323 results about "Scene labeling" patented technology

The invention provides an AI-based marketing video auditing method, and relates to the technical field of AI marketing video auditing, and the method comprises the steps: obtaining a multi-modal data original structure set, and extracting image semantic features, voice expression features, text semantic features and scene label information, and obtaining an image semantic feature set, a visual rhythm feature set, a voice expression feature set, a voice and picture synchronous association vector structure, a text semantic feature set and a subtitle semantic and image main body linkage relation graph. By constructing an image semantic feature set, a voice expression feature set, a text semantic feature set and a visual rhythm feature set and fusing the image semantic feature set, the voice expression feature set, the text semantic feature set and the visual rhythm feature set into a multi-modal content fusion feature tensor, unified modeling of an AI marketing video at visual, auditory and semantic levels can be realized; and subsequent microscopic consistency detection, compliance knowledge graph and emotion semantic conflict identification are effectively performed, so that full-link risk perception and accurate auditing of video contents are realized.

Provided is a method for training a scene recognition model. The scene recognition model includes a core feature extraction layer, a global information feature extraction layer, an LCS module of at least one level with an attention mechanism, and a fully-connected decision layer. The method includes: acquiring parameters of the core feature extraction layer and the global information feature extraction layer by training based on a first scene label of a sample image and a standard cross-entropy loss; training a weight parameter of the LCS module of each level, based on a loss value acquired by performing a pixel-by-pixel calculation on a feature map output from the LCS module of each level and the first scene label of the sample image; and acquiring a parameter of the fully-connected decision layer by training based on the first scene label of the sample image and the standard cross-entropy loss.

The invention discloses a smart park digital twinborn management platform and a virtual-real interaction method, and belongs to the technical field of park management. Structural states, equipment behaviors, environment disturbance and user activity data of entity objects are hierarchically collected according to frequency, and a plurality of digital twinborn models are synchronously constructed based on a scene labeling mechanism; the platform can generate a control strategy and mark a prediction channel according to the model evolution trajectory, then acquire entity response and perform graph structure comparison with a prediction path, identify response deviation and generate a deviation graph; the platform further constructs an error learning weight according to the deviation map, drives a twin model to perform progressive optimization adjustment of confidence parameters, and can trigger structure-level network evolution to realize dynamic enhancement of a model structure and prediction capability; compared with the prior art, the invention provides a smart park digital twinning implementation scheme supporting virtual and real closed loops, self-learning evolution and structure reconstruction.

The invention relates to the technical field of text classification, and provides a text classification method based on a large model and a rule engine, and the method comprises the steps: S1, storing multi-level rule classification labels, and constructing a classification rule template library; s2, receiving text data from various data sources, and preprocessing the text data; s3, performing rule matching on the text data based on the classification rule through a rule engine, and outputting a rule classification result; and S4, when any one of the following conditions is met, large language model classification is triggered: a, a classification rule is not matched; b, matching a classification rule, wherein the rule confidence is smaller than a rule confidence threshold; c, the text data length exceeds the preset text data length; d, matching a specific business scene label; outputting a model classification result; and S5, when the rule engine classification in the S3 and the large language model classification in the S4 are parallel, executing the strategy. The output reliability and the service continuity are guaranteed, and the method is suitable for scenes with high accuracy requirements such as financial compliance examination and the like.

The invention provides an intelligent power grid state monitoring method and system based on digital twinning, and the method comprises the steps: obtaining multi-source data in a power grid operation environment, and forming a scene feature vector and a corresponding operation scene label; selecting a basic twinborn model template according to the operation scene label, mapping the scene feature vector into a node parameter in the basic twinborn model template, and obtaining a corresponding parameter set and a digital twinborn model instance; according to the digital twinborn model instance, constructing a causal graph structure of the operation state between the devices so as to reflect a causal dependence path between current power grid state variables; and performing trend risk judgment according to the abnormal causal path set, and outputting structured early warning response information. According to the method, the problem of insufficient generalization ability caused by the static state of the traditional twin structure is effectively solved.

