Feature recognition and potential space time division multiplexing AIGC real-time copyright protection method, system and medium

Abstract

The application relates to an AIGC real-time copyright protection method and system based on feature recognition and potential space time division multiplexing and a medium, the method comprising the following steps: stopping sampling in a preset time step of denoising sampling of a diffusion model, extracting a noisy potential encoding code stream and decoding the noisy potential encoding code stream into a pixel space image, performing face detection on the pixel space image through MTCNN after denoising processing, and comparing and verifying face region facial embedding feature data set with a pre-trained copyright feature database; if the verification is a protected identity, an OSD mask is attached to the corresponding face region; the image after visual masking is re-encoded into a potential encoding code stream and noise is rematched to generate a modified potential encoding code stream; the modified potential encoding code stream is injected into the sampling process to restore the denoising sampling process, and a protected video frame is generated; the above steps are repeatedly performed on each frame of a video sequence to generate a complete protected video. Compared with the prior art, the application improves the real-time performance, accuracy and universality of AIGC copyright protection.

Description

Technical Field

[0001] This invention relates to the field of AI-generated content, and in particular to a method, system, and medium for real-time copyright protection of AIGC using feature recognition and potential spatial time-division multiplexing. Background Technology

[0002] With the rapid development of AI-generated technologies (such as diffusion models like Stable Diffusion), text-to-video (T2V) technology can now synthesize high-quality, highly realistic dynamic video content. The widespread adoption of this technology has greatly enriched content creation methods, but it also brings the risk of abuse, such as the unauthorized generation and dissemination of deepfake videos that infringe on the portrait rights or privacy rights of specific individuals (such as celebrities and public figures). Such abuse not only poses a serious threat to individual rights but may also trigger a crisis of social trust and legal disputes.

[0003] Currently, mainstream solutions for the security and copyright protection of Artificial Intelligence Generated Content (AIGC) mostly adopt a "post-event detection and filtering" model. These solutions typically identify infringing content using deep learning-based detection models after the content is generated and uploaded to the cloud. However, this method has significant limitations: first, it has high response latency, making real-time intervention at the source of content generation impossible; second, large-scale video content detection requires enormous cloud computing resources; and finally, the inability to intervene at the source means that infringing content may have already been generated and spread on a small scale, causing actual damage.

[0004] To intervene in the generation process, some technical solutions attempt to modify or control the generation model itself. However, these methods often require retraining or fine-tuning the underlying diffusion model, which is not only costly and complex, but may also damage the model's original generation capabilities and quality, making large-scale deployment in practical applications difficult. For example, patent application CN120746808A discloses a method and apparatus for copyright protection of a latent diffusion model based on frequency domain watermarking. This method verifies whether an image was generated by a specific model by embedding and extracting frequency domain watermarks into the generated image. It can only verify copyright ownership after infringing content is generated, but cannot identify and block specific illegal content such as portrait rights infringement in real time during the content generation process. Moreover, its implementation still requires fine-tuning of the underlying diffusion model, which poses risks of high cost and potential impact on generation capabilities.

[0005] Therefore, there is an urgent need for a technical solution that can be seamlessly integrated into the existing diffusion model generation process, without retraining, and can identify and process potential infringing content in real time and accurately during the generation process. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by providing a real-time AIGC copyright protection method, system, and medium that features recognition and potential spatial time-division multiplexing, thereby improving the real-time performance, accuracy, and versatility of AIGC copyright protection.

