A system and method for identifying and processing sensitive individuals in live streaming scenarios

Abstract

This application relates to a system and method for identifying and processing sensitive individuals in live streaming scenarios. The system includes: a task allocation module for setting sensitivity thresholds for target video streams to be detected based on the live streaming scenario, sorting all target video streams according to their sensitivity values, and then sorting the load of each cloud computing node in the detection module. Based on the sorting result, target video streams with high sensitivity values ​​are assigned to cloud computing nodes with lower loads for detection. A detection module is used to perform human detection on the target video streams and to perform sensitive content detection and sensitive area processing for different target subjects. This method can rationally allocate computing resources and reduce anomaly detection costs.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a system, method, computer equipment, and storage medium for identifying and processing sensitive individuals in a live streaming scenario. Background Technology

[0002] With the widespread use of smart terminal devices and the rapid development of mobile technology, live video streaming has risen rapidly. However, during this development, intentional or unintentional sensitive live streaming incidents have severely hampered the development of the live video streaming environment and had a negative impact on society. Currently, the supervision of live streaming platforms is mainly in a semi-manual state, relying on platform warnings and manual review. However, this work model, which relies on manual rule-based warnings and manual review, cannot adapt to the massive amount of live streaming content, and the separation of anomaly identification and handling technologies will seriously affect the efficiency of supervision. Therefore, reducing the human cost in live streaming review and supervision is crucial.

[0003] However, current traditional sensitive content recognition technologies primarily perform overall discrimination on video frames (images), extracting global features for anomaly detection. On the one hand, global features may overlook the fine-grained representation of local target subjects, leading to low detection accuracy. On the other hand, the anomaly handling model, which relies entirely on manual review, is too simplistic, consuming significant manpower and failing to meet the rapidly growing personalized needs of live streaming content. Therefore, a more integrated regulatory solution that customizes different anomaly detection and handling methods for different target subjects is more in line with the needs of real-world live streaming scenarios. Furthermore, considering the application of sensitive subject recognition and handling technologies in live streaming scenarios, processing every scenario meticulously would inevitably result in a significant waste of computational resources. Summary of the Invention

[0004] Therefore, it is necessary to provide a system, method, computer equipment, and storage medium for identifying and processing sensitive individuals in live streaming scenarios that can reasonably allocate computing resources and reduce the cost of anomaly detection, in order to address the above-mentioned technical problems.

[0005] A system for identifying and processing sensitive individuals in a live streaming scenario, the system comprising a task allocation module and a detection module; the detection module consists of several cloud computing nodes; the cloud computing nodes include an instant queue and a local queue; the task allocation module includes a global queue.

[0006] The task allocation module is used to set the sensitivity threshold of the target video stream to be detected according to the live streaming scenario, sort all the target video streams to be detected according to the sensitivity value, sort the load of each cloud computing node of the detection module, and deliver the target video stream with high sensitivity value to the cloud computing node with low load for detection according to the sorting result.

[0007] The detection module is used to perform human detection on the target video stream and to perform sensitive content detection and sensitive area processing for different target subjects.

[0008] In one embodiment, the task allocation module is further configured to obtain the length of the local queue of each cloud computing node and deliver the target video stream to be detected to the computing node with the shortest queue; if the length of the local queue of all cloud computing nodes is greater than a preset value, the target video stream to be detected is placed in the global queue; all cloud computing nodes prioritize processing the target video stream to be detected in the instant queue, and when there is no target video stream to be detected in the instant queue, they go to the local queue to obtain the target video stream to be detected.

[0009] The detection module is also used to mark the target video stream if it exceeds a sensitive threshold. When the cloud computing node obtains a marked target video stream for the second time, it writes the marked target video stream into the instant queue for priority detection and kicks the target video stream previously in the instant queue back to the local queue. If the local queue is full, it will be divided into two parts, and the second half will be written back to the global queue. When the local queue of the cloud computing node is empty, the node will first try to steal the target video stream from other nodes. If stealing the target video stream fails, the cloud computing node will go to the global queue to obtain the target video stream, obtaining half of the preset value of the local queue at a time.

[0010] In one embodiment, the detection module includes a target detection model and an anomaly detection model; the detection module is used to perform human detection on the target video stream to be detected, and to perform sensitive content detection and sensitive area processing for different target subjects, including:

[0011] The target detection model is used to detect the target in the video stream and obtain multiple output feature maps of the target.

[0012] Anomaly detection is performed on multiple output feature maps based on the anomaly detection model to obtain the sensitivity value of the target human figure.