The invention discloses a power market large language model evaluation method and device based on dynamic scene perception, and the method comprises the steps: building a large model evaluation system covering the three dimensions of understanding, generation and safety according to a power market application scene; the method comprises the following steps: constructing a test question bank, evaluating a power market large language model according to various evaluation methods, endowing three dimensions with different weights based on a large model evaluation system in the evaluation process, constructing a test question-scene label fine tuning data set, and performing scene classification fine tuning on the model by using a fine tuning method. Evaluating scene classification performance of the fine-tuned model, and distributing corresponding initial evaluation dimension weights for different scenes; according to the requirements and differences of different scenes, the initial evaluation dimension weight is adjusted, the scene is judged to be matched with the corresponding evaluation dimension weight according to the input content, weighted summation is carried out from the three dimensions of understanding, generation and safety, and an evaluation result is obtained. According to the method, accurate evaluation of the performance and reliability of the large language model in the power field can be realized.

The invention belongs to the technical field of machine room safety monitoring, and discloses an image review method fusing semantic understanding and visual recognition, which comprises the following steps: calculating visual / semantic feature dynamic credibility in real time through an exponential weighted moving average algorithm in combination with environment interference and equipment state parameters; resNet50 is adopted to extract visual features in global and local branches, and a BERT model is adopted to encode semantic features; correcting the feature correlation degree according to the scene, calculating a dynamic weight, carrying out heterogeneous calibration through an attention mechanism, and calling priority rules such as'physical security features are higher than behavior features' for arbitration during conflicts; the method is advantaged in that low-credibility feature interference fusion is avoided, abnormity identification accuracy in a complex scene of a machine room is improved, multi-modal data cooperation demands are adapted, scene labels are marked based on time, work orders and historical data dimensions, an exclusive sub-model is constructed for a high-density scene through DBSCAN density clustering, and low-frequency scene parameters are migrated to a similar model.

The invention belongs to the technical field of data centers, discloses an energy efficiency evaluation method of an artificial intelligence data center, and solves the problem that a static evaluation method is disjointed with the real-time operation state of the data center in the prior art by constructing a dynamic feature library and an AI adaptive weight model. The dynamic feature library collects and associates service scene labels and time-space coupling data in real time, and provides a scene basis for an AI adaptive weight model, so that weight adjustment is converted from passive response index mutation in the prior art to active adaptive service scene change. Nonlinear influences of dynamic factors such as server load fluctuation and environment adjustment on energy efficiency are effectively captured, so that an energy efficiency evaluation result better fits the actual operation state of the data center, energy efficiency factor relevance changes caused by service switching of the AI data center are adapted, scene-based self-adaptive adjustment of weights is achieved through a deep learning model, and energy efficiency evaluation accuracy is improved. The limitation that the fixed weight distribution ignores the difference of different service scenes in the prior art is solved.

The invention relates to a man-machine interaction method and device, computer equipment, a storage medium and a program product. The method comprises the following steps: acquiring interaction content input by a user; scene recognition is carried out on the interaction content through a large language model, and a scene label of the interaction is determined; determining a target long-term memory partition corresponding to the scene tag from a plurality of long-term memory partitions preset for the user; inputting target long-term memory content in the target long-term memory subarea and short-term memory content stored in a short-term memory area preset for the user into an intelligent agent; the intelligent agent is used for generating reply content for the interaction content according to the target long-term memory content and the short-term memory content. By adopting the method, effective association and collaborative calling of different scene memories can be realized, and the dynamic requirement of a user on coherent interaction in multi-scene switching is met.

The invention relates to an identification method for matching scene behaviors by using multi-modal features, and belongs to the technical field of scene behavior matching. The method comprises the following steps: acquiring multi-modal scene behavior data, and carrying out noise self-adaptive purification processing on the multi-modal scene behavior data to obtain a preprocessed scene image, scene audio data and a scene label text; secondly, performing feature collaborative extraction on the preprocessed data to obtain visual features, audio features and text features, inputting the cooperatively extracted features into a scene behavior matching network, and performing scene feature fusion and behavior feature fusion respectively to obtain a scene feature vector and a behavior feature vector; and constructing a bipartite graph according to the scene feature vector and the behavior feature vector, calculating the semantic similarity between nodes of the bipartite graph, dynamically updating the edge weight of the bipartite graph according to the semantic similarity between the nodes, and normalizing the updated bipartite graph to obtain a scene behavior matching result. According to the method, the association degree of the scene and the behavior can be accurately quantified, and the accuracy of a matching result is greatly improved.