[0007] The objective of this invention can be achieved through the following technical solutions: A real-time copyright protection method for AIGC that combines feature recognition and temporal multiplexing of latent space, applied to the content generation process of a diffusion model, includes the following steps: During the denoising sampling process of the diffusion model, sampling is interrupted at a preset time step, and the noisy latent coded bitstream at the current time step is extracted. The noisy latent coded bitstream is decoded into a pixel space image using the original decoder of the diffusion model; The pixel space image is denoised to obtain a denoised pixel space image. Face detection is performed on the denoised pixel space image using a pre-trained MTCNN, and the facial embedding feature dataset of the detected face region is compared and verified with a pre-built copyright feature database. If the identity is verified as protected, a predefined OSD mask will be attached to the corresponding facial area to achieve visual masking protection. The pixel space image with a predefined OSD mask attached is re-encoded into a latent coding stream, and noise re-matching is performed on the re-encoded latent coding stream to generate a modified latent coding stream that matches the noise statistical characteristics of the preset time step. The modified potential encoded bitstream is re-injected into the sampling process of the diffusion model, and the denoising sampling process is restored from the preset time step to generate a protected video frame. Repeat the above steps for each frame of the video sequence to generate a complete protected video.

[0008] Furthermore, the noisy latent coded bitstream is a latent feature that has not yet been fully converted into pixel representation, and its extraction location is located at the input end or intermediate bottleneck layer of the U-Net of the diffusion model.

[0009] Furthermore, the denoising process is related to the performance of face detection and includes strong denoising and lightweight denoising. The strong denoising process uses the denoising module of the diffusion model for denoising, while the lightweight denoising process uses an algorithm including bilateral filtering, nonlocal mean denoising, or a lightweight denoising network based on deep learning.

[0010] Furthermore, the pre-trained MTCNN adopts a three-level cascaded architecture, including a shallow proposal network, a refinement network, and an output network. The specific steps for face detection on the denoised pixel space image using the pre-trained MTCNN include: The denoised pixel space image is input into the shallow proposal network, and a candidate face window containing face probability and preliminary bounding box is generated through fast scanning and multi-scale pyramid processing. The candidate face windows are input into the refining network to optimize the bounding box coordinates and effectively eliminate a large number of falsely detected candidate face windows. The candidate face window, after being filtered by the refined network, is input into the output network for bounding box regression and facial feature point localization. The output includes one or more face bounding boxes and the coordinates of key facial feature points, including the left corner of the eye, the right corner of the eye, the tip of the nose, the left corner of the mouth, and the right corner of the mouth.

[0011] Furthermore, the pre-built copyright feature database includes facial feature vectors of multiple protected targets. The specific steps for comparing and verifying the facial embedding feature dataset of the detected face region with the pre-built copyright feature database include: Feature extraction is performed on the detected face regions, and a face embedding vector is generated through a deep convolutional network. Calculate the cosine similarity between the high-dimensional facial embedding vector and the facial feature vector of each protected target stored in the copyright feature database; When the cosine similarity is greater than the preset identity verification threshold, the identity is determined to be protected, triggering the subsequent visual occlusion protection process.

[0012] Furthermore, the predefined OSD mask forms include digital watermarking, blurring masking, and replacement masking. The digital watermarking type embeds copyright information, user ID, or tracking code into the facial area in an invisible or semi-transparent form. The blurring masking type performs Gaussian blur or pixelation on the facial area. The replacement masking type replaces the original facial features with a neutral facial texture or a cartoonish avatar.

[0013] Furthermore, when attaching the predefined OSD mask to the corresponding face region, for multiple face regions, the mask attachment is performed sequentially, and the coordinates of the occluded area are recorded for subsequent potential spatial alignment. The predefined OSD mask only covers the detected face bounding box and its outer region.

[0014] Further, the specific steps of re-encoding the pixel space image with the predefined OSD mask attached into a latent coding bitstream, and performing noise re-matching on the re-encoded latent coding bitstream to generate a modified latent coding bitstream that matches the noise statistical characteristics of the preset time step include: The pixel space image with the OSD mask attached is input into the variational autoencoder associated with the diffusion model, compressed and converted back to the latent space representation to obtain the re-encoded latent coded bitstream. The difference between the original latent coded bitstream and the recoded latent coded bitstream is calculated. According to the noise scheduling scheme of the diffusion model, noise is added back to the recoded latent coded bitstream, and it is ensured that the latent variables after adding noise are consistent with the noise statistical characteristics of the preset time step, so as to obtain the modified latent coded bitstream.

