A method, device, system and medium for content review of multi-path video

Abstract

This invention discloses a method, device, and medium for content moderation of multiple video streams. The method includes configuring a first content moderation model for each video stream; when a certain image of a video stream meets a first preset condition, storing the identifier of the video corresponding to the image in a candidate joint detection queue; when there are at least two identifiers in the candidate joint detection queue, stitching together the images of the videos corresponding to each identifier in the candidate joint detection queue according to the similarity between the images of the videos corresponding to the identifiers in the candidate joint detection queue to obtain a target stitched image; retrieving the identifiers of all videos corresponding to the target stitched image from the candidate joint detection queue and storing them in the target joint detection queue; and using a preset second content moderation model to detect the target stitched image. This invention solves the problem of low efficiency in content moderation of multiple video streams in the prior art.

Description

Technical Field

[0001] This invention relates to the field of video review, and more particularly to a method, device, system, and medium for reviewing content from multiple video streams. Background Technology

[0002] Existing multi-channel video auditing systems are used to audit multiple video streams simultaneously for illegal content. These systems include multiple content review models, each designed to review a single video stream. When multiple videos need to be reviewed jointly, staff typically manually configure the synchronization time and stitching settings for the images in each stream beforehand. Once the images are stitched together, the content review models are used for detection. Therefore, the efficiency of content review for multi-channel video in existing technologies is relatively low. Summary of the Invention

[0003] This invention provides a method, device, system, and medium for content review of multiple video streams, in order to solve the problem of low efficiency in content review of multiple video streams in the prior art.

[0004] This invention provides a method for content review of multiple video streams, the method comprising:

[0005] Step 100: Configure a first content review model for each video stream. Each first content review model is used to detect whether the corresponding video violates regulations.

[0006] Step 200: When a certain image of a certain video meets the first preset condition, the identifier of the video corresponding to the image is stored in a preset candidate joint detection queue.

[0007] Step 300: When there are at least two of the identifiers in the candidate joint detection queue, the images of the videos corresponding to each identifier in the candidate joint detection queue are stitched together according to the similarity between the images of the videos corresponding to the identifiers in the candidate joint detection queue to obtain the target stitched image.

[0008] Step 400: Remove the identifiers of all videos corresponding to the target stitched image from the candidate joint detection queue and perform association processing, and store the associated identifiers in the preset target joint detection queue;

[0009] Step 500: Using a preset second content review model, detect the target spliced ​​image corresponding to the associated processed identifier.

[0010] The aforementioned multi-video content moderation method, after performing individual content moderation for each video using a first content moderation model, automatically selects images from different videos with high similarity for stitching together using identifiers in the candidate joint detection queue to obtain a target stitched image. A second content moderation model then performs moderation on this target stitched image. The entire process requires no manual configuration, enabling automatic stitching of multiple images and subsequent content moderation, thus improving the efficiency of multi-video content moderation to a certain extent. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an application environment for a multi-channel video content review method according to an embodiment of the present invention;

[0013] Figure 2 This is a flowchart of a multi-channel video content review method according to an embodiment of the present invention;

[0014] Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] The content moderation method provided in this embodiment of the invention can be applied to, for example... Figure 1 The application environment shown. Specifically, this content moderation method is applied in a moderation system, which includes, for example,... Figure 1 The diagram shows a client and server that communicate over a network to enable joint review of multiple video streams. The client, also known as the user terminal, is the program that provides local services to the client, corresponding to the server. The client can be installed on, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be a standalone server or a server cluster consisting of multiple servers.

[0017] In one embodiment, such as Figure 2 As shown, this embodiment provides a method for content moderation of multi-channel video, which is applied to... Figure 1 Taking the server in the example, the content review method includes:

[0018] Step 100: Configure a first content review model for each video stream. Each first content review model is used to detect whether the corresponding video violates regulations.

[0019] The first content moderation model refers to an image recognition algorithm model specifically designed for content moderation. This model can utilize existing deep learning technologies, such as convolutional neural networks or YOLO-based object detection models. This image recognition algorithm model can transform continuous image information from videos into analyzable feature data, and then use intelligent algorithms to determine whether the content conforms to platform specifications. Compared to traditional manual moderation, it enables uninterrupted detection.

