Content auditing method and device, equipment and storage medium

Abstract

The invention provides a content auditing method and device, equipment and a storage medium, and the method comprises the steps: responding to a content auditing request, and obtaining a video image frame corresponding to the content auditing request; performing regional attention distribution on the video image frame, and determining an attention region corresponding to the video image frame; determining a positioning feedback parameter of each attention area, and determining a target violation area in the attention areas; and extracting violation region features in the target violation region, and outputting a video auditing result corresponding to the violation region features. According to the method and the device, multiple local observation can be automatically and selectively performed on the image frames during image content auditing, so that details of violation areas existing in the image frames can be better concerned, the auditing accuracy and the overall efficiency can be effectively improved, and the interpretability of violation content auditing can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a content review method, apparatus, device and storage medium. Background Art

[0002] At present, with the rapid development of the Internet and smart TV technology, more and more smart TVs can support increasingly rich and diverse video content. However, the quality of these video contents varies greatly, and there may be illegal, bad or sensitive content that needs to be reviewed and filtered. The existing video content review method is manual review, which requires a lot of manpower and has the problems of low review efficiency and low accuracy, resulting in poor video review results. Summary of the invention

[0003] The embodiments of the present application provide a content review method, apparatus, device and storage medium, which aim to solve the technical problems in the prior art of low efficiency and low review accuracy of video content review.

[0004] On the one hand, an embodiment of the present application provides a content review method, the content review method comprising the following steps:

[0005] In response to a content review request, obtaining a video image frame corresponding to the content review request;

[0006] Performing regional attention allocation on the video image frame to determine an attention region corresponding to the video image frame;

[0007] Determining positioning feedback parameters of each attention area, and determining a target violation area in the attention area;

[0008] Extract the illegal area features in the target illegal area, and output the video review results corresponding to the illegal area features.

[0009] In a possible implementation of this embodiment, performing regional attention allocation on the video image frame to determine the attention area corresponding to the video image frame includes:

[0010] Using a feature extraction module in a preset content audit model to extract features from the video image frame, to obtain image visual features corresponding to the video image frame;

[0011] Using the attention module in the content audit model to perform regional attention allocation on the image visual features, and determine the attention weight of each initial image region in the video image frame;

[0012] An attention area corresponding to the video image frame is determined according to the attention weight.

[0013] In a possible implementation of this embodiment, determining the attention area corresponding to the video image frame according to the attention weight includes:

[0014] Using the attention weights, weighting the regional image features corresponding to the initial image region to obtain weighted regional features;

[0015] If the weighted region feature is greater than a preset region feature threshold, the initial image region corresponding to the weighted region feature is determined as the attention region.

[0016] In a possible implementation of this embodiment, determining the positioning feedback parameters of each attention area and determining the target violation area in the attention area includes:

[0017] Using the reinforcement learning module in the preset internal audit model to strategically locate the attention area, and determine the positioning feedback parameters of each attention area;

[0018] If the positioning feedback parameter is the first feedback parameter, determining the attention area corresponding to the first feedback parameter as the target violation area;

[0019] If the positioning feedback parameter is the second feedback parameter, the attention area corresponding to the second feedback parameter is filtered.

[0020] In a possible implementation of this embodiment, the using of a reinforcement learning module in a preset internal audit model to strategically locate the attention area and determine the positioning feedback parameters of each attention area includes:

[0021] Utilizing the target execution network in the reinforcement learning module to perform strategy positioning according to the regional temporal characteristics and regional attention characteristics of the attention region, and determining an initial review region in the attention region;

[0022] The positioning feedback parameters of the initial review area are determined according to a preset reward function.

[0023] In a possible implementation of this embodiment, extracting the illegal area features in the target illegal area and outputting the video review results corresponding to the illegal area features include:

[0024] Extracting illegal area features in the target illegal area;

[0025] Decoding the illegal area features using a target decoder to obtain area violation information;

[0026] The violation information of each area is summarized to obtain the video violation information of the video image frame, and the video review result corresponding to the video violation information is output.

[0027] In a possible implementation of this embodiment, before extracting the illegal area features in the target illegal area and outputting the video review results corresponding to the illegal area features, the method further includes:

[0028] Obtain a source domain dataset of illegal images;

[0029] Using the illegal sample images in the illegal image source domain data set to train the initial audit model to obtain target training parameters;

[0030] Transfer learning is performed according to the target training parameters and the target audit task of the initial audit model to obtain a content audit model.

[0031] On the other hand, the present application provides a content review device, the content review device comprising:

[0032] A data acquisition module is configured to respond to a content review request and acquire a video image frame corresponding to the content review request;

[0033] A region identification module is configured to perform regional attention allocation on the video image frame and determine an attention region corresponding to the video image frame;

[0034] A violation positioning module is configured to determine positioning feedback parameters of each attention area and determine a target violation area in the attention area;

[0035] The content review module is configured to extract the violation area features in the target violation area and output the video review results corresponding to the violation area features.