[0013] If the sensitivity value is lower than the preset sensitivity threshold, the main measures will be to issue a warning and blur the image. The live scene corresponding to the target video stream will be marked in yellow and written to the log for easy analysis and viewing by administrators. Mosaic technology will be used to blur the location area of ​​the corresponding subject.

[0014] If the sensitivity value exceeds the preset sensitivity threshold, for the first occurrence, the main measures will be to issue a warning and blur the image, and mark the live scene corresponding to the target video stream in red. For live rooms that exceed the sensitivity threshold twice in a row, the live room will be banned, and the results will be fed back to the administrator for processing in real time.

[0015] In one embodiment, the process of constructing the object detection model includes:

[0016] Acquire different live videos and preprocess them into different video frame images. Label the target location of the collected video frame images to construct a target detection sample set. Divide the target detection sample set into a target detection training set, a target detection test set, and a target detection validation set according to a pre-set ratio.

[0017] Data augmentation is performed on the target detection training set, target detection test set, and target detection validation set using the Mosaic algorithm to obtain the augmented target detection training set, target detection test set, and target detection validation set.

[0018] The pre-built deep learning model for human object detection based on the improved YOLOv5 was trained using the enhanced object detection training set and object detection validation set to obtain the trained deep learning model for human object detection.

[0019] The trained deep learning model for human object detection is tested for correct recognition rate based on the enhanced object detection test set. If the correct recognition rate of the trained deep learning model for human object detection is lower than the recognition rate threshold required in the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the object detection model.

[0020] In one embodiment, the process of constructing a deep learning model for person object detection based on the improved YOLOv5 includes:

[0021] Obtain the YOLOv5 object detection model;

[0022] The three input channels of the YOLOv5 object detection model are compressed to obtain a single channel;

[0023] The single-channel input is fed into the backbone network CSPDarknet of the YOLOv5 object detection model to extract the shallow detailed features of the target person and obtain the feature map.

[0024] The single-channel input is fed into the upsampling module of the PANet feature fusion network of the YOLOv5 object detection model to obtain the upsampled feature map;

[0025] The final loss is calculated based on the feature map and the upsampled feature map. The gradient of the YOLOv5 object detection model is updated using the obtained final loss, resulting in a deep learning model for human object detection based on the improved YOLOv5.

[0026] In one embodiment, the process of constructing an anomaly detection model includes:

[0027] Acquire different live videos and preprocess them into different video frame images. Label the collected video frame images with target location labels and sensitive anomaly labels to construct an anomaly detection sample set. Then, divide the anomaly detection sample set into an anomaly detection training set, an anomaly detection test set, and an anomaly detection validation set according to a pre-set ratio.

[0028] Geometric and color transformation operations are performed on the anomaly detection training set, anomaly detection test set, and anomaly detection validation set to obtain the enhanced anomaly detection training set, anomaly detection test set, and anomaly detection validation set.

[0029] The pre-built sensitive content anomaly detection model based on ResNet and Focalloss is trained using the enhanced anomaly detection training set and anomaly detection validation set to obtain the trained sensitive content anomaly detection model.

[0030] The trained sensitive content anomaly detection model is tested for correct recognition rate based on the enhanced anomaly detection test set. If the correct recognition rate of the trained human target detection deep learning model is lower than the recognition rate threshold required in the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the anomaly detection model.

[0031] In one embodiment, constructing a sensitive content anomaly detection model based on ResNet and Focalloss includes:

[0032] The target person image is input into the ResNet backbone network model to extract the shallow detailed features of the target person and obtain the feature map of the target person.

[0033] The feature map is input into the output layer, and the loss is calculated using the FocalLoss loss function. The ResNet backbone network model is then updated based on the loss to obtain a sensitive content anomaly detection model based on ResNet and Focalloss.

[0034] A method for identifying and processing sensitive individuals in a live streaming scenario, the method comprising:

[0035] Set the sensitivity threshold of the target video stream to be detected according to the live streaming scenario, sort all the target video streams to be detected according to the sensitivity value, sort the load of each cloud computing node, and deliver the target video stream with high sensitivity value to the cloud computing node with low load for detection according to the sorting result.

[0036] Human detection is performed on the target video stream in the cloud computing node, and sensitive content detection and sensitive area processing are performed for different target subjects.

[0037] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0038] Set the sensitivity threshold of the target video stream to be detected according to the live streaming scenario, sort all the target video streams to be detected according to the sensitivity value, sort the load of each cloud computing node, and deliver the target video stream with high sensitivity value to the cloud computing node with low load for detection according to the sorting result.

[0039] Human detection is performed on the target video stream in the cloud computing node, and sensitive content detection and sensitive area processing are performed for different target subjects.