[0020] Step 200: When a certain image of a certain video meets the first preset condition, the identifier of the video corresponding to the image is stored in a preset candidate joint detection queue.

[0021] The first preset condition may include: determining that a video stream has undergone a significant change in its correlation with the preceding and following frames based on the difference between the images at a first time point and the images at an adjacent second time point. Specifically, a first semantic vector from the image at the first time point and a second semantic vector from the image at the second time point can be extracted. The similarity between the first and second semantic vectors can be calculated. If the similarity is not within a preset range, then it is determined that the video stream has undergone a significant change in its correlation with the preceding and following frames.

[0022] The identifiers in this step are used to mark the corresponding videos. For example, if there are n videos in total, identifier a1 represents the first video, identifier a2 represents the second video, and so on, with identifier an representing the nth video. The candidate joint detection queue is used to store the identifiers of videos where the correlation between the preceding and following segments changes significantly. Taking a total of n videos as an example, based on the judgment of the first preset condition, the identifiers of all videos that show significant changes in the correlation between the preceding and following segments at the same time can be stored in the candidate joint detection queue in chronological order for processing in subsequent steps.

[0023] Step 300: When there are at least two of the identifiers in the candidate joint detection queue, the images of the videos corresponding to each identifier in the candidate joint detection queue are stitched together according to the similarity between the images of the videos corresponding to the identifiers in the candidate joint detection queue to obtain the target stitched image.

[0024] Specifically, when the number of identifiers in the candidate joint detection queue is greater than or equal to 2, images of the videos corresponding to each identifier at the same time are acquired, edge images are extracted from each image, and the images are stitched together using the similarity between the edge images to obtain the target stitched image. If the similarity between the edge images of two images is lower than a preset value, these two images are not stitched together.

[0025] For example, the identifiers in the candidate joint detection queue include a1, a2, a3, a4, and a5. After comparing the similarity between the images of the videos corresponding to each identifier, the edge image similarity between the images of the videos corresponding to identifiers a1 and a2 is the highest, while the edge image similarity between the images of the videos corresponding to identifiers a1 and a3-a5 is relatively low, and the edge image similarity between the images of the videos corresponding to identifiers a2 and a3-a5 is also relatively low. Therefore, only the images of the videos corresponding to a1 and a2 need to be stitched together, which is the target stitched image. Similarly, the similarity between the images of the videos corresponding to identifiers a3, a4, and a5 can be compared to determine whether to stitch the images together.

[0026] Step 400: Remove the identifiers of all videos corresponding to the target stitched image from the candidate joint detection queue and perform association processing, and store the associated identifiers in the preset target joint detection queue;

[0027] The target joint detection queue is used to identify videos that can be stitched together. For example, in the candidate joint detection queue [a1, a2, a3, a4, a5], when only the images corresponding to a1 and a2 are stitched together to obtain the first target stitched image, the identifiers a1 and a2 are associated and stored in the target joint detection queue; when the images corresponding to a3 and a5 are stitched together to obtain the second target stitched image, the identifiers a3 and a5 are associated and stored in the target joint detection queue. For example, the target joint detection queue can be [a1-a2, a3-a5].

[0028] This step serves to identify and associate images. Following the stitching order indicated by the associated identifiers, the images are stitched together over a period of time to form multiple target stitched images.

[0029] Step 500: Using a preset second content review model, detect the target spliced ​​image corresponding to the associated processed identifier.

[0030] The second content moderation model is the same as or similar to the first content moderation model. It is also an image recognition algorithm model specifically used for content moderation, but it can adopt an image recognition algorithm model with a different structure than the first content moderation model.

[0031] The multi-channel video content moderation method in this embodiment performs individual content moderation for each video stream using a first content moderation model. Then, it uses identifiers in a candidate joint detection queue for indexing, automatically selecting images from different videos with high similarity for stitching together to obtain a target stitched image. A second content moderation model is then used to moderate the target stitched image. The entire process requires no manual configuration, enabling automatic stitching of multiple images and subsequent content moderation, thus improving the efficiency of multi-channel video content moderation to a certain extent.

[0032] In one embodiment, in step 200, the first preset condition includes: inputting an image from a certain video stream and an adjacent first image into a trained image association prediction model to obtain the similarity between the image and the first image, and the similarity is less than a preset minimum threshold; the image association prediction model is a neural network model used to predict the similarity between images.

