Multimedia content generation methods, apparatus, storage media and computer equipment

Abstract

The multimedia content generation method, apparatus, storage medium, and computer equipment provided in this application first involve a smart TV terminal performing voiceprint recognition on voice commands to obtain age-aware tags, and then converting the voice commands into text to be reviewed. The age-aware tags are then used to constrain text pre-review rules to obtain age-appropriate pre-review rules. The text to be reviewed is then detected, and the legally approved text and age-aware tags are uploaded to the cloud. The cloud then generates low-quality multimedia content with complete semantics based on the legally approved text using low-quality media parameters. This low-quality multimedia content is then detected according to compliance review rules. The detected low-quality multimedia content is then enhanced, and deep semantic review is performed simultaneously using age-aware tags. When the deep semantic review passes, high-quality multimedia content is output to the smart TV terminal. This achieves a step-by-step filtering of invalid tasks, reducing unnecessary consumption of computing resources during the multimedia content generation process.

Description

Technical Field

[0001] This application relates to the field of smart TV technology, and in particular to a multimedia content generation method, apparatus, storage medium, and computer equipment. Background Technology

[0002] With the development of AI-generated content technology, smart TVs have gradually acquired the ability to generate multimedia content such as images, videos, and audio. After a user initiates a content generation request through voice interaction or other means, the content is typically generated by a cloud-based system, and the result is returned to the smart TV for display.

[0003] Currently, multimedia content generation typically relies on large-scale generative models, consuming significant computing resources. This is especially true in image and video generation scenarios, where the process often involves complex inference calculations, consuming substantial GPU resources and processing time. However, existing solutions usually execute the entire content generation process directly based on user input. When the generated result fails to meet output requirements, the previously invested computing resources cannot be recovered, resulting in ineffective computation and resource waste. With the continuous growth of smart TV users and content generation demands, this problem will further exacerbate the cloud computing burden and reduce resource utilization efficiency. Therefore, the ineffective tasks generated during the current multimedia content generation process contribute to the consumption of computing resources. Summary of the Invention

[0004] The purpose of this application is to at least address one of the aforementioned technical defects, particularly the technical defect in the prior art where invalid tasks generated during the multimedia content generation process lead to the consumption of computing resources.

[0005] In a first aspect, this application provides a multimedia content generation method applied to a television system, the television system including a smart TV terminal and a cloud, the method comprising:

[0006] The smart TV terminal responds to voice commands, performs voiceprint recognition on the voice commands, obtains age perception tags, and converts the voice commands into text to be reviewed;

[0007] By using age-aware tags to constrain text pre-review rules, after obtaining age-adaptive pre-review rules, the text to be reviewed is detected, and the legal text that passes the detection and the age-aware tags are uploaded to the cloud.

[0008] The cloud platform generates low-quality multimedia content with complete semantics based on legal text and low-quality media parameters, and detects low-quality multimedia content according to compliance review rules.

[0009] The low-quality multimedia content that passes the detection is enhanced, and age-aware tags are used for deep semantic review. When the deep semantic review is passed, high-quality multimedia content is output to the smart TV terminal.

[0010] In one embodiment, age-aware labels are used to constrain text pre-screening rules to obtain age-appropriate pre-screening rules, including:

[0011] Determine the corresponding age category based on the age perception label;

[0012] Obtain the age constraint rules corresponding to the age category, and obtain the text pre-review rules;

[0013] By using age constraint rules, the sensitive content filtering conditions, semantic redirection conditions, and content generation constraints in the text pre-review rules are adjusted to obtain age-adaptive pre-review rules.

[0014] In one embodiment, detecting the text to be reviewed includes:

[0015] Based on age-appropriate pre-review rules, the text to be reviewed is detected for illegal content.

[0016] Based on age-appropriate pre-review rules, age-appropriateness checks are performed on the texts to be reviewed.

[0017] When the text to be reviewed passes the detection based on the results of the violation detection and the age suitability detection, the text to be reviewed will be determined as a legal text.

[0018] In one embodiment, generating semantically complete low-quality multimedia content based on legal text using low-quality media parameters includes:

[0019] Determine the content generation conditions based on the valid text;

[0020] Based on the preset low-quality generation strategy, determine the corresponding target media parameters;

[0021] While keeping the semantic information of the content generation conditions unchanged, configure low-quality media parameters according to the target media parameters.

[0022] By utilizing content generation conditions and low-quality media parameters, content generation is performed to obtain low-quality multimedia content with complete semantics.

[0023] In one embodiment, detecting low-quality multimedia content according to compliance audit rules includes:

[0024] Extract content features corresponding to low-quality multimedia content;

[0025] In accordance with compliance review rules, the content features are tested for illegal elements and cross-modal consistency.

[0026] The audit results are determined based on the results of the detection of non-compliant elements and the results of cross-modal consistency detection.

[0027] In one embodiment, low-quality multimedia content that passes detection is enhanced, and deep semantic review is performed simultaneously using age-aware tags, including:

[0028] Based on the target output quality, determine the enhancement configuration parameters, and perform quality enhancement processing on the low-quality multimedia content that has passed the detection based on the enhancement configuration parameters;

[0029] During the quality enhancement process, intermediate enhancement results are obtained;

[0030] Based on age-aware tags, determine the deep review rules for the corresponding age categories, and use the deep review rules to perform deep semantic review on the enhanced intermediate results;

[0031] When the deep semantic review results indicate that there is content that does not meet the requirements of the corresponding age category, an enhancement termination instruction is generated, and the quality enhancement process is interrupted using the enhancement termination instruction.

[0032] In one embodiment, deep semantic auditing of enhanced intermediate results is performed using deep auditing rules, including:

[0033] Perform implicit semantic analysis and sensitive content detection on the enhanced intermediate results;

[0034] Age appropriateness assessments were performed on the enhanced intermediate results based on age categories;

[0035] Based on the results of implicit semantic analysis, sensitive content detection, and age suitability assessment, a deep semantic review result is generated.

[0036] Secondly, this application provides a multimedia content generation device for use in a television system, the television system including a smart TV terminal and a cloud, the device comprising:

[0037] The text to be reviewed module is used by smart TV terminals to respond to voice commands, perform voiceprint recognition on the voice commands, obtain age perception tags, and convert the voice commands into text to be reviewed.

[0038] The text pre-review module is used to constrain the text pre-review rules with age-aware tags. After obtaining the age-adaptive pre-review rules, it detects the text to be reviewed and uploads the legal text and age-aware tags that pass the detection to the cloud.

[0039] The low-quality multimedia content generation module is used to generate semantically complete low-quality multimedia content in the cloud based on legal text and low-quality media parameters, and to detect low-quality multimedia content according to compliance review rules.

[0040] The high-quality multimedia content generation module is used to enhance low-quality multimedia content that has passed the detection, and simultaneously uses age-aware tags to perform deep semantic review. When the deep semantic review is passed, high-quality multimedia content is output to the smart TV terminal.

[0041] Thirdly, this application provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of any of the multimedia content generation methods described in the above embodiments.

[0042] Fourthly, this application provides a computer device, including: one or more processors, and a memory;

[0043] The memory stores computer-readable instructions, which, when executed by one or more processors, perform the steps of any of the multimedia content generation methods described in the above embodiments.