The invention discloses an ASR audio corpus generation method and device based on a multi-modal large model, and relates to the field of audio corpora. According to the method, a semantic vector and a condition vector are spliced into a joint vector, and first voice is generated; target noise is selected from a preset noise library according to the scene label, the target noise is superposed on the first voice to generate voice with noise, and adversarial noise is injected to generate second voice; performing noise labeling, text labeling, emotion labeling and speaker labeling on the second voice, and performing alignment to generate a multi-modal labeling file; and setting a word error rate threshold value and a semantic similarity threshold value according to the scene label, the noise type and the speaker information of the multi-modal labeling file, and screening a target corpus from the multi-modal labeling file according to the word error rate threshold value and the semantic similarity threshold value. By implementing the technical scheme provided by the invention, the audio corpus which is high in quality, meets specific requirements and is effectively screened can be generated.

An adaptive data management method for polymorphic addressing comprises the following steps: acquiring multi-scene data from the Internet of Things, preprocessing the multi-scene data, and constructing a standardized data set; and performing multi-dimensional classification on the multi-scene data based on the similarity among the key indexes in the standardized data set to obtain a classification table. And obtaining a polymorphic addressing scheme corresponding to each type of data in the classification table, and generating a polymorphic addressing mapping table in combination with the property parameters of each type of data. And storing the polymorphic addressing mapping table in a distributed storage end, carrying out addressing position distribution on each storage node in the distributed storage end through a network topology structure, and outputting a distributed storage strategy. And constructing a virtual node associated with each storage node, and when the distributed storage strategy is executed, splitting the polymorphic addressing mapping table so as to store the polymorphic addressing data with the first scene label to the virtual node. Thread blockage during reading and writing of the multi-scene data in the Internet of Things is avoided, and the data management efficiency is improved.

The invention relates to a real-time process monitoring method based on data analysis, and the method comprises the following steps: S1, building a three-dimensional component priority evaluation model based on a component operation scene type, a real-time resource occupancy rate and a data dependence degree, and dynamically matching an acquisition strategy; s2, a cleaning rule is adapted according to a data source, a feature extraction dimension is adjusted in combination with process dynamic features, and an improved time sequence decomposition algorithm is used for separating data trends, fluctuations and abnormal residual errors; s3, constructing an exclusive baseline sub-model by using an online learning algorithm according to a scene label, and establishing a scene switching mechanism; s4, in combination with component interaction anomaly features, an anomaly level is judged through a mixed detection model; s5, on the basis of exception processing and user feedback, constructing an error correction model optimization parameter; and S6, generating a report containing an abnormal propagation path, and triggering hierarchical collaborative response of the associated component. The invention aims to solve the problems of single acquisition dimension, no scene adaptability in preprocessing, incomplete abnormal detection and the like in the existing monitoring technology.

The invention relates to the related technical field of intelligent scene labeling, in particular to an intelligent driving scene collaborative labeling method and system based on multi-modal fusion, and the method comprises the steps: constructing a dynamic feature fusion model, carrying out the cross-source correlation analysis of a multi-modal scene data set, setting organization structure elements, and setting a multi-dimensional scene labeling model. And configuring a potential danger interaction path and a scene risk index, and generating an annotation instruction set to carry out collaborative annotation operation. The technical problems that under interference of severe weather such as sensor noise, rain and fog and strong light and abnormal traffic behavior scenes, the labeling accuracy is insufficient, and complex scenes and multi-modal data requirements are difficult to adapt are solved, labeling noise data are introduced into an adversarial sample configuration assembly, a multi-dimensional scene labeling model is constructed in combination with organizational structure elements, and the multi-dimensional scene labeling accuracy is improved. The technical effects of improving the training data reliability, adapting to complex environment labeling requirements, configuring potential risk interaction paths and scene risk indexes, carrying out risk-oriented priority labeling and improving the risk scene labeling precision are achieved.

The invention relates to an information processing method and device for man-machine practice and a medium, and belongs to the technical field of man-machine interaction, and the method comprises the following steps: obtaining interaction voice of a user for question reply in real time, and converting the interaction voice into text information with a timestamp; performing preprocessing and key semantic information extraction on the text information, and performing emotion recognition, mute recognition and speech speed recognition; evaluating the practice process in real time according to the quality inspection rule, and judging whether interaction needs to be interrupted or not according to an evaluation result; performing intention recognition based on the extracted key semantic information, mapping the recognized intention to a corresponding node of a knowledge graph, iteratively updating an associated question-chasing path according to the knowledge graph, generating a response question in combination with a scene label and a process rule, and sending the response question to a user; and obtaining the interactive voice of the user for the response question, and carrying out the next round of man-machine practice interaction. Compared with the prior art, the method has the advantages of being capable of achieving personalized practice and the like.

