Cross-platform video conference fusion method and system

By constructing a code-aware quality mapping model (PQMM) and adjusting code control parameters in real time, the problem of inconsistent visual quality caused by differences in encoding formats in cross-platform video conferencing is solved. Dynamic quality alignment is achieved in a distributed architecture, improving image consistency and user experience, and is suitable for large-scale multi-terminal collaborative scenarios.

CN121842392APending Publication Date: 2026-04-10HUIDIAN ZHIKE (GUANGDONG) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610041809.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-13
Publication Date
2026-04-10

AI Technical Summary

Technical Problem

In existing cross-platform video conferencing systems, inconsistent visual quality due to different encoding formats, especially when network conditions fluctuate, makes it difficult to ensure a balanced visual experience among participants. Furthermore, existing technologies have not effectively solved the problem of subjective quality consistency under multiple encoding formats.

Method used

A code-aware quality mapping model (PQMM) is constructed. By collecting terminal coding capability information, a coding capability configuration table is generated. An input feature space is constructed based on historical test datasets. Code control parameters are predicted and adjusted in real time to achieve quality alignment. Feedback correction is performed by combining SEI extended metadata to form a dynamic optimization model.

Benefits of technology

It achieves low-overhead, high-response dynamic quality alignment under a distributed SFU architecture, improving screen consistency and user experience in mixed terminal scenarios, balancing computational efficiency and system scalability, and is suitable for large-scale multi-terminal collaboration scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121842392A_ABST
    Figure CN121842392A_ABST
Patent Text Reader

Abstract

The invention discloses a cross-platform video conference fusion method and system, and the method comprises the steps: collecting the coding format, equipment type and network state of each terminal, generating a coding capability configuration table, constructing a coding perception quality mapping model in combination with historical subjective quality score data and video content features, and achieving the subjective quality prediction of multiple terminals in a conference scene; the system automatically triggers code control parameter optimization and code rate allocation dynamic adjustment according to the quality score difference of each terminal, a quantization matrix is optimized for low-quality nodes, code rate compression is implemented for high-quality nodes, and the overall visual quality consistency is improved; the prediction model is dynamically corrected based on real-time feedback data, the adaptability and robustness of the model are continuously improved through periodic incremental training, dynamic consistency of video quality of multiple coding nodes in a conference scene is effectively achieved, and collaborative experience is optimized.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blockchain smart contract and digital signature verification, and particularly relates to a cross-platform video conference fusion method and system. BACKGROUND

[0002] Current video conference systems have widely adopted multiple mainstream video coding standards, including H.264, H.265, VP9 and AV1, etc. These coding formats are optimized for different scenarios and hardware environments, and there are essential differences in compression efficiency, motion compensation accuracy and blocking effect suppression capability, etc. With the growing demand for cross-platform video conferencing, the system often needs to distribute video streams of multiple coding formats to different terminals (such as PC, smart phone, dedicated conference hardware, etc.) at the same time to compatible with the decoding ability and network environment of various devices; The current mainstream cross-platform video conference solution in the industry usually relies on two types of technical paths. One is to uniformly transcode all video streams at the server or edge node, and then distribute them to each terminal after converting them into a single coding format. This centralized transcoding method can maintain partial picture balance, but it significantly increases the server computing power consumption and brings additional transcoding delay, and it is difficult to adapt to large-scale distributed conference needs. The other solution uses a distributed media forwarding (such as SFU architecture), and each terminal selects the coding format and quality parameters according to its own conditions to achieve distributed coding and bandwidth optimization. This method improves the scalability and flexibility of the system, but due to the inconsistent subjective picture quality of different coding standards at the same code rate, it cannot effectively guarantee the visual experience balance among participants, and often there is a phenomenon of "great disparity in picture quality in the same conference". Especially when the network condition fluctuates, the adaptive behavior of different formats further exacerbates the difficulty of ensuring visual consistency; In the existing representative technical solutions, most of them only focus on the improvement of coding efficiency or bandwidth utilization, and do not systematically model and regulate the consistency of visual subjective quality under multiple coding formats. Some systems support terminal picture adaptation, layered coding or one-way quality hint signaling, but generally do not establish a subjective quality quantification, alignment and closed-loop feedback mechanism across coding formats, resulting in inherent bias in subjective perception such as compression residual, detail restoration capability under different standards. The existing technology also has limitations such as insufficient model generalization ability, not considering the influence of terminal type and network level on visual sensitivity, and lack of efficient real-time feedback and model continuous adaptation mechanism. SUMMARY

[0003] The present application provides a cross-platform video conference fusion method and system to solve the above technical problems.

[0004] The technical scheme of the present application is implemented as follows: a cross-platform video conference fusion method, comprising: S1: collecting the encoding capability information reported by different terminals in the cross-platform video conference system, including the supported video encoding format, device type identifier, and network state level, and generating an encoding capability configuration table; S2: constructing an input feature space of an encoding perceptual quality mapping model (PQMM) based on the texture complexity, motion intensity, and chroma distribution characteristics in the historical test data set, wherein the historical test data set contains subjective quality scores of H.264, H.265, VP9, and AV1 four encoding formats under different code control parameters; S3: matching the encoding format parameters in the encoding capability configuration table with the preset encoder characteristic parameter library, calling the corresponding PQMM model instance to perform quality prediction on the current code rate strategy, and generating an expected subjective quality score matrix of each participant terminal; S4: calculating the standard deviation of each element in the expected subjective quality score matrix, and if the standard deviation exceeds a preset quality difference threshold, executing a quality alignment decision process to generate a code control parameter adjustment instruction set for each encoding node; S5: according to the code control parameter adjustment instruction set, performing quantization matrix optimization processing on the encoding nodes with quality scores lower than the average value, and performing code rate compression processing on the encoding nodes with quality scores higher than the average value, to generate a dynamic quality aligned video stream; S6: collecting the quality feedback information in the SEI extension metadata of each encoding node in real time, including the decoding failure rate and Jitter Buffer statistical value, correcting the prediction results of the PQMM model, and generating updated quality mapping model parameters; S7: performing periodic model retraining based on the updated quality mapping model parameters, retaining the quality feedback data within the latest preset time period through a sliding window mechanism, and forming a dynamically optimized quality alignment model.

[0005] The present application also provides a cross-platform video conference fusion system, which adopts the above cross-platform video conference fusion method for cross-platform video conference fusion.

[0006] The cross-platform video conference fusion method provided by the present application has the following beneficial effects: (1) The application proposes a dynamic quality alignment mechanism based on a perception consistency model, which effectively overcomes the visual quality inconsistency problem caused by coding format differences in existing cross-platform video conference systems. By constructing a lightweight coding perception quality mapping model (PQMM), the content features and coding parameters are associated with quantifiable no-reference image quality scores, realizing the pre-judgment and horizontal comparison of the output quality of different encoders under the same code rate. On this basis, the visual fidelity gap caused by the compression efficiency difference of the coding standard itself can be actively identified and compensated, and the quality balance adjustment between multiple formats can be completed without relying on centralized transcoding, significantly improving the picture consistency and professionalism performance in the mixed terminal access scenario; (2) The application introduces a pre-compensation before encoding and feedback closed-loop quality regulation architecture, realizing the dynamic quality alignment capability with low overhead and high response under the distributed SFU architecture. Compared with the traditional method of static configuration code rate or fixed quantization parameter, the method can evaluate the predicted quality deviation of each end in real time during the conference operation, and implement differentiated code control adjustment according to the comparison results: for the terminal using inefficient coding format but facing quality disadvantage, the code rate allocation or optimization quantization matrix is moderately improved to enhance the detail retention; and for the redundant quality margin generated by the efficient coding format, the compression strength is allowed to be appropriately relaxed to release bandwidth resources for audio, auxiliary flow or network jitter buffer, thereby realizing the collaborative optimization of resource utilization and user experience. The process transmits instructions through SEI messages or extended metadata at the encoding layer, without changing the existing transmission protocol stack, and has strong compatibility and flexible deployment. At the same time, combined with the jitter buffer state, decoding success rate and other runtime feedback information of the receiving end, the PQMM model parameters are calibrated reversely, forming a closed-loop optimization path from prediction, adjustment to verification, greatly enhancing the robustness and adaptive ability of the system in the dynamic network and diverse terminal environment; (3) The perception consistency driven quality alignment system constructed by the application guarantees balanced cross-coding format visual experience while taking into account the computing efficiency and system scalability, and is suitable for large-scale, multi-terminal concurrent remote collaboration scenarios. The PQMM model design focuses on lightweight, relying only on texture complexity, motion intensity, chroma distribution and other basic content features as input, and can complete inference at the edge coding node with extremely low delay, avoiding the introduction of additional processing bottlenecks. The model supports hierarchical configuration by device type, network level or user role, facilitating fine-grained management between different service quality levels. The entire mechanism does not require centralized processing of video streams across the network, and is fully compatible with the current mainstream SFU forwarding architecture, reducing the computing burden of the center node while maintaining efficient end-to-end transmission paths. The resulting decentralized and scalable visual quality consistency guarantee solution not only solves the long-standing problem of picture imbalance in multi-coding coexistence, but also provides a universal interface and evaluation benchmark for the smooth introduction of future new coding standards (such as AV2 or neural compression). BRIEF DESCRIPTION OF DRAWINGS