[0033] The image correlation prediction model is a pre-trained neural network model, such as one built on a Siamese Network or a Triplet Network. Its input consists of two images, and its output is a similarity value representing the correlation between the content of the two images. During training, a large number of image pairs containing both strong correlations (e.g., consecutive action frames, different angles of the same scene) and weak correlations (e.g., scene transitions, irrelevant content) are collected as samples. The model is trained using a contrastive loss function or a triplet loss function, enabling the neural network model to accurately distinguish the degree of correlation between images. When the current image (a specific image) and its previous frame (the first image) in a video stream are input into this neural network model, if the output similarity value is lower than a preset minimum threshold, it indicates that the video stream has undergone significant content changes at the current moment, potentially revealing a new scene or event requiring joint review. In this case, the video's identifier is stored in the candidate joint detection queue.

[0034] The content review method in this embodiment can reliably determine whether a video has undergone a significant change in its correlation between different parts through an image correlation prediction model, which helps to improve the accuracy of the target spliced ​​image.

[0035] Given that existing multi-video auditing systems cannot determine whether the playback displays corresponding to multiple simultaneously connected video streams are placed together, and even if they know that multiple displays are placed together, they cannot know whether the currently playing multiple images are related (e.g., multiple displays each display a part of the image, which is combined to form a complete image), nor do they know the relative positions of the multiple images, they cannot accurately splice the images or conduct subsequent content auditing.

[0036] Therefore, in one embodiment, step 300 includes:

[0037] Step 301: Extract the first feature vector L, the second feature vector R, the third feature vector T, and the fourth feature vector D from each image of the video corresponding to the identifier. The first feature vector is a vector composed of the pixel values ​​of the i1th row on the left side of the image, where i1 is a positive integer. The second feature vector is a vector composed of the pixel values ​​of the i2th row on the right side of the image, where i2 is a positive integer. The third feature vector is a vector composed of the pixel values ​​of the i3th row on the top side of the image, where i3 is a positive integer. The fourth feature vector is a vector composed of the pixel values ​​of the i4th row on the bottom side of the image, where i4 is a positive integer.

[0038] Where i1, i2, i3, and i4 can be integers greater than or equal to 1, and i1 and i2 have the same value, and i3 and i4 have the same value, in order to ensure the accuracy of the similarity calculation in the subsequent step 303.

[0039] Step 302: Randomly select an image from the video corresponding to the identifier as the initial image, randomly select one from the remaining images corresponding to the identifiers as the image to be stitched, and use the remaining images as candidate images;

[0040] Step 303: Calculate the first similarity between the first feature vector of the initial image and the second feature vector of the image to be stitched, the second similarity between the second feature vector of the initial image and the first feature vector of the image to be stitched, the third similarity between the third feature vector of the initial image and the fourth feature vector of the image to be stitched, and the fourth similarity between the fourth feature vector of the initial image and the third feature vector of the image to be stitched.

[0041] Step 304: When the first similarity, the second similarity, the third similarity, and the fourth similarity are all less than a preset similarity threshold, proceed to step 305; otherwise, proceed to step 306.

[0042] Step 305: Replace one of the candidate images as the image to be stitched, and return to step 303; if there are no other candidate images to replace it at this time, output the current initial image as the target stitching image or candidate stitching image.

[0043] In this process, a candidate image is used as the image to be stitched, and the four similarities are recalculated and compared to see if stitching is possible. Once all candidate images have been sequentially compared for similarity, it is determined that there are no other replaceable candidate images, and the current initial image can be directly used as the target stitching image. Alternatively, the current initial image can be used as a candidate stitching image, which may or may not be selected as the target stitching image.

[0044] In this step, when the current initial image is used as the target stitching image, it is the output result of the entire step 300, and step 300 ends. When the current initial image is used as the candidate stitching image, it is necessary to perform subsequent steps 309 and 310 based on steps 301-308 to select the target stitching image from N candidate stitching images, which is the output result of the entire step 300.

[0045] Step 306: When the first similarity, the second similarity, the third similarity, or the fourth similarity is not less than a preset similarity threshold, determine the splicing position between the initial image and the image to be spliced ​​based on the maximum similarity among the first similarity, the second similarity, the third similarity, and the fourth similarity.

