Image processing method based on computing network

Through the image processing method based on computing power network, the problem that a single computing node is difficult to process large-scale high-concurrent image tasks in real time is solved. Through multi-grained reconstruction and virtual and real mapping technology, the rapid response and efficient execution of image processing tasks are achieved, and the improvement of the improvement of the image processing tasks is achieved through multi-grained reconstruction and virtual and real mapping technology. Resource utilization rate.

CN118628334BActive Publication Date: 2024-11-08NANJING BOYU IOT TECH CO LTD +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411098615.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-11-08
Estimated Expiration
2044-08-12

Smart Images

  • Figure CN118628334B_ABST
    Figure CN118628334B_ABST
Patent Text Reader

Abstract

The present invention discloses an image processing method based on a computing power network, which relates to the technical field related to image processing. The method includes: reading a distributed target computing power network to determine a computing power virtual architecture; separating a data plane from a control plane, and establishing an upper coordinator in combination with data-driven training; reading an image processing task to determine an image processing target; sensing the network computing power state and performing decision analysis to determine a virtual layer strategy; converting the virtual layer strategy based on a target computing power network to determine a physical layer strategy. The method solves the technical problems existing in existing image processing, such as being based on a single computing node, being difficult to perform large-scale, high-concurrency image processing in real time, being unable to accurately match image processing tasks with computing resources, and thus resulting in insufficient real-time data transmission for image processing, as well as low execution efficiency and resource utilization, thereby achieving rapid response and efficient execution of image processing tasks, and achieving the technical effect of improving image processing efficiency and resource utilization.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field related to image processing, and specifically to an image processing method based on a computing power network. Background Art

[0002] In the digital age, image processing, as a key technology in the field of computer vision and artificial intelligence, is facing unprecedented challenges and opportunities. With the widespread deployment of cloud computing, edge computing, and smart terminal devices, how to efficiently use these distributed computing resources to support complex and changeable image processing tasks has become an important link. With the rapid development of technologies such as big data, the Internet of Things, and 5G, the generation and processing of image data has exploded. The traditional centralized computing model can no longer meet the requirements of real-time, low latency, and high efficiency. In addition, the resources of edge computing nodes are limited, making it difficult to independently undertake large-scale and computing-intensive image processing tasks. Although cloud computing centers have powerful computing power, data transmission delays have become a limiting factor, affecting the execution efficiency of image processing tasks.

[0003] Therefore, in the current image processing related technologies, there are technical problems such as being based on a single computing node, difficulty in real-time large-scale, high-concurrency image processing, and inability to accurately match image processing tasks with computing resources, which in turn leads to insufficient real-time data transmission for image processing and low execution efficiency and resource utilization. Summary of the invention

[0004] This application provides an image processing method based on a computing network, and adopts technical means such as multi-granularity reconstruction, establishing virtual-real mapping relationships, establishing a higher-level coordinator, and strategy conversion to solve the technical problems existing in existing image processing, such as being based on a single computing node, difficulty in real-time large-scale, high-concurrency image processing, and inability to accurately match image processing tasks with computing resources, which in turn leads to insufficient real-time data transmission for image processing and low execution efficiency and resource utilization. It achieves rapid response and efficient execution of image processing tasks, and achieves the technical effect of improving image processing efficiency and resource utilization.

[0005] The present application provides an image processing method based on a computing power network, the method comprising: reading a distributed target computing power network, performing multi-granularity reconstruction deployment and computing power resource encoding based on the computing power type of image processing, determining a computing power virtual architecture, wherein a virtual-to-real mapping is established between the computing power virtual architecture and the target computing power network; based on the computing power virtual architecture, separating a data plane and a control plane, and establishing an upper coordinator in combination with data-driven training, wherein the upper coordinator is configured with a unified management rule based on a third perspective; reading an image processing task, performing task decomposition to determine a plurality of image processing nodes, coupling and binding the plurality of image processing nodes with a required computing power configuration, and determining an image processing target; based on the upper coordinator, sensing the network computing power state and performing decision analysis based on the image processing target, and determining a virtual layer strategy; based on the virtual-to-real mapping relationship, converting the virtual layer strategy based on the target computing power network, and determining a physical layer strategy, wherein the physical layer strategy includes a network transmission strategy and an image processing strategy; based on the physical layer strategy, executing the image processing task in response to a target computing power network node of the target computing power network.