[0007] Figure 1 A flowchart of a cross-platform video conference fusion method of the application; Figure 2 A sub-flowchart of a cross-platform video conference fusion method of the application; Figure 3 Another sub-flowchart of a cross-platform video conference fusion method of the application. DETAILED DESCRIPTION

[0008] Embodiments of the application are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the accompanying drawings are exemplary and are used only to explain the application, and cannot be understood as limiting the application.

[0009] The disclosure below provides many different embodiments or examples for implementing different structures of the application. In order to simplify the disclosure of the application, the components and arrangements of specific examples are described below. Of course, they are only examples and the purpose is not to limit the application. In addition, the application can repeatedly refer to numerals and / or letters in different examples, and such repetition is for the purpose of simplification and clarity, which itself does not indicate the relationship between the various embodiments and / or arrangements discussed.

[0010] As Figure 1 shown, the application provides a cross-platform video conference fusion method, specifically comprising: S1: Collect the encoding capability information reported by different terminals in the cross-platform video conference system, including supported video encoding formats, device type identification, and network state level, and generate an encoding capability configuration table; S2: Based on the texture complexity, motion intensity, and color distribution characteristics in the historical test data set, construct an input feature space of the encoding perceptual quality mapping model (PQMM), which contains subjective quality scores of H.264, H.265, VP9, and AV1 under different code control parameters; S3: Match the encoding format parameters in the encoding capability configuration table with the preset encoder characteristic parameter library, call the corresponding PQMM model instance to predict the quality of the current code rate strategy, and generate an expected subjective quality score matrix for each participant terminal; S4: Calculate the standard deviation of each element in the expected subjective quality score matrix. If the standard deviation exceeds the preset quality difference threshold, execute the quality alignment decision process and generate a code control parameter adjustment instruction set for each encoding node; S5: According to the code control parameter adjustment instruction set, perform quantization matrix optimization processing on the encoding nodes with quality scores lower than the average, and perform code rate compression processing on the encoding nodes with quality scores higher than the average, to generate a dynamic quality-aligned video stream; S6: Real-time collect quality feedback information in SEI extension metadata of each encoding node, including decoding failure rate and Jitter Buffer statistical value, correct the prediction results of the PQMM model, and generate updated quality mapping model parameters; S7: Based on the updated quality mapping model parameters, perform periodic model retraining, retain the quality feedback data within the latest preset time period through the sliding window mechanism, and form a dynamically optimized quality alignment model.

[0011] The step S1: Collect the encoding capability information reported by different terminals in the cross-platform video conference system, including supported video encoding formats, device type identification, and network state level, and generate an encoding capability configuration table. Specifically, it includes: S1.1: Based on the capability negotiation signaling reported by the terminal device during the conference access stage, obtain the supported video encoding format list, including one or more combinations of H.264, H.265, VP9, and AV1, to form an encoding format set; Based on the input conditions of the capability negotiation signaling analysis module of the terminal device during the conference access stage, extract and analyze the codec capability declaration field reported by the terminal, form an original encoding format identification data set, and use it for subsequent encoding format set generation processing; The protocol analysis algorithm (parameters: support SIP / SDP signaling analysis rule set, UTF-8 character decoding mapping table) is adopted to realize the identification of the "m=video" and the corresponding "a=rtpmap" field in the capability negotiation signaling, and the encoding format identification string is extracted, ensuring the distinction of different encoding types; Further, through the encoding format mapping table lookup algorithm (parameters: H.264, H.265, VP9, AV1 standard format identification and internal encoder index mapping relationship), the string obtained by signaling analysis is mapped to a unified internal encoding format enumeration value, and an encoding format enumeration vector is obtained; Further, the deduplication and validity detection algorithm (parameters: enumeration value legal range verification rule, repeated value removal strategy) is adopted to realize the repeated item cleaning and illegal format elimination of the encoding format enumeration vector, and an effective encoding format set verified is generated; Through the format set sorting and priority marking processing mode (parameters: internal encoding preference strategy file, network bandwidth constraint table), the effective encoding format set in the previous step is converted into a sorted list with priority labels, realizing the quick matching of the encoding format set in the subsequent conference-level capability integration; For example, in a multi-terminal access conference scenario, a mobile terminal includes the following fields in the capability negotiation signaling during the access stage: "m=video 9 RTP / AVP 96 97\na=rtpmap:96 H264 / 90000\na=rtpmap:97 VP9 / 90000". The analysis module identifies the encoding format identifiers "H264" and "VP9" according to the SDP rule, and then calls the encoding format mapping table to map them to internal enumeration values 1 and 3 respectively, forming an enumeration vector [1,3]. After deduplication and validity detection, it is confirmed that the set is valid and has no redundancy, and an effective encoding format set {H.264, VP9} is generated. According to the bandwidth constraint table (assuming that the current network is at a medium level, and the bandwidth limit is 3Mbps) and the preference strategy file (prefer to select the encoding with higher compression efficiency under medium bandwidth), VP9 is labeled with priority 1 and H.264 is labeled with priority 2, forming the final sorted list [VP9, H.264]. The output result will be combined with the device type and network level to form a terminal encoding capability description tuple when S1.4 is encapsulated, realizing efficient calling and quality prediction of the PQMM model in the subsequent stage; S1.2: Analyze the received terminal device identification information, extract the device type identification field, which includes mobile terminal, desktop terminal or professional conference terminal and other classification information, for subsequent quality model classification configuration; The received terminal device identification information is parsed by a protocol analysis method (parameters: terminal access signaling protocol type, field analysis rule set) to extract the original device identification data; Further, the original identification is converted into a device type code by a field mapping algorithm (parameters: identification field and device type classification table, mapping priority rule), and a device type code vector is obtained; Further, the device type code vector is verified and corrected by a regular matching and pattern recognition algorithm (parameters: regular expression set, pattern feature library), and a verified device type identification set is generated; Further, the device type identification set is processed by a device type classification logic (parameters: mobile terminal, desktop terminal, professional conference terminal classification weight), and a device type classification parameter set is generated; The device type classification parameter set of the previous step is converted into quality model classification configuration input data by a structured packaging processing method, and the correspondence between device type and model parameter is established; For example, in a certain cross-platform video conference system, the terminal device identification field is in the form of "DEV_MOB_Android_12_Huawei" or "DEV_DES_Windows_10_Dell", the protocol type is JSON structured message, and the field analysis rule set specifies that "DEV_MOB" is mapped to mobile terminal and "DEV_DES" is mapped to desktop terminal. Using the field mapping algorithm, "DEV_MOB" is parsed into type code 101 and "DEV_DES" is parsed into type code 102. The regular expression set ^DEV_(MOB|DES|PRO)_.*$ is used for pattern verification, "DEV_MOB_Android_12_Huawei" is successfully matched as a mobile terminal category, and "DEV_DES_Windows_10_Dell" is successfully matched as a desktop terminal category. According to the classification logic, the mobile terminal is assigned a weight parameter of 1.0, the desktop terminal is assigned a weight parameter of 0.8, and the professional conference terminal is assigned a weight parameter of 1.2, and finally the device type classification parameter set is formed as {terminal A: type code 101, weight 1.0; terminal B: type code 102, weight 0.8}. This parameter set is used as the quality model classification configuration input, effectively supporting the quality prediction and adaptation of the PQMM model for different device types, and significantly improving the consistency of the encoding quality in a multi-terminal environment; S1.3: Collect and quantify the real-time network bandwidth, packet loss rate and delay indicators reported by the network state monitoring module on the terminal to generate network state level parameters, which are divided into three levels: high, medium and low, for encoding strategy adaptation; The real-time network bandwidth data output by the network state monitoring module is subjected to sliding window mean filtering (parameters: window length w=5 seconds) to achieve short-time smoothing of the bandwidth sequence to eliminate transient peak interference and obtain a stable bandwidth estimate value; Further, the packet loss events counted by the transmission layer are subjected to normalization processing by a packet loss rate calculation unit (parameters: periodic sampling interval t=1 second) to generate a per-second packet loss rate sequence, and an exponential weighted moving average algorithm (parameters: decay coefficient a=0.3) is used to calculate the stability score of the packet loss rate; Further, for the delay indicator, a round-trip time (RTT) measurement method is used in combination with a time stamp difference algorithm (parameters: number of sampling packets n=50) to calculate the mean and variance of the delay, and the variance value is mapped to a delay fluctuation coefficient to represent the network delay stability; Further, based on the smoothed bandwidth estimate value, the stability score, and the delay fluctuation coefficient, a multi-indicator weighted scoring algorithm (weight settings: bandwidth weight 0.5, packet loss weight 0.3, delay weight 0.2) is executed to calculate the network comprehensive performance score by the following scoring formula :