[0036] On the other hand, the present application also provides a content review device, the content review device comprising:

[0037] one or more processors;

[0038] Memory; and

[0039] One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the steps of the content review method.

[0040] On the other hand, the present application also provides a computer-readable storage medium on which a computer program is stored, and the computer program is loaded by a processor to execute the steps in the content review method.

[0041] In this application, by responding to a content review request, a video image frame corresponding to the content review request is obtained; regional attention is allocated to the video image frame to determine the attention area corresponding to the video image frame; the positioning feedback parameters of each attention area are determined to determine the target violation area in the attention area; the violation area features in the target violation area are extracted, and the video review results corresponding to the violation area features are output. When the video image frame is subject to content review, the attention mechanism can be used to allocate attention to the video image frame, thereby preliminarily determining the attention area in the video image frame where there may be violation content, and further determining the target violation area in the attention area according to the positioning feedback parameters of each attention area, and outputting the video review results corresponding to the target violation area. When the image content is reviewed, multiple local observations can be automatically and selectively performed on the image frame to better focus on the details of the violation area in the image frame, which can effectively improve the accuracy and overall efficiency of the review, and can effectively improve the interpretability of the review of the violation content. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the 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 work.

[0043] Figure 1 A schematic diagram of a scenario of a content review method according to an embodiment of the present application;

[0044] Figure 2 A flowchart of an embodiment of a content review method in an embodiment of the present application;

[0045] Figure 3 A schematic diagram of the structure of an embodiment of a reinforcement learning module provided in an embodiment of the present application;

[0046] Figure 4 A flowchart of an embodiment of transfer learning in the content review method provided in the embodiment of the present application;

[0047] Figure 5 A schematic diagram of the structure of an embodiment of a content review device provided in an embodiment of the present application;

[0048] Figure 6 A structural diagram of an embodiment of a content review device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0050] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, which are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first" and "second" are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0051] In this application, the word "exemplary" is used to mean "serving as an example, illustration, or illustration." Any embodiment described in this application as "exemplary" is not necessarily to be construed as being preferred or advantageous over other embodiments. The following description is given to enable any person skilled in the art to implement and use the invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the invention can be implemented without using these specific details. In other instances, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed in this application.

[0052] At present, with the rapid development of the Internet and smart TV technology, more and more smart TVs can support increasingly rich and diverse video content. However, the quality of these video contents varies greatly, and there may be illegal, bad or sensitive content that needs to be reviewed and filtered. The existing video content review method is manual review, which requires a lot of manpower and has the problems of low review efficiency and low accuracy, resulting in poor video review results.

[0053] Based on this, the present application proposes a content review method, device, equipment and computer-readable storage medium to solve the technical problems in the prior art of low efficiency and low review accuracy of video content review.

[0054] The content review method in the embodiment of the present invention is applied to a content review apparatus, and the content review apparatus is arranged in a content review device, and the content review device is provided with one or more processors, a memory, and one or more application programs, wherein the one or more application programs are stored in the memory and configured to be executed by the processor to implement the content review method; wherein the content review device can be an intelligent terminal, such as a mobile phone, a tablet computer, a network device, and a smart computer, etc.; optionally, the content review device can also be a server, or a service cluster composed of multiple servers.

[0055] like Figure 1 As shown, Figure 1 This is a scenario diagram of the content review method in an embodiment of the present application. The content review scenario in the embodiment of the present invention includes a content review device 100 (a content review apparatus is integrated in the content review device 100), and a computer-readable storage medium corresponding to the content review method is running in the content review device 100 to execute the steps of the content review method.

[0056] Understandably, Figure 1 The content review equipment in the content review method scenario shown, or the devices included in the content review equipment, does not constitute a limitation on the embodiments of the present invention, that is, the number of devices and types of devices of the content review equipment included in the content review method scenario, or the number of devices and types of devices included in each device, do not affect the overall implementation of the technical solution in the embodiments of the present invention, and can all be regarded as equivalent replacements or derivatives of the technical solution claimed to be protected by the embodiments of the present invention.

[0057] The content review device 100 in the embodiment of the present invention is mainly used to: respond to a content review request and obtain a video image frame corresponding to the content review request;

[0058] Performing regional attention allocation on the video image frame to determine an attention region corresponding to the video image frame;

[0059] Determining positioning feedback parameters of each attention area, and determining a target violation area in the attention area;

[0060] Extract the illegal area features in the target illegal area, and output the video review results corresponding to the illegal area features.

[0061] The content review device 100 in the embodiment of the present invention can be an independent content review device, such as a smart terminal such as a mobile phone, a tablet computer, a network device, a server, and a smart computer, or it can be a content review network or content review cluster composed of multiple content review devices.