[0044] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0045] The multimedia content generation method, apparatus, storage medium, and computer equipment provided in this application are applied to a television system, which includes a smart TV terminal and a cloud. By performing voiceprint recognition on voice commands on the smart TV terminal to obtain age-aware tags and converting the voice commands into text to be reviewed, the terminal-side computing resources can be used to complete the perception and preprocessing of user age-related information, reducing the subsequent processing burden on the cloud. The age-aware tags constrain text pre-review rules, and the text to be reviewed is detected. Only legal text that passes the detection and age-aware tags are uploaded to the cloud, enabling the pre-filtering of unqualified tasks on the terminal side, reducing the consumption of network transmission resources and cloud computing resources by invalid tasks. Based on the legal text, the cloud uses low-quality media parameters to generate low-quality multimedia content with complete semantics and performs compliance review, enabling content validity verification at a lower computing cost and avoiding the direct execution of high-quality generation for non-compliant content. The approved low-quality multimedia content is enhanced and combined with age-aware tags for deep semantic review, enabling the generation of high-quality multimedia content while ensuring content compliance and age suitability. This high-quality multimedia content is then output to the smart TV terminal for presentation, thereby improving the user viewing experience. Thus, by collaborating between smart TV terminals and the cloud to complete content pre-screening, low-cost verification, and high-quality generation, invalid tasks are filtered out step by step, unnecessary computing resources are consumed in the process of multimedia content generation, and the overall resource utilization efficiency and content generation efficiency are improved. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating the multimedia content generation method provided in this application embodiment;

[0048] Figure 2 This is a schematic diagram of the structure of the multimedia content generation apparatus provided in the embodiments of this application;

[0049] Figure 3 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0050] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] This application provides a multimedia content generation method applied to a television system, which includes a smart TV terminal and a cloud platform. The smart TV terminal refers to a television device with audio and video playback, voice capture, and local computing capabilities. Its functions include lightweight computing tasks such as voice capture, voiceprint perception, and pre-screening of prompts. It utilizes the central processing unit, neural network processing unit, or digital signal processor of the television system chip to perform inference during video playback intervals and intercepts illegal instructions and sensitive voiceprint data locally. The cloud platform refers to a server cluster deployed in the backend of the television operator. Its functions include low-quality initial draft generation, rapid compliance review, iterative regeneration, and high-quality enhancement and deep semantic review, which can be achieved using lightweight graphics processor resources. The cloud platform also deploys a high-performance graphics processor cluster responsible for high-quality enhancement and deep semantic review, triggered only after the content passes the pre-compliance screening, ensuring the effective utilization of high-performance computing resources. Figure 1 As shown, the method includes the following steps:

[0052] S101: The smart TV terminal responds to voice commands, performs voiceprint recognition on the voice commands, obtains age perception tags, and converts the voice commands into text to be reviewed.

[0053] Voice commands are meaningful sound signals expressed by users to a smart TV terminal through spoken language. Voiceprint recognition is the technical process of extracting the speaker's unique acoustic features from the voice command and analyzing their age attributes accordingly. An age-aware tag is a data identifier output by voiceprint recognition that represents the speaker's age range or age characteristics. The text to be reviewed is the text content obtained by converting the voice command into text; this text can then be used for content review or semantic analysis.

[0054] In practice, the smart TV terminal continuously monitors ambient sounds. When the user utters a voice command, the smart TV terminal captures the audio data of that command through its built-in microphone array. In response to the voice command, the smart TV terminal simultaneously or sequentially performs two processing steps on the captured audio data: one step is voiceprint recognition to generate an age-aware tag, and the other step is speech-to-text conversion to generate text to be reviewed.

[0055] In the voiceprint recognition process, the smart TV terminal first extracts acoustic features reflecting the speaker's vocal cord structure from the audio data. One approach is to extract Mel-frequency cepstral coefficient features and input these features into a pre-trained Gaussian mixture model. This model establishes acoustic probability distributions for different age groups, such as 0-12 years old, 13-45 years old, and over 46 years old. The model outputs the age group with the highest probability as the age-aware label. Another approach is to construct a deep residual convolutional neural network. The original spectrogram of the speech command is directly fed into the network. The network automatically learns age-related acoustic patterns through multiple convolutional layers and finally outputs a discrete age group label, such as "child," "youth," or "elderly," as the age-aware label through a fully connected layer.

[0056] In the speech-to-text process, the smart TV terminal converts the same voice command into text. This can be achieved by the smart TV terminal compressing the audio data and sending it wirelessly to a cloud-based speech recognition service. The cloud service then decodes the Chinese character sequence based on an end-to-end attention model and returns it to the terminal, which saves this sequence as the text to be reviewed. Alternatively, a lightweight offline speech recognition engine can be integrated within the smart TV terminal. This engine uses a connection-time classification algorithm to directly map audio to text locally on the device, outputting the text to be reviewed without an internet connection.

[0057] After completing the above two processes, the smart TV terminal receives two outputs: an age-aware tag and text to be reviewed. Voiceprint recognition and speech-to-text can be performed in a pipeline manner: first, voiceprint recognition is completed to obtain the age-aware tag, and then this tag is used to assist in the selection of the acoustic model during the speech-to-text process. For example, a speech recognition model specifically adapted to children's pronunciation can be called for a child's age tag. Alternatively, the two processes can be executed in parallel, utilizing a multi-core processor to process audio data simultaneously, thereby shortening the overall response time.

[0058] It should be noted that by simultaneously or sequentially performing voiceprint recognition and voice conversion on the same voice command, the smart TV terminal can automatically obtain an age-aware tag without the user actively inputting any age information. This avoids the tedious operation of manually setting age modes and improves the convenience of human-computer interaction. Using the age-aware tag obtained from voiceprint recognition, subsequent processing can provide differentiated services based on the user's age group. For example, it can automatically filter suitable content for children's tags or increase the prompt volume for elderly tags, thus achieving a refined understanding and response to voice commands. Converting the voice command into text to be reviewed transforms the originally invisible voice content into storable, searchable, and reviewable text information, facilitating compliance checks or historical record queries of user commands and compensating for the difficulty of text analysis with pure voice data. Both steps are completed based on the same voice command, without adding additional data collection steps, thereby ensuring processing efficiency and reducing system power consumption. This implementation method allows the smart TV terminal to more intelligently and safely adapt to the voice interaction needs of users of different ages.

[0059] S102: Use age-aware tags to constrain text pre-review rules, obtain age-adaptive pre-review rules, detect the text to be reviewed, and upload the legal text that passes the detection and age-aware tags to the cloud.

[0060] Among them, the text pre-review rules are a set of conditions or filtering logic used to determine whether the text to be reviewed contains illegal content; the age-adapted pre-review rules are text pre-review rules that are dynamically adjusted according to age-aware tags and are applicable to specific age groups; and the legal text refers to the text that passes the pre-review rules and does not contain illegal content.

[0061] Specifically, after obtaining age-aware tags and text to be reviewed, the smart TV terminal uses the age-aware tags to adjust the text pre-review rules, thereby generating a set of age-appropriate pre-review rules that match the age tag. For example, when the age-aware tag indicates that the speaker is a child, the sensitivity threshold of the text pre-review rules is adjusted to the strictest level, and any words involving violence, terror, or adult topics are blocked; when the tag is youth, the rules are appropriately relaxed but still maintain high sensitivity to offensive language; when the tag is elderly, the rules focus on health and wellness rumors or fraudulent information. One way to achieve constraint adjustment is to establish a mapping table, which pre-stores the pre-review rule parameter set corresponding to different age tags, and directly looks up the corresponding age-appropriate pre-review rules based on the age-aware tag. Another way is to use a rule generator, which is a lightweight decision tree model. After inputting the age-aware tag, it outputs a set of dynamic rule weights and keyword lists. For example, for the child tag, it outputs a prohibited list containing words such as "fighting," and for the elderly tag, it outputs a warning list containing words such as "freebies."