[0012] wherein, B is the bandwidth normalized score, P is the packet loss rate stability score, L is the delay stability score; Further, the network comprehensive performance score calculated is mapped to high, medium, and low network state level parameters by a threshold classification algorithm (parameters: high-level threshold S≥0.8, medium-level threshold 0.5≤S<0.8, low-level threshold S<0.5); Through the above weighted scoring and classification algorithm, the smoothed bandwidth, the stabilized packet loss rate, and the delay fluctuation characteristics are converted into structured network state level parameters to achieve the goal of adaptive coding strategy; For example, during a multi-terminal video conference access process, the network state monitoring module collects the bandwidth sequence of a certain mobile terminal at the access instant as [4.8, 5.1, 4.9, 5.0, 4.7] Mbps, and after sliding mean filtering with a window length w=5 seconds, the bandwidth estimate value B=4.9 Mbps, and the normalized score B=0.82. The packet loss rate sequence is [0.9, 1.2, 1.0, 1.1, 0.8]%, and the stability score P=0.78 is calculated using an exponential weighted moving average with a=0.3. The mean is 63 ms and the variance is 5 ms, and the delay stability score L=0.85 is obtained after delay stability mapping. The above three scores are substituted into the comprehensive score formula: , the terminal network state level is determined to be high according to the classification threshold. This level parameter is packaged together with the coding format set and the device type identifier into the terminal coding capability description tuple, providing accurate network input conditions for subsequent PQMM model calls, significantly improving the reliability of coding quality prediction and alignment; S1.4: Structurally package the coding format set, device type identifier, and network state level parameter to generate a terminal coding capability description tuple, and upload it to the configuration management unit of the media processing module through the session control channel; S1.5: Based on the coding capability description tuples of all participating terminals, perform aggregation processing to generate a conference-level coding capability configuration table, which includes four-dimensional information of terminal identifier, coding format, device type, and network level, providing input basis for subsequent quality prediction model calls; Based on the set of terminal coding capability description tuples, a hash index aggregation method (parameters: terminal identifier key, coding format key, device type key, network level key) is used to realize fast grouping and deduplication of cross-terminal capability information; Further, through a multi-key association mapping algorithm (parameters: four-dimensional key set, association rule library), multi-dimensional cross combination of all participating terminal capability data in the same conference is realized, and a group mapping relationship matrix is obtained; Further, a feature consistency verification method (parameters: coding format set, device type classification table, network level mapping table) is used to realize field integrity and legality verification of the aggregated data, and generate a compliance data set that can be used for model input; Further, through a structured packaging algorithm (parameters: JSON schema definition, field serialization order), the compliance data set is uniformly encoded into a structured record to form the basic data structure of the conference-level coding capability configuration table; Further, an index optimization method (parameters: B+ tree structure, four-dimensional index key) is used to realize fast retrieval performance optimization of the conference-level configuration table, and generate an indexed data representation that can be directly called by the quality prediction model; Through the above aggregation and packaging processing method, the terminal capability tuple output in the previous step is converted into a conference-level coding capability configuration table containing four-dimensional information of terminal identifier, coding format, device type, and network level, realizing global consistent organization and efficient call of PQMM quality prediction input data; Exemplarily, in a cross-platform video conference scene containing 50 participating terminals, the terminal identification format is MAC address plus device serial number, the coding format set includes three combinations of H.264, H.265 and VP9, the device type is divided into two categories of mobile terminal and desktop terminal, and the network level is divided into three levels of high, medium and low. When the hash index aggregation method is used, the four-dimensional capability information of each terminal is mapped to the index key , wherein represents a connection operation, and realizes fast deduplication with O(1) time complexity. A four-dimensional relationship matrix is generated using a multi-key association mapping , wherein i represents the terminal index, j represents the coding format index, k represents the device type index, and l represents the network level index. Through feature consistency verification, it is detected that the network level fields of 2 records are empty, and the default level "medium" is applied for filling to ensure data legality. Structured packaging is converted into a unified JSON record, and each record is about 150 bytes. Through B+ tree index, the query is completed within . In this example, the generated configuration table can realize millisecond-level global capability matching in the PQMM calling stage, significantly improving the real-time performance and accuracy of cross-coding format quality prediction.

[0013] The step S2: based on the texture complexity, motion intensity and chroma distribution characteristics in the historical test data set, the input feature space of the encoding perceptual quality mapping model (PQMM) is constructed, and the historical test data set includes the subjective quality scores of H.264, H.265, VP9 and AV1 four encoding formats under different code control parameters. Specifically, it includes: S2.1: Perform content analysis-based feature extraction processing on the original video frame sequence in the historical video conference data set to obtain texture complexity feature parameters; the texture complexity feature parameters are calculated based on the joint calculation of local binary pattern (LBP) and gray level co-occurrence matrix (GLCM), and are used to quantify the repetition and edge density of local structure of an image; The local binary pattern (LBP) feature extraction method (parameters: radius r=1, neighborhood sampling point number p=8) is used on the original video frame sequence in the historical video conference data set to realize binary coding calculation on the pixel neighborhood in each video frame; Further, through the gray level co-occurrence matrix (GLCM) method (parameters: gray level g=256, offset vector ), the statistical relationship modeling of the video frame is realized, and the energy, contrast, correlation and other statistical quantities are obtained; Further, by weighting and fusing the local pattern histogram of the LBP feature map and the GLCM statistics (the weight coefficients are set according to the cross-validation results, for example, the LBP weight is 0.6 and the GLCM weight is 0.4), multi-dimensional representation of texture complexity is realized, and a comprehensive index for quantifying the local structure repeatability and edge density of the image is generated; Further, a complexity coefficient calculation formula based on variance normalization is used to standardize each dimension of the fused feature, so as to eliminate the influence of the dimension of different feature values, and generate a standardized feature value . The formula is as follows:

[0014] Among them, is the current feature value, is the mean value of the feature dimension, is the standard deviation of the feature dimension; Further, the mean value and the standard deviation of all video frame feature coefficients are globally statistically analyzed to obtain a texture complexity feature parameter set, which is used for PQMM input feature space construction; Through the above LBP and GLCM joint algorithm, the local structure pattern and the gray distribution relationship of the original video frame sequence are converted into standardized texture complexity feature parameters, realizing the quantification and description of the structure repeatability and edge density of the video image; For example, for a historical conference video data set with a resolution of 1920x1080, LBP feature extraction is used for each frame, the radius r is set to 1, the sampling point p is set to 8, and a 256-dimensional pattern histogram is generated; GLCM feature calculation is used, the gray level is 256, the offset vector , and four statistics, i.e., energy, contrast, correlation and entropy, are calculated to obtain 16-dimensional GLCM features. The 256-dimensional LBP features and the 16-dimensional GLCM features are weighted and fused (weight 0.6:0.4) to form a 272-dimensional fused feature vector. For each dimension of the fused feature, the variance normalization formula is used for standardization to obtain a feature distribution with a mean value of about 0 and a standard deviation of 1. The texture complexity coefficient obtained by global statistical calculation ranges from -2.3 to 2.7, and the edge density index is significantly improved. Finally, a 272-dimensional texture complexity feature parameter set is formed, which provides texture representation input for PQMM and improves the consistency and stability of the encoding quality prediction; S2.2: performing motion intensity feature extraction processing based on the optical flow method on the original video frame sequence to obtain motion vector amplitude distribution statistics; the motion intensity feature extraction is based on the Horn-Schunck optical flow algorithm to estimate the pixel displacement between consecutive frames, and outputs the mean value and the variance of the inter-frame motion vector amplitude as the motion intensity representation parameters; S2.3: performing HSV color space-based chroma distribution feature extraction processing on the original video frame sequence to obtain chroma histogram statistical features; the chroma distribution feature extraction is based on histogram statistics on the H channel after converting the RGB image to the HSV space, and outputs chroma distribution entropy and dominant color tone concentration indicators; S2.4: performing feature normalization and fusion processing on the texture complexity feature parameters, motion intensity representation parameters, and chroma distribution entropy indicators to generate a standardized multi-dimensional content feature vector; the feature normalization is based on Z-score method to standardize each dimension feature, and principal component analysis (PCA) is used for feature dimension reduction, and a low-dimensional compact feature representation is output; S2.5: performing feature space division processing based on the multi-dimensional content feature vector and the pre-labeled subjective quality score dataset to generate an input feature space of the encoding perceptual quality mapping model (PQMM); the input feature space is divided into multiple feature subspaces, each corresponding to a specific content type and an encoding response mode, for supporting subsequent model quality prediction adaptation under different encoding formats.

[0015] As shown in Figure 2 the step S3: matching the encoding format parameters in the encoding capability configuration table with the preset encoder characteristic parameter library, calling the corresponding PQMM model instance to perform quality prediction on the current code rate strategy, and generating an expected subjective quality score matrix of each participating terminal. Specifically, it includes: S3.1: parsing the encoding format parameters in the encoding capability configuration table, extracting the video encoding format identifiers supported by each participating terminal to determine the matching index in the corresponding encoder characteristic parameter library; The encoding format parameters of the encoding capability configuration table are structured read, and a field parsing method (parameters: field separator type, character encoding method, data length limit) is used to realize the item-by-item disassembly and encoding format field positioning of the encoding capability entry of each participating terminal; Further, through the encoding format normalization algorithm (parameters: format string regular pattern, abnormal character filtering rule), the mapping from string to standard encoding format identifier set is realized, and the encoding format identifier list after format verification is obtained.

[0016] Further, the encoding format legality verification method (parameters: supported encoding format full set={H.264, H.265, VP9, AV1}) is used to realize the comparison between the identifier list and the support set, and generate a legal encoding format identifier sequence; Further, the encoding format identifier sequence is converted into a unique matching index value in the encoder characteristic parameter library by using an encoding format to characteristic parameter library index mapping algorithm (parameters: index hash function type, conflict resolution strategy), and a matching index table is generated; Through the above analysis, normalization, verification, and index mapping processing mode, the original encoding format parameters in the encoding capability configuration table are converted into matching index data that can directly drive the PQMM model call, realizing the expected technical effect of accurate matching of the encoder model. For example, in a conference scenario, the encoding format field of a certain participant terminal in the encoding capability configuration table is "H.264|VP9". The field parsing method (field separator type is "|", character encoding method is UTF-8) can decompose two initial format strings "H.264" and "VP9". After processing by the encoding format normalization algorithm (regular pattern matching video encoding format naming rules), the formatted identifier set {H.264, VP9} is obtained. By comparing with the support set {H.264, H.265, VP9, AV1} using the legality verification method, it is confirmed that both are legal formats. The index values are generated using the hash mapping function h(fmt)=hash(fmt) , where fmt is the format identifier string. For example, H.264 is mapped to index 10 and VP9 is mapped to index 25. The final output matching index table is {H.264:10, VP9:25}, which can be directly input into the PQMM model instance loading process to realize accurate matching of the model parameters of the two encoding formats. In the test and verification, this processing mode significantly improves the call efficiency and matching success rate of the PQMM quality prediction module; S3.2: Based on the encoding format identifier, retrieve the model call interface of the corresponding encoder from the preset encoder characteristic parameter library, which contains parameter configuration information of H.264, H.265, VP9, and AV1 four encoding formats, to realize accurate matching of the PQMM model instance; Based on the parsed encoding format identifier, an index matching retrieval method (parameters: encoding format identifier set, encoder characteristic parameter library index structure) is used to realize fast positioning of the corresponding encoder entry in the preset encoder characteristic parameter library; Further, by using a hash mapping retrieval method (parameters: encoding format identifier hash code, characteristic library hash table), constant time access of the encoder entry in the characteristic library is realized, and a complete configuration set of the encoder model, inherent encoding efficiency parameters, and adjustable code control parameters is obtained; Further, an interface analysis and binding algorithm (parameters: encoder calling interface description file, PQMM model interface adaptation rule) is adopted to realize automatic analysis and function binding of the retrieved encoder calling interface, to generate an interface control object that can be directly used for PQMM model instance input, and to generate a confirmation flag of successful interface binding; Further, a parameter consistency verification method (parameters: encoder inherent parameter set, PQMM model required input parameter format) is adopted to realize matching verification between inherent parameters and model input format, and to generate a parameter mapping table, for ensuring that the input data structure in the PQMM calling process is consistent with the actual output structure of the encoder; Through the above retrieval and interface binding processing mode, the encoding format identifier of the previous step is converted into an encoder interface object that can be accurately called and matched with the PQMM model instance, to realize seamless docking of the encoder characteristics to the model prediction module. For example, in a conference scenario, terminal A supports the H.265 format, and the encoding format identifier is "H265_Main10". The corresponding entry in the characteristic parameter library records the quantization parameter range QP=[20, 45] of the encoder at the Main10 level, the code rate control mode CBR / VBR, and the motion compensation accuracy 1 / 4 pixel. The system first calculates the hash code of the identifier, and locates the encoder entry with index ID=1024 in the library hash table, to read all inherent parameters and dynamically adjustable parameters. In the interface analysis stage, the system analyzes the input signature of the calling interface, including the texture complexity parameter (float), the motion intensity parameter (float), the chroma distribution parameter (vector <float>), the parameter consistency verification process, the generated parameter mapping table indicates that the inherent parameter "maximum QP" is mapped to the model input field "qp_max", and the inherent parameter "motion compensation precision" is mapped to the model input field "mc_precision". Finally, the encoder interface object CO_H265_Main10 is successfully bound to the PQMM model instance MI_H265_Main10, and MI_H265_Main10 is directly called in the subsequent quality prediction stage to evaluate the picture quality score of terminal A under the current code rate strategy. The whole process ensures the consistency and integrity of the encoder characteristic parameters and the PQMM input, thereby significantly improving the prediction accuracy and real-time response ability; S3.3: Use the PQMM model instance to predict the encoding quality of the code rate strategy under the current conference scene. The input includes the texture complexity, motion intensity and color distribution characteristics of the current video content. The output is the expected subjective quality score matched with the encoding format and content characteristics; S3.4: Perform PQMM model prediction on different encoding formats used by each participating terminal under the same video source to obtain the subjective quality score value of each encoding node under the current code rate strategy, and form a quality score vector; Based on the unified video source and the encoding format identifier of each participating terminal obtained in the current conference scene, the matched PQMM model instance is called (parameters: texture complexity characteristics, motion intensity characteristics, color distribution characteristics and code rate strategy) to realize the subjective quality prediction of a single terminal under a specific encoding format; Further, through the batch model inference mechanism (parameters: terminal encoding format set, unified video source content feature vector), the quality prediction of all participating terminals under different encoding formats is realized, and the corresponding single-point quality score value is obtained; Further, the vectorization result aggregation method (parameters: each encoding node single-point quality score value is clustered according to the terminal) is used to realize the ordered collection of quality scores and generate a preliminary score set indexed by the terminal; Further, through the data consistency checking algorithm (parameters: data integrity and format legality threshold in the score set), the effectiveness of the predicted score vector is verified, and abnormal scores that do not meet the prediction constraints are removed; Through the quality score merging processing method, the single-node score value of the previous step is converted into a structured quality score vector, realizing the ordered data input required for subsequent matrix organization; Exemplarily, in a cross-platform video conference containing four participating terminals, the uniform video source extracts texture complexity mean value of 0.68, motion intensity mean value of 3.2 pixels / frame, and color distribution entropy of 1.45 through the content analysis module. The four terminals support H.264, H.265, VP9, and AV1 encoding formats respectively, and the code rate strategy is set to 1500kbps. For the encoding format of each terminal, the corresponding PQMM model instance is called for inference, and the model input is the above-mentioned three-dimensional content features and the code rate parameter. The output single-point quality score values are 78.6, 82.3, 76.9, and 80.5 respectively. Using vector aggregation processing, these scores are organized into a quality score vector [78.6, 82.3, 76.9, 80.5]. Through a consistency check algorithm with a threshold of ±5 points, it is confirmed that all scores are within the acceptable range, forming a final valid quality score vector that will be used as input for matrix organization to support the generation of a complete quality score matrix in the subsequent S3.5 step, realizing the basis for unified quality analysis across encoding formats; S3.5: Organize the quality score vectors of each participating terminal into a matrix to generate an expected subjective quality score matrix, with the rows corresponding to different participating terminals and the columns corresponding to the quality scores under their encoding formats, to support subsequent quality consistency judgment and dynamic alignment processing; Using the quality score vectors formed by each participating terminal in S3.4 as input parameters, a matrix organization method is used (parameters: terminal identifier, encoding format type, quality score value) to structure the scattered quality score data into a two-dimensional matrix form and establish a corresponding relationship for matrix row and column indexing; Further, through an index mapping algorithm (parameters: terminal identifier mapping table, encoding format sequence), the matrix rows correspond to different participating terminals and the columns correspond to the encoding formats used by each terminal, and the corresponding quality score values are filled in the matrix cells, thereby ensuring that the matrix structure and the elements of the encoding format set correspond one-to-one; Further, a sparse matrix filling strategy (parameters: quality score missing marker, default fill value) is used to fill the unused encoding formats in the matrix with default scores or null markers, ensuring the integrity and parsability of the matrix and avoiding abnormal subsequent calculations due to data missing; Further, a matrix normalization processing algorithm (parameters: maximum score, minimum score) is used to linearly normalize each score value in the matrix, making the scores of different encoding formats uniform in dimension and eliminating calculation bias caused by different score scales; By matrix organization and normalization processing, the quality score vector set of the previous step is converted into an expected subjective quality score matrix, realizing the unified bearing of quality data and providing high-structured and high-computable input basis for the quality consistency judgment and dynamic alignment processing of the subsequent S4 step; Exemplarily, in a cross-platform video conference scene containing 4 participating terminals (T1, T2, T3, T4), terminal T1 supports H.264 and H.265, T2 supports VP9 and AV1, T3 supports H.264 single format, and T4 supports H.265 and AV1. In the quality score vector output by the S3.4 step, the score of T1 is [4.5, 4.8], the score of T2 is [4.2, 4.7], the score of T3 is [4.6], and the score of T4 is [4.9, 4.4]. Using the matrix organization method, the matrix row label is mapped to T1 to T4, and the column label is mapped to the encoding format set [H.264, H.265, VP9, AV1]. The matrix cells are filled by using the index mapping algorithm, such as writing to the T1 row H.264 column, writing to the T1 row H.265 column, and filling null values to the T1 row VP9 column and the AV1 column. After the sparse matrix filling strategy processing, the default score is filled for the null value. The linear normalization is performed on the non-null values in the matrix by using the normalization algorithm: let the maximum score be and the minimum score be , then the H.264 score normalization calculation formula of T1 is , and the result is . After normalization, the complete expected subjective quality score matrix is obtained, which provides accurate input for the standard deviation calculation and threshold judgment of the S4 step.