[0015] According to another aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, enables the AIGC real-time copyright protection method of feature recognition and latent spatial time-division multiplexing as described above.

[0016] According to another aspect of the present invention, an AIGC real-time copyright protection system with feature recognition and latent spatial time-division multiplexing is provided, comprising: The noisy latent coded bitstream extraction module is used to interrupt sampling at a preset time step during the denoising sampling process of the diffusion model and extract the noisy latent coded bitstream at the current time step. The decoding module is used to decode the noisy latent coded bitstream into a pixel space image using the original decoder of the diffusion model; A denoising module is used to denoise the pixel space image to obtain a denoised pixel space image. The face comparison and verification module is used to perform face detection on the denoised pixel space image using the pre-trained MTCNN, and to compare and verify the facial embedding feature dataset of the detected face region with the pre-built copyright feature database. The visual masking protection module is used to attach a predefined OSD mask to the corresponding face area when the detected facial features are verified as a protected identity, thereby achieving visual masking protection. The recoding and noise rematching module is used to recode the pixel space image with a predefined OSD mask attached into a latent coding bitstream, and to perform noise rematching on the recoded latent coding bitstream to generate a modified latent coding bitstream that matches the noise statistical characteristics of the preset time step. The denoising sampling recovery module is used to re-inject the modified latent coded bitstream into the sampling process of the diffusion model, recover the denoising sampling process from the preset time step, and generate the final protected video frame. The protected video generation module is used to repeat the above steps for each frame of the video sequence to generate a complete protected video.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention interrupts sampling at a preset time step during the denoising sampling process of the diffusion model, extracts the noisy latent coded bitstream and decodes it into a pixel space image. A pre-trained MTCNN is then used to perform face detection on the denoised pixel space image. The detected face features are compared and verified against a pre-built copyright feature database. Face features verified as protected are visually masked. This allows for real-time intervention during the generation process without retraining the underlying model, solving the problem of high costs and compromised generation quality caused by retraining or fine-tuning existing intervention schemes. This improves the accuracy and versatility of AIGC real-time copyright protection.

[0018] 2. This invention achieves synchronous intervention at the source of content generation by decoding the potential encoded bitstream in real time at the intermediate time step of denoising sampling, performing face recognition and processing, and synchronizing the results through a noise rematching step in the backsampling process. This solves the problems of high response delay and the potential spread of infringing content in traditional post-event detection schemes, thus improving the timeliness of copyright protection and the effectiveness of source blocking.

[0019] 3. This invention ensures that the modified latent encoded bitstream and the noise statistical characteristics of the sampling process are consistent by performing masking operations and noise rematching in the latent space, so that the OSD mask can be naturally integrated by the subsequent denoising process. This solves the problem that external intervention can easily lead to visual artifacts and inconsistent styles in the generated content, and improves the visual naturalness and overall quality of the protected generated results. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing proposed in this invention. Figure 2 This is a schematic diagram of the structure of an AIGC real-time copyright protection system that combines feature recognition and latent spatial time-division multiplexing, as proposed in this invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0022] The following English abbreviations are involved: Multi-Task Convolutional Neural Network (MTCNN) Screen display: On-Screen Display, OSD Example 1 This embodiment provides a real-time copyright protection method for AIGC that combines feature recognition and latent spatial time-division multiplexing, applied to the content generation process of a diffusion model, such as... Figure 1 As shown, it includes the following steps: S1. During the denoising sampling process of the diffusion model, sampling is interrupted at a preset time step, and the noisy potential coded bitstream of the current time step is extracted.

[0023] This embodiment uses Stable Diffusion as an example. When the Stable Diffusion U-Net denoising network reaches time step t, the noisy latent coded bitstream of the current time step is extracted from the latent space of the U-Net through a hook mechanism or intermediate layer interception technique. The noisy latent coded bitstream consists of latent features that have not yet been fully converted into pixel representations, and its extraction location is at the input end of the U-Net of the diffusion model or at the intermediate bottleneck layer.