[0062] The embodiments of the present application provide a content review method, apparatus, device, and computer-readable storage medium, which are described in detail below.

[0063] It can be understood by those skilled in the art that Figure 1 The application environment shown in is only one of the application scenarios related to the present application solution and does not constitute a limitation on the application scenario of the present application solution. Other application environments may also include Figure 1 More or fewer content review devices shown, or content review network connection relationships, such as Figure 1 Only one content review device is shown. It can be understood that the scenario of the content review method can also include one or more content review devices, which are not specifically limited here. The content review device 100 can also include a memory for storing video image data and other data.

[0064] It should be noted that Figure 1 The scenario diagram of the content review method shown is only an example. The scenario of the content review method described in the embodiment of the present invention is to more clearly illustrate the technical solution of the embodiment of the present invention, and does not constitute a limitation on the technical solution provided by the embodiment of the present invention.

[0065] Based on the scenario of the above-mentioned content review method, various embodiments of the content review method disclosed in the present invention are proposed.

[0066] like Figure 2 As shown, Figure 2 This is a flow chart of an embodiment of a content review method in an embodiment of the present application. The content review method includes the following steps 201 to 204:

[0067] 201. Respond to a content review request and obtain a video image frame corresponding to the content review request;

[0068] The content review method in this embodiment is applied to content review devices, and the type and number of content review devices are not specifically limited. That is, the content review device can be one or more intelligent terminals or servers for performing content review on the video image frames to be delivered.

[0069] Specifically, during operation, the content review device responds to a content review request for a video image frame to be delivered, wherein the content review request is an operation instruction to drive the content review device to conduct a content review on the video image frame, thereby detecting whether there are any illegal, bad or sensitive content areas in the video image frame. The content review request may be automatically triggered by the content review device. For example, an automatic review process is pre-set in the content review device, and after receiving the video image frame to be reviewed, the content review request for the video image frame is actively triggered. The video image frame is a video image sequence to be delivered to a designated TV terminal or video application. The video image frame contains at least one or more image key frames carrying video content information.

[0070] Specifically, after obtaining the content review request, the content review device parses the content review request, obtains the video image frame corresponding to the content review request, and uses the preset content review model to perform content review on the video image frame, thereby determining whether there is any illegal content in the video image frame. The video image frame is a video image sequence to be delivered to a designated TV terminal or video application. The video image frame contains at least one or more image key frames carrying video content information.

[0071] Optionally, in a specific embodiment, the content audit model includes a feature extraction module, an attention module, a reinforcement learning module and a target decoder. Among them, the feature extraction module is a model network that extracts features from video image frames. In a specific embodiment, the feature extraction module can be a convolutional neural network. The attention module is an attention network that uses an attention mechanism to select image areas that may violate the rules in the input video image frame and sets the attention weight of the image area accordingly. The reinforcement learning module is a deep reinforcement learning module that can generate a region selection strategy and perform secondary screening and auditing on each attention area in the video image frame. In a specific embodiment, the reinforcement learning module can be a deep deterministic policy gradient (DDPG, Deep Deterministic Policy Gradient) module. The target decoder module is a decoder module that can extract the features of the violation area of ​​the target violation area. In a specific embodiment, the target decoder can be an LSTM (Long Short-Term Memory) long short-term memory network.

[0072] 202. Performing regional attention allocation on the video image frame to determine an attention region corresponding to the video image frame;

[0073] Specifically, after acquiring the video image frame, the content audit device also uses the attention module in the content audit model to perform regional attention allocation on the video image frame, thereby determining the attention area in the video image frame where illegal content may exist, and the attention weight corresponding to each attention area. Among them, the attention area is the video image area in the video image frame where illegal content may exist. Each attention area corresponds to an attention weight, and the attention weights of different attention areas can be different attention weight values.

[0074] Specifically, after acquiring the video image frame, the content review device also first uses the feature extraction module in the content review video to perform feature extraction on the video image frame, thereby obtaining the image visual features of the video image frame. That is, the content review device obtains the video timing information corresponding to the video image frame, and inputs the video image frame into the feature extraction module according to the video timing information, and uses the feature extraction module to perform feature extraction on the video image frame, thereby obtaining the image visual features corresponding to the video image frame.

[0075] Specifically, after obtaining the image visual features corresponding to the video image frame, the content review device also drives the attention module in the content review model to use the spatial attention allocation mechanism to perform regional attention allocation on each initial image area corresponding to the image visual features, thereby determining the attention weights of each initial image area of ​​the video image frame in different video timing information, and determining the attention area corresponding to the video image frame based on the attention weight, thereby determining the image content area that needs to be paid attention to in the corresponding video timing information of the video image frame, and determining whether the image content area contains illegal image content in subsequent steps.

[0076] Optionally, in order to improve content review efficiency, after obtaining each initial image area and the attention weight corresponding to each initial image area, the content review device also determines the attention area corresponding to the video image frame based on the attention weight.