[0017] As shown in Figure 3 , the step S4: calculating the standard deviation of each element in the expected subjective quality score matrix, if the standard deviation exceeds the preset quality difference threshold, executing the quality alignment decision process to generate the code control parameter adjustment instruction set of each encoding node. Specifically comprising: S4.1: based on the quality score data of each participating terminal in the expected subjective quality score matrix, calculating the standard deviation of the quality score to quantify the visual quality dispersion degree under different encoding formats in the current conference; S4.2: comparing the calculated standard deviation with the preset quality difference threshold to judge whether there is a visual quality inconsistency problem caused by the difference in encoding formats in the current conference; The standard deviation calculation result of the expected subjective quality score matrix is used as input data, and the threshold comparison algorithm (parameter: quality difference threshold ε) is used to realize the quantitative comparison between the standard deviation value and the quality difference threshold. Further, by floating-point precision unification processing (parameters: double precision 64-bit, rounding mode is even rounding), the calculation precision consistency of the standard deviation value and the threshold value is realized, and a set of comparison variables that can directly participate in logical judgment is obtained; Further, by using a double conditional branch decision algorithm (parameters: condition one is standard deviation > ε, condition two is standard deviation ≤ ε), the binary classification result output of the visual quality consistency state in the current conference is realized, and the state identification code corresponding to the decision result is generated; Further, by using a preset state mapping table (parameters: state code 0 corresponds to consistency satisfied, state code 1 corresponds to consistency not satisfied), the mapping of the state identification code to the quality consistency semantic label is realized, and readable state information for triggering the subsequent decision process is obtained; By conditional judgment and state mapping processing, the standard deviation value of the previous step is converted into a quality consistency judgment result, and the existence of visual quality inconsistency problems caused by coding format differences is quickly identified; For example, in a cross-platform video conference scene containing four participants, the expected subjective quality scores of each participant are 7.8, 7.5, 8.2, and 6.9, respectively. The average value and standard deviation of the subjective quality scores are calculated, and the average value is 7.6 and the standard deviation is about 0.55. When the threshold value ε is set to 0.5, after floating-point precision unification processing, the comparison result is 0.55>0.5, and the corresponding state code is 1, which is mapped to "visual quality inconsistency". This result triggers the quality alignment subsequent process, so that the system adjusts the code control parameters in the next cycle, and the dispersion degree of the conference video quality is significantly reduced; S4.3: If the standard deviation exceeds the preset quality difference threshold, generate a quality alignment trigger signal as a control input condition for starting the subsequent coding parameter adjustment process; S4.4: Based on the quality scores of each participant in the quality score matrix, calculate the deviation direction of each coding node relative to the overall score average, and determine the node set that needs to improve or reduce the coding quality; S4.5: According to the deviation direction and deviation amplitude of the coding node, generate a corresponding code control parameter adjustment instruction set, including quantization matrix optimization instructions and code rate compression instructions, which are used to drive the subsequent encoder to perform dynamic quality alignment operation; The input conditions are the deviation direction and deviation amplitude data of each coding node relative to the overall score average, which are output by the S4.4 sub-step and include coding node identification, deviation direction (positive or negative), and deviation amplitude value; By using a deviation amplitude normalization method (parameters: maximum deviation value, minimum deviation value), the normalization processing of the deviation amplitude of different coding nodes is realized, so as to determine the code control adjustment ratio under the same dimension. Further, by quantizing matrix optimization strategy mapping algorithm (parameters: encoding format identification, device type, network level), the precise matching of the coding node with a negative deviation and the preset optimization template is realized, and a set of quantization parameter adjustment rules is obtained; Further, by using a code rate compression strategy matching algorithm (parameters: encoding format identification, current code rate, target compression ratio, network bandwidth constraint), the code rate compression template retrieval of the coding node with a positive deviation is realized, and a set of code control parameter compression rules is generated; Further, based on the instruction set construction model (parameters: quantization matrix adjustment rule set, code rate compression rule set), the above rule set is structured and encapsulated into a code control parameter adjustment instruction set, which contains the adjustment type, specific parameter value and execution priority of each coding node; By using an instruction sequence optimization algorithm (parameters: execution priority, dependency table), the instruction set of the previous step is converted into an execution sequence that can directly drive the encoder control interface, and the dynamic alignment control of the coding quality of each coding node is realized; For example, in a cross-platform conference scene containing four coding nodes (using H.264, H.265, VP9 and AV1 encoding formats respectively), the deviation amplitude data is: node 1 (H.264) =-4.5, node 2 (H.265) =-2.0, node 3 (VP9) =+3.0, and node 4 (AV1) =+1.5. The normalization processing uses the minimum-maximum normalization formula:

[0018] wherein, is the normalized deviation amplitude, is the original deviation amplitude, is the minimum deviation value (-4.5), For the maximum deviation value (+3.0), the negative amplitudes of node 1 and node 2 after normalization are 0.0 and 0.3125, respectively, and the positive amplitudes of node 3 and node 4 are 1.0 and 0.875, respectively. The nodes with negative deviations adopt the high-frequency detail enhancement template in the quantization matrix optimization strategy library, and the high-frequency QP reduction amplitude is set to be the normalized deviation x 2.0, and the low-frequency QP increase amplitude is the normalized deviation x 1.0; the nodes with positive deviations adopt the code rate compression template, and the code rate compression ratio is set to be the normalized deviation x 5%, and the bandwidth release ratio is the normalized deviation x 3%. The finally generated instruction set includes: node 1 executes the quantization optimization instruction of high-frequency QP reduction by 9 and low-frequency QP increase by 4; node 2 executes the quantization optimization instruction of high-frequency QP reduction by 6 and low-frequency QP increase by 2; node 3 executes the compression instruction of code rate compression by 5% and bandwidth release by 3%; and node 4 executes the compression instruction of code rate compression by 4.375% and bandwidth release by 2.625%. These instructions are packaged into an encoding interface execution sequence according to priority, and in actual tests, the quality scores of the nodes tend to be consistent, and the global visual quality of the conference is significantly improved.

