Deep Learning-Based Real-Time QR Code Interactive System for Digital Humans

Abstract

The present invention discloses a two-dimensional code real-time digital human interaction system based on deep learning, which relates to the field of two-dimensional code digital human interaction technology. The system includes an identity recognition and extraction module, a model screening and construction module, a trigger key evaluation module, a sequence control execution module, and an interaction feedback optimization module; the model screening and construction module filters out all deep learning models loaded in parallel in the current interaction from the deep learning model list and establishes a multi-dimensional interaction loading model set; the trigger key evaluation module obtains the loading control information of each deep learning model in the multi-dimensional interaction loading model set, evaluates the interaction trigger criticality of each deep learning model based on the loading control information, and classifies each deep learning model according to the evaluation results. The present invention solves the problem of inaccurate judgment of the loading priority of multiple models and realizes intelligent control and efficient response of the loading order of digital human interaction.

Description

Technical Field

[0001] The present invention relates to the field of two-dimensional code digital human interaction technology, and in particular to a two-dimensional code real-time digital human interaction system based on deep learning. Background Art

[0002] With the rapid development of artificial intelligence, particularly deep learning, virtual digital humans have made significant progress in speech recognition, emotion understanding, image generation, and natural language processing, bringing a more natural and immersive experience to human-computer interaction systems. Furthermore, QR codes, as an efficient information portal, have been widely used in scenarios such as identity recognition, data indexing, and triggering interactive portals. This has led to the development of a hybrid technology approach that integrates deep learning and QR code recognition: deep learning-based real-time digital human interaction with QR codes. This technology deploys a unique QR code identifier on a physical carrier or user terminal. Users can trigger the system portal by scanning the code. The system then invokes a pre-set or matched digital human model, presenting a personalized virtual interactive avatar in real time. This supports various human-computer interaction modes, including voice dialogue, graphic content output, and dynamic expression generation. The entire process leverages the deep learning model's ability to process multimodal information, including speech, image, and language, to achieve personalized digital human responses, providing technical support for digital presentation and emotional connection in multiple scenarios.

[0003] Existing deep learning-based QR code digital human interaction systems typically rely on a combination of user scanning, cloud-based model scheduling, and multimodal information processing to build a virtual interactive platform that supports customized responses. A typical implementation process includes: First, a unique QR code is set on the target entity or medium, and the user triggers the system access portal by scanning the QR code with a mobile device. Subsequently, the system calls the associated digital human model based on the identity information or index parameters carried by the QR code. This model, built using deep learning technology, integrates multiple features such as voice, image, and text to generate a virtual interactive avatar with personalized expression. Next, the system calls modules such as natural language understanding models, emotion recognition algorithms, and semantic graphs to parse user input and identify intent. Finally, a deep generative model outputs real-time voice, text, or action responses, completing the dynamic presentation and multi-dimensional interaction of the digital human on the terminal. The entire process encompasses multiple closely coordinated deep learning-supported links, including QR code parsing, identity information matching, model loading and scheduling, input analysis, response generation, and front-end rendering.

[0004] The existing technology has the following deficiencies:

[0005] When a user scans a QR code on their mobile device to access the interactive system interface, the system launches multiple deep learning models in parallel in the background to support the digital human's voice generation, text generation, and image rendering functions. This "parallel model loading" can easily lead to some models loading first. Because the current model scheduling mechanism prioritizes model size and average response speed, lacking a comprehensive assessment of each model's functional necessity during the initial interaction phase, image-driven models, such as those with smaller sizes and faster loading times, are prioritized, while the text generation model, which actually carries the core interactive content, has not yet finished loading. This can lead to the digital human displaying animated movements but failing to simultaneously output speech or text, giving users the illusion of "system unresponsiveness" or "program anomalies." Existing deep learning-based real-time digital human interaction technologies for QR codes cannot dynamically adjust the model loading order based on the criticality of interaction triggers when multiple models are loaded in parallel. This results in a lack of a task-criticality assessment mechanism for model scheduling, leading to delayed loading of core interaction models. This ultimately severely impacts users' perception of system stability and interaction continuity, and can easily lead to user abandonment during initial use.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0007] The purpose of the present invention is to provide a two-dimensional code real-time digital human interaction system based on deep learning to solve the problems in the above-mentioned background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solutions: a QR code real-time digital human interaction system based on deep learning, comprising an identity recognition and extraction module, a model screening and construction module, a trigger key evaluation module, a sequence control execution module, and an interaction feedback optimization module;

[0009] The identity extraction module parses the identity identifier passed in by the user scanning the QR code and extracts the list of deep learning models used to implement the digital human interaction function based on the identity identifier;

[0010] The model screening construction module filters out all deep learning models loaded in parallel in the current interaction from the deep learning model list and establishes a multi-dimensional interactive loading model set;

[0011] The triggering criticality assessment module obtains the loading control information of each deep learning model in the multi-dimensional interactive loading model set, evaluates the interaction trigger criticality of each deep learning model based on the loading control information, and classifies each deep learning model according to the evaluation results;

[0012] The sequential control execution module pre-sorts the loading order of each deep learning model according to the classification results, executes the loading operation according to the pre-sorted order, and dynamically controls the loading order during the loading process based on the current resource status and execution progress;

[0013] The interactive feedback optimization module executes the digital human interactive response after loading is completed and records the data during the loading process to optimize the subsequent loading sequence control strategy.

[0014] Preferably, in the model screening and construction module, by comparing the function labels of each deep learning model in the deep learning model list with the function call requirements of the current interaction process, the deep learning models that need to be loaded in parallel in the current interaction are screened out, and the screened deep learning models are grouped according to the speech generation class, text generation class and image rendering class, and the grouping results are used to construct a multi-dimensional interactive loading model set.

[0015] Preferably, in the trigger critical evaluation module, the loading control information of each deep learning model in the multidimensional interactive loading model set is obtained and preprocessed after acquisition; the structural interaction topology information and resource load index information are extracted from the preprocessed loading control information of each deep learning model, and analyzed after extraction to generate the topological coupling coefficient and load impact index of each deep learning model respectively; an interactive trigger criticality evaluation model is constructed for the topological coupling coefficient and load impact index of each generated deep learning model, and the interactive critical coefficient of each deep learning model is generated by weighted summation; a pre-set interactive critical coefficient threshold range is determined, and after determination, it is compared with the interactive critical coefficient of each generated deep learning model, and the interactive trigger criticality of each deep learning model is evaluated according to the comparison result, and each deep learning model is divided into a loading-first model, a loading-parallel model and a loading-delayed model according to the evaluation result.