[0077] That is, after the content review device obtains the initial image area and the attention weight of each initial image area, it uses the attention weight to weight the regional image feature corresponding to the initial image area to obtain a weighted regional feature, and compares the weighted regional feature with a preset regional feature threshold to determine whether the initial image area is an attention area suspected of containing illegal content. The regional feature threshold is a feature detection threshold used to preliminarily detect whether there is illegal content in the initial image area.

[0078] Optionally, if the weighted area feature is greater than the area feature threshold, the content review device determines that there is a high probability that the initial image area corresponding to the weighted area feature may contain illegal content, and determines the initial image feature corresponding to the weighted area feature as the attention area to be reviewed.

[0079] 203. Determine positioning feedback parameters of each attention area, and determine a target violation area in the attention area;

[0080] Specifically, after the content audit device obtains the attention area in the video image frame, in order to avoid the situation where only the attention mechanism is used to locate the image area in the video image frame and the regional details of the video image frame to be audited may not be paid detailed attention, the content audit device also uses the reinforcement learning module in the content audit model to further locate and select the attention area, so as to better pay attention to the regional details in the video image frame to determine the target violation area in the attention area. The target violation area is a video image area that characterizes the presence of illegal content in the video image frame. The illegal content is the violation detection content set by the content audit device according to relevant regulations. In a specific embodiment, the reinforcement learning module is a deep deterministic policy gradient (DDPG) network.

[0081] Specifically, the content audit device can use the reinforcement learning module to strategically locate the attention module output by the attention module, select the initial audit area in the attention area according to the regional positioning strategy in the deep learning reinforcement module, and calculate the positioning feedback parameter of the initial audit area, and determine whether the initial audit area is the target violation area with illegal content through the positioning feedback parameter. Among them, the positioning feedback parameter is an action feedback parameter corresponding to the action of locating the specified attention area in the regional positioning strategy.

[0082] Specifically, the content review device uses the deep learning reinforcement module to perform a positioning selection action on a specific attention area according to the target positioning features specified in the regional positioning strategy, determines the selected initial review area in the attention area, and calculates the action feedback parameters corresponding to the positioning selection action according to the preset reward function, and determines the action feedback parameters as the positioning feedback parameters of the attention area. Among them, the target positioning feature is an image feature that characterizes the matching and positioning of the regional image features of the attention area. For example, the target positioning feature can be any one or more of color features, motion features, object recognition features, face label features, text features, and voice features.

[0083] The color feature is a color feature of a video image used to detect different areas and contents.

[0084] Motion features are features used to capture dynamic changes of objects in video image frames to determine whether there are abnormal events.

[0085] Object recognition features are feature parameters used to detect and classify object information in video image frames.

[0086] The face label feature is a feature parameter used to detect whether there is any illegal face behavior / label in the video image frame.

[0087] The text feature and / or voice feature is a feature parameter used to detect whether there is any illegal text content / voice content in the video image frame.

[0088] Specifically, after determining the target positioning feature, the content review device compares the target positioning feature with the timing feature and / or regional attention feature of the attention area, thereby performing strategic positioning according to the regional timing feature and regional attention feature of the attention area, determining the initial review area in the attention area, and providing positioning feedback parameters for the initial review area. The regional attention feature is feature information carried in the attention area that is of the same feature type as the target positioning feature.

[0089] Specifically, after determining the positioning feedback parameters of each attention area, the content review device also judges whether the initial review area corresponding to the positioning feedback parameters is an image area carrying illegal content based on the positioning feedback parameters, and then determines the target illegal area in the attention area, so as to further improve the regional review intensity of the video image frame, thereby improving the review accuracy.

[0090] Optionally, if the positioning feedback parameter is greater than a preset feedback threshold, the content review device determines that the positioning feedback parameter is a first feedback parameter, determines that the initial review area corresponding to the positioning feedback parameter is an image area carrying illegal content, and determines that the attention area corresponding to the positioning feedback parameter is a target illegal area. In a specific embodiment, the preset feedback threshold is 0, and the first feedback parameter is a positioning feedback parameter with a positive value, that is, when the reinforcement learning module locates the attention area carrying illegal content, it can calculate a positioning feedback parameter with a positive value according to the preset reward function.

[0091] Optionally, if the positioning feedback parameter is less than the preset feedback threshold, the content review device determines that the positioning feedback parameter is a second feedback parameter, determines that the initial review area corresponding to the positioning feedback parameter is an image area that does not carry illegal content, and filters the attention area corresponding to the second feedback parameter.

[0092] Optionally, before using the reinforcement learning module to locate and select the violation area in each attention area, the content review device also trains the reinforcement learning module so that the reinforcement learning module can perform a secondary screening of each attention area to determine the target violation area in the attention area.