[0019] The step S5: according to the code control parameter adjustment instruction set, performing quantization matrix optimization processing on the encoding nodes with quality scores lower than the average, and performing code rate compression processing on the encoding nodes with quality scores higher than the average, to generate a video stream after dynamic quality alignment. Specifically comprising: S5.1: based on the encoding format, device type identification and current network state level of the encoding nodes with quality scores lower than the average, matching the corresponding optimization template from the preset quantization matrix optimization strategy library, wherein the optimization template contains quantization parameter (QP) adjustment rules and weight distribution strategies for different encoders; S5.2: locally adjusting the original quantization matrix of the encoding nodes by using the matched quantization matrix optimization template, specifically including: reducing the QP value of the coefficients in the high-frequency texture area to enhance the detail retention capability, and increasing the QP value of the coefficients in the low-frequency flat area to save code rate resources, to generate an optimized quantization matrix parameter set; The input condition is the matched quantization matrix optimization template, and the template presets the quantization parameter adjustment rules and weight distribution strategies of different areas according to the encoding format, device type identification and current network state level of the encoding nodes; The matrix partition mapping method (parameters: area division rules and weight coefficient matrix in the optimization template) is adopted to realize the positioning and marking of the coefficients in the high-frequency texture area of the original quantization matrix; Further, the quantization parameter decrement algorithm (parameter: high-frequency area QP adjustment step size ΔQPf, set according to the optimization template) is adopted to realize the reduction of the quantization step value corresponding to the high-frequency texture area, to enhance the detail retention capability, and obtain an adjusted high-frequency area QP coefficient set; Further, the quantization step value corresponding to the low-frequency flat region is increased by a quantization parameter increment algorithm (parameter: low-frequency region QP adjustment step size AQPl) to compress redundant information and save code rate resources, and an adjusted low-frequency region QP coefficient set is obtained. Further, a matrix reorganization operation (parameter: high-frequency region adjustment coefficient set and low-frequency region adjustment coefficient set) is used to realize local replacement and overall reconstruction of the original quantization matrix, and an optimized quantization matrix parameter set is generated. A quantization matrix consistency checking method (parameter: global code rate constraint and visual quality tolerance threshold in the optimization template) is used to verify the compliance of the parameter range and code rate prediction value of the optimized quantization matrix, and the final parameter matrix that can be used by the encoder in real time is obtained. Through matrix regional adjustment and parameter reconstruction processing, the quantization matrix optimization instruction set of the previous step is converted into optimized parameters that can be directly written into the encoder register, the high-frequency detail restoration degree is improved, and the bandwidth is saved in the low-frequency flat region, and the dynamic quality alignment technology effect under different encoding formats is realized. For example, for a professional conference terminal supporting H.265 encoding, the device type identifier is a desktop terminal, the network status level is medium, the high-frequency region QP adjustment step size AQPf in the matched quantization matrix optimization template is set to , and the low-frequency region QP adjustment step size AQPl is set to . The high-frequency region coefficient range of the original quantization matrix is , and the low-frequency region coefficient range is . Through the matrix partition mapping method, the high-frequency region is located, and the quantization parameter decrement algorithm is used to adjust the coefficient range to , and the low-frequency region is adjusted by the quantization parameter increment algorithm to . The matrix reorganization operation completes local replacement and generates an optimized quantization matrix, the code rate prediction value is reduced from the original Mbps to Mbps, the high-frequency detail fidelity score is significantly improved, while meeting the global code rate constraint and visual quality tolerance threshold. The optimized quantization matrix is successfully loaded into the encoder, the conference picture presents clearer and sharper edges in high-texture regions, and smooth transitions in low-texture regions, thereby realizing subjective quality consistency improvement under different encoding formats. S5.3: For the encoding node with a quality score higher than the average, based on the current code rate, target compression ratio and bandwidth constraint conditions, a dynamic code rate compression algorithm is used to adjust the code control parameters of the encoder, the dynamic code rate compression algorithm includes a sliding window-based code rate allocation model and a content complexity-based rate-distortion optimization function. S5.4: Inject the optimized quantization matrix parameter set and the adjusted code control parameter into the corresponding encoding node respectively, write into the parameter register of the video encoder through the encoder control interface, to realize real-time intervention and dynamic regulation and control of the encoding process, and generate the preliminary quality-aligned video encoding stream; S5.5: Real-time quality evaluation is performed on the preliminary quality-aligned video encoding stream, SEI extension metadata, key frame interval and quantization parameter statistical value in the video stream are extracted, combined with the no-reference quality evaluation model (such as NIQE or BRISQUE), the subjective quality score after alignment is calculated, and the quality feedback information after dynamic quality alignment is generated; For the preliminary quality-aligned video encoding stream, the SEI extension metadata analysis method (parameters: frame-level transport stream, extension information identifier) is used to realize the structured extraction of the coding quality related signaling embedded in the video stream; Further, by using the key frame statistical algorithm (parameters: GOP structure, timestamp sequence), the accurate calculation of the key frame interval is realized, and the time interval data reflecting the frame refresh frequency is obtained, which is used for comprehensive evaluation of the ability of the encoder to retain motion and texture details in the video stream; Further, by using the quantization parameter sampling method (parameters: macroblock-level QP value distribution, frame average QP value), the global and local statistics of the code control strength in the quantization matrix execution process are realized, and the quantization parameter statistical value matrix is generated, which provides quantitative basis for the subsequent quality evaluation on the encoding control side; Further, the SEI extension metadata, key frame interval data, and quantization parameter statistical value extracted above are used as inputs, and the no-reference quality evaluation model (such as NIQE, BRISQUE) is used to score the quality of the video stream. For the NIQE model, a low-dimensional feature space is constructed based on the statistical characteristics of natural scenes, the model matching error is calculated, and the natural statistical deviation degree of the video stream is calculated by the following mean square error formula:

[0020] wherein, is the no-reference feature score of the i-th content segment, is the expected feature score, is the total number of content segments; Further, the weighted fusion algorithm (parameters: SEI metadata weight, GOP interval weight, QP statistical weight) is used to weight and sum the NIQE and BRISQUE model scores to generate the aligned comprehensive subjective quality score, forming the quality feedback information after dynamic quality alignment, which is used for subsequent quality mapping model bias correction; The preliminary quality alignment video stream of the previous step is converted into a quantifiable comprehensive subjective quality score index through real-time quality evaluation and score generation processing, and the expected technical effect of dynamic quality consistency monitoring is achieved. For example, in a cross-platform video conference scenario, the SEI extension metadata of the video stream is parsed to obtain an encoding quality identification value of 85, the key frame interval data extraction uses a GOP length of 60 frames and a time interval of 2 seconds, and the quantization parameter statistical value matrix shows that the global average QP value is 28 and the QP value of the high-frequency texture region is reduced to 24. The above data is input into the NIQE model, and the natural statistical deviation is calculated to be 0.054 based on the statistical characteristics of the natural scene.

[0021] The calculation result is 0.054. The score is weighted and fused with the score 34.2 obtained by the BRISQUE model, and the fusion weight configuration is that the SEI metadata score weight is 0.4, the GOP interval score weight is 0.3, and the QP statistical score weight is 0.3. The final comprehensive subjective quality score is 36.8. Under this score result, it is verified that the subjective visual experience of different terminal participants tends to be consistent, the high-frequency details are obviously retained, and the video stream playing stability is significantly improved.