[0024] The dimensions of the noisy latent coded bitstream are typically 64×64×4 or 32×32×4, depending on the version of StableDiffusion, and include compressed image semantic information and residual noise. The extraction operation must maintain the integrity of the computation graph to avoid disrupting the backpropagation path and ensure that subsequent sampling processes can be recovered normally. The extracted noisy latent coded bitstream is backed up and stored for sampling recovery in step S7.

[0025] S2. The existing decoder of the diffusion model is used to decode the noisy latent coded bitstream into a pixel space image.

[0026] The noisy latent coded bitstream obtained in step S1 is processed using the existing decoder (VAE Decoder) of Stable Diffusion. Convert to pixel space image The decoder employs a variational autoencoder (VAE) structure, which upsamples the latent spatial features to the original image resolution (such as 512×512 or 1024×1024) through deconvolutional layers.

[0027] The decoding process follows the standard SD inference flow:

[0028] Because the noisy potential encoded bitstream is in the middle time step Decoded pixel space image The image presents a partially denoised state, preserving the original semantic structure while containing visible noise and texture; the decoded pixel space image. It uses RGB three-channel format for storage, with a bit depth of 8 bits or 16 bits.

[0029] S3. Perform denoising processing on the pixel space image to obtain the denoised pixel space image.

[0030] The denoising process is related to the performance of face detection, including strong denoising and lightweight denoising. Strong denoising is performed by calling the denoising module of the diffusion model, while lightweight denoising uses algorithms such as bilateral filtering, nonlocal mean denoising, or a lightweight denoising network based on deep learning.

[0031] Noise reduction intensity controlled at Within the specified range, avoid excessive smoothing that could lead to loss of facial features.

[0032] This step is an optional optimization step, when the time step... Later in the day (when noise levels are low), this step can be skipped to reduce computational overhead.

[0033] The pixel space image after denoising is denoted as Its edge information and texture details are enhanced, which is beneficial for MTCNN to detect small or low-contrast faces.

[0034] S4. Perform face detection on the denoised pixel space image using the pre-trained MTCNN, and compare and verify the facial embedding feature dataset of the detected face regions with the pre-built copyright feature database.

[0035] The pre-trained MTCNN employs a three-stage cascaded architecture, including a shallow proposal network, a refinement network, and an output network. The specific steps for face detection on the denoised pixel space image using the pre-trained MTCNN include: The denoised pixel space image is input into a shallow proposal network, and candidate face windows containing face probabilities and preliminary bounding boxes are generated through fast scanning and multi-scale pyramid processing. The candidate face windows are input into the refining network to optimize the bounding box coordinates and effectively eliminate a large number of false positive candidate face windows. The candidate face windows, refined through network filtering, are input into the output network for bounding box regression and facial feature point localization. The output includes one or more face bounding boxes and the coordinates of key facial feature points, including the left and right corners of the eyes, the tip of the nose, and the left and right corners of the mouth. These coordinates are used to construct a convex hull mask, accurately outlining the facial contours. This precise masking minimizes the impact of subsequent OSD mask application on non-facial areas of the image, thus ensuring privacy while maintaining the overall visual integrity of the generated content. The detection confidence threshold is set to... Only high-confidence detection results are retained; the set of face region coordinates is output. ,in , indicating the first The bounding box parameters of an individual's face.

[0036] The pre-built copyright feature database includes facial feature vectors of multiple protected targets. The specific steps for comparing and verifying the facial embedding feature dataset of detected face regions with the pre-built copyright feature database include: For the detected face region Feature extraction is performed, and facial embedding vectors are generated using a deep convolutional network.

[0037] Calculate the cosine similarity between the high-dimensional facial embedding vector and the facial feature vector of each protected target stored in the copyright feature database. .