[0006] In a possible implementation, the computing power virtual architecture is determined by performing the following processing: based on the computing power type, the distributed target computing power network is clustered and segmented to determine the first computing power; based on the distributed location and computing power type, the first computing power is deployed in a hierarchical distributed network to determine the initialization reconstruction network; based on image processing elements, the initialization reconstruction network is traversed to perform multi-level granularity division of each reconstructed grid to determine multiple groups of computing power grids, wherein each computing power network node corresponds to a group of computing power grids; based on the initialization reconstruction network and the multiple groups of computing power grids, the computing power virtual architecture is constructed.

[0007] In a possible implementation, the distributed target computing power network is read and the following processing is performed: the target computing power network has a unique distributed code, and the code sequence is scalable; if there is an entity change in the target computing power network, the computing power virtual architecture is synchronously updated.

[0008] In a possible implementation, the establishment of the upper-level coordinator also performs the following processing: guided by the decision control based on image processing, a decision control plane is constructed based on the computing power virtual architecture; based on the demand of network communication transmission, an image transmission plane is constructed based on the computing power virtual architecture; the decision control plane and the image transmission plane are combined, and sample data supervised training is called until convergence to generate the upper-level coordinator.

[0009] In a possible implementation, the sample data is called for supervised training until convergence to generate the upper-level coordinator, and the following processing is also performed: historical image processing records are retrieved and called as training samples, and the training samples include input samples and output samples divided by mapping; the decision control plane and the image transmission plane are supervised trained based on the training samples, and verification samples are screened to obtain verification outputs, wherein the verification samples are random input samples of a first number; based on the verification output and the mapped output samples, error compensation training is performed to obtain the constructed upper-level coordinator.

[0010] In a possible implementation, the upper coordinator is configured with unified management rules based on a third perspective, and also performs the following processing: traverse the target computing power network to determine distinctive network node characteristics, wherein the network node characteristics include computing power resource heterogeneity and network protocol differences; based on the network node characteristics, perform unified perspective operation and maintenance conversion based on global image processing to determine unified management rules based on a third perspective; and configure the upper coordinator based on the unified management rules.

[0011] In a possible implementation, the image processing method based on the computing power network also performs the following processing: if the image processing task is a multi-concurrent task, multiple groups of image processing nodes are dissociated and determined; the multiple groups of image processing nodes are integrated based on computing power homology, computing power collision analysis is performed based on phase constraints, and collision image processing nodes are determined; the target computing power network performs computing power resource perception, performs virtual-real interaction, traverses the collision image processing nodes for scheduling and avoidance, and determines the collision coordination strategy.

[0012] In a possible implementation, the image processing method based on the computing power network also performs the following processing: reading the image processing records of a predetermined time zone, and mining the image processing rules based on the time phase cycle; based on the image processing rules, identifying the computing power resource coding and the node computing power ratio, and generating a deterministic network; configuring a resource reservation mechanism based on image processing for the deterministic network.