[0040] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0041] Set the sensitivity threshold of the target video stream to be detected according to the live streaming scenario, sort all the target video streams to be detected according to the sensitivity value, sort the load of each cloud computing node, and deliver the target video stream with high sensitivity value to the cloud computing node with low load for detection according to the sorting result.

[0042] Human detection is performed on the target video stream in the cloud computing node, and sensitive content detection and sensitive area processing are performed for different target subjects.

[0043] The aforementioned system, method, computer equipment, and storage medium for identifying and processing sensitive individuals in a live streaming scenario first establishes a task allocation module that sets sensitivity thresholds for the target video streams to be detected based on the live streaming scenario. All target video streams are then sorted according to their sensitivity values. Next, the load of each cloud computing node in the detection module is sorted, and high-sensitivity target video streams are assigned to lower-load cloud computing nodes for detection. Real-time and local queues are set up in the cloud computing nodes, and a global queue is set up in the task allocation module. The detection module performs human detection on the target video streams and performs sensitive content detection and sensitive area processing for different target subjects. This application is tailored to the specific scenario. By setting sensitivity thresholds for the target video streams to be detected and employing multi-level queues for allocation and processing, the system efficiently allocates cloud computing resources, prioritizing high-sensitivity tasks and balancing the load across the entire cloud computing system. This effectively improves detection efficiency and reduces computational costs. When detecting anomalies, the system first detects the target subject and then performs sensitive content detection and sensitive area processing for different target subjects. This effectively extracts more refined semantic features of people and improves anomaly detection accuracy. The system achieves live sensitive content detection and excellent cloud computing resource scheduling and allocation, improving the regulatory efficiency of the live streaming platform, effectively reducing operational manpower costs, and thus greatly meeting the actual needs of the rapid growth of live streaming content. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the structure of a target subject sensitive person identification and processing system in a live streaming scenario, as shown in one embodiment.

[0045] Figure 2 This is a schematic diagram of a multi-level queue structure for processing tasks on a single cloud node in one embodiment.

[0046] Figure 3 This is a flowchart illustrating a method for identifying and processing sensitive individuals in a live streaming scenario, as described in another embodiment.

[0047] Figure 4 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0049] In one embodiment, such as Figure 1As shown, a system for identifying and processing sensitive individuals in a live streaming scenario is provided. The system includes a task allocation module and a detection module. The detection module consists of several cloud computing nodes. The cloud computing nodes include an instant queue and a local queue. The task allocation module includes a global queue.

[0050] The task allocation module is used to set the sensitivity threshold of the target video stream to be detected according to the live streaming scenario, sort all the target video streams to be detected according to the sensitivity value, sort the load of each cloud computing node of the detection module, and deliver the target video stream with high sensitivity value to the cloud computing node with low load for detection according to the sorting result.

[0051] The detection module is used to perform human detection on the target video stream and to perform sensitive content detection and sensitive area processing for different target subjects.

[0052] In the aforementioned system for identifying and processing sensitive individuals in a live streaming scenario, a task allocation module first sets sensitivity thresholds for the target video streams to be detected based on the live streaming scenario. All target video streams are then sorted according to their sensitivity values. Next, the load of each cloud computing node in the detection module is sorted, and high-sensitivity target video streams are assigned to lower-loaded cloud computing nodes for detection. Each cloud computing node processes the target video streams in parallel, improving processing efficiency. Real-time and local queues are set up in the cloud computing nodes, and a global queue is set up in the task allocation module. The detection module performs human detection on the target video streams and performs sensitive content detection and sensitive area processing for different target subjects. This application sets the sensitivity threshold of the target video stream to be detected according to the actual scenario, and adopts a multi-level queue for the allocation and processing of the target video stream to be detected in a way that is appropriate to local conditions. It efficiently allocates the cloud computing resources of the entire system, allowing high-sensitivity tasks to be detected first, and also balancing the load of the entire cloud computing system, effectively improving detection efficiency and reducing computing costs. When performing target anomaly detection, the target subject is detected first, and then sensitive content detection and sensitive area processing are performed for different target subjects. This can effectively extract more refined semantic features of people, effectively improve the accuracy of anomaly detection, realize the detection of sensitive content in live broadcasts, and achieve excellent cloud computing resource scheduling and allocation. This improves the supervision efficiency of the live broadcast platform, effectively reduces the operating manpower cost, and thus greatly meets the actual needs of the rapid growth of live broadcast content.

[0053] In one embodiment, the task allocation module is further configured to obtain the length of the local queue of each cloud computing node and deliver the target video stream to be detected to the computing node with the shortest queue; if the length of the local queue of all cloud computing nodes is greater than a preset value, the target video stream to be detected is placed in the global queue; all cloud computing nodes prioritize processing the target video stream to be detected in the instant queue, and when there is no target video stream to be detected in the instant queue, they go to the local queue to obtain the target video stream to be detected.