[0062] After obtaining the age-appropriate pre-screening rules, these rules are used to examine each line of the previously converted text to be reviewed. The detection process involves matching the text to be reviewed against the keyword library or regular expression patterns in the age-appropriate pre-screening rules. This can be achieved by a rule engine sequentially scanning every word in the text; if any word matches the prohibited keyword list, the text is immediately deemed invalid and the detection process terminates. Alternatively, a pattern matching algorithm based on a finite state automaton can be used. This algorithm compiles the entire text into a state machine at once and matches multiple patterns simultaneously. After matching, a compliance score is output; only when the score is higher than a preset threshold is the text considered valid.

[0063] For legitimate text that passes the detection and contains no illegal content, the legitimate text and its corresponding age-aware tag are assembled into a data packet and then uploaded to the cloud. The upload operation is only performed after the text is determined to be legitimate; illegitimate text is discarded without being uploaded. The upload to the cloud can be achieved by using a message queue protocol to send the data packet to the cloud's receiving interface, while simultaneously initiating a timeout retransmission mechanism to ensure data transmission reliability. Alternatively, the legitimate text can be losslessly compressed first, and then the compressed data is concatenated with the age-aware tag before being uploaded via a secure transmission protocol. This approach saves bandwidth while protecting data integrity.

[0064] It's important to note that by using age-aware tags to constrain text pre-screening rules, the strictness and focus of these rules automatically adapt to the expressive characteristics and security needs of users of different ages. This avoids the false positives or false negatives that can result from applying the same rigid set of rules to all age groups. Detecting text under review after obtaining age-adapted pre-screening rules significantly improves the accuracy and targeting of identifying inappropriate content. Uploading only valid text with age-aware tags to the cloud, while blocking and discarding inappropriate text, reduces the waste of cloud resources processing invalid data and lowers the risk of sensitive information leakage at the source, thus protecting user privacy.

[0065] S103: The cloud generates low-quality multimedia content with complete semantics based on legal text and low-quality media parameters, and detects low-quality multimedia content according to compliance review rules.

[0066] Among them, low-quality media parameters are a set of configuration parameters with low values ​​used to control the resolution, frame rate, bit rate and model iteration steps during the multimedia content generation process; low-quality multimedia content refers to multimedia data generated according to low-quality media parameters that is relatively rough in terms of image quality, smoothness or detail richness but still maintains semantic integrity; compliance review rules are pre-set detection conditions or classification standards used to determine whether multimedia content contains illegal information.

[0067] Specifically, after receiving legitimate text and age-aware tags uploaded by the smart TV terminal, the cloud first generates low-quality multimedia content based on the legitimate text. During the generation process, the cloud intentionally selects low-quality media parameters, such as setting the image resolution to 320 x 240 pixels, the frame rate to 8 frames per second, and compressing the video bitrate to 200 kilobits per second. It also uses a lightweight generation model and reduces its iteration steps. While the generated video or animation may appear visually rough, it can fully express the storyline, action sequences, and semantic information described in the legitimate text. One way to achieve this low-quality generation is for the cloud to call a lightweight Transformer decoder. This decoder contains only two layers of attention networks. After receiving legitimate text as input, it directly outputs a keyframe sequence, with each keyframe stored as a low-precision floating-point number, thus quickly obtaining a low-resolution animation with basic semantic coherence. Another approach is to use a rule-based template splicing method in the cloud to break down legal text into several semantic units. Each unit corresponds to a pre-stored low-poly model animation. These animation segments are then spliced ​​together in text order to form a complete low-quality multimedia clip. No deep neural network rendering is enabled throughout the process.

[0068] After generating low-quality multimedia content, the cloud immediately performs preliminary detection on this content according to compliance review rules. These rules can be a set of sensitive content detection logic for images and audio content, such as identifying whether violent or dangerous actions appear in the image, or detecting whether abusive language is present in the audio. One implementation for detecting low-quality multimedia content is to run a lightweight convolutional binary classification network in the cloud. This network takes each frame of the low-quality multimedia content down to 64x64 pixels as input and outputs a violation probability. If the probability of any frame exceeds 0.5, the content is deemed non-compliant. Another implementation is to use a fast comparison method based on feature hashing in the cloud. A violation content feature library is pre-built, and the perceptual hash values ​​of keyframes of the low-quality multimedia content are extracted and compared one by one with the hash values ​​in the library. If the similarity is found to be higher than a preset threshold, a rejection mechanism is triggered. By generating low-quality content first and then quickly detecting it, the cloud can filter out non-compliant content in advance with less computational resources, avoiding unnecessary enhancement processing of invalid content later.

[0069] It's understandable that generating semantically complete low-quality multimedia content using low-quality media parameters in the cloud can quickly produce a first draft with minimal computational cost and extremely short timeframes, avoiding the high latency and costs associated with directly generating high-quality content. Then, by checking this low-quality multimedia content against compliance review rules, violations can be detected and blocked early, preventing them from progressing to the subsequent high-quality enhancement stage. Performing low-quality generation and compliance checks sequentially significantly reduces the waste of cloud computing resources and ensures that only semantically correct and compliant content is retained, thereby improving the overall response efficiency and security of the multimedia generation system.

[0070] S104: Enhance the low-quality multimedia content that has passed the detection, simultaneously perform deep semantic review using age-aware tags, and output high-quality multimedia content to the smart TV terminal when the deep semantic review is passed.

[0071] Among them, enhancement refers to the process of improving the resolution, frame rate, color depth, or detail richness of multimedia content; deep semantic review refers to the process of deeply understanding the semantic level of multimedia content and judging whether there is any illegal or inappropriate information; high-quality multimedia content refers to multimedia data with high resolution, high smoothness, and rich detail after enhancement processing.

[0072] In practice, the cloud performs enhancement operations on low-quality multimedia content that passes detection, while simultaneously conducting deep semantic review using age-aware tags during this enhancement process. Specifically, the cloud feeds previously generated rough animations or low-resolution videos into a super-resolution and detail synthesis network. This network improves image quality while analyzing character behavior, scene attributes, and dialogue meaning in each frame to ensure that the final output content is both clear and meets the viewing safety requirements of users in that age group. One approach is for the cloud to use a diffusion-based enhancement framework. This model receives age-aware tags as conditional input at each step of the reverse denoising process. Based on the child or elderly tags, the model actively suppresses the rendering of violent or terrifying elements. For example, when generating high-resolution images, it automatically blurs weapon shapes or removes gory details. At the same time, it uses an internal attention map to perform real-time review of semantic regions, immediately interrupting the diffusion process if any semantic violations are detected. Another approach is to use a combination of video frame interpolation and style transfer network in the cloud. This network first enlarges each frame of low-quality content by 4 times, and then selects the corresponding style filter based on age-aware tags. For example, it retains the realistic style for the youth tag and adds a cartoonish filter for the children tag. At the same time, a review node based on a visual language model is embedded in the enhancement pipeline. This node extracts high-level semantic features frame by frame and compares them with a violation concept library. The comparison process runs in parallel with the enhancement process without adding extra waiting time.

[0073] When enhancement and deep semantic review are performed simultaneously, the cloud also needs to ensure that the review results are fed back to the enhancement process in real time. A dual-path parallel mechanism can be designed, with one path performing pixel-level enhancement and the other performing semantic-level understanding. The two paths exchange data after each frame is processed. One implementation is to call a high-performance graphics processing unit cluster in the cloud to run two neural networks simultaneously. The enhancement network is responsible for upsampling and detail generation, while the review network is responsible for object detection and scene graph construction. The compliance score output by the review network directly determines whether the current frame of the enhancement network is retained. If the compliance score of a frame is lower than the threshold, the entire generation task is terminated immediately. Another implementation is to use an end-to-end joint model. This model uses age-aware labels as auxiliary supervision signals and adds a review branch to the intermediate layer of the decoder. This branch outputs a binary classification result. Only when the review results of all intermediate layers are passed does the model output the final high-quality frame. Otherwise, it outputs a preset replacement image and prompts the smart TV terminal that the review has failed.