[0038] When the cosine similarity is greater than the preset authentication threshold, the authentication threshold is set. ,when When the identity is determined to be protected, the subsequent visual occlusion protection process is triggered.

[0039] The copyright feature database supports batch comparison of multiple identities and uses vector retrieval engines such as FAISS or Milvus to accelerate large-scale feature matching, with a single query latency controlled within 10ms.

[0040] S5. If the identity is verified as protected, a predefined OSD mask will be attached to the corresponding face area to achieve visual masking protection.

[0041] The predefined OSD mask types include digital watermarking, blur masking, and replacement masking. Digital watermarking embeds copyright information, user ID, or tracking code into the facial area in an invisible or semi-transparent form. Blur masking applies Gaussian blur or pixelation to the facial area. Replacement masking replaces the original facial features with a neutral facial texture or a cartoon avatar.

[0042] When attaching a predefined OSD mask to the corresponding face region, for multiple face regions, the mask attachment is performed sequentially, and the coordinates of the occluded area are recorded for subsequent potential spatial alignment. The mask attachment follows the principle of minimum intervention, covering only the detected face bounding box and its outward expansion region (outward expansion coefficient). (Make sure to cover the entire face).

[0043] The attachment process uses Alpha hybridization technology:

[0044] in, For pixel-space images with predefined OSD masks attached, Image of a face mask. The mixing coefficient, This is the pixel space image after noise reduction.

[0045] S6. Re-encode the pixel space image with the predefined OSD mask attached into a latent coding stream, and perform noise re-matching on the re-encoded latent coding stream to generate a modified latent coding stream that matches the noise statistical characteristics of the preset time step.

[0046] The specific steps of re-encoding a pixel-space image with a predefined OSD mask attached into a latent coding bitstream, and then performing noise re-matching on the re-encoded latent coding bitstream to generate a modified latent coding bitstream that matches the noise statistical characteristics of a preset time step include: The pixel-space image with the OSD mask attached is input into a variational autoencoder associated with the diffusion model, compressed, and converted back to the latent space representation to obtain the re-encoded latent code stream. , .

[0047] Calculate the original latent coded bitstream With the re-encoded potential bitstream To address the differences, noise is re-added to the re-coded potential bitstream according to the noise scheduling scheme of the diffusion model (such as DDPM, DDIM, or Euler scheduling). It is ensured that the latent variables after adding noise are consistent with the noise statistical characteristics at the preset time step, resulting in the modified potential bitstream. , ,in For time steps The cumulative noise figure, This is resampling noise.

[0048] Noise consistency constraint: Ensures the consistency of the modified potential coded bitstream after adding noise. Consistent with the statistical characteristics of the original sampling path, this avoids introducing additional randomness that could cause flickering between video frames.

[0049] S7. Re-inject the modified potential encoded bitstream into the sampling process of the diffusion model, restore the denoising sampling process from the preset time step, and generate protected video frames.

[0050] The modified latent variables obtained in step S6 Re-inject into the U-Net sampling pipeline, from time step Restore denoised sampling, to As a new starting point, continue executing the remaining steps. The denoising step involves calling U-Net to predict noise and update latent variables: ,in For U-Net noise prediction network, For diffusion sampling schedulers.

[0051] Since the occlusion operation is completed before the latent space is re-injected, in order to maintain semantic consistency, the subsequent denoising process treats the OSD mask as inherent content of the image, ensuring that the mask blends naturally with the lighting and texture style of the background.

[0052] S8. Repeat the above steps for each frame of the video sequence to generate a complete protected video.

[0053] Repeat steps S1-S7 for each frame of the video sequence to generate a complete protected video.

[0054] By employing an inter-frame latent coding bitstream caching mechanism, duplicate detection is skipped for static background regions, and local updates are performed only for dynamic face regions, reducing computational overhead by more than 30%.