[0013] The image processing method based on computing power network proposed in this application is proposed to read the distributed target computing power network, perform multi-granularity reconstruction deployment and computing power resource encoding based on the computing power type of image processing, determine the computing power virtual architecture, and establish a virtual-to-real mapping between the computing power virtual architecture and the target computing power network; based on the computing power virtual architecture, separate the data plane and the control plane, and establish an upper coordinator in combination with data-driven training, and the upper coordinator is configured with a unified management rule based on the third perspective; read the image processing task, perform task decomposition to determine multiple image processing nodes, couple and bind the multiple image processing nodes with the required computing power configuration, and determine the image processing target; based on the upper coordinator, perceive the network computing power status and perform decision analysis based on the image processing target to determine the virtual layer strategy; based on the virtual-to-real mapping relationship, convert the virtual layer strategy based on the target computing power network to determine the physical layer strategy, and the physical layer strategy includes a network transmission strategy and an image processing strategy; based on the physical layer strategy, respond to the target computing power network node of the target computing power network and execute the image processing task. It solves the technical problems of existing image processing, such as being based on a single computing node, difficulty in real-time large-scale, high-concurrency image processing, and inability to accurately match image processing tasks with computing resources, which in turn leads to insufficient real-time data transmission for image processing and low execution efficiency and resource utilization. It achieves rapid response and efficient execution of image processing tasks, achieving the technical effect of improving image processing efficiency and resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solution of the embodiment of the present disclosure, the accompanying drawings of the embodiment of the present disclosure will be briefly introduced below. A flow chart is used in the present application to illustrate the operations performed by the system according to the embodiment of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps can be processed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or more operations can be removed from these processes.

[0015] Figure 1 A schematic diagram of a flow chart of an image processing method based on a computing network provided in an embodiment of the present application;

[0016] Figure 2 A schematic diagram of the process of determining a computing power virtual architecture in a computing power network-based image processing method provided in an embodiment of the present application. DETAILED DESCRIPTION

[0017] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0018] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0019] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms "first\second" involved are merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "including" and "having" and any variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by technicians in the technical field of this application. The terms used herein are for the purpose of describing the embodiments of the present application only.

[0020] The present application embodiment provides an image processing method based on a computing network, such as Figure 1 As shown, the method includes:

[0021] Step S100, read the distributed target computing power network, perform multi-granularity reconstruction deployment and computing power resource encoding based on the computing power type of image processing, and determine the computing power virtual architecture. The computing power virtual architecture and the target computing power network establish a virtual-real mapping. Discover, identify and locate all available computing resources in the network. These resources may include various servers, data centers, cloud resources, edge computing nodes, etc., which are distributed in different geographical locations, different data centers or edge computing nodes. These computing resources together constitute a distributed target computing power network. Perform multi-granularity reconstruction deployment based on the computing power type of image processing. According to the specific needs of the image processing task, the collected computing resources are reconstructed and deployed at multiple granularities, including determining which resources are suitable for processing which type of image processing tasks (such as image compression, enhancement, filtering, segmentation, recognition, etc.), and how to combine these resources to form an optimal computing environment. Multi-granularity reconstruction deployment refers to the reconstruction of resources at different levels from a single computing node to the entire data center to meet image processing tasks of different scales and complexities. Computing power resource encoding refers to encoding the reconstructed computing resources so that the system can identify and schedule these resources. The encoding may include resources. The identifier, location information, performance parameters (such as CPU, GPU, memory and other specifications), availability, etc. of the source are used to achieve standardized management and efficient scheduling of computing resources, and then a virtual computing power architecture is constructed. This architecture is an abstraction and mapping of actual physical computing power resources. The computing power virtual architecture encapsulates physical computing power resources into an independent virtual computing environment. Each environment has specific computing capabilities and resource configurations. The virtual architecture has the characteristics of flexibility, scalability and manageability. It can be dynamically adjusted and optimized according to the needs of image processing tasks. After determining the computing power virtual architecture, a virtual-to-real mapping relationship is established between it and the actual target computing power network, ensuring that each virtual computing environment in the virtual architecture can find the corresponding physical computing resources to realize its function, realizing seamless connection and efficient collaboration between virtual computing environments and physical computing resources, realizing effective integration and efficient utilization of distributed computing resources, and providing powerful computing power support for image processing tasks.