[0074] Once the deep semantic review is approved, the cloud packages the enhanced, complete, high-quality multimedia content and pushes it to the smart TV terminal via network protocols. The smart TV terminal receives the content and decodes and plays it directly, providing the user with high-quality visuals that match their voice commands and are age-appropriate. One push method involves the cloud using an adaptive streaming media protocol to transmit high-quality content in chunks, allowing the smart TV terminal to receive and buffer the content simultaneously, reducing waiting time. Another method involves the cloud compressing the entire video file at once and sending it via a reliable transmission protocol. The smart TV terminal receives the file, verifies its integrity, and then plays it locally. This method is suitable for shorter content or scenarios with stable network conditions.

[0075] It's important to note that the cloud-based enhancement of low-quality multimedia content that passes initial testing transforms rough drafts into smooth, high-quality visuals, significantly improving the viewing experience. Simultaneously, age-aware tags are used for deep semantic review, ensuring the review process is no longer independent of the generation process. This allows for dynamic adjustments to review standards and violation criteria based on user age groups during enhancement, preventing content that, while visually appealing, is unsuitable for specific age groups or contains hidden violations. High-quality multimedia content is only output to smart TV terminals after passing deep semantic review, guaranteeing that the final content presented on the terminal possesses excellent visual effects and is fully compliant at the semantic level. This avoids the wasted computing power and increased latency caused by enhancing everything before subsequent review, achieving a balance between quality, security, and efficiency.

[0076] In the above embodiments, by performing voiceprint recognition on voice commands on the smart TV terminal to obtain age-aware tags and converting the voice commands into text to be reviewed, the terminal-side computing resources can be used to complete the perception and preprocessing of user age-related information, reducing the burden of subsequent cloud processing. By using age-aware tags to constrain text pre-review rules and detecting the text to be reviewed, only legal text that passes the detection and age-aware tags are uploaded to the cloud. This allows for the early filtering of tasks that do not meet the requirements on the terminal side, reducing the consumption of network transmission resources and cloud computing resources by invalid tasks. Based on the legal text, the cloud uses low-quality media parameters to generate low-quality multimedia content with complete semantics and performs compliance review. This can complete the content validity verification at a lower computing cost and avoid directly generating high-quality content for non-compliant content. The low-quality multimedia content that passes the review is enhanced and combined with age-aware tags for deep semantic review. This can generate high-quality multimedia content while ensuring content compliance and age adaptability, and output the high-quality multimedia content to the smart TV terminal for presentation, thereby improving the user viewing experience. Thus, by collaborating between smart TV terminals and the cloud to complete content pre-screening, low-cost verification, and high-quality generation, invalid tasks are filtered out step by step, unnecessary computing resources are consumed in the process of multimedia content generation, and the overall resource utilization efficiency and content generation efficiency are improved.

[0077] In one embodiment, age-aware labels are used to constrain text pre-screening rules to obtain age-appropriate pre-screening rules, including:

[0078] Determine the corresponding age category based on the age perception label;

[0079] Obtain the age constraint rules corresponding to the age category, and obtain the text pre-review rules;

[0080] By using age constraint rules, the sensitive content filtering conditions, semantic redirection conditions, and content generation constraints in the text pre-review rules are adjusted to obtain age-adaptive pre-review rules.

[0081] Among them, age category is a specific age grouping based on age-aware tags; age constraint rules are a set of pre-set restrictions on content safety and generation direction for a specific age category; text pre-review rules are a set of filtering logic used to determine whether the text to be reviewed contains illegal content; sensitive content filtering conditions are used to identify and block inappropriate words in the text; semantic redirection conditions are used to convert the original intention of the text into a more appropriate expression direction; and content generation constraint conditions are used to add additional security or style restrictions to the generation process.

[0082] After obtaining the age-perceived labels, the first step is to determine the speaker's age category. One approach is to directly output the age group with the highest confidence among the labels as the category; for example, if the probability is 0.82 for children, 0.15 for teenagers, and 0.03 for adults, then the children category is selected. Another approach is to make a smooth decision based on historical interaction records. If the previous three interactions were all with teenagers, even if the probability of teenagers is slightly lower this time, the teenager category is maintained to avoid frequent jumps.

[0083] After determining the age category, obtain the age constraint rules corresponding to that category, as well as the basic text pre-screening rules. Age constraint rules define the safety boundaries specific to that age group; for example, horror descriptions are prohibited in the children's category. These rules can be obtained by reading from a local memory rule table or by pulling the latest version from the operator's backend via the network.

[0084] Next, age-based constraints are used to adjust the three types of conditions in the text pre-screening rules. For sensitive content filtering conditions, the sensitive word library is expanded based on age category, and for the children's category, words unsuitable for children are added to the general word library. One approach is to take the union of the two word libraries to generate a new sensitive word set; another is to use rule overriding, with age-based constraints having higher priority, and in case of conflict, the age-based constraints take precedence. For semantic redirection conditions, the mapping target of certain intentions in the text is changed. For example, the original text "draw a monster" is mapped to "draw a cute little monster" for the children's category. One approach is to directly replace it using a redirection mapping table; another is to call a lightweight semantic rewriting model, inputting the original text and age category, and outputting the redirected text. For content generation constraints, additional constraint prefixes are added based on age category. For example, for the children's category, "cartoon style, no horror elements" is forcibly added. This can be done through string concatenation or template filling, filling the style, no-horror, etc. slots with content corresponding to the age group. After completing the above three adjustments, the age-adaptive pre-screening rules are obtained, which cover sensitive filtering, semantic redirection, and generation constraints.

[0085] It's understandable that determining the age category first and then obtaining the corresponding rules ensures that adjustments are precisely matched to the user's age group, avoiding a one-size-fits-all approach that might result in insufficient protection for children or excessive restrictions for adults. By adjusting sensitive filtering, semantic redirection, and generation constraints separately, the pre-review rules undergo multi-dimensional refinement, with these three elements working together to ensure the rules truly adapt to age. The resulting age-appropriate pre-review rules can handle instructions from users of different ages more safely and appropriately, reducing the probability of children accessing inappropriate content while preserving reasonable expression space for adults.

[0086] In one embodiment, detecting the text to be reviewed includes:

[0087] Based on age-appropriate pre-review rules, the text to be reviewed is detected for illegal content.

[0088] Based on age-appropriate pre-review rules, age-appropriateness checks are performed on the texts to be reviewed.

[0089] When the text to be reviewed passes the detection based on the results of the violation detection and the age suitability detection, the text to be reviewed will be determined as a legal text.

[0090] After obtaining the age-appropriate pre-screening rules, the text to be reviewed is first detected for violations according to the sensitive content filtering conditions in these rules. The smart TV terminal runs a lightweight text classification model locally using the neural network processing unit of the system chip, scanning every word and semantic combination in the text to be reviewed. One approach is to use a fast matching method based on a keyword blacklist, comparing the text to be reviewed one by one with a pre-set sensitive word database; if any word is matched, it is considered a violation. Another approach is to run a compressed DistilBERT classification network, which directly outputs a binary probability of whether the text is a violation or not; a probability exceeding 0.8 indicates a violation.