[0016] Preferably, the specific acquisition logic of the topological coupling coefficient of each deep learning model is as follows:

[0017] The structural interaction topology information is extracted from the pre-processed loading control information of each deep learning model, including the number of output interfaces of each deep learning model, the number of dependent models, and the number of calls in the preset historical time period, and marked as 、 and , Indicates the The number of output interfaces of a deep learning model, Indicates the The number of models that a deep learning model depends on, Indicates the The number of times a deep learning model is called within a preset historical time period, , is a positive integer;

[0018] Calculate the topological coupling coefficient of each deep learning model The specific calculation method is: The number of times a deep learning model is called within a preset historical time period Take the square root to get the first part of the calculation value; The number of output interfaces of a deep learning model and the number of dependent models Multiply them, add one to the product, and then take the natural logarithm of the value to get the second part of the calculation value; add the first part of the calculation value and the second part of the calculation value to get the first part of the calculation value. Topological coupling coefficient of deep learning models .

[0019] Preferably, the specific logic for obtaining the load impact index of each deep learning model is as follows:

[0020] The resource load index information is extracted from the pre-processed loading control information of each deep learning model, including the model file volume, memory usage peak and static allocation capacity of each deep learning model on the GPU memory, and calibrated as 、 and , Indicates the The model file size of a deep learning model, Indicates the The peak memory usage of deep learning models, Indicates the The static allocation capacity of a deep learning model on GPU memory, , is a positive integer;

[0021] Calculate the load impact index of each deep learning model The specific calculation method is: Model file size of a deep learning model , peak memory usage Statically allocated capacity on GPU memory Sum the natural logarithms after adding one to get the Load impact index of deep learning models .

[0022] Preferably, the topological coupling coefficients of each deep learning model generated are and load impact index Construct an interaction trigger criticality evaluation model and generate the interaction criticality coefficients of each deep learning model through weighted summation .

[0023] Preferably, a preset interaction key coefficient threshold interval is determined , and after determination, the interaction key coefficients of each deep learning model generated A comparison is performed, and the interaction triggering criticality of each deep learning model is evaluated based on the comparison results. Based on the evaluation results, each deep learning model is divided into loading-first model, loading-parallel model, and loading-delay model. The specific comparison analysis and classification are as follows:

[0024] like , the interaction triggering criticality of the deep learning model is low, and the deep learning model is classified as a loading delay model;

[0025] like ,The interaction trigger criticality of the deep learning model is medium, and the deep learning model is classified as a loading parallel model;

[0026] like , the interaction trigger criticality of the deep learning model is high, and the deep learning model is classified as a loading-first model.

[0027] Preferably, in the sequential control execution module, the loading order of each deep learning model is pre-sorted according to the classification results, and the loading operation is performed according to the pre-sorted order: the loading precedent class models are sorted from high to low according to the interaction key coefficient, and the loading operations are performed in sequence; in the process of loading the precedent class models, the loading parallel class models are sorted from low to high according to the resource occupancy level, and the loading operations are started in sequence; after the loading of the preceding class model and the loading of the parallel class model is completed, the loading of the delayed class model is executed in sequence.

[0028] Preferably, in the sequential control execution module, the loading order is dynamically controlled during the loading process in combination with the current resource status and execution progress. Specifically, during the loading process, the current available memory capacity, GPU video memory remaining amount and CPU occupancy rate are continuously obtained, and the loading order is dynamically controlled in combination with the loading progress of each deep learning model and the loading order of the models that have not yet been loaded, specifically including: pausing the subsequent loading of delayed class models, adjusting the remaining loading order of parallel class models, and giving priority to loading deep learning models whose resource occupancy is lower than the set threshold.

[0029] In the above technical solution, the technical effects and advantages provided by the present invention are:

[0030] 1. This invention constructs two types of evaluation parameters, the topological coupling coefficient and the load impact index, and integrates them to generate the interaction key coefficient, thereby achieving a dual quantitative evaluation of the functional importance and resource burden level of each deep learning model in the interaction process, thereby scientifically and accurately assessing the interaction triggering criticality of each model. The loading classification and sequence pre-sorting based on this evaluation result no longer rely on traditional single static parameters such as model volume or average loading time. This effectively solves the problem that the existing technology cannot distinguish between "high priority but slow loading" models and "low priority but fast loading" models, and greatly improves the rationality and pertinence of loading scheduling.

[0031] 2. The present invention further introduces a dynamic control mechanism based on resource status and execution progress during the model loading and execution phase. By continuously monitoring system operating parameters such as memory, video memory, and CPU usage, and combining the current loading queue status and execution progress, the loading order is adjusted in real time. In resource-constrained scenarios, models with high resource adaptability are loaded first, low-priority tasks are suspended, and the loading path is rearranged. This mechanism significantly enhances the responsiveness of the interactive system in complex environments, avoids model loading blockages or core model response delays caused by resource congestion, ensures that key semantic responses and multimodal presentation modules can be completed synchronously at the start of the interaction, and significantly improves the user's sense of system stability and trust in the first interaction.

[0032] 3. This invention also establishes a closed-loop mechanism from loading execution to data recording and then to policy optimization through an interactive feedback optimization module. After the interaction is completed, core operational data from the loading process is collected and used to update subsequent loading sorting rules and control strategies, enabling the system to continuously learn and dynamically adjust loading logic during long-term operation. By analyzing the actual loading performance of the model and user interaction behavior, the system can automatically optimize the loading priority configuration, gradually improving the environmental adaptability and user scenario matching of the scheduling strategy, thereby realizing a highly intelligent, personalized, and adaptive deep learning model scheduling system with excellent scalability and long-term evolution potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction to the drawings required for use in the embodiments will be given below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0034] Figure 1 This is a module diagram of the deep learning-based real-time digital human interaction system using QR codes according to the present invention.