The invention relates to the crossing field of artificial intelligence technology and e-government, and discloses an e-government enterprise service intelligent response method based on AI cue word engineering, comprising: constructing and maintaining a government knowledge graph and a scenarized cue word template library; receiving and analyzing a service request input by an enterprise, and extracting an enterprise feature tag and a business intention; selecting a target cue word template based on the matching degree of the enterprise feature tag and the scene tag of the template in the scene cue word template library, and generating an adapted cue word template; retrieving associated knowledge data from the government affair knowledge graph according to the service request, and filling the adapted cue word template with the retrieved associated knowledge data to generate a structured cue word; inputting the structured cue word into an AI model to drive the AI model to generate a government affair service response; and collecting multi-dimensional evaluation feedback of government affair service response, and carrying out iterative optimization. The demand understanding precision is improved, and the risk of core information misjudgment is reduced.

The invention discloses a double-arm clothes folding robot control method and device, equipment and a storage medium. The method comprises the following steps: determining three-dimensional point cloud data of a target workbench according to a collected image through a visual perception assembly; determining a scene label graph and a desktop edge line set through a scene division assembly according to the three-dimensional point cloud data, determining a clothing mask of the to-be-processed clothing according to the scene label graph through a clothing classification assembly, and determining a clothing description vector according to the clothing mask; determining a low-risk action set and a protection area mask from the candidate clothes folding action set through a safety protection component according to the scene label graph, the desktop edge line set and the clothes mask; determining a target clothes folding action sequence through an action planning component according to the clothes description vector, the low-risk action set and the protection area mask; and folding the to-be-processed clothes through the robot control assembly according to the target clothes folding action sequence. According to the scheme, the clothes folding efficiency and reliability of the robot are improved.

The invention belongs to the technical field of Internet data security, and discloses a multi-dimensional Internet data security fusion processing system, which is characterized in that an acquisition module depends on a three-dimensional decision model, combines reinforcement learning dynamic allocation tasks, only acquires threat associated key data in a low-computing-power and high-threat scene, and starts full-amount lightweight acquisition in a high-computing-power and low-threat scene; the data preprocessing module adds scene labels for data, filters logic and unifies formats according to scene design, and eliminates data isomerism. The identification module builds a basic threat library and a scene adaptation layer, adapts a new scene through transfer learning, calculates a risk value in combination with a four-dimensional quantitative model such as propagation probability, generates an attack link map, and effectively eliminates a monitoring blind area caused by edge computing power limitation; the decision-making module constructs a five-dimensional trust evaluation model of cooperation duration, compliance records and the like, calculates weights by using an analytic hierarchy process, establishes a compliance and trust linkage engine, synchronizes industry specifications in real time and adapts to scene adjustment rules.

The invention provides an automatic driving system reliability self-diagnosis method and platform based on multi-modal sensor fusion. The method comprises the following steps: acquiring image modal data, performing feature extraction by using a visual Transform model, generating an environment scene label, and representing the type of a current environment; carrying out statistics on key modal characteristics in actual automatic driving operation data to obtain a modal score vector corresponding to each environment scene label to reflect the prior credibility of each sensor modal in the environment, and calculating a modal trust weight through softmax to obtain a confidence weight vector; performing weighted fusion on the key modal features based on the confidence weight vector to obtain fusion features; generating a modal reconstruction feature according to the fusion feature, and generating a modal consistency error vector based on the modal reconstruction feature; and in combination with the confidence weight vector and the modal consistency error vector, determining an early warning signal of the current system for judging whether the system has a fusion degradation or modal failure risk.

The invention relates to a peripheral feedback effect control method and system, a storage medium and electronic equipment, and relates to the technical field of peripheral control, and the method comprises the following steps: determining whether a current target user has a target scene label in a target game according to actual game state data; if the target scene label exists, determining a first light parameter and a first vibration parameter of the actual peripheral according to the target scene label; optimizing the first light parameter and the first vibration parameter according to the actual electric quantity to obtain a final light parameter and a final vibration parameter, and controlling the feedback effect of the actual peripheral according to the final light parameter and the final vibration parameter; and if the target scene label does not exist, determining a second feedback parameter of the actual peripheral according to the user game state data, and controlling a feedback effect of the actual peripheral according to the second feedback parameter. The method and the device have the effect of improving the peripheral feedback effect and the user adaptability.