[0091] Simultaneously or subsequently, based on the semantic redirection and content generation constraints within the same age-appropriate pre-review rule, age-appropriateness checks are performed on the text to be reviewed. Age-appropriateness checks focus on whether the text content matches the cognitive maturity of users in that age group. For example, for children's categories, the check detects whether the text contains abstract, complex adult emotions or excessively frightening imagery. One implementation method is to use a rule-based template matching approach, predefining a list of inappropriate topics for each age group; if these words appear in the text to be reviewed, it is deemed inappropriate. Another implementation method is to call a lightweight age-appropriateness regression model. This model takes the text to be reviewed and the age category as input and outputs an appropriateness score between 0 and 1; a score below 0.6 is considered inappropriate.

[0092] After obtaining the results of the violation detection and age suitability detection, a comprehensive judgment is made on whether the text to be reviewed passes the inspection. The text is considered to have passed only if both the violation detection result is "no violation found" and the age suitability detection result is "age-appropriate." One judgment logic uses a logical AND operation, outputting a pass flag only when both conditions are true. Another judgment logic uses a weighted scoring method, with violation detection weighted at 0.7 and age suitability detection weighted at 0.3. A pass is only granted if the total score is above the threshold and each violation detection item is true. Once the inspection passes, the text to be reviewed is determined to be legal and prepared for uploading to the cloud; if any detection fails, the text is discarded without further uploading, and a user-friendly prompt is displayed on the television screen.

[0093] Understandably, simultaneously conducting violation content detection and age-appropriateness detection not only blocks illegal and harmful information but also filters out content that, while not violating regulations, exceeds the cognitive level or psychological tolerance of the target age group, thus achieving dual security. Combining the results of both detections before determining the legality of the text avoids misjudgments or omissions caused by a single detection dimension. For example, a text that does not contain prohibited words but is overly adult in nature may still be blocked by age-appropriateness detection. Through this dual detection mechanism, the final legal text ensures both content compliance and matching the speaker's age characteristics, thereby providing safe and age-appropriate input for subsequent multimedia content generation at the source.

[0094] In one embodiment, generating semantically complete low-quality multimedia content based on legal text using low-quality media parameters includes:

[0095] Determine the content generation conditions based on the valid text;

[0096] Based on the preset low-quality generation strategy, determine the corresponding target media parameters;

[0097] While keeping the semantic information of the content generation conditions unchanged, configure low-quality media parameters according to the target media parameters.

[0098] By utilizing content generation conditions and low-quality media parameters, content generation is performed to obtain low-quality multimedia content with complete semantics.

[0099] Among them, the content generation conditions are the core semantic descriptions or instructions used to guide the generation of multimedia content; the preset low-quality generation strategy is a predefined parameter configuration scheme that produces auditable content at a lower computing cost; the target media parameters are the specific values ​​specified in the strategy, such as resolution, frame rate, bit rate, etc.; and the low-quality media parameters are the generation parameters actually configured according to the target media parameters.

[0100] In practice, after receiving legitimate text, the cloud first parses or extracts content generation conditions from the text. These conditions determine the subject, action, or plot that the multimedia will ultimately express. This can be done by directly using the legitimate text itself as the generation condition, without any additional processing. Alternatively, the legitimate text can be subjected to lightweight semantic compression, extracting the core triple consisting of subject, predicate, and object as the generation condition. For example, "draw a cute little monster wearing clothes" can be compressed into "the little monster is wearing clothes and is cute."

[0101] Next, based on the preset low-quality generation strategy, the target media parameters to be used in this generation are determined. The low-quality generation strategy predefines a set of parameter templates. For example, for image generation, the resolution is limited to 480 x 270 pixels and the color depth to 8 bits; for video generation, the resolution is limited to 854 x 480 pixels, the frame rate is 15 frames per second, and the audio uses mono with a sampling rate of 16 kHz. One implementation method is to read a fixed set of parameters from a local configuration file, with different parameter sets corresponding to different content types such as images, videos, and audio. Another implementation method is to dynamically calculate the target parameters, elastically adjusting the resolution and frame rate within a preset range according to the current load of the cloud-based lightweight graphics processor. When the load is high, the parameters are further reduced to ensure response speed.

[0102] After determining the target media parameters, these parameters are configured as low-quality media parameters while ensuring that the semantic information carried by the content generation conditions is not altered. In other words, although the resolution, frame rate, or audio quality is reduced, the generated image is still clear enough for reviewers or automated review models to discern the main actions and identify the scene composition. One approach is to modify only the output size and sampling steps of the generation model, preserving the original encoding of the input semantic text. Another approach is to skip some detail reconstruction layers during the generation process, but retain outline and structural information, thus maintaining semantic integrity while reducing quality.

[0103] Finally, the content generation process is executed using content generation conditions and low-quality media parameters. A lightweight diffusion model or Transformer decoder is invoked in the cloud to control the spatiotemporal resolution of the output using low-quality media parameters. One approach is to run a denoising diffusion model with fewer iterations; the standard model requires 50 iterations, while low-quality generation only requires 10, resulting in a lower-resolution image with discernible object outlines. Another approach is to use a keyframe sequence generation method. Instead of generating all frames of the complete video, a keyframe is generated every few frames, and then simple interpolation is used to fill in the intermediate frames. While the resulting low-quality video is not smooth, the narrative structure and character movements are still recognizable. The final generated content is semantically complete but has downgraded audiovisual technical indicators, becoming low-quality multimedia content, which can serve as a draft for subsequent compliance review and enhancement processing.

[0104] It's important to note that determining content generation conditions based on legal text ensures consistency between the generation direction and user instructions. Then, determining and configuring target media parameters according to a preset low-quality generation strategy significantly reduces the computational cost of a single generation while preserving the core semantics. Utilizing low-quality media parameters for generation yields a semantically complete but visually rough initial draft. This allows for the rapid production of content samples suitable for review with extremely low computational costs, avoiding the high latency and resource consumption associated with directly generating high-quality content. This creates an efficient foundation for subsequent compliance screening and on-demand enhancement.

[0105] In one embodiment, detecting low-quality multimedia content according to compliance audit rules includes:

[0106] Extract content features corresponding to low-quality multimedia content;

[0107] In accordance with compliance review rules, the content features are tested for illegal elements and cross-modal consistency.

[0108] The audit results are determined based on the results of the detection of non-compliant elements and the results of cross-modal consistency detection.

[0109] Among them, content features are a set of computable vectors or labels extracted from low-quality multimedia content to represent visual images or audio text information; violation element detection is the process of identifying whether the content contains prohibited elements; cross-modal consistency detection is the process of verifying whether the semantics between video images and audio text match; and the review result is the output identifier that ultimately determines whether the content is compliant, non-compliant, or has ambiguous boundaries.

[0110] In practice, after obtaining low-quality multimedia content, the first step is to extract content features suitable for review. These features can include the color histogram of an image, edge contours, object category probabilities, or the text sequence transcribed from the audio. One approach is to compress each frame of the low-quality video into a small 64x64 pixel image, input it into a lightweight convolutional neural network, and the network outputs a 1024-dimensional feature vector that condenses the visual semantic information in that frame. Another approach is to perform real-time speech recognition on the audio track of the low-quality video, using the recognized text along with the timestamp as audio features, while simultaneously extracting the perceptual hash values ​​of the video keyframes as visual feature indices.