[0035] Figure 2This is a system mind map of the deep learning-based QR code real-time digital human interaction system of the present invention. DETAILED DESCRIPTION

[0036] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these example embodiments are provided so that the description of this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0037] The present invention provides Figure 1 and Figure 2 The deep learning-based real-time digital human interaction system shown in the figure includes an identity recognition and extraction module, a model screening and construction module, a trigger key evaluation module, a sequence control execution module, and an interaction feedback optimization module;

[0038] The identity extraction module parses the identity identifier passed in by the user scanning the QR code and extracts the list of deep learning models used to implement the digital human interaction function based on the identity identifier;

[0039] Parsing the identity information passed in by the user scanning the QR code can be achieved through software methods that combine the QR code recognition algorithm with identity information parsing rules. Specifically, after the user scans the QR code with a mobile terminal device, the QR code recognition module calls an image decoding algorithm (such as one based on OpenCV or ZBar) to restore the encoded content in the QR code into a structured data string. This string usually carries identity parameters in the form of a URL or encrypted string, such as a unique user number or an associated database index code. The system extracts fields from the decoding results through the configured parsing rules (for example, based on JSON or Token segmentation format), extracts the key identity value used to associate the digital human interaction model, and caches it or passes it as a retrieval parameter to the subsequent model scheduling process.

[0040] After obtaining the identity identifier, the system uses the identifier as a query key to retrieve the model list corresponding to the identity from the pre-built model index database or configuration mapping table. The model list is stored in the form of a data structure, which usually contains the unique identifiers, functional descriptions and loading paths (such as speech synthesis model ID, image-driven model ID, etc.) of several models required for digital human interaction. In software implementation, key-value matching queries can be performed by docking with databases (such as MySQL, Redis, etc.) or configuration files (such as YAML, JSON), extracting all necessary model resource information bound to the identity and generating a standardized data set structure for use by subsequent model screening modules. This process does not involve physical model entity calls, only generates a logical list, and does not increase system load.

[0041] By parsing the user's scanned QR code to obtain the identity identifier, and based on the identifier, extracting the model list required for digital human interaction, it has a high degree of technical necessity and system integrity. The QR code scene has the personalized feature of "one person to one identification". The digital human image, language characteristics and interactive content corresponding to each user may be different, so the system must dynamically bind the model based on the unique identity identifier. Through the identity-driven model extraction mechanism, the system can implement a model loading strategy of "on-demand call, personalized scheduling, and resource streamlining", which not only avoids the waste of computing resources caused by redundant loading of global models, but also ensures the contextual accuracy of interactive behaviors and the consistency of individual characteristics of digital human responses. This method can be fully implemented through software structure, has good scalability and adaptability, and is the core pre-link for realizing personalized and lightweight digital human interaction processes.

[0042] The model screening construction module filters out all deep learning models loaded in parallel in the current interaction from the deep learning model list and establishes a multi-dimensional interactive loading model set;

[0043] In this embodiment, in the model screening and construction module, by comparing the function labels of each deep learning model in the deep learning model list with the function call requirements of the current interaction process, the deep learning models that need to be loaded in parallel in the current interaction are screened out, and the screened deep learning models are grouped according to the speech generation class, text generation class and image rendering class, and the grouping results are used to construct a multi-dimensional interactive loading model set.

[0044] The model screening module uses a configurable function matching mechanism to filter deep learning models that need to be loaded in parallel during the current interaction from the deep learning model list. Specifically, the system presets a function call requirement mapping table, which records the function nodes that need to be triggered during the startup phase of the current interaction process, such as voice broadcast, text generation, initial image presentation, and other functions, and assigns a unique function identifier to each function item. Furthermore, each deep learning model in the model list is accompanied by a function label field that characterizes the service function corresponding to the model, such as "speech synthesis," "image drive," "text generation," etc. During the screening process, the system traverses the model list and compares each model's function label with the set of function identifiers activated in the current interaction process. If a mapping relationship exists between the function label and the requirement identifier, the model is marked as a model that needs to be loaded and included in the initial parallel loading set. Furthermore, after the screening is complete, the system divides the result set into three subsets: voice, text, and image, based on the functional dimension to which the model belongs, forming a set of interactive loading models with a multi-dimensional structure. This process can be achieved through basic data operations such as conditional screening, key-value matching, and set grouping in the software. It does not require hardware participation and has good logical clarity and computing efficiency.

[0045] After completing the initial screening of deep learning models, the model screening and construction module divides the screening results into multi-dimensional structures through a grouping mechanism based on the functional type field. The specific implementation method is as follows: each screened model object contains a preset functional type field in the model list. The value of this field comes from the functional type dictionary uniformly set by the system, such as functional categories such as "speech generation", "text generation", and "image rendering". After the screening is completed, the system automatically calls the clustering grouping logic to classify and group the model collection according to the value of the functional type field. This process can be implemented at the software level through hash mapping, key-value classification, or logical branch control. That is, when traversing the model collection, the system maps the model into the subset under the corresponding category key according to its functional type value, and finally forms three types of model subsets with functional category as the dimension label. After construction is completed, the three subsets are uniformly encapsulated into a structured data object to form a multi-dimensional interactive loading model collection, in which each dimension corresponds to a core functional model group required to start the interaction. This collection provides a structural carrier for subsequent evaluation, sorting and regulation, and has technical advantages such as clear hierarchy, clear dimensions, and easy indexing and dynamic scheduling. The entire process can be efficiently implemented in the software platform through data structure operations and collection processing logic.

[0046] The triggering criticality assessment module obtains the loading control information of each deep learning model in the multi-dimensional interactive loading model set, evaluates the interaction trigger criticality of each deep learning model based on the loading control information, and classifies each deep learning model according to the evaluation results;

[0047] In this embodiment, in the trigger critical evaluation module, the loading control information of each deep learning model in the multidimensional interactive loading model set is obtained and preprocessed after acquisition; the structural interaction topology information and resource load index information are extracted from the preprocessed loading control information of each deep learning model, and analyzed after extraction to generate the topological coupling coefficient and load impact index of each deep learning model respectively; an interactive trigger criticality evaluation model is constructed for the topological coupling coefficient and load impact index of each generated deep learning model, and the interactive critical coefficient of each deep learning model is generated by weighted summation; a pre-set interactive critical coefficient threshold interval is determined, and after determination, it is compared with the interactive critical coefficient of each generated deep learning model, and the interactive trigger criticality of each deep learning model is evaluated according to the comparison result, and each deep learning model is divided into a loading-first type model, a loading-parallel type model and a loading-delayed type model according to the evaluation result.