The invention relates to the technical field of video content analysis, and discloses a video scene content label determination method based on a knowledge graph. The method comprises the following steps: inputting a multi-modal feature sequence formed by visual, audio and text features extracted from a video stream into a knowledge graph inference engine comprising an entity relationship network and a semantic association rule base; performing node mapping through an engine to generate an initial scene entity set; performing hierarchical reasoning on the set based on a semantic association rule base to obtain a scene semantic topological structure; screening core entities according to entity weight distribution in the topological structure, and generating a candidate tag set; performing time sequence consistency verification on the candidate tag set, and correcting tag time sequence offset in combination with timestamp information; performing cross-modal disambiguation on the corrected label set by using an entity relationship network to eliminate semantic conflicts; and according to a disambiguation result, constructing a scene label knowledge sub-graph containing entity attributes and relation path constraints.

The invention relates to the technical field of psychological state assessment, and particularly discloses a psychological state assessment method and system based on environmental perception, and a readable storage medium, and the method comprises the following steps: collecting an environment scene and an assessed object behavior image through an RGB camera, and extracting an object category, a spatial layout and a dynamic event through a semantic segmentation model; acquiring an environment voiceprint through a microphone, and recognizing a voice category by using a voice recognition model; synchronously acquiring facial micro-expression feature points, voice rhythm features and head, neck and shoulder vibration frequencies of the evaluated object; and generating a semantic vector from the object / scene label through a semantic embedding model. According to the psychological state assessment method based on environmental perception, through three core innovations of quantitative modeling of environmental semantics, causal reasoning and a dynamic attention mechanism and personalized adaptive calibration, the neglect limitation of traditional psychological assessment on environmental factors is broken through, and the accuracy, robustness and practicability of psychological state assessment are remarkably improved.

The invention provides a dynamic protection control method for a compressor and an electronic starter controller, and the method comprises the steps: carrying out the scene labeling of the operation data of the compressor, dividing various load scenes, and setting a basic critical value offset coefficient; parameters of the compressor are collected in real time, the real-time load intensity is calculated through a weighting algorithm, and a buffer area is determined; the method comprises the following steps: acquiring data of a compressor, converting the data into a time sequence feature sequence, performing normalization processing, then inputting the time sequence feature sequence into a pre-trained neural network model of a double-branch network structure, and outputting a trend level and a PID parameter interval; according to the trend level and the PID parameter interval, determining a basic PID parameter, and introducing a risk deviation coefficient to determine a final PID parameter; and smooth transition amplitude limiting is set for the final output quantity of the PID parameters, the compressor is operated under the smooth transition amplitude limiting, and the problems of time delay and curve fluctuation of traditional critical value control are solved.

The invention provides an intelligent quality inspection method and device based on RAG and a dual-enhancement mechanism, and the method comprises the steps: determining a retrieval text based on a transcriptional text of voice data to be subjected to quality inspection and an initial quality inspection result; based on the semantics of the retrieval text and the scene label of the voice data, retrieving from a preset knowledge base to obtain matched entries and the relevancy of the matched entries; whether an IDK enhancement mechanism is triggered or not is judged based on the matching items, if not, a direct preference optimization DPO model is called to conduct preference matching degree scoring on the initial quality inspection result, and the DPO matching degree of the initial quality inspection result is obtained; and determining the accuracy and confidence of the initial quality inspection result based on the correlation degree of the matched items, the confidence of the large model, the DPO matching degree and the accuracy of the historical similar cases, and determining a final quality inspection result of the voice data based on the accuracy and confidence. According to the method and the device provided by the invention, the result accuracy, the illusion resistance and the service adaptation efficiency of intelligent quality inspection are remarkably improved.

The invention relates to the technical field of health monitoring, in particular to a health state alarm response system based on multi-dimensional monitoring, which comprises a dynamic threshold module, a physiological monitoring module, an anomaly positioning module, an inducement output module, an anomaly intervention module and an early warning response module, according to the method, multi-dimensional original data of a user is collected, meanwhile, a timestamp alignment and scene label binding technology is adopted, an exclusive physiological index normal fluctuation interval library is constructed for different large classes of scenes, then the problem of misjudgment of a static threshold value is fundamentally solved, the cause of'abnormity 'is deeply and accurately positioned, and the accuracy of the abnormal condition is improved. The abnormity-scene-behavior-cause full-link association not only enables the user to clearly know the cause of the abnormity, but also provides data support for the medical personnel to formulate a targeted intervention scheme, and reasonably formulates the targeted intervention scheme based on the intervention list, thereby avoiding the ineffectiveness of general suggestions, enabling resources to be inclined to high-risk users, and improving the user experience. The complication occurrence rate is effectively reduced.