[0046] Specifically, when the maximum similarity is the first similarity, the splicing position is determined to be the left side of the initial image and the right side of the image to be spliced; when the maximum similarity is the second similarity, the splicing position is determined to be the right side of the initial image and the left side of the image to be spliced; when the maximum similarity is the third similarity, the splicing position is determined to be the top edge of the initial image and the bottom edge of the image to be spliced; and when the maximum similarity is the fourth similarity, the splicing position is determined to be the bottom edge of the initial image and the top edge of the image to be spliced.

[0047] Step 307: When the remaining edge of the image to be stitched has a second image with an adjacent edge, if the fifth similarity between the feature vector corresponding to the remaining edge of the image to be stitched and the feature vector corresponding to the adjacent edge of the second image is not greater than the similarity threshold, then proceed to step 305; if the fifth similarity between the feature vector corresponding to the remaining edge of the image to be stitched and the feature vector corresponding to the adjacent edge of the second image is greater than the similarity threshold, then proceed to step 308.

[0048] The second image is one of the images contained in the initial image (which may already be a composite image of at least two images). One edge of the second image is adjacent to a remaining edge of the current image to be stitched. After determining a stitching position in step 306, stitching is not performed immediately. Instead, it is determined whether the adjacent edges of the second image, which has other adjacent relationships with the image to be stitched, can also be stitched with the image to be stitched. If the fifth similarity between the feature vector corresponding to the remaining edge of the image to be stitched and the feature vector corresponding to the adjacent edge of the second image is greater than the similarity threshold, then it is determined that the image to be stitched can be stitched with the initial image, and step 308 is executed. If the fifth similarity between the feature vector corresponding to the remaining edge of the image to be stitched and the feature vector corresponding to the adjacent edge of the second image is not greater than the similarity threshold, then it is determined that the image to be stitched cannot be stitched with the initial image, and the process returns to step 305 to replace the candidate image and continue the next round of stitching evaluation for the new image to be stitched.

[0049] Step 308: Use the image obtained by stitching the initial image and the image to be stitched together as the initial image, and return to step 302.

[0050] The content moderation method in this embodiment extracts feature vectors from image edges and calculates multi-directional similarity to accurately determine potential splicing relationships between different video images. It then determines the specific splicing position (e.g., left-right or top-bottom splicing) based on the maximum similarity. Furthermore, by performing secondary verification on the remaining edges of the spliced ​​image and other candidate images, it effectively avoids mis-sponging caused by high similarity in a single direction. This multi-dimensional splicing logic based on feature vectors automatically adapts to different image combination methods in various scenarios without requiring manual preset of display layouts or video association rules. This significantly improves the integrity and accuracy of the target spliced ​​image, providing a high-quality analysis object for subsequent joint detection by the second content moderation model.

[0051] For example, when there are video images labeled a6, a7, and a8 in the candidate joint detection queue, if the right feature vector of the a6 image has the highest similarity to the left feature vector of the a7 image (the second highest similarity), then the a6 and a7 images are first stitched together horizontally. Subsequently, if the lower feature vector of the stitched image (a6+a7) has a similarity (the fourth similarity) to the upper feature vector of the a8 image greater than the threshold, and the other sides of the a8 image (such as the left and right sides) have a similarity to the adjacent sides of the stitched image (such as the right side of the a7 image) but not exceeding the threshold, then the a8 image can be further stitched below the (a6+a7) image to form a target stitched image containing three video images, thereby achieving effective capture of complex related scenes.

[0052] In one embodiment, when the current initial image is output as a candidate stitching image in step 305, step 300 further includes:

[0053] Step 309: Take the image of the video corresponding to each of the aforementioned identifiers as the initial image, and proceed with steps 301 to 308 to obtain N candidate stitched images, where N is the number of identifiers in the candidate joint detection queue.

[0054] Step 310: Select the image with the most identical images among the N candidate stitching images as the target stitching image.

[0055] The content moderation method in this embodiment uses all images corresponding to the identifiers in the candidate joint detection queue as initial images, performs a complete stitching process to generate N candidate stitched images, and then determines the candidate stitched image with the most identical images as the final target stitched image by counting the number of identical images in each candidate stitched image. This method can effectively avoid the local optimum problem caused by different initial image selections, ensuring that the final target stitched image covers as many related video frames as possible.