[0054] The detection module is also used to mark the target video stream if it exceeds a sensitive threshold. When the cloud computing node obtains a marked target video stream for the second time, it writes the marked target video stream into the instant queue for priority detection and kicks the target video stream previously in the instant queue back to the local queue. If the local queue is full, it will be divided into two parts, and the second half will be written back to the global queue. When the local queue of the cloud computing node is empty, the node will first try to steal the target video stream from other nodes. If stealing the target video stream fails, the cloud computing node will go to the global queue to obtain the target video stream, obtaining half of the preset value of the local queue at a time.

[0055] In a specific embodiment, such as Figure 2 As shown, the instant queue points to the next detection target, the local queue contains detection tasks delivered by the task allocation module, and the global queue is used to place new tasks when a local queue is full. The instant and local queues reside within the cloud computing nodes; there are N instant and local queues for every N cloud computing nodes. The global queue, however, is located within the task allocation module, and there is only one global queue across the entire system.

[0056] The "steal" function effectively reduces the frequency with which cloud computing nodes retrieve tasks from the global queue. Currently, there is only one global queue for all nodes. Every time a cloud computing node retrieves a task from the global queue, it causes "lock waiting" for other computing nodes that are also trying to retrieve tasks, consuming significant computer resources.

[0057] In one embodiment, the detection module includes a target detection model and an anomaly detection model; the detection module is used to perform human detection on the target video stream to be detected, and to perform sensitive content detection and sensitive area processing for different target subjects, including:

[0058] The target detection model is used to detect the target in the video stream and obtain multiple output feature maps of the target.

[0059] Anomaly detection is performed on multiple output feature maps based on the anomaly detection model to obtain the sensitivity value of the target human figure.

[0060] If the sensitivity value is lower than the preset sensitivity threshold, the main measures will be to issue a warning and blur the image. The live scene corresponding to the target video stream will be marked in yellow and written to the log for easy analysis and viewing by administrators. Mosaic technology will be used to blur the location area of ​​the corresponding subject.

[0061] If the sensitivity value exceeds the preset sensitivity threshold, for the first occurrence, the main measures will be to issue a warning and blur the image, and mark the live scene corresponding to the target video stream in red. For live rooms that exceed the sensitivity threshold twice in a row, the live room will be banned, and the results will be fed back to the administrator for processing in real time.

[0062] In a specific embodiment, video images are captured based on a video stream sampling time of approximately 1 second for detecting sensitive content in online live streaming. For a specific live video stream, three consecutive video frames of the main subject image are sampled and input into the anomaly detection model. The predicted sensitivity confidence is input, and the average value of three or fewer identical subjects is used as the sensitivity value of the target subject of the live stream within that time interval.

[0063] The mosaic technique is used to blur the location area of ​​the corresponding subject. Specifically, the target area of ​​the person to be blurred is divided into multiple small blocks, and then the average value of each small block is taken to achieve the blurring of the sensitive area.

[0064] In one embodiment, the process of constructing the object detection model includes:

[0065] Acquire different live videos and preprocess them into different video frame images. Label the target location of the collected video frame images to construct a target detection sample set. Divide the target detection sample set into a target detection training set, a target detection test set, and a target detection validation set according to a pre-set ratio.

[0066] Data augmentation is performed on the target detection training set, target detection test set, and target detection validation set using the Mosaic algorithm to obtain the augmented target detection training set, target detection test set, and target detection validation set.

[0067] The pre-built deep learning model for human object detection based on the improved YOLOv5 was trained using the enhanced object detection training set and object detection validation set to obtain the trained deep learning model for human object detection.

[0068] The trained deep learning model for human object detection is tested for correct recognition rate based on the enhanced object detection test set. If the correct recognition rate of the trained deep learning model for human object detection is lower than the recognition rate threshold required in the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the object detection model.

[0069] In a specific embodiment, a training set for human target detection is obtained by collecting different live videos and preprocessing them into different video frame images. The labelMe tool is used to label each image target, marking the location region of the target human and anomaly category labels. Based on the location labels, the training set required for the target detection model can be constructed, and the training set, test set, and validation set are randomly divided in a 6:2:2 ratio. The established target detection training set, validation set, and test set are preprocessed, and the Mosaic algorithm is used to perform data augmentation on the human target detection images to increase the model's generalization ability.

[0070] First, set the parameters required for the deep learning model for human object detection based on the improved YOLOv5, including a series of model hyperparameters such as learning rate, batch size, optimizer type, number of iterations, and activation function. Then, according to the input image size set by the network, scale the target human image training set and the target detection validation set images for model training.