[0048] Obtaining the loading control information of each deep learning model in the multi-dimensional interactive loading model set can be achieved through the model monitoring and resource registration interface preset in the software system. Specifically, after initializing the multi-dimensional model set, the system establishes a data mapping relationship with the model resource configuration library, the system performance monitoring module, and the task call log database, and extracts the loading-related raw data of each model in turn according to the model's unique identifier. These data include model file volume, memory usage peak, video memory request size, output interface configuration, structural dependency chain, historical call records, etc., all of which are stored in the model management index table or resource status snapshot table as structured fields. Through standardized data extraction processes, such as key-value matching, field merging, and resource pull interface calls, the system can pull the current loading control information of each deep learning model item by item to form a complete model control information set for use in subsequent evaluation links.

[0049] The purpose of preprocessing the loaded regulatory information is to ensure that data from different sources and of different types are uniform, comparable, and efficient in the subsequent analysis and modeling process. Since some loaded regulatory information may have problems such as unit differences, data redundancy, missing fields, or inconsistent formats, the data must be standardized and cleaned through software. Specific preprocessing includes: unifying the units of numerical fields (such as MB, number, etc.); setting default values or backfilling approximate values for missing fields; encoding or mapping text labels to standard fields (such as mapping functional type fields to three-category dimensional indexes); and normalizing or logarithmically compressing continuous values to eliminate magnitude differences. The preprocessing logic can be called immediately after data extraction, and automatically processes the data through the set data cleaning rule chain and field mapping table to generate a high-quality regulatory data set with consistent structure and unified format, ensuring the accuracy of subsequent parameter generation and the robustness of the calculation model.

[0050] After loading and preprocessing the control information, the system extracts structural interaction topology information and resource load indicator information through field classification and multidimensional mapping. Specifically, the system first semantically categorizes each data field in the preprocessed information set based on a field type identification mechanism. Fields related to model structure relationships (such as the number of model output interfaces, the number of dependent models, and the number of historical calls) are automatically grouped into the structural interaction topology information field group; fields related to resource usage (such as model file size, peak memory usage, and requested graphics memory size) are grouped into the resource load indicator information field group. This process relies on field name matching rules, field value feature analysis, and automatic association of field annotation tables, eliminating the need for manual intervention. During the extraction process, the system performs structural mapping on each field group, packaging them into two information entity objects indexed by model ID. Each object corresponds to the topology information structure and resource indicator structure of a deep learning model, respectively. This entire process is automated through the software's field filters, data grouping functions, and model object generation mechanism, enabling efficient and standardized information extraction and providing a stable data foundation for subsequent parameter calculations.

[0051] The threshold intervals for interaction key coefficients are automatically set and dynamically adjusted in the software during system initialization by combining statistical analysis of historical interaction data with rule configuration strategies. Specifically, the system first utilizes historical deep learning model loading and interaction behavior data stored in a database to construct a sample set of key coefficient distributions based on time period, user type, or scenario category. By performing mathematical analysis (such as quantile statistics, density distribution estimation, and cluster center extraction) on the generated interaction key coefficients in this sample set, the software automatically extracts two key thresholds: a lower threshold for the maximum value of the delayed loading model, an upper threshold for the minimum value of the advanced loading model, and an intermediate threshold for the parallel loading model. These threshold intervals can be written into the system configuration table and periodically re-evaluated and updated based on data drift during operation. The entire determination process is automatically executed in the software through a data sampling module, a parameter distribution modeler, and an interval calculation module, eliminating the need for manual fixed threshold settings. This ensures that the evaluation intervals are data-driven and dynamically adaptable, providing a stable reference standard for classification decisions.

[0052] In this embodiment, the specific logic for obtaining the topological coupling coefficient of each deep learning model is as follows:

[0053] The structural interaction topology information is extracted from the pre-processed loading control information of each deep learning model, including the number of output interfaces of each deep learning model, the number of dependent models, and the number of calls in the preset historical time period, and marked as 、 and , Indicates the The number of output interfaces of a deep learning model, Indicates the The number of models that a deep learning model depends on, Indicates the The number of times a deep learning model is called within a preset historical time period, , is a positive integer;

[0054] The number of output interfaces, the number of dependent models, and the number of calls within a preset historical time period for each deep learning model can all be automatically extracted by software from the model configuration data structure, the system dependency graph, and task call logs. Specifically, the number of output interfaces refers to the number of data output terminals exposed by each deep learning model within the system. This is typically defined in the model's configuration file or network structure description (such as ONNX or TensorFlowGraphDef). The software automatically counts and records these by parsing the number of output nodes or output tensors in the model graph. The number of dependent models refers to the number of other models in the current interactive process topology that directly reference the model's outputs as inputs. This can be obtained by constructing a model dependency graph (DAG) or reading a pre-set model dependency mapping table in the system. The system automatically backtracks and counts the number of downstream dependencies for each model when it was an upstream node based on the connection relationships. The number of calls within a preset historical time period refers to the actual frequency of model calls over a period of time (e.g., the last 30 days). This can be obtained by reading model service call logs, API access counters, or the system call monitoring module. During log parsing, the software aggregates the statistical results by model ID and archives them by time window. These three types of data all have clear physical meanings and quantifiable characteristics, and can truly reflect the degree of coupling of the model in the structural network and the task activity during the operation phase. Moreover, their collection process can be automatically completed within the software platform through the data parser, log analysis module and structure recognition engine, without relying on manual operation or external system support, and has good stability and engineering feasibility.

[0055] Calculate the topological coupling coefficient of each deep learning model The specific calculation method is: The number of times a deep learning model is called within a preset historical time period Take the square root to get the first part of the calculation value; The number of output interfaces of a deep learning model and the number of dependent models Multiply them, add one to the product, and then take the natural logarithm of the value to get the second part of the calculation value; add the first part of the calculation value and the second part of the calculation value to get the first part of the calculation value. Topological coupling coefficient of deep learning models The specific calculation formula is as follows:

[0056]

[0057] Where, For the Topological coupling coefficient of a deep learning model.

