0044] Step S1: Video Preprocessing and Frame-Level High-Level Semantic Feature Extraction. The input video V to be detected is preprocessed. First, a sequence of video frames of length T (e.g., T=8) is extracted from the video using a uniform sampling method, denoted as {I1, I2,..., I...}. T}. Frame extraction can be performed using uniform sampling. Subsequently, each frame image I in the sequence is... t The input is fed into a pre-trained CLIP model's visual encoder (e.g., ViT-B / 32). This encoder processes each frame of image I... t This is mapped to a high-dimensional (e.g., 512-dimensional) semantic feature vector f. t These feature vectors constitute a frame-level semantic feature sequence F = {f1, f2, ..., f...} T}, which serves as the input for the subsequent two parallel analysis branches.

[0045]

[0046]

[0047] Step S2: Global Semantic Structure Modeling (GST Branch) The frame-level semantic feature sequence F obtained in Step S1 is input into the Global Semantic Modeling (GST) branch. The core of this branch is a multi-layer Transformer encoder (e.g., containing 2 to 6 layers).

[0048] First, the input semantic feature sequence F is combined with positional encoding to introduce temporal information.

[0049] The processed sequence is then fed into the Transformer encoder. Through its internal multi-head self-attention mechanism, the model is able to learn the semantic dependencies and organization patterns of the video frame sequence on a global scale.

[0050] The Transformer encoder outputs a context-aware feature sequence. The sequence is then subjected to mean pooling over the time dimension to obtain a fixed-length representation vector that represents the global semantics of the entire video. .

[0051] Finally, vectors The video is fed into a multilayer perceptron (MLP) classification head and ultimately passes through a sigmoid activation function, outputting a probability value representing the video's global semantic level as generated by AI. .

[0052]

[0053]

[0054]

[0055]

[0056]

[0057] Step S3: Inter-frame semantic temporal consistency modeling (IFSC branch) In parallel with step S2, the frame-level semantic feature sequence F is input into the inter-frame semantic consistency modeling branch (IFSC).

[0058] The feature sequence F is fed into a bidirectional long short-term memory network (BiLSTM, e.g., containing 1 to 3 layers). The BiLSTM learns long-term semantic dependencies between frames from both forward and backward time dimensions, outputting the hidden state at each time step. This state integrates semantic information from the preceding and following context.

[0059] BiLSTM output sequence It is fed into a frame-level attention mechanism. This attention mechanism uses a learnable weight matrix. sum vector Calculate the attention weights for each frame. This allows for the allocation of higher weights to frames with significant semantic mutations.

[0060] By aggregating the weighted features, a representation vector that comprehensively reflects the degree of local semantic consistency between video frames is obtained. .

[0061] Similar to the GST branch, vector It is also fed into an MLP classification head, which outputs a probability value representing the video's inter-frame semantic consistency at the level of AI generation. It is worth noting that this branch models consistency within a high-level semantic space. Experiments show that even when the video frame sequence is randomly shuffled, the detection accuracy of this branch remains almost unchanged. This indicates that its discrimination mechanism is independent of the physical temporal order, but rather focuses on the rationality of the content within the semantic framework. This is the key difference between this invention and traditional temporal modeling methods.

[0062]

[0063]

[0064]

[0065]

[0066] Step S4: Dynamic fusion of dual-branch discrimination information. To integrate the detection capabilities of the GST branch and the IFSC branch, this invention designs a dynamic fusion module.

[0067] The initial discrimination probability from the GST branch and the initial discrimination probability from the IFSC branch By concatenating, we get .

[0068] The concatenated vector Input a lightweight fusion network (Attn_fuse), which consists of a small MLP followed by a Softmax layer. This network is trained to output two dynamic fusion weights. and ,and + = 1.

[0069] These two weights are input-aware, meaning they are dynamically calculated based on the features of the current input video, rather than being fixed global values.

[0070] Final fusion discrimination score It is obtained by weighted summation of the output probabilities of the two branches.

[0071]

[0072]

[0073]

[0074] Step S5: Final classification decision - The fused discrimination score obtained in step S4 will be used for the final classification decision. Compare with a preset threshold (e.g., 0.5). If If the value exceeds the threshold, the video is determined to be an AI-generated video; otherwise, it is determined to be a real video.