[0071] In one embodiment, the process of constructing a deep learning model for person object detection based on the improved YOLOv5 includes:

[0072] Obtain the YOLOv5 object detection model;

[0073] The three input channels of the YOLOv5 object detection model are compressed to obtain a single channel;

[0074] The single-channel input is fed into the backbone network CSPDarknet of the YOLOv5 object detection model to extract the shallow detailed features of the target person and obtain the feature map.

[0075] The single-channel input is fed into the upsampling module of the PANet feature fusion network of the YOLOv5 object detection model to obtain the upsampled feature map;

[0076] The final loss is calculated based on the feature map and the upsampled feature map. The gradient of the YOLOv5 object detection model is updated using the obtained final loss, resulting in a deep learning model for human object detection based on the improved YOLOv5.

[0077] In one embodiment, the process of constructing an anomaly detection model includes:

[0078] Acquire different live videos and preprocess them into different video frame images. Label the collected video frame images with target location labels and sensitive anomaly labels to construct an anomaly detection sample set. Then, divide the anomaly detection sample set into an anomaly detection training set, an anomaly detection test set, and an anomaly detection validation set according to a pre-set ratio.

[0079] Geometric and color transformation operations are performed on the anomaly detection training set, anomaly detection test set, and anomaly detection validation set to obtain the enhanced anomaly detection training set, anomaly detection test set, and anomaly detection validation set.

[0080] The pre-built sensitive content anomaly detection model based on ResNet and Focalloss is trained using the enhanced anomaly detection training set and anomaly detection validation set to obtain the trained sensitive content anomaly detection model.

[0081] The trained sensitive content anomaly detection model is tested for correct recognition rate based on the enhanced anomaly detection test set. If the correct recognition rate of the trained human target detection deep learning model is lower than the recognition rate threshold required in the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the anomaly detection model.

[0082] In a specific embodiment, an anomaly detection training set is obtained: training data for different subject target images are obtained using the labeled location tags and anomaly category tags of the target person, constructing the training set required for sensitive content anomaly detection, and dividing the training set, test set, and validation set in a 6:2:2 ratio. Since the vast majority of live video content consists of normal images (positive samples), there are few negative samples of sensitive images, and manual collection is time-consuming and costly, which cannot fully meet the needs of model training. Therefore, it is necessary to preprocess the negative samples in the established anomaly detection training set, validation set, and test set, using the following data augmentation methods:

[0083] Geometric transformations: These mainly involve performing geometric transformations on images, including image flipping, rotation, cropping, deformation, scaling, etc.

[0084] Color transformation: Image processing through blurring, color transformation, erasing, filling, etc.

[0085] First, set the parameters required for the sensitive content anomaly detection model based on ResNet and Focalloss, including a series of model hyperparameters such as learning rate, batch size, optimizer type, number of iterations, and activation function. Then, according to the input image size set by the network, scale the images of the target person image anomaly detection training set and anomaly detection validation set for model training.

[0086] In one embodiment, constructing a sensitive content anomaly detection model based on ResNet and Focalloss includes:

[0087] The target person image is input into the ResNet backbone network model to extract the shallow detailed features of the target person and obtain the feature map of the target person.

[0088] The feature map is input into the output layer, and the loss is calculated using the FocalLoss loss function. The ResNet backbone network model is then updated based on the loss to obtain a sensitive content anomaly detection model based on ResNet and Focalloss.

[0089] In one embodiment, such as Figure 3 As shown, a method for identifying and processing sensitive individuals in a live streaming scenario is provided, including:

[0090] Step 302: Set the sensitivity threshold of the target video stream to be detected according to the live streaming scenario, sort all the target video streams to be detected according to the sensitivity value, sort the load of each cloud computing node, and deliver the target video stream with high sensitivity value to the cloud computing node with low load for detection according to the sorting result.

[0091] Step 304: Perform human detection on the target video stream to be detected in the cloud computing node, and perform sensitive content detection and sensitive area processing for different target subjects.

[0092] In one embodiment, the length of the local queue of each cloud computing node is obtained, and the target video stream to be detected is delivered to the computing node with the shortest queue; if the length of the local queue of all cloud computing nodes is greater than a preset value, the target video stream to be detected is placed in the global queue; all cloud computing nodes prioritize processing the target video stream to be detected in the real-time queue, and when there is no target video stream to be detected in the real-time queue, they go to the local queue to obtain the target video stream to be detected.