[0111] Next, based on compliance review rules, the extracted content features are subjected to violation element detection and cross-modal consistency detection. Violation element detection focuses on whether sensitive objects or scenes appear in the image. For example, a binary classification network is used to determine whether each frame contains exposed skin areas or weapon shapes. One implementation is to use a lightweight detector based on MobileNet, which maintains a high recall rate even on low-resolution images. After inputting a frame, it outputs a confidence vector for the violation category. If the confidence of the violence category exceeds 0.7, it is determined to contain violent elements. Another implementation is to use feature comparison. A violation feature library is pre-built, and the cosine similarity of the feature vector of the frame to be detected with each feature in the library is calculated. When the highest similarity exceeds a preset threshold, a violation is determined. Cross-modal consistency detection verifies whether the lip movements and actions of people in the video match the audio text, and whether the content appearing in the image contradicts the audio description. One approach is to calculate the alignment score between the nouns in the audio text and the object labels detected in the visual features. For example, if the audio says "cat" but there is no cat-like object in the image, an inconsistency is determined. Another approach is to use a pre-trained multimodal contrastive model to project the features of video frames and audio text into the same semantic space, calculate the cosine distance between them, and if the distance is too large, it indicates cross-modal inconsistency.

[0112] After completing the above two checks, the final review result is determined based on the results of the violation element detection and cross-modal consistency detection. If the violation element detection finds any explicit violation element, or the cross-modal consistency detection finds a serious inconsistency that points to malicious deception, the review result is "failed." If neither detection finds any issues, the review result is "passed." If the violation detection is in a state of ambiguity (e.g., confidence level between 0.4 and 0.6), or the cross-modal consistency detection score is in the critical zone, the review result is "ambiguous boundary." One decision logic uses a decision tree: first, determine if there is an explicit violation; if so, output "failed" directly; otherwise, determine if the boundary is ambiguous; if so, output "ambiguous boundary"; otherwise, output "passed." Another decision logic uses a weighted voting method: the two detection results are scored separately and then added together; if the total score is below a threshold, it fails; if it is above the threshold, it passes; the area in between is considered "ambiguous boundary." The review result will determine the subsequent process: failure terminates the task and destroys the low-quality multimedia content; ambiguity triggers iterative regeneration; and passing enters the enhancement phase.

[0113] It's understandable that extracting content features from low-quality multimedia content and then performing detection based on compliance review rules allows the review process to move away from relying on raw high-resolution data, enabling rapid screening under low computing power. Simultaneously, detecting illegal elements and cross-modal consistency of content features not only intercepts sensitive content directly appearing in the video but also prevents cross-modal attacks where the video is compliant but the audio is illegal, forming a double line of defense. The review result is determined by combining the results of these two tests, avoiding misjudgments based on a single dimension. It can distinguish between clear violations, ambiguous boundaries, and complete compliance, providing accurate decision-making basis for subsequent iterative regeneration or high-quality enhancement. This allows for early, low-cost interception of illegal content, avoiding the waste of high computing resources.

[0114] In one embodiment, low-quality multimedia content that passes detection is enhanced, and deep semantic review is performed simultaneously using age-aware tags, including:

[0115] Based on the target output quality, determine the enhancement configuration parameters, and perform quality enhancement processing on the low-quality multimedia content that has passed the detection based on the enhancement configuration parameters;

[0116] During the quality enhancement process, intermediate enhancement results are obtained;

[0117] Based on age-aware tags, determine the deep review rules for the corresponding age categories, and use the deep review rules to perform deep semantic review on the enhanced intermediate results;

[0118] When the deep semantic review results indicate that there is content that does not meet the requirements of the corresponding age category, an enhancement termination instruction is generated, and the quality enhancement process is interrupted using the enhancement termination instruction.

[0119] Among them, the target output quality is the desired technical level of the final multimedia product, such as high resolution or high frame rate; the enhancement configuration parameters are a set of specific values ​​used to control enhancement operations such as super-resolution, frame interpolation, and audio restoration; the enhancement intermediate results are some picture frames or audio segments that have been generated but not yet completed during the quality improvement process; the deep review rules are a set of fine-grained conditions for judging whether the content is appropriate at a deep semantic level for a specific age category; and the enhancement termination command is a control signal used to immediately stop the current enhancement task and release resources.

[0120] In practice, after determining the target output quality, the enhancement parameters are first configured accordingly. For example, if the desired output is 4K resolution video at 60 frames per second, the enhancement parameters would be set to a super-resolution multiplier of 4x, a frame interpolation multiplier of 2x, and an audio sampling rate increased to 48 kHz. One approach is to read parameters from a preset quality level table; for example, HD and UHD correspond to different resolutions and bitrates. Another approach is dynamic calculation, automatically determining the super-resolution multiplier based on the ratio between the original resolution of the low-quality content and the target resolution, while simultaneously adjusting the batch size by referencing the available video memory of the current high-performance cloud graphics processor cluster.

[0121] Next, based on these enhancement configuration parameters, quality enhancement processing is performed on the low-quality multimedia content that has passed the detection. This process feeds the low-resolution draft into a super-resolution and frame interpolation network, gradually generating high-resolution, high-frame-rate images. While the enhancement process is running, intermediate enhancement results are acquired periodically or continuously. These intermediate results can be feature maps of a specific layer of the super-resolution network, or they can be partially generated high-resolution frames. One implementation method is to extract data in time slices, reading the pixel data of 5 frames as intermediate results after each 5-frame rendering is completed. Another implementation method is to monitor the output buffer of the enhancement network, immediately copying it as an intermediate result whenever the buffer is full, without waiting for the entire enhancement process to complete.

[0122] During the acquisition of enhanced intermediate results, in-depth review rules for the corresponding age category are simultaneously determined based on the age-aware label. If the age-aware label indicates "children," the in-depth review rules will pay additional attention to whether there are potential terrifying images or subtle inappropriate hints in the image, such as whether a particular color scheme or a character's expression is too ferocious. If the label indicates "teenagers," the rules will focus more on identifying subtle sexual innuendo or violent aesthetics. One approach is to pre-build a deep review rule library for each age category and directly index the corresponding rule set based on the label. Another approach is to use a rule adapter, taking the age-aware label as a conditional input to a lightweight classification network, which outputs a list of review concerns specific to that age group.

[0123] Then, these deep review rules are used to perform deep semantic review on the enhanced intermediate results. The review process goes beyond simple object detection, aiming to understand the behavior of people in the scene, the atmosphere of the scene, and the underlying meaning. One approach is to feed the intermediate frames into a visual language model, which outputs a semantic description of the frame and determines whether there are elements that do not conform to the age category. For example, under the children's rule, if it finds "a child is crying and there are shadows around him," it is determined to be inappropriate for the age. Another approach is to use attention heatmap analysis to extract the self-attention weights in the augmentation network, check whether the model is over-focusing on the violation area, and trigger a violation alarm if it is found that the model is trying to reinforce a sensitive area.

[0124] Once the deep semantic review result finds that the current enhanced intermediate result contains content that does not meet the age category requirements, an enhancement termination instruction is immediately generated. This instruction is sent to the enhancement processing thread to forcibly interrupt the remaining calculations. One approach is to set a global interrupt flag, and the enhancement thread checks the flag before processing each frame. If the flag is true, it directly breaks out of the loop and releases the video memory. Another approach is to directly send a termination signal to the operating system to kill the enhancement process and reclaim all intermediate caches. After the interruption, all partially generated high-resolution data is destroyed, and a prompt indicating that the content cannot be generated is returned to the smart TV terminal, without continuing any enhancement operations. Only when the deep semantic review continues to pass until all enhancement processing is completed, the final high-quality multimedia content will be output.