[0093] Specifically, Figure 3 As shown, Figure 3 A schematic diagram of the structure of an embodiment of the reinforcement learning module provided in the present application. The reinforcement learning module includes an experience replay pool, and at least one execution network and an evaluation network. Figure 3 In the illustrated embodiment, the reinforcement learning module includes an experience replay pool 301, a first execution network 302, a first evaluation network 303, a second execution network 304, and a second evaluation network 305. The first evaluation network and the second evaluation network are used to select an initial review area in the positioning attention area according to specific target positioning features, and the first evaluation network and the second evaluation network are used to output positioning feedback parameters corresponding to the selected positioning operation of the corresponding first action network and the corresponding second action network.

[0094] Specifically, the content review device pre-establishes the first execution network u(s,a|θ u ) and the first evaluation network Q(s,a|θ Q ), and establish the second learning

[0095] ”

[0096] The second execution network u'(s,a|θ in the learning network u ) and the second evaluation network Q'(s,a|θ Q ). Among them, θ Q and θ u

[0097] ”

[0098] are the network weights of the first evaluation network and the first execution network respectively. θ Q and θ u Second Rating Network

[0099] ”

[0100] The network weights of the network and the second execution network, and θ Q Approaching θ Q ,θ u Approaching θ u In a specific embodiment, the first learning network is a core network, and the second learning network is a target network.

[0101] Specifically, the content review device also establishes an experience replay pool corresponding to the reinforcement learning module, which is an empty experience replay pool, and uses the training positioning features in the preset training execution strategy to control the first execution network and the second execution network to perform positioning training actions, and uses the first evaluation network and the second evaluation network to calculate the training feedback parameters of the positioning training action respectively, and stores the training positioning features, positioning training actions and training feedback parameters, as well as updated training positioning features as training parameter combinations in the experience replay pool. After multiple cycles, the training parameter combination samples corresponding to N training parameter combinations are taken out from the experience replay pool, and the sample mean corresponding to the training parameter combination samples is calculated. The reward function in the first execution network and the second execution network is updated by the preset update function and the sample mean, and the positioning training action is updated by the loss function and the gradient descent function.

[0102] Among them, the expression formula of the preset update function is:

[0103] y i =r i +λQ′(s i+1 ,a′(s i+1 |θ u′ |θ Q′ ))

[0104] Among them, r i is the training feedback parameter, s i+1 Position features for update training.

[0105] Among them, the expression of the loss function is:

[0106]

[0107] Among them, L is the loss function of the reinforcement learning module, s i is for training positioning features, and ai is for positioning training actions.

[0108] Among them, the expression of updating the positioning training action using the gradient descent function is as follows:

[0109]

[0110] Specifically, after training the reinforcement learning module, the content review device updates the first learning network and / or the second learning network in the reinforcement learning module using the updated training actions obtained from the training until the training error is less than the preset error threshold, then exits the training cycle to obtain a trained reinforcement learning module.

[0111] 204. Extract the illegal area features in the target illegal area, and output the video review results corresponding to the illegal area features.

[0112] Specifically, after obtaining each target violation area corresponding to the video image frame, the content review device also performs content review on the target violation area, thereby extracting the violation area features in the target violation area, and outputting the video review results corresponding to the violation area features.

[0113] Specifically, the content review device uses the target decoder in the content review model to decode the violation area features in the target violation area, obtains the regional violation information in the target violation area, and uses the regional violation information as the review violation reason of the target violation area as the video review result, and outputs the video review result.

[0114] Specifically, after determining each target violation area corresponding to the video image frame, the content review device extracts the violation area features in the target violation area, and uses the target decoder to decode the violation area features, that is, decodes the violation area features from a vector into violation information text to obtain the area violation information.

[0115] Specifically, after obtaining the violation information of each area, the content review device also summarizes the violation information of each area, obtains the video violation information of the video image frame, and outputs the video review result corresponding to the video violation information, so that the user can intuitively determine the violation cause of the video image frame, thereby improving the explainability of the content review.

[0116] In this embodiment, the content audit device obtains the video image frame corresponding to the content audit request by responding to the content audit request; performs regional attention allocation on the video image frame to determine the attention area corresponding to the video image frame; determines the positioning feedback parameters of each attention area to determine the target violation area in the attention area; extracts the violation area features in the target violation area, and outputs the video audit result corresponding to the violation area features. When performing content audit on the video image frame, the attention mechanism can be used to allocate attention to the video image frame, thereby preliminarily determining the attention area in the video image frame where the violation content may exist, and further determining the target violation area in the attention area according to the positioning feedback parameters of each attention area, and outputting the video audit result corresponding to the target violation area. When auditing the image content, the image frame can be automatically and selectively observed locally in multiple places, so as to better focus on the details of the violation area in the image frame, which can effectively improve the accuracy and overall efficiency of the audit, and can effectively improve the interpretability of the violation content audit.