[0056] For example, if only a6+a7 (used here to represent the spliced ​​image) is obtained when the image corresponding to the video identified by a6 is used as the initial image, and a7+a6+a8 is obtained when the image corresponding to the video identified by a7 is used as the initial image, and only a8 is obtained when the image corresponding to the video identified by a8 is used as the initial image, then three candidate spliced ​​images will be generated through step 309. Among them, a7+a6+a8 (which is the same image combination as a6+a7+a8) will be selected as the target spliced ​​image, thereby capturing the complex correlation content between multiple videos more comprehensively and further improving the effectiveness of subsequent joint review.

[0057] In one embodiment, step 200 further includes: if the identifier stored in the candidate joint detection queue is not assigned to the target joint detection queue for a first preset time, then the corresponding identifier in the candidate joint detection queue is cleared.

[0058] The content review method in this embodiment effectively avoids the problem of redundant data accumulation caused by prolonged non-participation in stitching by setting a residence time limit for identifiers in the candidate joint detection queue. Specifically, when an identifier is stored in the candidate joint detection queue, if it fails to be successfully stitched with images corresponding to other identifiers and enter the target joint detection queue within a consecutive first preset time (e.g., 5s, 10s), the system will automatically remove the identifier from the candidate joint detection queue. This mechanism ensures that the identifiers in the candidate joint detection queue always reflect video images with potential stitching value, reduces interference from invalid data to subsequent stitching algorithms, and lowers the system's storage and computing resource consumption. Simultaneously, video identifiers that are mistakenly identified as entering the candidate joint detection queue due to brief content changes can also be automatically removed through time threshold filtering, further improving the dynamic adaptability and resource utilization efficiency of the entire review process. For example, if a video stream is misjudged as having undergone significant content changes due to momentary image flicker and enters the candidate joint detection queue, but no other video images that can be stitched together appear within the subsequent first preset time, the identifier will be automatically cleared to prevent it from occupying queue resources for a long time and affecting the stitching process of other valid identifiers.

[0059] In one embodiment, step 300, before stitching together the images of the videos corresponding to each identifier in the candidate joint detection queue, further includes: detecting whether the size of each image is a preset size and whether the resolution of each image is consistent; when the size of each image is not a preset size and / or the resolution of each image is inconsistent, preprocessing each image to make the size of each image the preset size and the resolution of each image consistent.

[0060] The content moderation method in this embodiment first performs a consistency check on the basic attributes of the images before stitching together the images of the videos corresponding to each identifier in the candidate joint detection queue. Specifically, it automatically detects whether the size of each image to be stitched conforms to a preset standard size and whether the resolution of all images is uniform. If it detects that the size of an image does not conform to the preset size, or that there is a situation where the resolution is inconsistent with other images, an image preprocessing mechanism will be triggered.

[0061] For example, images can be scaled or stretched using interpolation algorithms to adjust them to a preset size and uniform resolution. For cases where significant resolution differences may lead to image distortion, edge padding or cropping can be combined to normalize the size while ensuring the integrity of key content. This preprocessing step effectively avoids feature vector extraction deviations caused by differences in image size or resolution, ensuring the comparability of pixel value vectors at the edges of different images, thereby improving the accuracy of subsequent similarity calculations and laying the foundation for reliable stitching position determination. If the size and resolution consistency requirements still cannot be met after preprocessing, such as severe distortion after image stretching, the image will be temporarily marked as abnormal, excluded from the current stitching process, and relevant information will be recorded for subsequent manual verification or algorithm optimization.

[0062] In one embodiment, step 300, before stitching together the images of the videos corresponding to each identifier in the candidate joint detection queue, further includes: detecting whether the direction of each image is a preset direction; if the direction of an image is not a preset direction, then rotating the corresponding image.