[0125] It should be noted that determining the enhancement configuration parameters according to the target output quality and executing the enhancement processing ensures that the final content can reach the expected technical level. During the enhancement process, intermediate results are obtained in real time and deep semantic review based on age-aware tags is performed synchronously, so that the review no longer lags behind generation, and non-compliant elements can be identified as soon as they show signs, rather than waiting until the complete 4K video is rendered for review. Once content that does not meet the age category requirements is found, a termination instruction is immediately generated to interrupt the enhancement processing, which avoids the huge waste of computing power caused by discarding all calculations after they are completed on a high-performance graphics processor, and also prevents non-compliant content from being fully generated. This linkage mechanism of simultaneous enhancement, review, and interruption moves the security control forward to the interior of the enhancement pipeline, achieving efficient unification of quality improvement and content security.

[0126] In one embodiment, performing deep semantic review on the enhanced intermediate result using deep review rules includes:

[0127] Performing implicit content semantic analysis and sensitive content detection on the enhanced intermediate result;

[0128] Performing age suitability assessment on the enhanced intermediate result according to the age category;

[0129] Generating a deep semantic review result according to the implicit content semantic analysis result, the sensitive content detection result, and the age suitability assessment result.

[0130] Wherein, the implicit content semantic analysis refers to the process of understanding and analyzing the potential cultural meaning, value orientation and indirectly expressed intention in the enhanced intermediate result; the sensitive content detection is the process of identifying whether the enhanced intermediate result contains explicit non-compliant elements; the age suitability assessment is the process of judging whether the content in the enhanced intermediate result is suitable for the cognitive and psychological bearing capacity of users of the corresponding age group; the deep semantic review result is a conclusion output after integrating the above analyses, which is used to characterize whether the content has passed the review.

[0131] Specifically, after obtaining the enhanced intermediate results, the first step is to perform implicit semantic analysis and sensitive content detection on the incomplete visuals or audio. Implicit semantic analysis focuses on meanings not directly depicted in the visuals but which can be inferred; for example, a character smiling and holding a gift might imply friendliness or surprise. Sensitive content detection directly scans the visuals for weapons, bloodstains, or inappropriately exposed areas. One implementation involves feeding several consecutive frames from the enhanced intermediate results into a multimodal large model. This model outputs a natural language description and an additional implicit intent label; for example, "This clip implies the character is about to encounter danger" is considered a negative implicit semantic. Another implementation uses two parallel networks: one for sensitive content detection and the other for semantic embedding extraction. The extracted embedding vectors are then matched against a pre-built implicit semantic database. If a match is found for patterns such as "deception" or "discrimination," it is marked as an implicit risk.

[0132] Simultaneously, based on the previously determined age categories, an age-appropriateness assessment is performed on the same batch of enhanced intermediate results. This assessment does not simply determine whether the content violates regulations, but rather considers whether the content is too complex for a specific age group or may cause psychological discomfort. For example, a story about "a broken and reconciled friendship" may contain too many negative emotions for children, while it is a normal theme for adolescents' growth. One approach is to construct an age-appropriateness regression model, taking the feature vector of the frame sequence and the age category as input, and outputting an appropriateness score from 0 to 1; a score below 0.4 is considered age-inappropriate. Another approach is to use a rule-based expert system, with a rule base storing scene types that are inappropriate for each age group. For example, abstract emotional expressions such as "lonely back view" or "despair expression" are inappropriate for children. These scene patterns are matched frame by frame, and once matched, the content is considered age-inappropriate.

[0133] After obtaining the results of implicit semantic analysis, sensitive content detection, and age suitability assessment, these three results are combined to generate the final deep semantic review result. The review result can be "pass," "fail," or "requires manual review." One decision-making method is to use priority judgment: if sensitive content detection finds any explicit violations, it is directly output as "fail"; otherwise, if the age suitability assessment is inappropriate or implicit semantic analysis finds seriously negative values, it is also output as "fail"; only when all three are good is it output as "pass." Another decision-making method is weighted scoring, quantifying the three results into scores and then summing them by weight, with implicit semantic analysis weighted at 0.3, sensitive content detection weighted at 0.5, and age suitability weighted at 0.2. If the total score is below the threshold, it is "fail." When the score is in the critical zone, it outputs "requires manual review." The generated deep semantic review result will be used in subsequent processes. If it fails, an enhancement termination instruction is immediately triggered, interrupting the quality enhancement process; if it passes, enhancement can continue and the final content can be output.

[0134] It's important to note that simultaneously conducting implicit semantic analysis and sensitive content detection allows the review process to not only intercept explicit violations but also uncover potentially harmful values ​​or misleading intentions conveyed in the visuals, filling the blind spots of traditional surface-level detection. Combined with age-based suitability assessments, this further ensures that the content is not only compliant but also genuinely appropriate for the cognitive level and psychological tolerance of users in that age group, avoiding confusion or fear caused by overly abstract or emotionally complex content. Finally, the combined results of these three analyses generate a deep semantic review outcome, achieving multi-layered security oversight from surface to deep, from general to personalized, significantly reducing the false negative rate and providing an accurate and reliable basis for subsequent enhanced termination decisions.

[0135] The multimedia content generation apparatus provided in the embodiments of this application is described below. The multimedia content generation apparatus described below can be referred to in correspondence with the multimedia content generation method described above. For example... Figure 2 As shown, this application provides a multimedia content generation device for use in a television system, which includes a smart TV terminal and a cloud platform. The device includes:

[0136] The text to be reviewed module 201 is used by the smart TV terminal to respond to voice commands, perform voiceprint recognition on the voice commands, obtain age perception tags, and convert the voice commands into text to be reviewed.

[0137] The text pre-review module 202 is used to constrain the text pre-review rules with age-aware tags. After obtaining the age-adaptive pre-review rules, it detects the text to be reviewed and uploads the legal text and age-aware tags that pass the detection to the cloud.

[0138] The low-quality multimedia content generation module 203 is used to generate low-quality multimedia content with complete content semantics based on legal text and low-quality media parameters in the cloud, and to detect low-quality multimedia content according to compliance review rules.

[0139] The high-quality multimedia content generation module 204 is used to enhance the low-quality multimedia content that has passed the detection, simultaneously perform deep semantic review using age-aware tags, and output high-quality multimedia content to the smart TV terminal when the deep semantic review is passed.

[0140] In one embodiment, the text pre-screening module 202 includes:

[0141] The age category determination unit is used to determine the corresponding age category based on the age-perceived label.

[0142] The age constraint rule acquisition unit is used to acquire age constraint rules corresponding to age categories, as well as text pre-review rules;

[0143] The age-appropriate pre-screening rule determination unit is used to adjust the sensitive content filtering conditions, semantic redirection conditions, and content generation constraints in the text pre-screening rules using age constraint rules, so as to obtain age-appropriate pre-screening rules.

[0144] In one embodiment, the text pre-screening module 202 includes:

[0145] The violation detection unit is used to detect violation content in the text to be reviewed based on age-appropriate pre-review rules.

[0146] The age suitability detection unit is used to perform age suitability detection on the text to be reviewed based on the age suitability pre-review rules.

[0147] The legitimate text determination unit is used to determine the text to be reviewed as legitimate text when the text to be reviewed passes the detection based on the results of the violation content detection and the age suitability detection.

[0148] In one embodiment, the low-quality multimedia content generation module 203 includes:

[0149] The content generation condition determination unit is used to determine the content generation conditions based on the legal text.

[0150] The target media parameter determination unit is used to determine the corresponding target media parameters according to the preset low-quality generation strategy.

[0151] The low-quality media parameter configuration unit is used to configure low-quality media parameters according to the target media parameters while keeping the content semantic information corresponding to the content generation conditions unchanged.

[0152] The low-quality multimedia content determination unit is used to perform content generation using content generation conditions and low-quality media parameters to obtain low-quality multimedia content with complete content semantics.