[0022] In one possible implementation, Figure 2As shown, step S100 further includes step S110, based on the computing power type, clustering and segmenting the distributed target computing power network to determine the first computing power. The computing power type generally refers to the capability characteristics of computing resources. The computing power type can be classified according to different standards, such as the type of computing device (CPU, GPU, ASIC, etc.), computing accuracy (FP64, FP32, FP16, etc.) or the type of computing task (general computing, dedicated acceleration, etc.). Each computing power type has different advantages and efficiencies when processing different types of tasks. Cluster analysis is performed according to the computing power type of each node in the computing power network, and nodes with similar computing power types in the distributed target computing power network are divided into the same cluster. After the clustering and segmentation is completed, the nodes in each cluster will share similar computing power characteristics, thereby determining the first computing power, which refers to any one of the multiple computing power types.

[0023] Step S120, based on the distributed location and computing power type, hierarchical distributed network deployment is performed on the first computing power to determine the initialization and reconstruction network. The distributed location, i.e., the geographical location of the node, affects the delay of data transmission, bandwidth utilization, and network reliability. Specifically, based on the distributed location and computing power type of the node, a hierarchical network deployment scheme is designed, which usually includes multiple levels, such as edge layer, aggregation layer, and core layer. Each level is responsible for different network functions and data processing tasks. Through hierarchical distributed network deployment, an initialized reconstructed network is constructed. On the basis of maintaining the original computing power resources, the network efficiency and reliability are improved by optimizing the connection and resource configuration between nodes.

[0024] Step S130, based on the image processing elements, traverse the initialization reconstruction network, perform multi-level granularity division of each reconstruction grid, and determine multiple groups of computing power grids, wherein each computing power network node corresponds to a group of computing power grids. Image processing elements refer to various elements related to image processing tasks, such as image resolution, processing algorithm, computational complexity, etc., which will affect the performance and efficiency of the computing power network when processing image processing tasks. The constructed initialization reconstruction network is traversed to understand the status and performance of each node and connection. According to the image processing elements and node performance, each grid in the reconstruction network is divided into multi-level granularity, that is, the network is divided into regions of different sizes (called computing power grids). The nodes in each grid have similar computing power and processing performance. Through multi-level granularity division, multiple groups of computing power grids are obtained, and the nodes in each group of grids will jointly undertake part of the work in the image processing task to achieve efficient allocation and execution of tasks.

[0025] Step S140, constructing the computing power virtual architecture based on the initialized reconstructed network and the multiple groups of computing power grids. Based on the initialized reconstructed network and the multiple groups of computing power grids, constructing the computing power virtual architecture, specifically, abstracting and encapsulating the physical computing power resources, including encapsulating the computing power, storage capacity, network bandwidth and other resources of the physical nodes into a virtual computing power resource pool, and uniformly managing and scheduling through virtualization management software, through virtualization technology, physical servers, storage devices and network devices and other resources can be pooled to form a resource pool, and dynamically allocated to different virtual machines and containers as needed, so as to improve resource utilization and flexibility, after completing the computing power resource abstraction and encapsulation, constructing the computing power virtual architecture based on the initialized reconstructed network and the multiple groups of computing power grids, including virtual nodes, virtual links, virtual services and other elements, together forming a scalable and configurable virtual computing power network, after the construction is completed, it is also necessary to optimize and adjust the computing power virtual architecture, including dynamically adjusting and optimizing the configuration of computing power resources according to the needs of actual application scenarios; and monitoring and managing the performance and stability of the computing power virtual architecture in real time and evaluating it.