[0055] Example 2 This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it can realize the AIGC real-time copyright protection method with feature recognition and latent spatial time-division multiplexing as provided in Embodiment 1.

[0056] The rest is the same as in Example 1.

[0057] Example 3 This embodiment provides an AIGC real-time copyright protection system that combines feature recognition and latent spatial time-division multiplexing, such as... Figure 2 As shown, it includes: The noisy latent coded bitstream extraction module is used to interrupt sampling at a preset time step during the denoising sampling process of the diffusion model and extract the noisy latent coded bitstream at the current time step. The decoding module is used to decode the noisy latent coded bitstream into a pixel space image using the original decoder of the diffusion model; The denoising module is used to denoise the pixel space image to obtain the denoised pixel space image. The face comparison and verification module is used to perform face detection on the denoised pixel space image using the pre-trained MTCNN, and to compare and verify the facial embedding feature dataset of the detected face region with the pre-built copyright feature database. The visual masking protection module is used to attach a predefined OSD mask to the corresponding face area when the detected facial features are verified as a protected identity, thereby achieving visual masking protection. The recoding and noise rematching module is used to recode the pixel space image with a predefined OSD mask attached into a latent coding bitstream, and to perform noise rematching on the recoded latent coding bitstream to generate a modified latent coding bitstream that matches the noise statistical characteristics of a preset time step. The denoising sampling recovery module is used to re-inject the modified potential coded bitstream into the sampling process of the diffusion model, recover the denoising sampling process from the preset time step, and generate the final protected video frame. The protected video generation module is used to repeat the above steps for each frame of the video sequence to generate a complete protected video.

[0058] The rest is the same as in Example 1.

[0059] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A real-time copyright protection method for AIGC that combines feature recognition and latent spatial time-division multiplexing, applied to the content generation process of a diffusion model, characterized in that... Includes the following steps: During the denoising sampling process of the diffusion model, sampling is interrupted at a preset time step, and the noisy latent coded bitstream at the current time step is extracted. The noisy latent coded bitstream is decoded into a pixel space image using the original decoder of the diffusion model; The pixel space image is denoised to obtain a denoised pixel space image. Face detection is performed on the denoised pixel space image using a pre-trained MTCNN, and the facial embedding feature dataset of the detected face region is compared and verified with a pre-built copyright feature database. If the identity is verified as protected, a predefined OSD mask will be attached to the corresponding facial area to achieve visual masking protection. The pixel space image with a predefined OSD mask attached is re-encoded into a latent coding stream, and noise re-matching is performed on the re-encoded latent coding stream to generate a modified latent coding stream that matches the noise statistical characteristics of the preset time step. The modified potential encoded bitstream is re-injected into the sampling process of the diffusion model, and the denoising sampling process is restored from the preset time step to generate a protected video frame. Repeat the above steps for each frame of the video sequence to generate a complete protected video.

2. The AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing according to claim 1, characterized in that, The noisy latent coded bitstream is a latent feature that has not yet been fully converted into pixel representation, and its extraction location is at the input end or intermediate bottleneck layer of the U-Net of the diffusion model.

3. The AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing according to claim 1, characterized in that, The denoising process is related to the performance of face detection and includes strong denoising and lightweight denoising. The strong denoising process uses the denoising module of the diffusion model for denoising. The lightweight denoising process uses an algorithm including bilateral filtering, nonlocal mean denoising, or a lightweight denoising network based on deep learning.

4. The AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing according to claim 1, characterized in that, The pre-trained MTCNN adopts a three-level cascaded architecture, including a shallow proposal network, a refinement network, and an output network. The specific steps for face detection on the denoised pixel space image using the pre-trained MTCNN include: The denoised pixel space image is input into the shallow proposal network, and a candidate face window containing face probability and preliminary bounding box is generated through fast scanning and multi-scale pyramid processing. The candidate face windows are input into the refining network to optimize the bounding box coordinates and effectively eliminate a large number of falsely detected candidate face windows. The candidate face window, after being filtered by the refined network, is input into the output network for bounding box regression and facial feature point localization. The output includes one or more face bounding boxes and the coordinates of key facial feature points, including the left corner of the eye, the right corner of the eye, the tip of the nose, the left corner of the mouth, and the right corner of the mouth.