[0093] When a target video stream is detected and its sensitivity threshold is exceeded, it is marked. When a cloud computing node acquires a marked target video stream for the second time, it immediately writes the marked target video stream into the instant queue for priority detection and kicks the target video stream previously in the instant queue back to the local queue. If the local queue is full, it will be divided into two parts, and the latter half will be written back to the global queue. When the local queue of a cloud computing node is empty, the node will first try to steal the target video stream from other nodes. If stealing the target video stream fails, the cloud computing node will go to the global queue to obtain the target video stream, obtaining half of the preset value of the local queue at a time.

[0094] In one embodiment, human detection is performed on the target video stream to be detected, and sensitive content detection and sensitive region processing are performed for different target subjects, including:

[0095] The target detection model is used to detect the target in the video stream and obtain multiple output feature maps of the target.

[0096] Anomaly detection is performed on multiple output feature maps based on the anomaly detection model to obtain the sensitivity value of the target human figure.

[0097] If the sensitivity value is lower than the preset sensitivity threshold, the main measures will be to issue a warning and blur the image. The live scene corresponding to the target video stream will be marked in yellow and written to the log for easy analysis and viewing by administrators. Mosaic technology will be used to blur the location area of ​​the corresponding subject.

[0098] If the sensitivity value exceeds the preset sensitivity threshold, for the first occurrence, the main measures will be to issue a warning and blur the image, and mark the live scene corresponding to the target video stream in red. For live rooms that exceed the sensitivity threshold twice in a row, the live room will be banned, and the results will be fed back to the administrator for processing in real time.

[0099] In one embodiment, the process of constructing the object detection model includes:

[0100] Acquire different live videos and preprocess them into different video frame images. Label the target location of the collected video frame images to construct a target detection sample set. Divide the target detection sample set into a target detection training set, a target detection test set, and a target detection validation set according to a pre-set ratio.

[0101] Data augmentation is performed on the target detection training set, target detection test set, and target detection validation set using the Mosaic algorithm to obtain the augmented target detection training set, target detection test set, and target detection validation set.

[0102] The pre-built deep learning model for human object detection based on the improved YOLOv5 was trained using the enhanced object detection training set and object detection validation set to obtain the trained deep learning model for human object detection.

[0103] The trained deep learning model for human object detection is tested for correct recognition rate based on the enhanced object detection test set. If the correct recognition rate of the trained deep learning model for human object detection is lower than the recognition rate threshold required in the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the object detection model.

[0104] In one embodiment, the process of constructing a deep learning model for person object detection based on the improved YOLOv5 includes:

[0105] Obtain the YOLOv5 object detection model;

[0106] The three input channels of the YOLOv5 object detection model are compressed to obtain a single channel;

[0107] The single-channel input is fed into the backbone network CSPDarknet of the YOLOv5 object detection model to extract the shallow detailed features of the target person and obtain the feature map.

[0108] The single-channel input is fed into the upsampling module of the PANet feature fusion network of the YOLOv5 object detection model to obtain the upsampled feature map;

[0109] The final loss is calculated based on the feature map and the upsampled feature map. The gradient of the YOLOv5 object detection model is updated using the obtained final loss, resulting in a deep learning model for human object detection based on the improved YOLOv5.

[0110] In one embodiment, the process of constructing an anomaly detection model includes:

[0111] Acquire different live videos and preprocess them into different video frame images. Label the collected video frame images with target location labels and sensitive anomaly labels to construct an anomaly detection sample set. Then, divide the anomaly detection sample set into an anomaly detection training set, an anomaly detection test set, and an anomaly detection validation set according to a pre-set ratio.

[0112] Geometric and color transformation operations are performed on the anomaly detection training set, anomaly detection test set, and anomaly detection validation set to obtain the enhanced anomaly detection training set, anomaly detection test set, and anomaly detection validation set.

[0113] The pre-built sensitive content anomaly detection model based on ResNet and Focalloss is trained using the enhanced anomaly detection training set and anomaly detection validation set to obtain the trained sensitive content anomaly detection model.

[0114] The trained sensitive content anomaly detection model is tested for correct recognition rate based on the enhanced anomaly detection test set. If the correct recognition rate of the trained human target detection deep learning model is lower than the recognition rate threshold required in the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the anomaly detection model.

[0115] In one embodiment, constructing a sensitive content anomaly detection model based on ResNet and Focalloss includes:

[0116] The target person image is input into the ResNet backbone network model to extract the shallow detailed features of the target person and obtain the feature map of the target person.

[0117] The feature map is input into the output layer, and the loss is calculated using the FocalLoss loss function. The ResNet backbone network model is then updated based on the loss to obtain a sensitive content anomaly detection model based on ResNet and Focalloss.