[0117] like Figure 4 As shown, Figure 4This is a flow chart of an embodiment of transfer learning in the content review method provided in the embodiment of the present application. Specifically, in this embodiment, the content review method further includes steps 401 to 403:

[0118] 401. Obtain a source domain dataset of illegal images;

[0119] 402. Train the initial audit model using the illegal sample images in the illegal image source domain data set to obtain target training parameters;

[0120] 403. Perform transfer learning according to the target training parameters and the target audit task of the initial audit model to obtain a content audit model.

[0121] Based on the above embodiments, in this embodiment, since the existing video content review tasks require a large amount of training data for training, and the existing training data samples are relatively small, the content review device also pre-transfers the content review model before performing content review on the video image frames to obtain a trained content review model.

[0122] Specifically, the content audit device obtains a source domain dataset of illegal images, wherein the illegal source domain image dataset contains several frames of annotated illegal sample images. In addition, the content audit device also obtains video data with illegal video content audit information as a target domain dataset. Optionally, in a specific embodiment, the content audit device also uses the illegal image source domain dataset and the target domain dataset to divide into a training set, a validation set, and a test set, thereby ensuring the quality and balance of the dataset.

[0123] Specifically, after obtaining the illegal image source domain data set, the content review device also uses the illegal sample images in the illegal image source domain data set to train the initial review model, and uses the feature extraction module in the initial review model to perform image classification task training on the illegal sample images to obtain the target training parameters of the initial review model.

[0124] Specifically, after obtaining the target training parameters, the content audit device performs transfer learning based on the target training parameters and the target audit task of the initial audit model to obtain the content audit model. That is, the content audit device freezes part or all of the target training parameters in the initial audit model, and transfers the target training parameters to the target audit task corresponding to the target domain data set, and continues to perform transfer learning on the initial audit model using the target domain data set and the target training parameters, and uses the target domain data set to fine-tune the parameters of the initial audit model to generate a content audit model that can complete the target audit task. Among them, the target audit task is a training task for auditing illegal content in video image frames.

[0125] In this embodiment, the content audit device obtains a source domain dataset of illegal images, which contains labeled illegal sample images; trains the initial audit model using the illegal sample images in the source domain dataset to obtain target training parameters; and performs transfer learning based on the target training parameters and the target audit task of the initial audit model to obtain a content audit model. Transfer learning training of the model using the source domain dataset of illegal images is implemented, thereby improving the efficiency of model training, and can enhance the performance and generalization ability of the model, so as to better cope with video audit tasks.

[0126] In order to better implement the content audit method in the embodiment of the present application, based on the content audit method, the embodiment of the present application also provides a content audit device, such as Figure 5 As shown, Figure 5 The schematic diagram of the structure of the content audit device provided in the embodiment of the present application is as follows. Specifically, the content audit device 500 includes:

[0127] The data acquisition module 501 is configured to respond to a content review request and acquire a video image frame corresponding to the content review request;

[0128] A region identification module 502 is configured to perform regional attention allocation on the video image frame and determine an attention region corresponding to the video image frame;

[0129] The violation positioning module 503 is configured to determine positioning feedback parameters of each attention area and determine a target violation area in the attention area;

[0130] The content review module 504 is configured to extract the illegal area features in the target illegal area and output the video review results corresponding to the illegal area features.

[0131] In a possible implementation of this embodiment, the content review device performs regional attention allocation on the video image frame to determine the attention area corresponding to the video image frame, including:

[0132] Using a feature extraction module in a preset content audit model to extract features from the video image frame, to obtain image visual features corresponding to the video image frame;

[0133] Using the attention module in the content audit model to perform regional attention allocation on the image visual features, and determine the attention weight of each initial image region in the video image frame;

[0134] An attention area corresponding to the video image frame is determined according to the attention weight.

[0135] In a possible implementation of this embodiment, the content review device determines the attention area corresponding to the video image frame according to the attention weight, including:

[0136] Using the attention weights, weighting the regional image features corresponding to the initial image region to obtain weighted regional features;

[0137] If the weighted region feature is greater than a preset region feature threshold, the initial image region corresponding to the weighted region feature is determined as the attention region.

[0138] In a possible implementation of this embodiment, the content audit device determines the positioning feedback parameters of each attention area and determines the target violation area in the attention area, including:

[0139] Using the reinforcement learning module in the preset internal audit model to strategically locate the attention area, and determine the positioning feedback parameters of each attention area;

[0140] If the positioning feedback parameter is the first feedback parameter, determining the attention area corresponding to the first feedback parameter as the target violation area;

[0141] If the positioning feedback parameter is the second feedback parameter, the attention area corresponding to the second feedback parameter is filtered.