5. The AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing according to claim 4, characterized in that, The pre-built copyright feature database includes facial feature vectors of multiple protected targets. The specific steps for comparing and verifying the facial embedding feature dataset of the detected face region with the pre-built copyright feature database include: Feature extraction is performed on the detected face regions, and a face embedding vector is generated through a deep convolutional network. Calculate the cosine similarity between the high-dimensional facial embedding vector and the facial feature vector of each protected target stored in the copyright feature database; When the cosine similarity is greater than the preset identity verification threshold, the identity is determined to be protected, triggering the subsequent visual occlusion protection process.

6. The AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing according to claim 1, characterized in that, The predefined OSD mask types include digital watermarking, blurring masking, and replacement masking. The digital watermarking type embeds copyright information, user ID, or tracking code into the facial area in an invisible or semi-transparent form. The blurring masking type performs Gaussian blur or pixelation on the facial area. The replacement masking type replaces the original facial features with a neutral facial texture or a cartoonish avatar.

7. The AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing according to claim 1, characterized in that, When attaching a predefined OSD mask to the corresponding face region, for multiple face regions, the mask is attached sequentially, and the coordinates of the occluded area are recorded for subsequent potential spatial alignment. The predefined OSD mask only covers the detected face bounding box and its outer region.

8. The AIGC real-time copyright protection method based on feature recognition and latent spatial time-division multiplexing according to claim 1, characterized in that, The specific steps of re-encoding a pixel-space image with a predefined OSD mask attached into a latent coding bitstream, and performing noise re-matching on the re-encoded latent coding bitstream to generate a modified latent coding bitstream that matches the noise statistical characteristics of the preset time step include: The pixel space image with the OSD mask attached is input into the variational autoencoder associated with the diffusion model, compressed and converted back to the latent space representation to obtain the re-encoded latent coded bitstream. The difference between the original latent coded bitstream and the recoded latent coded bitstream is calculated. According to the noise scheduling scheme of the diffusion model, noise is added back to the recoded latent coded bitstream, and it is ensured that the latent variables after adding noise are consistent with the noise statistical characteristics of the preset time step, so as to obtain the modified latent coded bitstream.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, enables the AIGC real-time copyright protection method of feature recognition and latent spatial time-division multiplexing as described in any one of claims 1 to 8.

10. A real-time copyright protection system for AIGC that combines feature recognition and latent spatial time-division multiplexing, characterized in that, include: The noisy latent coded bitstream extraction module is used to interrupt sampling at a preset time step during the denoising sampling process of the diffusion model and extract the noisy latent coded bitstream at the current time step. The decoding module is used to decode the noisy latent coded bitstream into a pixel space image using the original decoder of the diffusion model; A denoising module is used to denoise the pixel space image to obtain a denoised pixel space image. The face comparison and verification module is used to perform face detection on the denoised pixel space image using the pre-trained MTCNN, and to compare and verify the facial embedding feature dataset of the detected face region with the pre-built copyright feature database. The visual masking protection module is used to attach a predefined OSD mask to the corresponding face area when the detected facial features are verified as a protected identity, thereby achieving visual masking protection. The recoding and noise rematching module is used to recode the pixel space image with a predefined OSD mask attached into a latent coding bitstream, and to perform noise rematching on the recoded latent coding bitstream to generate a modified latent coding bitstream that matches the noise statistical characteristics of the preset time step. The denoising sampling recovery module is used to re-inject the modified latent coded bitstream into the sampling process of the diffusion model, recover the denoising sampling process from the preset time step, and generate the final protected video frame. The protected video generation module is used to repeat the above steps for each frame of the video sequence to generate a complete protected video.

Images