[0063] The content moderation method in this embodiment requires a standardization check of the image orientation before stitching together the images of the videos corresponding to each identifier in the candidate joint detection queue. Specifically, it can detect whether the display orientation of each image to be stitched conforms to a preset orientation (e.g., horizontal, vertical, or a specific angle). For example, if the preset orientation is horizontal (width greater than height), and an image is vertical (height greater than width) or rotated by an abnormal angle such as 90 degrees or 180 degrees due to shooting angle or transmission error, an orientation correction mechanism will be triggered. For example, key feature points in the image (such as text direction, person's posture, etc.) can be extracted through image recognition algorithms or orientation markers in image metadata can be read to automatically determine the correct orientation of the image and adjust it to the preset orientation through rotation or flipping operations. If the image has no obvious orientation features and the metadata is missing, it can be combined with the user's historical settings or the conventional orientation of similar video images for auxiliary judgment, or it can be temporarily marked as having an unknown orientation and images with a clear orientation can be processed first.

[0064] This step avoids misalignment of edge feature vectors due to incorrect image orientation (such as misidentifying the original top edge as the left edge), ensuring that the first to fourth feature vectors accurately correspond to the left, right, top, and bottom edges of the image. This guarantees the effectiveness of subsequent similarity calculations and stitching position determination. For example, if an image to be stitched is actually horizontal but has been rotated to vertical due to transmission issues, its original left edge (first feature vector) might be mistakenly extracted as the top edge (third feature vector), causing a deviation in similarity calculations with other images. After orientation correction, the extraction position of the edge feature vectors is restored to the correct position, effectively improving stitching accuracy.

[0065] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0066] In one embodiment, such as Figure 3 As shown, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the content moderation method described in the above embodiments, for example... Figure 1 Steps 100 to 500 shown are omitted here to avoid repetition.

[0067] In one embodiment, a content moderation system is provided, the content moderation system including: a monitoring device and a computer device as described in the previous embodiment, the computer device communicating with the monitoring device to obtain data sent by the monitoring device, i.e., video frames.

[0068] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program implements the control method for the AGV picking vehicle described in the above embodiment, for example... Figure 1 Steps 100 to 500 shown are omitted here to avoid repetition.

[0069] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the methods described above. Furthermore, any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0070] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0071] While specific embodiments of the invention have been described in detail by way of example, those skilled in the art should understand that the examples are for illustrative purposes only and not intended to limit the scope of the invention. It should also be understood that various modifications can be made to the embodiments without departing from the scope and spirit of the invention. The scope of this invention is defined by the appended claims.

Claims

1. A method for content review of multiple video streams, characterized in that, The content review methods include: Step 100: Configure a first content review model for each video stream. Each first content review model is used to detect whether the corresponding video violates regulations. Step 200: When a certain image of a certain video meets the first preset condition, the identifier of the video corresponding to the image is stored in a preset candidate joint detection queue. The first preset condition includes: determining that a video has undergone a significant change in its correlation with the preceding and following frames based on the difference between the images at a first time point and the images at an adjacent second time point; calculating the similarity between the first semantic vector extracted from the image at the first time point and the second semantic vector extracted from the image at the second time point; if the similarity is not within a preset range, determining that the video has undergone a significant change in its correlation with the preceding and following frames; and, based on the determination of the first preset condition, storing the identifiers of all videos that have undergone significant changes in their correlation with the preceding and following frames at the same time point in the candidate joint detection queue in chronological order. Step 300: When there are at least two of the identifiers in the candidate joint detection queue, the images of the videos corresponding to each identifier in the candidate joint detection queue are stitched together according to the similarity between the images of the videos corresponding to the identifiers in the candidate joint detection queue to obtain the target stitched image. When the number of identifiers in the candidate joint detection queue is greater than or equal to 2, the images of the videos corresponding to each identifier at the same time are obtained, the edge images in each image are extracted, and the images are stitched together using the similarity between the edge images to obtain the target stitched image; when the similarity between the edge images of two images is lower than a preset value, the two images are not stitched together. Step 400: Remove the identifiers of all videos corresponding to the target stitched image from the candidate joint detection queue and perform association processing, and store the associated identifiers in the preset target joint detection queue; Step 500: Using a preset second content review model, detect the target spliced ​​image corresponding to the associated processed identifier.

2. The content review method for multiple video streams according to claim 1, characterized in that, In step 200, the first preset condition includes: An image from a video stream and its adjacent first image are input into a trained image association prediction model to obtain the similarity between the image and the first image, and the similarity is less than a preset minimum threshold; the image association prediction model is a neural network model used to predict the similarity between images.