The invention provides an intention recognition method, device and equipment and a computer program product, and relates to the technical field of computers and Internet. The method comprises the following steps: in multiple rounds of sessions, obtaining session content expressed by a target object in a current round; obtaining a dynamic memory matrix, a plurality of preset intention categories and a dialogue scene label, the dynamic memory matrix comprising at least one group of historical intention conversion information, wherein one group of historical intention conversion information comprises an intention conversion pair, a semantic similarity between two intentions in the intention conversion pair and an intention transition probability corresponding to the intention conversion pair; and performing intention recognition on the session content according to the session content, the dynamic memory matrix, the plurality of preset intention categories and the session scene label to determine a corresponding intention category in the session content of the target object. The intention recognition method and device can improve intention recognition accuracy.

The invention relates to the technical field of medical data analysis, in particular to a clinical team cooperation ability assessment method and system based on multi-modal data dynamic fusion, and the method comprises the steps: collecting physiological signal data based on wearable equipment, collecting audio / video data through an environment sensor, and collecting physical operation data through simulated human equipment; adding a timestamp and a scene label; carrying out sampling frequency synchronization and time axis alignment on the multi-modal data by taking the first-aid event triggering time as a reference; calling an attention mechanism based on the first-aid scene type, and automatically adjusting the weight of each modal data; and integrating the technical operation score and the non-technical cooperation feature to generate an individual and team visual evaluation report. According to the invention, the problem of blank collection of team non-technical skill data in traditional training is solved, and the effects of quantitative evaluation and real-time feedback of collaboration capability are realized.

The present invention discloses a method, device, equipment and medium for model training and scene recognition. When training a scene recognition model, first, the parameters of the core feature extraction layer and the global information feature extraction layer are trained through the first scene label of the sample image and the standard cross-entropy loss. Then, according to the feature maps output by the LCS modules at each level and the loss value calculated pixel by pixel with the first scene label of the sample image, the weight parameters of the LCS modules at each level are trained. Finally, the parameters of the fully connected decision layer of the scene recognition model are trained. This enables the scene recognition model to have the ability to extract features with high richness, and based on the scene recognition model for scene recognition, the accuracy of scene recognition is greatly improved.

The invention relates to an automatic data security grading method based on machine learning, and the method comprises the steps: obtaining the text description of a data element tag and a scene tag from a preset knowledge base, generating a combined text, mapping the combined text into a semantic vector through a text embedding model, and constructing an initial semantic vector set; according to the initial semantic vector set in combination with the annotation data, adopting regression model training to generate an initial security score set; calculating a difference value between a score value and an integer value according to the initial safety score set, marking fuzzy entries and constructing a fuzzy entry set; performing priority ranking on the fuzzy entry set, extracting feature data for comparative analysis, and generating a feature matching report; adjusting a score value according to the feature matching report, and constructing an intermediate safety score set; secondary comparison is carried out on the intermediate security score set, a fuzzy mark state is reserved, and a final item set needing to be rechecked is constructed; and generating and storing a structured prediction data report according to the final entry set needing to be rechecked.

The invention discloses an audio and video processing system and method supporting an AI intelligent repair technology, and relates to the technical field of audio and video data processing, and the method comprises the following steps: carrying out the scene recognition based on a playing source type, a user interaction form, an audio and video code rate, an audio sampling rate and network transmission time delay, and generating live broadcast, film watching and singing scene labels; determining a time delay upper limit, a resolution target, a detail reduction level, an audio fidelity index and a beat alignment threshold according to the scene label, and forming a performance constraint vector; according to the invention, a closed-loop mechanism of scene identification, performance constraint setting, repair link arrangement, cloud and terminal task allocation, real-time adaptive adjustment and parameter iterative optimization is constructed for different requirements of live broadcast, film watching, singing and other scenes, so that low delay, high image quality, high tone quality and computing power adaptation are realized; and the stability and the continuous optimization capability of the system are improved through versioning and staged release.