[0153] In one embodiment, the low-quality multimedia content generation module 203 includes:

[0154] The content feature extraction unit is used to extract content features corresponding to low-quality multimedia content.

[0155] The compliance review unit is used to detect illegal elements and perform cross-modal consistency checks on content features according to compliance review rules.

[0156] The audit result determination unit is used to determine the audit result based on the results of the detection of non-compliant elements and the results of cross-modal consistency detection.

[0157] In one embodiment, the high-quality multimedia content generation module 204 includes:

[0158] The enhancement configuration parameter determination unit is used to determine the enhancement configuration parameters according to the target output quality, and perform quality enhancement processing on the low-quality multimedia content that has passed the detection based on the enhancement configuration parameters;

[0159] An enhanced intermediate result acquisition unit is used to acquire enhanced intermediate results during the quality enhancement process.

[0160] The deep semantic review unit is used to determine the deep review rules for the corresponding age category based on the age-aware tags, and to use the deep review rules to perform deep semantic review on the enhanced intermediate results.

[0161] The enhanced termination instruction generation unit is used to generate an enhanced termination instruction when the deep semantic review result representation contains content that does not meet the requirements of the corresponding age category, and to interrupt the quality enhancement process using the enhanced termination instruction.

[0162] In one embodiment, the deep semantic review unit includes:

[0163] The content implicit semantic analysis subunit is used to perform content implicit semantic analysis and sensitive content detection on the enhanced intermediate results;

[0164] The age suitability assessment subunit is used to assess the age suitability of enhanced intermediate results based on age categories;

[0165] The deep semantic review result generation subunit is used to generate deep semantic review results based on the results of implicit semantic analysis of content, the results of sensitive content detection, and the results of age suitability assessment.

[0166] In one embodiment, this application also provides a storage medium storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the multimedia content generation method as described in any of the above embodiments.

[0167] In one embodiment, this application also provides a computer device storing computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the multimedia content generation method as described in any of the above embodiments.

[0168] Indicatively, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a computer device 300 provided in an embodiment of this application. The computer device 300 can be provided as a server. (Refer to...) Figure 3 The computer device 300 includes a processing component 302, which further includes one or more processors, and memory resources represented by memory 301 for storing instructions, such as application programs, that can be executed by the processing component 302. The application programs stored in memory 301 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 302 is configured to execute instructions to perform the multimedia content generation method of any of the above embodiments.

[0169] The computer device 300 may also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate on an operating system stored in memory 301, such as Windows Server™, Mac OS X™, Unix™, Linux™, Free BSD™, or similar.

[0170] Those skilled in the art will understand that Figure 3 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.

[0171] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the related listed items.

[0172] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0173] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating multimedia content, characterized in that, Applied to a television system, the television system including a smart TV terminal and a cloud, the method includes: The smart TV terminal responds to voice commands, performs voiceprint recognition on the voice commands to obtain age perception tags, and converts the voice commands into text to be reviewed; After obtaining age-appropriate pre-review rules by constraining the text with the age-aware tags, the text to be reviewed is detected, and the legal text that passes the detection and the age-aware tags are uploaded to the cloud. The cloud platform generates low-quality multimedia content with complete semantics based on the legal text using low-quality media parameters, and detects the low-quality multimedia content according to compliance review rules. The low-quality multimedia content that passes the detection is enhanced, and deep semantic review is performed simultaneously using the age-aware tag. When the deep semantic review is passed, high-quality multimedia content is output to the smart TV terminal.

2. The method according to claim 1, characterized in that, The process of using the age-aware tags to constrain text pre-screening rules to obtain age-appropriate pre-screening rules includes: Based on the age-perceived tags, determine the corresponding age category; Obtain the age constraint rules corresponding to the age category, and obtain the text pre-review rules; By using the age constraint rules, the sensitive content filtering conditions, semantic redirection conditions, and content generation constraints in the text pre-review rules are adjusted to obtain age-adaptive pre-review rules.

3. The method according to claim 1, characterized in that, The detection of the text to be reviewed includes: Based on the age-appropriate pre-screening rules, the text to be reviewed is subjected to violation content detection; Based on the age-appropriate pre-review rules, the text to be reviewed is subjected to age-appropriateness detection. When the text to be reviewed passes the detection based on the results of the violation detection and the age suitability detection, the text to be reviewed is determined to be legal text.

4. The method according to claim 1, characterized in that, The step of generating semantically complete low-quality multimedia content based on the legal text using low-quality media parameters includes: Based on the legal text, determine the content generation conditions; Based on the preset low-quality generation strategy, determine the corresponding target media parameters; While keeping the content semantic information corresponding to the content generation conditions unchanged, low-quality media parameters are configured according to the target media parameters. By utilizing the aforementioned content generation conditions and the aforementioned low-quality media parameters, content generation is performed to obtain low-quality multimedia content with complete content semantics.

5. The method according to claim 1, characterized in that, The detection of low-quality multimedia content according to compliance review rules includes: Extract the content features corresponding to the low-quality multimedia content; According to the compliance review rules, the content features are subjected to violation element detection and cross-modal consistency detection; The audit results are determined based on the results of the detection of non-compliant elements and the results of cross-modal consistency detection.

6. The method according to claim 1, characterized in that, The enhancement of low-quality multimedia content that passes detection, and the simultaneous use of the age-aware tags for deep semantic review, include: Based on the target output quality, the enhancement configuration parameters are determined, and based on the enhancement configuration parameters, the low-quality multimedia content that has passed the detection is subjected to quality enhancement processing. During the quality enhancement process, intermediate enhancement results are obtained; Based on the age-aware tags, determine the deep review rules for the corresponding age categories, and use the deep review rules to perform deep semantic review on the enhanced intermediate results; When the deep semantic review results indicate that there is content that does not meet the requirements of the corresponding age category, an enhancement termination instruction is generated, and the quality enhancement process is interrupted by the enhancement termination instruction.

7. The method according to claim 6, characterized in that, The step of performing deep semantic auditing on the enhanced intermediate results using the deep auditing rules includes: The enhanced intermediate results are subjected to implicit semantic analysis and sensitive content detection. An age appropriateness assessment is performed on the enhanced intermediate results based on the age category; The in-depth semantic review results are generated based on the results of implicit semantic analysis, sensitive content detection, and age suitability assessment.

8. A multimedia content generation device, characterized in that, Applied to a television system, the television system including a smart TV terminal and a cloud, the device includes: The text to be reviewed module is used by the smart TV terminal to respond to a voice command, perform voiceprint recognition on the voice command, obtain an age perception tag, and convert the voice command into text to be reviewed; The text pre-review module is used to constrain the text pre-review rules with the age-aware tag, obtain the age-adaptive pre-review rules, detect the text to be reviewed, and upload the legal text that passes the detection and the age-aware tag to the cloud. The low-quality multimedia content generation module is used to generate semantically complete low-quality multimedia content in the cloud based on the legal text using low-quality media parameters, and to detect the low-quality multimedia content according to compliance review rules. A high-quality multimedia content generation module is used to enhance low-quality multimedia content that has passed the detection, simultaneously perform deep semantic review using the age-aware tags, and output high-quality multimedia content to the smart TV terminal when the deep semantic review is passed.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions that, when executed by one or more processors, cause the one or more processors to perform the steps of the multimedia content generation method as described in any one of claims 1 to 7.

10. A computer device, characterized in that, include: One or more processors, and memory; The memory stores computer-readable instructions, which, when executed by the one or more processors, perform the steps of the multimedia content generation method as described in any one of claims 1 to 7.

Images