[0026] In a possible implementation, step S100 further includes that the target computing power network has a unique distributed code, and the coding sequence is scalable. Distributed coding refers to a special processing of data in a distributed system for the security, integrity or efficient transmission of data, which is used to ensure the consistency and identifiability of data when it is transmitted or stored between multiple nodes. Uniqueness means that in the computing power network, each distributed code is unique, and can clearly identify or distinguish different data blocks, tasks or nodes to ensure the accuracy, integrity and traceability of the data, and avoid data confusion or errors. The scalability of the coding sequence means that the coding sequence can be flexibly adjusted as the scale of the system expands to support more data blocks, tasks or nodes. For example, according to the size of the data block or the complexity of the task, the length of the code is dynamically adjusted, such as the subsequent grid division of the network node, which can be continued to be attached after the coding to distinguish the grid. It also includes that if there is a physical change in the target computing power network, the computing power virtual architecture is updated synchronously. Through sensors and log analysis tools deployed in the network, the operating status and performance indicators of the target computing network are monitored in real time, including the computing power, storage status, network bandwidth, etc. of the nodes. Data analysis and machine learning algorithms are used to process and analyze the monitored data, identify entity changes in the network, such as new nodes, node failures, link congestion, etc., and reconfigure the resources in the virtual architecture according to the entity changes. For example, if a node fails, the virtual architecture will automatically migrate the tasks of the node to other available nodes; if a new node is added, the virtual architecture will include it in the resource pool and assign tasks as needed. According to the impact of the entity changes and the current status of the virtual architecture, the computing power allocation, load balancing, fault recovery and other strategies are adjusted to ensure the stability and efficiency of the computing service.

[0027] Step S200, based on the computing power virtual architecture, separate the data plane and the control plane, and establish an upper-level coordinator in combination with data-driven training. The upper-level coordinator is configured with unified management rules based on the third perspective. Based on the determined computing power virtual architecture, the data plane and the control plane in the computing power network are separated. Specifically, the separation of the data plane and the control plane is the key to achieving efficient management and flexible scheduling. The data plane is mainly responsible for the actual forwarding and processing of data. In image processing computing tasks, the data plane is responsible for receiving input data (such as image data), executing corresponding computing tasks, and outputting processing results. It includes a series of processing units (such as CPU, GPU, etc.). These processing units process input data according to the instructions and policies of the control plane. At the same time, the data plane also handles data transmission and storage to ensure that data can flow efficiently and accurately in the network; the control plane is mainly responsible for decision-making, scheduling and management, and collects network status. The control plane collects information, computing resource status, and task requirements, and then formulates scheduling strategies and control instructions based on this information and sends them to the data plane for execution. It makes decisions on the allocation of computing resources and the scheduling of tasks. Based on the decision results, the control plane allocates computing resources to the corresponding tasks and monitors the execution of tasks. It is also responsible for the management and maintenance of the entire computing network, including optimal resource allocation, fault detection and recovery, etc. By separating the data plane from the control plane, it can respond to network changes and resource fluctuations more flexibly. The control plane can perform intelligent scheduling and management based on task priority and resource status to ensure efficient use of computing resources. The data plane can reduce network latency and storage costs by optimizing data transmission and storage strategies. Data-driven training refers to the use of data-driven technologies such as machine learning and deep learning to train and optimize the decision-making algorithm of the upper coordinator through the analysis of a large amount of historical data and real-time data. That is, by collecting and analyzing information from different data sources (such as network status, task execution efficiency, computing resource status, image processing task requirements, etc.), the upper coordinator can more accurately predict network status and resource requirements, thereby formulating more effective scheduling strategies and management rules. Among them, the third perspective refers to examining and managing the entire system from a global perspective. The upper coordinator is configured with unified management rules based on the third perspective, which means that it can transcend the limitations of a single data plane or control plane and formulate management rules and scheduling strategies from the perspective of the entire system. The unified management rules may include resource allocation strategies, task scheduling algorithms, fault recovery mechanisms, etc., which together constitute the core management logic of the upper coordinator and are the basis for the upper coordinator to make decisions and schedule, ensuring the collaborative work and resource sharing between different parts of the system.