[0058] The calculation of the topological coupling coefficient uses a composite mathematical model that combines structural dependency and historical call frequency, aiming to comprehensively evaluate the structural importance and call activity of each deep learning model during the startup of the interactive system. It is the The square root of the number of times a deep learning model has been called in a preset historical period is used. The purpose of using the square root is to maintain the trend of the increase in the number of calls while avoiding a sharp increase in the value, so that the frequently called model will show higher interactive activity in the score but will not dominate the overall score. It is a joint expression of the complexity of the model output structure and the strength of system dependence, where is the number of output interfaces, representing the number of structural channels for the model to provide data to the outside. The number of dependent models indicates how many other models in the system rely on the model's output. The product of the two quantifies the degree of structural coupling of the model within the system topology, while taking the natural logarithm of the two is used to compress the range and enhance stability, preventing extreme distributions of inter-model dependency differences due to excessive coupling. The overall structure achieves a dual evaluation of "frequency" and "structurality" through summation, effectively capturing the key role of the model in the interaction initiation process and demonstrating good mathematical interpretability and technical rationality.

[0059] The topological coupling coefficient of each deep learning model reflects the model's dependency strength and frequency of use within the interaction process structure. Therefore, its value is directly positively correlated with the model's interaction triggering criticality. Specifically, a higher topological coupling coefficient indicates that the model has more output interfaces and a higher degree of structural coupling within the system topology—that is, its outputs are dependent on multiple other models. Furthermore, the model is frequently called in historical interactions, indicating that it plays a high role in function triggering in actual operation. Therefore, when evaluating the interaction triggering criticality of each deep learning model, the topological coupling coefficient can serve as a core metric to measure its "dependency" and "interaction activity." A higher value indicates that the model is more likely to be a pre-process or core model during interaction initiation, and its loading priority should be increased accordingly to ensure the integrity of the interaction process and consistent response. By introducing the topological coupling coefficient, the system can quantitatively identify the structural roles of models, thereby precisely regulating their loading order.

[0060] In this embodiment, the specific logic for obtaining the load impact index of each deep learning model is as follows:

[0061] The resource load index information is extracted from the pre-processed loading control information of each deep learning model, including the model file volume, memory usage peak and static allocation capacity of each deep learning model on the GPU memory, and calibrated as 、 and , Indicates the The model file size of a deep learning model, Indicates the The peak memory usage of deep learning models, Indicates the The static allocation capacity of a deep learning model on GPU memory, , is a positive integer;

[0062] The model file size, peak memory usage, and statically allocated GPU memory capacity of each deep learning model are automatically collected by software during the model loading and runtime lifecycle, with a clear engineering data source and acquisition path. First, the model file size refers to the model's on-disk storage size. This can be directly obtained by calling the system's model loading path or reading the model file's physical size field in the model configuration index. This value represents the basic requirements for disk I / O and cache scheduling during the model loading phase. Second, peak memory usage refers to the maximum amount of main memory (RAM) used by the model during initialization or inference. During runtime, the system monitors the dynamic memory usage of each model in real time using the resource monitoring module or runtime memory allocation logging. After the model completes its execution, the peak value is recorded in the performance metrics log as a stable value. Finally, statically allocated GPU memory capacity refers to the memory block size pre-allocated by deep learning frameworks (such as TensorFlow, PyTorch, and ONNX Runtime) when the model is deployed to the GPU device. This information is automatically read by the software during the model loading initialization phase by calling the model compiler or the framework's memory allocation interface. These three types of data are all objective numerical information automatically generated when the system is running. They do not require manual definition or external input. They are characterized by strong stability, low collection cost, and standardized modeling. They are the basic supporting elements for constructing resource load indicator information and can provide reliable and real-time resource occupancy descriptions for subsequent evaluation logic.

[0063] Calculate the load impact index of each deep learning model The specific calculation method is: Model file size of a deep learning model , peak memory usage Statically allocated capacity on GPU memory Sum the natural logarithms after adding one to get the Load impact index of deep learning models The specific calculation formula is as follows:

[0064]

[0065] Where, For the Load impact index of deep learning models.

[0066] The load impact index is calculated using a weighted superposition structure, which aims to comprehensively measure the overall level of system resource usage of each deep learning model during loading and running. Indicates the The file size of a deep learning model directly reflects the basic storage resources required to load the model from disk to memory. The larger the file size, the more I / O operations required for loading, and the more obvious the impact on startup efficiency. The peak memory usage of the model The square root transformation of Indicates the maximum RAM usage of the model during runtime. Using square root operations can maintain the trend of increasing values as resource requirements increase, while preventing high-memory models from dominating the overall results, reflecting the compression stability of the values. It is to statically request capacity for the model on the GPU memory. The logarithmic transformation of The larger the value, the more GPU resources the model occupies during inference, affecting the parallel execution capability. Logarithmic transformation is used to compress the memory difference within a reasonable range while maintaining its sensitivity to the overall evaluation. The three components are directly added to avoid the asymmetric amplification or suppression effects that may be introduced by the multiplication and division structure, making the evaluation results more balanced in expressing the consumption of various resources. Ultimately, this calculation logic makes A higher value means that the total pressure on storage, memory, and video memory during model loading and running is greater. It is suitable for evaluating the impact of the model on the system load and providing a quantitative basis for model loading priority decisions.

[0067] The load impact index of each deep learning model is used to quantify the comprehensive level of system key resources (storage, memory, and video memory) occupied by the model during loading and running. Therefore, its value is closely related to the loading feasibility of the model during the interaction initialization phase. When evaluating the criticality of interaction triggering, the load impact index is negatively correlated with the criticality, that is, The larger the value, the more resources the model requires and the higher the load on the system. If such a model is loaded at the initial stage of interaction, it may occupy too much memory or video memory, causing other core models in the overall interaction chain to be unable to load or run in time, thus affecting the user's first round of interactive response experience. On the contrary, Models with lower values typically have lighter resource requirements and are suitable for priority loading or parallel scheduling, which facilitates the construction of responsive, continuous, and smooth interaction processes. Therefore, when jointly evaluating the interaction triggering criticality of each deep learning model, the load impact index, as a system resource impact factor, can be used to reduce the priority of resource-intensive models. This allows the evaluation mechanism to not only consider the importance of the model's functional structure, but also balance the overall resource allocation efficiency, thereby achieving a more stable and efficient interaction initiation process.

[0068] In this embodiment, the topological coupling coefficients of each deep learning model generated are and load impact index Construct an interaction trigger criticality evaluation model and generate the interaction criticality coefficients of each deep learning model through weighted summation , the specific calculation formula is as follows:

[0069]

[0070] Where, For the The interaction key coefficients of the deep learning model, and are the topological coupling coefficients of each deep learning model and load impact index The reciprocal of The non-zero weight coefficient of .