[0142] In a possible implementation of this embodiment, the content audit device uses the reinforcement learning module in the preset internal audit model to strategically locate the attention area and determine the positioning feedback parameters of each attention area, including:

[0143] Utilizing the target execution network in the reinforcement learning module to perform strategy positioning according to the regional temporal characteristics and regional attention characteristics of the attention region, and determining an initial review region in the attention region;

[0144] The positioning feedback parameters of the initial review area are determined according to a preset reward function.

[0145] In a possible implementation of this embodiment, the content review device extracts the illegal area features in the target illegal area and outputs the video review results corresponding to the illegal area features, including:

[0146] Extracting illegal area features in the target illegal area;

[0147] Decoding the illegal area features using a target decoder to obtain area violation information;

[0148] The violation information of each area is summarized to obtain the video violation information of the video image frame, and the video review result corresponding to the video violation information is output.

[0149] In a possible implementation of this embodiment, before the content review device extracts the illegal area features in the target illegal area and outputs the video review result corresponding to the illegal area features, it also includes:

[0150] Obtain a source domain dataset of illegal images;

[0151] Using the illegal sample images in the illegal image source domain data set to train the initial audit model to obtain target training parameters;

[0152] Transfer learning is performed according to the target training parameters and the target audit task of the initial audit model to obtain a content audit model.

[0153] In this embodiment, the content audit device obtains the video image frame corresponding to the content audit request by responding to the content audit request; performs regional attention allocation on the video image frame to determine the attention area corresponding to the video image frame; determines the positioning feedback parameters of each attention area to determine the target violation area in the attention area; extracts the violation area features in the target violation area, and outputs the video audit result corresponding to the violation area features. When performing content audit on the video image frame, the attention mechanism can be used to allocate attention to the video image frame, thereby preliminarily determining the attention area in the video image frame where the violation content may exist, and further determining the target violation area in the attention area according to the positioning feedback parameters of each attention area, and outputting the video audit result corresponding to the target violation area. When auditing the image content, the image frame can be automatically and selectively observed locally in multiple places, so as to better focus on the details of the violation area in the image frame, which can effectively improve the accuracy and overall efficiency of the audit, and can effectively improve the interpretability of the violation content audit.

[0154] The embodiment of the present invention also provides a content review device, such as Figure 6 As shown, Figure 6 A schematic diagram of the structure of an embodiment of the content review device provided in the embodiments of the present application.

[0155] The content review device integrates any one of the content review devices provided by the embodiments of the present invention, and the content review device includes:

[0156] one or more processors;

[0157] Memory; and

[0158] One or more applications, wherein the one or more applications are stored in the memory and are configured so that the processor executes the steps of the content review method described in any of the above-mentioned content review method embodiments.

[0159] Specifically, the content audit device may include one or more processing core processors 601, one or more computer-readable storage media memories 602, a power supply 603, an input unit 604, and other components. Those skilled in the art will understand that Figure 6 The content audit device structure shown in the figure does not constitute a limitation on the content audit device, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently. Among them:

[0160] The processor 601 is the control center of the content audit device. It uses various interfaces and lines to connect the various parts of the entire content audit device. By running or executing software programs and / or modules stored in the memory 602, and calling the data stored in the memory 602, it performs various functions of the content audit device and processes data, thereby monitoring the content audit device as a whole. Optionally, the processor 601 may include one or more processing cores; preferably, the processor 601 may integrate an application processor and a modem processor, wherein the application processor mainly processes the operating system, user interface, and application programs, etc., and the modem processor mainly processes wireless communications. It is understandable that the above-mentioned modem processor may not be integrated into the processor 601.

[0161] The memory 602 can be used to store software programs and modules. The processor 601 executes various functional applications and data processing by running the software programs and modules stored in the memory 602. The memory 602 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the content review device, etc. In addition, the memory 602 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage devices. Accordingly, the memory 602 may also include a memory controller to provide the processor 601 with access to the memory 602.

[0162] The content audit device also includes a power supply 603 for supplying power to each component. Preferably, the power supply 603 can be logically connected to the processor 601 through a power management system, so as to manage charging, discharging, and power consumption through the power management system. The power supply 603 can also include any components such as one or more DC or AC power supplies, recharging systems, power failure detection circuits, power converters or inverters, and power status indicators.

[0163] The content review device may also include an input unit 604, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal input related to user settings and function control.

[0164] Although not shown, the content audit device may also include a display unit, etc., which will not be described in detail here. Specifically in this embodiment, the processor 601 in the content audit device will load the executable files corresponding to the processes of one or more applications into the memory 602 according to the following instructions, and the processor 601 will run the application stored in the memory 602 to achieve various functions, as follows:

[0165] In response to a content review request, obtaining a video image frame corresponding to the content review request;

[0166] Performing regional attention allocation on the video image frame to determine an attention region corresponding to the video image frame;

[0167] Determining positioning feedback parameters of each attention area, and determining a target violation area in the attention area;

[0168] Extract the illegal area features in the target illegal area, and output the video review results corresponding to the illegal area features.