[0075] The reason why the method of this invention exhibits good robustness and cross-model generalization ability is that: First, it relies on high-level semantic features extracted by the CLIP model. These features are more resistant to video compression (such as pixel-level low-level artifacts) than pixel-level low-level artifacts. Figure 5 As shown), blurry (such as) Figure 6 As shown), scaling (as shown) Figure 7This invention avoids information loss caused by common interferences such as those shown in the figure. Secondly, by modeling the global semantic structure (GST branch) and inter-frame local semantic temporal logic (IFSC branch) of the video content in parallel, this invention can capture the deep semantic patterns inherent in real video content that are difficult to perfectly and continuously replicate by diverse AI generation models. This focus on the essential semantic attributes of the content makes the detection system less susceptible to the superficial and volatile artifacts introduced by specific generation models. Therefore, it can maintain high detection performance even when facing unseen generation models or complex and varied video content, as shown in its leading performance on the GVD and GVF datasets (see Figure 1). Figure 3 and Figure 4 ).

[0076] Furthermore, to verify the effectiveness of the core mechanism of this method, key experiments were conducted. For example... Figure 8 As shown, even when the video frame sequence is randomly shuffled, the detection accuracy of this method remains almost unaffected. This result strongly demonstrates that the criterion for the inter-frame semantic consistency branch (IFSC) is the consistency of semantic logic, rather than the continuity of physical temporal sequence. This is fundamentally different from traditional detection methods that rely on motion artifacts, and is also the key to the strong generalization ability of this invention.

[0077] at last, Figure 9 The ablation experiment results shown indicate that both the Global Semantic Branch (GST) and the Inter-Frame Semantic Consistency Branch (IFSC) make significant contributions to the final performance, and the dynamic fusion module can effectively integrate the advantages of both. The combination achieves the best detection effect, verifying the rationality and necessity of the dual-branch structure design of this invention.

[0078] Example 2

[0079] This invention also provides an AI-generated video detection system that integrates global semantic awareness and inter-frame semantic consistency to implement the above-mentioned method. The system mainly includes the following modules:

[0080] 1. Feature Extraction Module

[0081] This module is responsible for preprocessing and feature extraction of the input video to be detected. Its workflow is as follows: First, the module extracts a fixed-length sequence of video frames from the input video using uniform sampling. Then, it inputs each frame into a pre-trained CLIP model visual encoder (e.g., an encoder based on the Vision Transformer architecture). This encoder transforms each frame into a high-dimensional feature vector rich in high-level semantic information. Finally, the module outputs a sequence of semantic feature vectors from all frames, providing a foundation for subsequent analysis.

[0082] 2. Global Semantic Modeling Module

[0083] This module receives frame-level semantic feature sequences from the feature extraction module, focusing on analyzing the overall narrative structure and macro-semantic coherence of the video. At its core is an encoder based on the Transformer architecture. During processing, the module first adds positional encoding to the feature sequences to incorporate temporal information. Then, it utilizes the Transformer's multi-head self-attention mechanism to capture deep dependencies between semantics across the entire sequence. Next, the module performs average pooling on the processed sequence along the temporal dimension, generating a fixed-dimensional global semantic summary vector. Finally, this vector is fed into a fully connected classifier, outputting a preliminary judgment on whether the video's global semantics are abnormal.

[0084] 3. Inter-frame consistency modeling module

[0085] This module works in parallel with the global semantic modeling module, also receiving frame-level semantic feature sequences from the feature extraction module. However, its core task is to perform fine-grained temporal logic analysis to detect semantic jumps and inconsistencies between adjacent or nearby frames. The module first utilizes a bidirectional long short-term memory (BiLSTM) network to learn the forward and backward dependencies of inter-frame semantics. Then, it introduces a frame-level attention mechanism that automatically evaluates and highlights the frames or inter-frame transition regions with the most significant semantic incoherence in the sequence. By aggregating the weighted sequence information, the module forms an inter-frame consistency representation vector, and, similarly, outputs a preliminary judgment result regarding the semantic consistency between video frames through a fully connected classifier.

[0086] 4. Dynamic Fusion Module

[0087] This module serves as the system's decision-making hub, designed to adaptively integrate analytical signals from both global and local dimensions. It receives two preliminary discrimination results (typically classification probability vectors) from the global semantic modeling module and the inter-frame consistency modeling module. Internally, the module contains a lightweight neural network (such as a small multilayer perceptron) that dynamically calculates an optimal set of fusion weights based on the specific circumstances of the two discrimination results for the current video. This set of weights flexibly adjusts the contribution ratio of global semantic information and local consistency information to the final decision. Subsequently, the module uses this set of weights to perform a weighted sum of the two preliminary discrimination results, generating a fused, more reliable final discrimination score.

[0088] 5. Judgment Module

[0089] This module serves as the system's output interface, responsible for making binary classification decisions based on the final discrimination score generated by the dynamic fusion module. Internally, the module has a preset adjustable decision threshold (typically set to 0.5). It compares the final discrimination score with this threshold: if the score is higher than the threshold, the input video is determined to be AI-generated content; if the score is lower than or equal to the threshold, it is determined to be a real video. This module outputs the final judgment result for use by upper-layer applications (such as content moderation platforms).