[0071] Interaction key coefficient The generation process is achieved by building a configurable weighted fusion model in the software system. Specifically, the system obtains the topological coupling coefficient of each deep learning model. and load impact index Then, a weighted sum structure is used to jointly model the two types of indicators, where Retain as a positive indicator, reflecting the structural dependence and usage frequency of the model, and Take the reciprocal form , as the reverse factor of resource consumption, reflects the principle that the smaller the resource occupation, the higher the priority. The two indicators are multiplied by non-zero weight coefficients respectively. and ,in Used to regulate the weight of structural importance in the overall assessment, Used to regulate the weight of the impact of resource burden on loading priority. The sum of the two is set to 1 to ensure that the composite coefficient has a normalized effect. The weight coefficient can be dynamically set by the system according to the actual interaction scenario, or it can be empirically set by offline analysis of historical model loading effects and combined with the scheduling strategy configuration table. The system can load the weight parameters under the current interaction scenario through the configuration file or the model scheduling strategy interface to ensure the adaptability and stability of the evaluation model. The final generated It can serve as a direct input for loading prioritization and regulation decisions of various deep learning models.

[0072] In this embodiment, the preset interaction key coefficient threshold interval is determined , and after determination, the interaction key coefficients of each deep learning model generated A comparison is performed, and the interaction triggering criticality of each deep learning model is evaluated based on the comparison results. Based on the evaluation results, each deep learning model is divided into loading-first model, loading-parallel model, and loading-delay model. The specific comparison analysis and classification are as follows:

[0073] like , the interaction triggering criticality of the deep learning model is low, and the deep learning model is classified as a loading delay model;

[0074] This situation indicates that the model has neither significant structural dependencies nor resource-lightweight models in the current interactive process. In other words, the model may be responsible for non-basic response tasks, and its topological coupling is low, that is, its output has little impact on other models. At the same time, the resources required to load or run it are relatively high, and the cost-effectiveness of the system prioritizing loading this model is low. Therefore, this type of model is classified as a delayed loading model, and its loading behavior will not be carried out until the system resources are initially released or the main interactive link is established, so as to prevent it from occupying too many initial resources and affecting the normal construction of the interactive link. This strategy helps to ensure the response efficiency and resource scheduling stability during the system initialization phase, and avoid scheduling conflicts such as "secondary models occupying primary resources."

[0075] like ,The interaction trigger criticality of the deep learning model is medium, and the deep learning model is classified as a loading parallel model;

[0076] This situation shows that the model is relatively balanced between structural dependency and resource consumption, that is, there is a certain degree of calling relationship in the functional chain, and the resource usage is also in a medium range. This type of model does not belong to the core node of the interactive startup chain, but may undertake tasks such as scene switching, image updates, and auxiliary responses, and its loading timing has a certain degree of flexibility. Therefore, classifying this type of model as a loading parallel model means that the system can use idle computing resources to asynchronously load the prior model while the main thread is loading it, thereby improving the overall system concurrency and loading throughput. This strategy helps to achieve a reasonable combination of primary and secondary tasks, and improve the system fluency and versatility perceived by users.

[0077] like , the interaction trigger criticality of the deep learning model is high, and the deep learning model is classified as a loading-first model.

[0078] This indicates that the model combines high structural coupling with low resource usage within the interaction process. Its output is relied upon by multiple other models, yet it is inherently lightweight and technically capable of rapid loading and execution. Therefore, this model is assessed as a key triggering node in the current interaction chain, and the system should prioritize its loading to ensure a timely response to the user's first interaction, thereby establishing a closed-loop interaction. Classifying this model as a load-first model allows the system to prioritize the construction of key task paths during the resource-limited initialization phase, improving the hit rate and response time of the first load, significantly enhancing the user's initial interaction experience, and enhancing system stability and trust.

[0079] The sequential control execution module pre-sorts the loading order of each deep learning model according to the classification results, executes the loading operation according to the pre-sorted order, and dynamically controls the loading order during the loading process based on the current resource status and execution progress;

[0080] In this embodiment, in the sequential control execution module, the loading order of each deep learning model is pre-sorted according to the classification results, and the loading operation is performed according to the pre-sorted order: the loading priority class models are sorted from high to low according to the interaction key coefficient, and the loading operations are performed in sequence; in the process of loading the priority class models, the resource occupancy level of the loading parallel class models is also sorted from low to high, and the loading operations are started in sequence; after the loading of the loading priority class model and the loading parallel class model is completed, the loading of the delayed class model is executed in sequence.

[0081] In the sequential control execution module, the loading order of each deep learning model is pre-sorted based on the completed model classification results, and loading operations are executed according to a hierarchical control strategy. This is implemented at the software level through the priority scheduling logic and resource allocation instruction set embedded in the scheduling engine. Specifically, after receiving the set of models in the "pre-load" category, the system sorts the models in the set in descending order based on their corresponding interaction key coefficients. This sorting process is based on a model-coefficient mapping table built in memory and can be performed using a standard comparator function. After sorting, a loading queue is generated and loading is immediately initiated in sequence. Simultaneously, the system processes the "parallel load" category in a separate concurrent thread, extracting resource usage metrics corresponding to each model (such as model file size, peak memory usage, and video memory request value). Based on these metrics, a resource usage ranking rule is constructed, ranking the models from low to high, and allocating idle resource channels for sequential loading to avoid resource preemption conflicts. Furthermore, once all the models in the "pre-load" and "parallel load" categories have been loaded, the system transfers scheduling control to the "delay load" category, which then loads each model in a pre-set order through a unified execution interface. Throughout the entire process, the hierarchical execution strategy of loading order ensures that high-priority models obtain computing resources before low-priority models, thereby improving the system's responsiveness during the interactive startup phase. This hierarchical sorting and staged execution approach is adopted because the response experience during the interactive process is highly sensitive to the initial loading order. Prioritizing the loading of models with high interaction criticality and low resource overhead can maximize the rapid availability of core functions. At the same time, by rationally allocating resources, competition between models can be avoided, thereby improving the overall loading efficiency and operational stability of the system. All sorting operations, generation of priority queues, and distribution of loading commands are completed in the software through the collaborative work of the task scheduler, resource status mapper, and loading instruction control logic, without relying on external hardware control, to achieve a clear, implementable, and maintainable path.