[0022] The step S6: real-time collection of quality feedback information in the SEI extension metadata of each encoding node, including decoding failure rate and Jitter Buffer statistical value, deviation correction of the prediction result of the PQMM model, and generation of updated quality mapping model parameters. Specifically, it includes: S6.1: Collect and normalize the decoding failure rate reported by each encoding node in the SEI extension metadata to obtain a standardized decoding anomaly index sequence, which is used to reflect the decoding stability problems caused by encoding format or network fluctuations in actual transmission; S6.2: Based on the collected Jitter Buffer statistical value, calculate the video frame buffer delay fluctuation of each encoding node at the receiving end, and use the sliding window mean filtering algorithm to smooth the fluctuation to obtain the encoding transmission delay stability feature, which is used to evaluate the indirect influence of network state on video quality; S6.3: Feature fusion of the decoding anomaly index sequence and the video frame buffer delay stability feature to generate a multi-dimensional quality feedback feature vector as the input basis for deviation correction of the PQMM model to reflect the system error between model prediction and real quality in actual running environment; S6.4: Based on the multi-dimensional quality feedback feature vector and the original predicted quality score of the PQMM model, the quality prediction deviation value of each encoding node in the current period is calculated, and the weighted least squares method is used for regression fitting to obtain a model prediction error compensation function; S6.5: The output score of the PQMM model is dynamically corrected using the model prediction error compensation function, and updated quality mapping model parameters are generated, so that the PQMM model has stronger environmental adaptability and prediction robustness in subsequent quality alignment decisions; The model prediction error compensation function parameters obtained by weighted least squares regression fitting are corrected dynamically (parameters: error compensation coefficient vector, feature weight matrix) to realize real-time value correction of the PQMM model output score; Further, by using a piecewise linear adjustment algorithm (parameters: prediction deviation threshold grouping, quality score upper and lower limit constraint), the compensation amplitude is adaptively distributed for different quality score intervals, and a correction coefficient mapping table is obtained; Further, using the correction coefficients corresponding to each encoding node in the mapping table, the output score matrix of the PQMM model in the current decision period is element-wise weighted and corrected to generate a quality correction matrix; Further, by matrix fusion operation (parameters: original score matrix, quality correction matrix, fusion weight), the compensated score data and the original score are smoothly synthesized to generate an updated score matrix, reducing the influence of transient network fluctuations on score stability; Further, using a parameter backwriting mechanism, the updated score matrix is back-propagated to the PQMM model parameter vector to correct the coefficients of the texture complexity, motion intensity and chroma distribution weights in the model, so that the PQMM model has stronger environmental adaptability and prediction robustness in subsequent decisions; Through the dynamic correction processing mode, the prediction error compensation result of the previous step is converted into PQMM model update parameters, realizing the expected technical effect of improving the quality prediction stability for different network states and terminal characteristics; For example, in a cross-platform video conference scenario, the original PQMM prediction score matrix of a certain encoding node is [[72.4, 69.8], [75.3, 73.5]], and the error compensation function obtained by weighted least squares fitting of the quality feedback feature vector is

[0023] wherein, The original score bias is 0.85, the slope parameter is 0.85, and the intercept parameter is 5.2. In the correction process, the mapping table assigns a compensation of 7.5 points to the interval with a score lower than 70, a compensation of 5.0 points to the interval with a score between 70 and 75, and no compensation to the interval with a score higher than 75. In this case, the compensation is applied to the score 69.8:

[0024] The result is 77.3. The compensation is applied to the score 72.4:

[0025] The result is 77.4. After matrix-based element-wise weighted fusion (fusion weight: 0.6 original score, 0.4 compensation score), the updated score matrix is obtained as [[74.84, 75.32], [75.3, 73.5]]. When the updated score matrix is back-propagated to the PQMM model weight coefficients, the texture complexity weight increases from the original 0.28 to 0.33, the motion intensity weight decreases from the original 0.42 to 0.39, and the chroma distribution weight remains unchanged at 0.30. Finally, in the quality alignment decision-making of the next cycle, the updated model parameters achieve significant improvement in score stability under weak network conditions, and the participant's image quality consistency is significantly improved.

[0026] The step S7: based on the updated quality mapping model parameters, performing periodic model retraining, retaining the quality feedback data in the recent preset time length through a sliding window mechanism to form a dynamically optimized quality alignment model. Specifically, it includes: S7.1: based on the updated quality mapping model parameters, constructing a model retraining input data set, the input data set containing quality feedback data in the recent preset time window, the quality feedback data including decoding failure rate, Jitter Buffer statistical value, and corresponding encoding format, device type, and network state level; In the first sub-step of performing periodic model retraining, the updated quality mapping model parameters are used as the basic configuration input for algorithm execution, and the quality feedback data set stored in the system running monitoring module is called. The quality feedback data set is limited to the sampling range within the recent preset time window, which contains the decoding failure rate sequence, Jitter Buffer statistical value sequence, and associated encoding format, device type identifier, and network state level parameters of each encoding node in different transmission cycles; A data index matching algorithm (parameter: timestamp index range = preset window length) is used to realize time period filtering of the feedback records stored in the running monitoring module to ensure that the extracted data set only covers recent valid samples; Further, through the feedback data parser (parameters: field mapping table including decoding failure rate, buffer delay fluctuation, encoding format, device type, network level), the structured mapping of the original record is realized, and a multi-dimensional data record set classified by encoding nodes is obtained; Further, by using the feature labeling algorithm (parameters: encoding format mapping matrix, device type priority table, network level weight coefficient), the feature label supplement processing of each data record is realized, and the encoding parameter environment features of each quality feedback sample are explicitly associated in the record set, generating a labeled data set with environmental context; Further, by using the data consistency checking method (parameters: packet loss rate threshold = 0.05, delay fluctuation threshold = 50ms), the samples with abnormal values or missing values in the labeled data set are removed, and a purified data set meeting the consistency requirements is generated, ensuring the quality stability of the subsequent training input; Further, by using the data structure reorganization algorithm (parameters: feature field sorting strategy = encoding format → device type → network level → decoding failure rate → buffer delay), the feature fields of each sample in the purified data set are arranged in a predetermined order, and a unified format model retraining input data set structure is generated; Through the above index matching, field parsing, feature labeling, consistency checking and structure reorganization processing methods, the updated quality mapping model parameters obtained in the previous step and the quality feedback samples in the recent window are converted into structured input data that can be directly used for PQMM model training, realizing the efficient self-adaptive retraining ability of the model in the dynamic environment; Exemplarily, in a cross-platform video conference system comprising 12 encoding nodes, the preset time window length is 30 minutes, the system running monitoring module records feedback samples of each node once per second, and the total number of samples is 21600. When the index matching algorithm is executed, only the sample records with timestamps within the last 30 minutes are retained, obtaining 21600 data. The parser maps each record to five fields of decoding failure rate (unit: failed frames / total frames), jitter buffer delay fluctuation (unit: milliseconds), encoding format (such as H.264), device type (such as desktop terminal), and network level (such as medium level). In the feature labeling process, the format code F1 is added for the H.264 format according to the encoding format mapping matrix, the device code D2 is added for the desktop terminal, the network code N2 is added for the medium level network, and the comprehensive feature label T(F1, D2, N2) is embedded in the record. The consistency check eliminates a total of 152 abnormal samples with delay fluctuation exceeding 50 ms or packet loss rate exceeding 0.05, and the remaining 21448 valid data. The structure reorganization algorithm restructures the valid samples into a vector form according to the preset field order, forms an input matrix of 21448x5, and directly inputs the input matrix into the PQMM model for incremental training, so as to realize the parameter optimization of the model in the current multi-encoding mixed environment and prepare the basis data for the model; S7.2: The quality feedback data is truncated in the time dimension by using a sliding window mechanism to retain sample data within a preset time length, and the sample data is normalized to obtain a standardized model training input feature vector; S7.3: Based on the standardized model training input feature vector, the current PQMM model parameters are updated incrementally by using a weighted least squares method to reduce the residual sum of squares between the model prediction value and the actual feedback quality score, and an updated model parameter set is obtained; S7.4: The updated model parameter set is subjected to convergence judgment, if the parameter change is less than a preset convergence threshold, it is judged that the model has converged, otherwise the next round of model iteration training is continued to obtain higher precision of quality prediction ability; S7.5: The updated model parameters after convergence are deployed to the PQMM model inference unit of each encoding node to realize unified update of the model version, and the model update timestamp and version identification are recorded to form a traceable model iteration log.