3. The content review method for multi-channel video according to claim 1, characterized in that, Step 300 includes: Step 301: Extract the first feature vector L, the second feature vector R, the third feature vector T, and the fourth feature vector D from each image of the video corresponding to the identifier. The first feature vector is a vector composed of the pixel values ​​of the i1th row on the left side of the image, where i1 is a positive integer. The second feature vector is a vector composed of the pixel values ​​of the i2th row on the right side of the image, where i2 is a positive integer. The third feature vector is a vector composed of the pixel values ​​of the i3th row on the top side of the image, where i3 is a positive integer. The fourth feature vector is a vector composed of the pixel values ​​of the i4th row on the bottom side of the image, where i4 is a positive integer. Step 302: Randomly select an image from the video corresponding to the identifier as the initial image, randomly select one from the remaining images corresponding to the identifiers as the image to be stitched, and use the remaining images as candidate images; Step 303: Calculate the first similarity between the first feature vector of the initial image and the second feature vector of the image to be stitched, the second similarity between the second feature vector of the initial image and the first feature vector of the image to be stitched, the third similarity between the third feature vector of the initial image and the fourth feature vector of the image to be stitched, and the fourth similarity between the fourth feature vector of the initial image and the third feature vector of the image to be stitched. Step 304: When the first similarity, the second similarity, the third similarity, and the fourth similarity are all less than a preset similarity threshold, proceed to step 305; otherwise, proceed to step 306. Step 305: Replace one of the candidate images as the image to be stitched, and return to step 303; if there are no other candidate images to replace it at this time, output the current initial image as the target stitching image or candidate stitching image. Step 306: When the first similarity, the second similarity, the third similarity, or the fourth similarity is not less than a preset similarity threshold, determine the splicing position between the initial image and the image to be spliced ​​based on the maximum similarity among the first similarity, the second similarity, the third similarity, and the fourth similarity. Step 307: When the remaining edge of the image to be stitched has a second image with an adjacent edge, if the fifth similarity between the feature vector corresponding to the remaining edge of the image to be stitched and the feature vector corresponding to the adjacent edge of the second image is not greater than the similarity threshold, then proceed to step 305; if the fifth similarity between the feature vector corresponding to the remaining edge of the image to be stitched and the feature vector corresponding to the adjacent edge of the second image is greater than the similarity threshold, then proceed to step 308. Step 308: Use the image obtained by stitching the initial image and the image to be stitched together as the initial image, and return to step 302.

4. The content review method for multi-channel video according to claim 3, characterized in that, When step 305 outputs the current initial image as a candidate stitching image, step 300 further includes: Step 309: Take the image of the video corresponding to each of the aforementioned identifiers as the initial image, and proceed with steps 301 to 308 to obtain N candidate stitched images, where N is the number of identifiers in the candidate joint detection queue. Step 310: Select the image with the most identical images among the N candidate stitching images as the target stitching image.

5. The content review method for multiple video streams according to claim 1, characterized in that, Step 200 further includes: if the identifier stored in the candidate joint detection queue is not assigned to the target joint detection queue for a first preset time, then the corresponding identifier in the candidate joint detection queue is cleared.

6. The content review method for multiple video streams according to claim 1, characterized in that, In step 300, before stitching together the images of the videos corresponding to each identifier in the candidate joint detection queue, the method further includes: detecting whether the size of each image is a preset size and whether the resolution of each image is consistent; When the size of each image is not the preset size and / or the resolution of each image is inconsistent, each image is preprocessed to make the size of each image the preset size and the resolution of each image consistent.

7. The content review method for multi-channel video according to claim 6, characterized in that, In step 300, before stitching together the images of the videos corresponding to each identifier in the candidate joint detection queue, the method further includes: detecting whether the direction of each image is a preset direction; if the direction of an image is not a preset direction, then rotating the corresponding image.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the content review method for multi-channel video according to any one of claims 1 to 7.

9. A content moderation system, characterized in that, The content review system includes: a monitoring device and a computer device as described in claim 8, wherein the computer device communicates with the monitoring device to obtain data sent by the monitoring device.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the content review method for multi-channel video as described in any one of claims 1 to 7.

Images