[0082] In this embodiment, in the sequential control execution module, the loading order is dynamically controlled in combination with the current resource status and execution progress during the loading process. Specifically, during the loading process, the current available memory capacity, GPU video memory remaining amount and CPU occupancy rate are continuously obtained, and the loading order is dynamically controlled in combination with the loading progress of each deep learning model and the loading order of the models that have not yet been loaded. Specifically, it includes: pausing the subsequent loading of delayed class models, adjusting the remaining loading order of parallel class models, and giving priority to loading deep learning models whose resource occupancy is lower than the set threshold, so as to maintain the smoothness and stability of the loading process under resource-constrained conditions.

[0083] During the loading process, the loading order of the deep learning model is dynamically regulated in combination with the current resource status and execution progress. This can be achieved through the software's embedded resource monitoring sub-module and the real-time reconstruction mechanism of the scheduling priority. Specifically, at the startup phase of the loading task, the control module will establish a resource status acquisition thread, which calls the operating system or driver interface at a fixed frequency to obtain key operating parameters such as the current available memory capacity, the remaining amount of GPU video memory, and the CPU occupancy rate, and writes the acquisition results to the scheduling status buffer. At the same time, the control module also maintains a loading progress status table for each model being loaded or to be loaded, and records in real time whether its loading is complete, the loading percentage, the initialization time estimate and other indicators, so that the adaptability between system resources and model queues can be dynamically judged before executing the loading operation. During the judgment phase, the software uses a defined resource threshold judgment strategy (for example, setting a "resource-constrained state" when the remaining video memory falls below a certain percentage) to determine whether a resource bottleneck exists. Once a resource shortage is determined, the system immediately executes three preset dynamic control behaviors: First, it pauses unloaded tasks in deferred models to prevent low-priority models from competing for limited resources. Second, it adjusts the loading order based on the estimated resource usage of the remaining parallel models, prioritizing low-resource models. Third, it performs a resource threshold screening on all unloaded models, identifying models with resource requirements below the set upper limit and inserting them into a priority queue, thus forming a resource adaptability optimization path. This dynamic control mechanism is designed to avoid the phenomenon of "high-resource models blocking the loading channel of medium- and low-resource models" in parallel loading scenarios. In particular, when computing resources are limited, it can significantly improve the loading efficiency of critical path models and maintain the continuity of the loading response rhythm, thereby enhancing the overall system stability and real-time perception performance during the real-world interactive startup phase. The entire dynamic control process does not require human intervention. All monitoring, judgment, sorting and reconstruction operations are automatically completed by the state perception engine and scheduling strategy execution unit in the software based on data-driven methods. It has high automation, real-time response and resource adaptation capabilities.

[0084] During the loading process, the dynamic control behavior of "pausing the subsequent loading of delayed class models, adjusting the remaining loading order of parallel class models, and giving priority to loading deep learning models with resource usage lower than the set threshold" can be executed. The refined issuance of control instructions can be achieved by constructing a model loading task status table and a resource scheduling policy interface. First, when the system monitors in real time that the available memory or GPU video memory is lower than the preset safety threshold (such as: available memory is less than 300MB, and the remaining video memory is less than 500MB), the control module will query the task identifier of the delayed class model that has not yet been loaded, and mark its status in the loading task queue as "suspended". The loading thread of the model will be terminated and enter the waiting pool to release the scheduling window; at the same time, the system will re-sort the remaining tasks of loading parallel class models according to their resource usage estimates (such as weighted calculation of model volume, the most recent loading peak record), and give priority to loading models with a size less than 10MB and a memory peak less than 100MB. Lightweight models ensure that tasks enter the loading process smoothly within resource constraints. For example, if parallel models A, B, and C are to be loaded, with A expected to require 300MB of video memory, B 180MB, and C only 90MB, the control module will adjust the loading order to C → B → A. Furthermore, among all models to be loaded, the control module invokes threshold screening logic to prioritize models whose current resource requirements are significantly below the system's remaining resource threshold (e.g., resource requirements less than 20%). For example, if model D requires only 80MB of video memory and can load within 0.2 seconds, it will be inserted and executed before any interrupted tasks. By pausing the loading of non-critical models, rescheduling parallel tasks, and prioritizing lightweight models with high resource compatibility, this effectively prevents resource crowding and loading bottlenecks, ensuring that critical models load on time and improving usability and stability during the initial interactive phase. Throughout this process, control operations are automatically triggered by the software task scheduler based on the resource mapping table and task priority rules, eliminating the need for manual intervention and ensuring real-time and robust scheduling logic.

[0085] The interactive feedback optimization module executes the digital human interactive response after loading is completed and records the data during the loading process to optimize the subsequent loading sequence control strategy.

[0086] After loading is complete, the digital human interactive response is executed, and data from the loading process is recorded to optimize the implementation of subsequent loading sequence control strategies. This is accomplished through the collaborative work of the software's interactive response engine and data log collection mechanism. Specifically, once all preceding and parallel models have been loaded and initialized and verified, the interactive response engine automatically activates, invoking the digital human voice, image, and text generation models associated with the user's identity, initiating the digital human's visualization and semantic response behaviors, thus completing the human-computer interaction process. During this time, the interactive feedback optimization module monitors the entire loading phase in real time through a task listener. Key metrics such as the loading start time, completion time, resource usage, loading status, and user interaction latency (such as first voice response time) for each model are recorded in the system operation log and stored in a structured "Loading History Performance Database." This data collection process is automatically completed by the software's resource traceback interface and model call recorder, forming a complete closed loop from loading to interaction to recording.

[0087] The core purpose of establishing this data recording and optimization strategy is to support the continuous evolution and adaptive improvement of the loading sequence control strategy based on the continuous accumulation of loading behavior samples. Specifically, by analyzing the loading time distribution, loading success rate, resource usage peaks of different models in historical interactions, and their correlation with the interactive response experience, the system can automatically identify which models should have higher or lower loading priority in different scenarios. For example, in multiple user scan code interactions, if a resource-intensive model frequently fails to load in the initial stage and has little impact on user perception, the system will lower the model's pre-ranking weight based on the accumulated data. Conversely, if a lightweight model plays a key role in the first sentence of the interaction, its interaction key coefficient will be increased to prioritize loading. Ultimately, this mechanism can form a "self-optimizing loading control curve" based on historical operation data, achieving the ability to generalize scheduling strategies across users and multiple scenarios, improving the overall system response performance and consistency with the digital human experience. Moreover, the entire process is executed by software, which is highly maintainable and has intelligent adjustment capabilities.