[0090] System Workflow Overview

[0091] In actual operation, the video to be detected flows through the five modules mentioned above in sequence. After the feature extraction module completes its initialization, the global semantic modeling module and the inter-frame consistency modeling module process the feature sequence in parallel, forming a complementary analytical perspective. Their outputs are intelligently integrated by the dynamic fusion module, and the final conclusion is given by the decision module. This modular design makes the system not only highly accurate and robust, but also clear in structure, easy to deploy and maintain, and able to effectively serve various practical application scenarios.

Claims

1. An AI-generated video detection method that fuses global and inter-frame semantics, characterized in that, The method comprises the following steps: S1, high-level semantic feature extraction of video frame sequence: frame sequence extraction is performed on the video to be detected to obtain a video frame sequence; each video frame in the sequence is input into an image encoder of a pre-trained multi-modal contrastive language-image model on a large amount of text-image pairs, high-level semantic feature vectors of the frame image are extracted, and a frame-level semantic feature sequence is obtained; S2, video global semantic structure modeling: the frame-level semantic feature sequence is input into a global semantic modeling branch; the branch adopts a Transformer sequence model architecture to deeply model the semantic feature sequence of the entire video; the output sequence of the Transformer model is pooled to form a global semantic representation vector of the video, and a classifier is input to obtain a preliminary global semantic discrimination result; S3, video inter-frame semantic temporal consistency modeling: in parallel with step S2, the frame-level semantic feature sequence is input into an inter-frame semantic consistency modeling branch; the branch uses a bidirectional long short-term memory network combined with a frame-level attention mechanism to model the evolution of inter-frame semantic features; through attention mechanism weighted aggregation, an inter-frame semantic consistency evaluation result of the video is formed, and a classifier is input to obtain a preliminary inter-frame consistency discrimination result; S4, dynamic fusion of double-branch discrimination results and final determination: the preliminary global semantic discrimination result from step S2 and the preliminary inter-frame consistency discrimination result from step S3 are input into a dynamic fusion module; The module predicts a set of dynamic fusion weights for the two branch outputs for the current input video sample through a lightweight neural network; the dynamic fusion weights are used to weight and fuse the two discrimination results to obtain a final discrimination score after fusion; S5, comparing the final discrimination score with a preset threshold to output a final determination result of whether the video to be detected is AI-generated content.

2. The method of claim 1, wherein, In step S1, the frame sequence extraction adopts a uniform sampling method.

3. The method of claim 1, wherein, In step S1, the multi-modal contrastive language-image model is a CLIP model, and the image encoder thereof is a visual Transformer architecture.

4. The method of claim 1, wherein, In step S2, the Transformer sequence model is a multi-layer Transformer encoder; when the frame-level semantic feature sequence is input, the position encoding is added first to introduce the time sequence information.

5. The method of claim 1, wherein, In step S2, the pooling is an average pooling operation in the time dimension.

6. The method of claim 1, wherein, In step S3, the bidirectional long short-term memory network is a BiLSTM containing multiple hidden layers; the frame-level attention mechanism assigns attention weights to each time step in the frame sequence by calculating a weight vector.

7. The method of claim 1, wherein, In step S4, the dynamic fusion module performs the following operations: the two preliminary discrimination results are spliced, and the spliced vector is input into the lightweight neural network, which is followed by a Softmax layer to output two dynamic fusion weights.

8. An AI generated video detection system fusing global and inter-frame semantics, for implementing the method of any one of claims 1-7. The method comprises: a feature extraction module configured to extract a frame sequence from an input video and obtain a frame-level high-level semantic feature sequence through a pre-trained image encoder; The global semantic modeling module models a global semantic structure of the feature sequence based on a Transformer architecture, and outputs a global semantic discrimination result; The inter-frame consistency modeling module models inter-frame semantic consistency of the feature sequence based on a BiLSTM and an attention mechanism, and outputs an inter-frame consistency discrimination result; The dynamic fusion module is configured to receive the two discrimination results, predict a fusion weight through a neural network, and perform weighted fusion, and output a final detection probability. The determination module is configured to compare the final probability with a threshold value, and output a detection result.

9. The system of claim 8, wherein, The image encoder used by the feature extraction module is a visual encoder of a CLIP model.

10. The system of claim 8, wherein, The dynamic fusion module includes a multi-layer perception and a Softmax layer, and is configured to generate an input-adaptive fusion weight.

Images