[0027] The application also provides a cross-platform video conference fusion system which adopts the above-mentioned cross-platform video conference fusion method to perform cross-platform video conference fusion.

[0028] So far, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical solutions after the changes or replacements will all fall within the protection scope of the present application.

[0029] The above only describes the preferred embodiments of the present application and is not intended to limit the present application; the present application can have various changes and variations for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included in the protection scope of the present application.< / float>

Claims

1. A cross-platform video conference fusion method, characterized in that, The method comprises the following steps: S1: Collecting the encoding capability information reported by different terminals in the cross-platform video conference system, and generating an encoding capability configuration table; S2: Based on the texture complexity, motion intensity and color distribution characteristics in the historical test data set, an input feature space of the encoding perceptual quality mapping model is constructed, and the historical test data set contains subjective quality scores of H.264, H.265, VP9 and AV1 four encoding formats under different code control parameters; S3: Matching the encoding format parameters in the encoding capability configuration table with the preset encoder characteristic parameter library, calling the corresponding encoding perceptual quality mapping model instance to predict the quality of the current code rate strategy, and generating an expected subjective quality score matrix of each participant terminal; S4: Calculating the standard deviation of each element in the expected subjective quality score matrix, if the standard deviation exceeds the preset quality difference threshold, a quality alignment decision process is executed, and a code control parameter adjustment instruction set of each encoding node is generated; S5: According to the code control parameter adjustment instruction set, the quantization matrix optimization processing is performed on the encoding nodes with quality scores lower than the average, and the code rate compression processing is performed on the encoding nodes with quality scores higher than the average, and a dynamic quality aligned video stream is generated; S6: Real-time collection of quality feedback information in SEI extension metadata of each encoding node, including decoding failure rate and Jitter Buffer statistical value, deviation correction of the prediction result of the encoding perceptual quality mapping model, and generation of updated encoding perceptual quality mapping model parameters.

2. The cross-platform video conference fusion method of claim 1, wherein, The step S6 further comprises: S7: Based on the updated encoding perceptual quality mapping model parameters, periodic model retraining is performed, and quality feedback data in the latest preset time period is retained through a sliding window mechanism to form a dynamically optimized quality alignment model.

3. The cross-platform video conference fusion method of claim 1, wherein, The step S1 specifically comprises: Based on the capability negotiation signaling reported by the terminal device during the conference access stage, the supported video encoding format list is obtained, including one or more combinations of H.264, H.265, VP9 and AV1, forming an encoding format set; The received terminal device identification information is parsed, and the device type identification field is extracted, including classification information such as mobile terminal, desktop terminal or professional conference terminal; The real-time network bandwidth, packet loss rate and delay indicators output by the network state monitoring module reported by the terminal are collected and quantized to generate network state level parameters, which are divided into three levels of high, medium and low for encoding strategy adaptation; The encoding format set, the device type identification and the network state level parameter are structured and packaged to generate a terminal encoding capability description tuple, and uploaded to the configuration management unit of the media processing module through the session control channel; Based on the encoding capability description tuples of all participant terminals, aggregation processing is performed to generate a conference-level encoding capability configuration table, which includes four-dimensional information of terminal identification, encoding format, device type and network level.

4. The cross-platform video conference fusion method of claim 3, wherein, The step S1 further comprises parsing the supported encoding format field in the capability negotiation signaling by a protocol parsing algorithm, generating an effective encoding format set by a deduplication and validity detection algorithm, and generating a device type identifier by combining device type mapping, regular pattern verification and hierarchical weight configuration; the network state is monitored and scored by multi-index weighting, and is classified into high, medium and low levels, and is structured and packaged.

5. The cross-platform video conference fusion method of claim 1, wherein, The step S2 specifically comprises: Performing content analysis-based feature extraction processing on the original video frame sequence in the historical video conference data set to obtain a texture complexity feature parameter, which is calculated based on a joint calculation of a local binary pattern and a gray level co-occurrence matrix; Performing motion intensity feature extraction processing based on an optical flow method on the original video frame sequence to obtain a motion vector amplitude distribution statistic, wherein the motion intensity feature extraction is based on Horn-Schunck optical flow algorithm to estimate the pixel displacement between continuous frames, and outputs the mean and variance of the inter-frame motion vector amplitude as motion intensity representation parameters; Performing color distribution feature extraction processing based on the HSV color space on the original video frame sequence to obtain color histogram statistical features, wherein the color distribution feature extraction is based on histogram statistics on the H channel after converting the RGB image to the HSV space, and outputs the color distribution entropy and the dominant color tone concentration index; Performing feature normalization and fusion processing on the texture complexity feature parameter, the motion intensity representation parameter, and the color distribution entropy and dominant color tone concentration index, generating a standardized multi-dimensional content feature vector, and performing feature dimension reduction by principal component analysis, and outputting a low-dimensional compact feature representation; Performing feature space division processing based on the multi-dimensional content feature vector and a pre-labeled subjective quality score data set to generate an input feature space of the encoding perceptual quality mapping model.

6. The cross-platform video conference fusion method of claim 1, wherein, The step S3 specifically comprises: Parsing the encoding format parameters in the encoding capability configuration table, extracting the video encoding format identifiers supported by each participating terminal, and determining the matching index in the corresponding encoder characteristic parameter library; Based on the video encoding format identifier, retrieving the model calling interface of the corresponding encoder from the pre-set encoder characteristic parameter library, wherein the encoder characteristic parameter library contains parameter configuration information of four encoding formats of H.264, H.265, VP9 and AV1; Using the encoding perceptual quality mapping model instance to predict the encoding quality of the code rate strategy under the current conference scene, inputting the texture complexity, motion intensity and color distribution features of the current video content, and outputting the expected subjective quality score matching the encoding format and content features; Performing the encoding perceptual quality mapping model prediction on different encoding formats used by each participating terminal under the same video source to obtain the subjective quality score values of each encoding node under the current code rate strategy, and forming a quality score vector; Matrixing the quality score vector to generate an expected subjective quality score matrix.

7. The cross-platform video conference fusion method of claim 6, wherein, The rows of the expected subjective quality score matrix correspond to different participating terminals, and the columns correspond to the quality scores under their encoding formats.

8. The cross-platform video conference fusion method of claim 1, wherein, The step S4 specifically comprises: calculating a standard deviation of the quality scores based on the quality score data of each participant in the expected subjective quality score matrix; comparing the standard deviation with a preset quality difference threshold to determine whether there is a visual quality inconsistency problem caused by the difference in encoding format in the current conference; if the standard deviation exceeds the preset quality difference threshold, generating a quality alignment trigger signal as a control input condition for starting a subsequent encoding parameter adjustment process; calculating the deviation direction of each encoding node relative to the overall score mean based on the quality score data of each participant in the expected subjective quality score matrix, and determining a node set that needs to improve or reduce the encoding quality; generating a corresponding code control parameter adjustment instruction set according to the deviation direction and deviation amplitude of the encoding node, including a quantization matrix optimization instruction and a code rate compression instruction.

9. The cross-platform video conference fusion method of claim 1, wherein, The step S5 specifically includes: based on the encoding format, device type identifier, and current network state level of the encoding node with a quality score lower than the mean, matching a corresponding optimization template from a preset quantization matrix optimization strategy library; using the optimization template to locally adjust the original quantization matrix of the encoding node to generate an optimized quantization matrix parameter set; for the encoding node with a quality score higher than the mean, adjusting the code control parameters of the encoder based on the current code rate, target compression ratio, and bandwidth constraint conditions using a dynamic code rate compression algorithm to generate adjusted code control parameters; injecting the optimized quantization matrix parameter set and the adjusted code control parameters into the corresponding encoding node, respectively, and writing them into the parameter register of the video encoder through the encoder control interface to generate a preliminary quality-aligned video encoding stream; performing real-time quality evaluation on the preliminary quality-aligned video encoding stream, extracting the SEI extension metadata, key frame interval, and quantization parameter statistical value in the video stream, and combining a no-reference quality evaluation model to calculate the aligned subjective quality score to generate dynamic quality-aligned quality feedback information.

10. A cross-platform video conference fusion system, characterized in that: The cross-platform video conference fusion method of any one of claims 1-9 is used for cross-platform video conference fusion.