[0088] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0089] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0090] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0091] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0093] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0094] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0095] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A QR code real-time digital human interaction system based on deep learning, characterized by: It includes identity recognition and extraction module, model screening and construction module, trigger key evaluation module, sequence control execution module and interactive feedback optimization module; The identity extraction module parses the identity identifier passed in by the user scanning the QR code and extracts the list of deep learning models used to implement the digital human interaction function based on the identity identifier; The model screening construction module filters out all deep learning models loaded in parallel in the current interaction from the deep learning model list and establishes a multi-dimensional interactive loading model set; The triggering criticality assessment module obtains the loading control information of each deep learning model in the multi-dimensional interactive loading model set, evaluates the interaction trigger criticality of each deep learning model based on the loading control information, and classifies each deep learning model according to the evaluation results; In the trigger key evaluation module, the loading control information of each deep learning model in the multi-dimensional interactive loading model set is obtained and preprocessed after acquisition; Extract structural interaction topology information and resource load index information from the preprocessed loading control information of each deep learning model, and analyze them after extraction to generate the topological coupling coefficient and load impact index of each deep learning model respectively; construct an interaction trigger criticality assessment model for the generated topological coupling coefficient and load impact index of each deep learning model, and generate the interaction critical coefficient of each deep learning model by weighted summation; determine a pre-set interaction critical coefficient threshold range, and compare it with the generated interaction critical coefficient of each deep learning model after determination, evaluate the interaction trigger criticality of each deep learning model based on the comparison results, and divide each deep learning model into a loading-first model, a loading-parallel model, and a loading-delayed model based on the evaluation results; The specific logic for obtaining the topological coupling coefficients of each deep learning model is as follows: The structural interaction topology information is extracted from the pre-processed loading control information of each deep learning model, including the number of output interfaces of each deep learning model, the number of dependent models, and the number of calls in the preset historical time period, and marked as 、 and , Indicates the The number of output interfaces of a deep learning model, Indicates the The number of models that a deep learning model depends on, Indicates the The number of times a deep learning model is called within a preset historical time period, , is a positive integer; Calculate the topological coupling coefficient of each deep learning model The specific calculation method is: The number of times a deep learning model is called within a preset historical time period Take the square root to get the first part of the calculation value; The number of output interfaces of a deep learning model and the number of dependent models Multiply them, add one to the product, and then take the natural logarithm of the value to get the second part of the calculation value; add the first part of the calculation value and the second part of the calculation value to get the first part of the calculation value. Topological coupling coefficient of deep learning models ; The specific logic for obtaining the load impact index of each deep learning model is as follows: The resource load index information is extracted from the pre-processed loading control information of each deep learning model, including the model file volume, memory usage peak and static allocation capacity of each deep learning model on the GPU memory, and calibrated as 、 and , Indicates the The model file size of a deep learning model, Indicates the The peak memory usage of deep learning models, Indicates the The static allocation capacity of a deep learning model on GPU memory, , is a positive integer; Calculate the load impact index of each deep learning model The specific calculation method is: Model file size of a deep learning model , peak memory usage Statically allocated capacity on GPU memory Sum the natural logarithms after adding one to get the Load impact index of deep learning models ; The sequential control execution module pre-sorts the loading order of each deep learning model according to the classification results, executes the loading operation according to the pre-sorted order, and dynamically controls the loading order during the loading process based on the current resource status and execution progress; The interactive feedback optimization module executes the digital human interactive response after loading is completed and records the data during the loading process to optimize the subsequent loading sequence control strategy.

2. The deep learning-based real-time digital human interaction system based on QR codes according to claim 1, characterized in that: In the model screening and construction module, by comparing the function labels of each deep learning model in the deep learning model list with the function call requirements of the current interaction process, the deep learning models that need to be loaded in parallel in the current interaction are screened out, and the screened deep learning models are grouped according to the speech generation class, text generation class and image rendering class, and the grouping results are used to construct a multi-dimensional interactive loading model set.

3. The deep learning-based real-time digital human interaction system for QR codes according to claim 2, characterized in that: Topological coupling coefficients of each generated deep learning model and load impact index Construct an interaction trigger criticality evaluation model and generate the interaction criticality coefficients of each deep learning model through weighted summation .

4. The deep learning-based real-time digital human interaction system for QR codes according to claim 3 is characterized in that: Determine the pre-set interaction key coefficient threshold range , and after determination, the interaction key coefficients of each deep learning model generated A comparison is performed, and the interaction triggering criticality of each deep learning model is evaluated based on the comparison results. Based on the evaluation results, each deep learning model is divided into loading-first model, loading-parallel model, and loading-delay model. The specific comparison analysis and classification are as follows: like , the interaction triggering criticality of the deep learning model is low, and the deep learning model is classified as a loading delay model; like ,The interaction trigger criticality of the deep learning model is medium, and the deep learning model is classified as a loading parallel model; like , the interaction trigger criticality of the deep learning model is high, and the deep learning model is classified as a loading-first model.

5. The deep learning-based real-time digital human interaction system for QR codes according to claim 4 is characterized in that: In the sequential control execution module, the loading order of each deep learning model is pre-sorted according to the classification results, and the loading operation is performed according to the pre-sorted order. Specifically: the loading priority class models are sorted from high to low according to the interaction key coefficient, and the loading operations are performed in sequence; in the process of loading the priority class models, the resource occupancy level of the loading parallel class models is also sorted from low to high, and the loading operations are started in sequence; after the loading of the loading priority class model and the loading parallel class model is completed, the loading of the delayed class model is executed in sequence.

6. The deep learning-based real-time digital human interaction system for QR codes according to claim 5, characterized in that: In the sequential control execution module, the loading order is dynamically controlled during the loading process in combination with the current resource status and execution progress. Specifically, during the loading process, the current available memory capacity, GPU video memory remaining amount and CPU occupancy rate are continuously obtained, and the loading order is dynamically controlled in combination with the loading progress of each deep learning model and the loading order of the models that have not yet been loaded. Specifically, the loading order is suspended for the subsequent loading of delayed models, the loading order of the remaining parallel models is adjusted, and the deep learning models with resource occupancy below the set threshold are prioritized.

Images