[0169] To this end, an embodiment of the present invention provides a computer-readable storage medium, which may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc. A computer program is stored thereon, and the computer program is loaded by a processor to execute the steps in any content review method provided in an embodiment of the present invention. For example, the computer program may be loaded by a processor to execute the following steps:

[0170] In response to a content review request, obtaining a video image frame corresponding to the content review request;

[0171] Performing regional attention allocation on the video image frame to determine an attention region corresponding to the video image frame;

[0172] Determining positioning feedback parameters of each attention area, and determining a target violation area in the attention area;

[0173] Extract the illegal area features in the target illegal area, and output the video review results corresponding to the illegal area features.

[0174] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the detailed description of other embodiments above, and will not be repeated here.

[0175] In specific implementation, the above units or structures can be implemented as independent entities, or can be arbitrarily combined to be implemented as the same or several entities. The specific implementation of the above units or structures can refer to the previous method embodiments, which will not be repeated here.

[0176] The specific implementation of the above operations can be found in the previous embodiments, which will not be described in detail here.

[0177] The above is a detailed introduction to a content review method provided in an embodiment of the present application. Specific embodiments are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A content review method, characterized in that: The content review method includes: In response to a content review request, obtaining a video image frame corresponding to the content review request; Performing regional attention allocation on the video image frame to determine an attention region corresponding to the video image frame; Determining positioning feedback parameters of each attention area, and determining a target violation area in the attention area; Extract the illegal area features in the target illegal area, and output the video review results corresponding to the illegal area features.

2. The content review method according to claim 1, characterized in that: The performing regional attention allocation on the video image frame to determine the attention area corresponding to the video image frame includes: Using a feature extraction module in a preset content audit model to extract features from the video image frame, to obtain image visual features corresponding to the video image frame; Using the attention module in the content audit model to perform regional attention allocation on the image visual features, and determine the attention weight of each initial image region in the video image frame; An attention area corresponding to the video image frame is determined according to the attention weight.

3. The content review method according to claim 2, characterized in that: The determining the attention area corresponding to the video image frame according to the attention weight includes: Using the attention weights, weighting the regional image features corresponding to the initial image region to obtain weighted regional features; If the weighted region feature is greater than a preset region feature threshold, the initial image region corresponding to the weighted region feature is determined as the attention region.

4. The content review method according to claim 1, characterized in that: The determining of the positioning feedback parameters of each attention area and the determining of the target violation area in the attention area include: Using the reinforcement learning module in the preset internal audit model to strategically locate the attention area, and determine the positioning feedback parameters of each attention area; If the positioning feedback parameter is the first feedback parameter, determining the attention area corresponding to the first feedback parameter as the target violation area; If the positioning feedback parameter is the second feedback parameter, the attention area corresponding to the second feedback parameter is filtered.

5. The content review method according to claim 4, characterized in that: The using of the reinforcement learning module in the preset internal audit model to strategically locate the attention area and determine the positioning feedback parameters of each attention area includes: Utilizing the target execution network in the reinforcement learning module to perform strategy positioning according to the regional temporal characteristics and / or regional attention characteristics of the attention region, and determining an initial review region in the attention region; The positioning feedback parameters of the initial review area are determined according to a preset reward function.

6. The content review method according to claim 1, characterized in that: The extracting the illegal area features in the target illegal area and outputting the video review results corresponding to the illegal area features include: Extracting illegal area features in the target illegal area; Decoding the illegal area features using a target decoder to obtain area violation information; The violation information of each area is summarized to obtain the video violation information of the video image frame, and the video review result corresponding to the video violation information is output.

7. The content review method according to any one of claims 1 to 6, characterized in that: Before extracting the illegal area features in the target illegal area and outputting the video review results corresponding to the illegal area features, the method further includes: Obtaining a source domain dataset of illegal images, wherein the source domain dataset of illegal images contains annotated illegal sample images; Using the illegal sample images in the illegal image source domain data set to train the initial audit model to obtain target training parameters; Transfer learning is performed according to the target training parameters and the target audit task of the initial audit model to obtain a content audit model.

8. A content review device, characterized in that: The content review device comprises: A data acquisition module is configured to respond to a content review request and acquire a video image frame corresponding to the content review request; A region identification module is configured to perform regional attention allocation on the video image frame and determine an attention region corresponding to the video image frame; A violation positioning module is configured to determine positioning feedback parameters of each attention area and determine a target violation area in the attention area; The content review module is configured to extract the violation area features in the target violation area and output the video review results corresponding to the violation area features.

9. A content review device, characterized in that: The content review device comprises: one or more processors; Memory; and One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the processor to implement the steps of the content review method described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program is loaded by a processor to execute the steps of the content review method described in any one of claims 1 to 7.

Images