[0118] In the aforementioned method for identifying and processing sensitive individuals in a live streaming scenario, the following steps are taken: First, a sensitivity threshold for the target video stream to be detected is set according to the live streaming scenario. All target video streams are then sorted based on their sensitivity values. Next, the load of each cloud computing node is sorted, and based on the sorting results, the target video streams with high sensitivity values ​​are delivered to cloud computing nodes with lower loads for detection. Each cloud computing node processes the target video stream in parallel, which improves processing efficiency. Real-time queues and local queues are set in the cloud computing nodes, and a global queue is set in the task allocation module. Human detection is performed on the target video stream, and sensitive content detection and sensitive area processing are performed for different target subjects. This application is based on the actual scenario... The system sets sensitive thresholds for the target video streams to be detected and uses multi-level queues to allocate and process them, efficiently allocating cloud computing resources across the entire system. This prioritizes high-sensitivity tasks and balances the load on the entire cloud computing system, effectively improving detection efficiency and reducing computing costs. When detecting anomalies, the system first detects the target subject and then performs sensitive content detection and sensitive area processing for different target subjects. This effectively extracts more refined semantic features of people and improves anomaly detection accuracy. The system achieves live sensitive content detection and excellent cloud computing resource scheduling and allocation, improving the regulatory efficiency of the live streaming platform, effectively reducing operational manpower costs, and thus greatly meeting the actual needs of the rapid growth of live streaming content.

[0119] It should be understood that, although Figure 3 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 3 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0120] For specific limitations regarding the method for identifying and processing sensitive individuals in a live streaming scenario, please refer to the limitations of the system for identifying and processing sensitive individuals in a live streaming scenario described above, which will not be repeated here. Each module in the aforementioned system for identifying and processing sensitive individuals in a live streaming scenario can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0121] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a system for identifying and processing sensitive individuals in a live streaming scenario. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0122] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0123] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0124] In one embodiment, a computer storage medium is provided that stores a computer program, which, when executed by a processor, implements the steps of the method described above.

[0125] 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. The 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 above methods. 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. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0126] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A system for identifying and processing sensitive individuals in a live streaming scenario, characterized in that, The system includes a task allocation module and a detection module; the detection module consists of several cloud computing nodes; the cloud computing nodes include instant queues and local queues; the task allocation module includes a global queue. The task allocation module is used to set the sensitivity threshold of the target video stream to be detected according to the live streaming scenario, sort all the target video streams to be detected according to the sensitivity value, sort the load of each cloud computing node of the detection module, and deliver the target video stream with high sensitivity value to the cloud computing node with low load for detection according to the sorting result. The detection module is used to perform human detection on the target video stream to be detected, and to perform sensitive content detection and sensitive area processing for different target subjects; The task allocation module is also used to obtain the length of the local queue of each cloud computing node and deliver the target video stream to be detected to the computing node with the shortest queue; if the length of the local queue of all cloud computing nodes is greater than the preset value, the target video stream to be detected is placed in the global queue. All cloud computing nodes prioritize processing the target video stream to be detected in the real-time queue. When there is no target video stream to be detected in the real-time queue, they go to the local queue to obtain the target video stream to be detected. The detection module is also used to mark the target video stream if it detects a value exceeding a sensitive threshold. When the cloud computing node obtains a marked target video stream for the second time, it directly writes the marked target video stream into the instant queue for priority detection and kicks the target video stream previously in the instant queue back to the local queue. If the local queue is full, it will be divided into two parts, and the latter half will be written back to the global queue. When the local queue of the cloud computing node is empty, the node will first try to steal the target video stream from other nodes. If stealing the target video stream fails, the cloud computing node will go to the global queue to obtain the target video stream, obtaining half of the preset value of the local queue at a time.

2. The system according to claim 1, characterized in that, The detection module includes a target detection model and an anomaly detection model; the detection module is used to perform human detection on the target video stream to be detected, and to perform sensitive content detection and sensitive area processing for different target subjects, including: The target detection model is used to perform target detection on the video stream of the target to be detected, and multiple output feature maps of the target subject are obtained; Anomaly detection is performed on the multiple output feature maps based on the anomaly detection model to obtain the sensitivity value of the target human figure. If the sensitivity value is lower than the preset sensitivity threshold, the main measures will be to issue a warning and blur the image. The live scene corresponding to the target video stream to be detected will be marked in yellow and written to the log for easy analysis and viewing by management personnel. Mosaic technology will be used to blur the location area of ​​the corresponding subject. If the sensitivity value exceeds the preset sensitivity threshold, for the first occurrence, the main measures will be to issue a warning and blur the image, marking the live scene corresponding to the target video stream in red. For live rooms that exceed the sensitivity threshold twice consecutively, the live room will be banned, and the results will be fed back to the administrator for processing in real time.

3. The system according to claim 2, characterized in that, The process of building an object detection model includes: Acquire different live videos and preprocess them into different video frame images. Label the target location of the collected video frame images to construct a target detection sample set. Divide the target detection sample set into a target detection training set, a target detection test set, and a target detection validation set according to a pre-set ratio. Data augmentation is performed on the target detection training set, target detection test set, and target detection validation set using the Mosaic algorithm to obtain the augmented target detection training set, target detection test set, and target detection validation set. The enhanced target detection training set and target detection validation set are used to train the pre-built deep learning model for human target detection based on the improved YOLOv5, resulting in a trained deep learning model for human target detection. The trained deep learning model for human object detection is tested for correct recognition rate based on the enhanced target detection test set. If the correct recognition rate of the trained deep learning model for human object detection is lower than the recognition rate threshold requirement of the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the target detection model.

4. The system according to claim 3, characterized in that, The process of building a deep learning model for human object detection based on the improved YOLOv5 includes: Obtain the YOLOv5 object detection model; The three input channels of the YOLOv5 target detection model are compressed to obtain a single channel; The single channel is input into the backbone network CSPDarknet of the YOLOv5 object detection model to extract the shallow detailed features of the target person and obtain the feature map. The single-channel input is fed into the upsampling module of the PANet feature fusion network of the YOLOv5 object detection model to obtain the upsampled feature map; The final loss is calculated based on the feature map and the upsampled feature map. The gradient of the YOLOv5 object detection model is updated using the obtained final loss to obtain a deep learning model for human object detection based on improved YOLOv5.

5. The system according to claim 3, characterized in that, The process of building an anomaly detection model includes: Different live videos are acquired and preprocessed into different video frame images. Target location labels and sensitive anomaly labels are labeled on the collected video frame images to construct an anomaly detection sample set. The anomaly detection sample set is then divided into an anomaly detection training set, an anomaly detection test set, and an anomaly detection validation set according to a pre-set ratio. Geometric transformation and color transformation operations are performed on the anomaly detection training set, anomaly detection test set, and anomaly detection verification set to obtain enhanced anomaly detection training set, anomaly detection test set, and anomaly detection verification set; The enhanced anomaly detection training set and anomaly detection validation set are used to train the pre-built sensitive content anomaly detection model based on ResNet and Focalloss to obtain the trained sensitive content anomaly detection model. The trained sensitive content anomaly detection model is tested for correct recognition rate based on the enhanced anomaly detection test set. If the correct recognition rate of the trained human target detection deep learning model is lower than the recognition rate threshold requirement of the actual scene, the model is further optimized and trained, and the overall parameters of the model are tuned until the required recognition rate is achieved, thus obtaining the anomaly detection model.

6. The system according to claim 5, characterized in that, The construction of a sensitive content anomaly detection model based on ResNet and Focalloss includes: The target person image is input into the ResNet backbone network model to extract the shallow detailed features of the target person and obtain the feature map of the target person. The feature map is input into the output layer, and the loss is calculated using the FocalLoss loss function. The ResNet backbone network model is then updated based on the loss to obtain a sensitive content anomaly detection model based on ResNet and Focalloss.

7. A method for identifying and processing sensitive individuals as target subjects in a live streaming scenario, characterized in that, The method includes: Based on the live streaming scenario, a sensitivity threshold is set for the target video streams to be detected. All target video streams are sorted according to their sensitivity values, and then the load of each cloud computing node is also sorted. Based on the sorting result, target video streams with high sensitivity values ​​are delivered to cloud computing nodes with lower loads for detection. The length of the local queue on each cloud computing node is obtained, and the target video stream is submitted to the computing node with the shortest queue. If the length of the local queue on all cloud computing nodes is greater than a preset value, the target video stream is placed in a global queue. All cloud computing nodes prioritize processing target video streams in the real-time queue. When there are no target video streams in the real-time queue, they retrieve them from the local queue. Video stream; if a sensitive threshold exceeding that of the target video stream is detected, the target video stream is marked. When the cloud computing node obtains a marked target video stream for the second time, it directly writes the marked target video stream into the instant queue for priority detection and kicks the target video stream previously in the instant queue back to the local queue. If the local queue is full, it will be divided into two parts, and the latter half will be written back to the global queue. When the local queue of the cloud computing node is empty, the node will first try to steal the target video stream from other nodes. If stealing the target video stream fails, the cloud computing node will go to the global queue to obtain the target video stream, obtaining half of the preset value of the local queue at a time. Human detection is performed on the target video stream in the cloud computing node, and sensitive content detection and sensitive area processing are performed for different target subjects.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method of claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method of claim 7.

Images