[0028] In a possible implementation, step S200 further includes step S210, which is guided by decision control based on image processing and based on the computing power virtual architecture to construct a decision control plane. The decision control plane focuses on processing image data, uses image processing technology and machine learning algorithms to analyze images, makes decisions based on the analysis results, and finally generates control signals to guide the devices in the actual application scenarios to perform corresponding actions. The computing power virtual architecture is the basis for building the decision control plane. Through virtualization technology, computing resources, storage resources and network resources are encapsulated into an independent virtual environment to provide the required computing power support for the decision control plane. The steps of building the decision control plane include clarifying the application scenarios, functional requirements and performance indicators of the decision control plane, designing the system architecture, algorithm flow and interactive interface of the decision control plane according to the demand analysis results, deploying image processing algorithms and machine learning models in the computing power virtual architecture, realizing image data processing, feature extraction and decision analysis functions, and at the same time, developing modules for control signal generation and transmission, integrating modules such as image processing, decision analysis and control signal transmission into the decision control plane, and performing joint debugging and testing to ensure the collaborative work between the modules, and finally obtaining the decision control plane.

[0029] Step S220, based on the computing power virtual architecture, an image transmission plane is constructed with network communication transmission as the demand. Depending on the application scenario, the size and transmission frequency of the image data will also vary. The required network bandwidth and storage capacity are evaluated and determined. For application scenarios with high real-time requirements (such as video surveillance, telemedicine, etc.), it is necessary to strictly control the transmission delay of the image data. According to the reliability requirements of the application scenario, an appropriate error detection and correction mechanism is selected to ensure the integrity of the image data during transmission. Based on the computing power virtual architecture, for example, according to the changes in the demand for image data transmission, the allocation of network resources (such as bandwidth, IP addresses, etc.) is dynamically adjusted. Through redundant design, load balancing and other mechanisms, the high availability and stability of the image transmission plane are improved to ensure that the image data can be transmitted normally in the event of a single point of failure. The construction of the transmission plane, specifically, according to the results of the demand analysis of network communication transmission, designs a network architecture suitable for image data transmission, including selecting a suitable network topology, configuring network equipment (such as switches, routers, etc.), and planning network addresses, etc. According to the characteristics of image data and transmission requirements, select a suitable transmission protocol, such as TCP (Transmission Control Protocol) and UDP (User Datagram Protocol), etc. Then, according to the characteristics of image data, optimize the transmission algorithm to reduce transmission delay and bandwidth occupancy. For example, use a compression algorithm to compress image data to reduce the amount of transmitted data. Finally, obtain the image transmission plane, integrate the image transmission plane with the computing power virtual architecture, and realize the dynamic allocation and automation of resource management.

[0030] Step S230, combining the decision control plane and the image transmission plane, calling sample data supervised training until convergence, and generating the upper coordinator. The decision control plane and the image transmission plane are integrated through standard network communication protocols and interfaces to achieve seamless data transmission and sharing, calling sample data supervised training, specifically, preparing a large amount of representative sample data, including image data and corresponding labels (such as target categories, location information, etc.), preprocessing the sample data, and using a deep learning model to train the sample data. During the training process, the model learns the feature representation of the image data and establishes a mapping relationship between the feature and the label, receives the real label from the sample data as a supervisory signal, and minimizes the difference between the predicted label and the real label by continuously adjusting the model parameters until the model converges to generate an upper coordinator, which is responsible for receiving the image data and processing results from the image transmission plane, generating control signals according to the decision analysis results of the decision control plane, and coordinating the work of each execution device or system.

[0031] In a possible implementation, step S230 further includes step S231, retrieving and calling historical image processing records as training samples, wherein the training samples include input samples and output samples divided by mapping. Retrieve and call historical image processing records from a database, which records may include the results of various image processing tasks, such as image classification, object detection, image segmentation, etc., including image data before and after processing and corresponding processing results or labels, and organize the retrieved historical image processing records into a training sample set, which should include input samples (i.e., original images to be processed) and output samples (i.e., processed images or corresponding labels).

[0032] Step S232, supervised training is performed on the decision control plane and the image transmission plane based on the training samples, and verification samples are screened to obtain verification output, wherein the verification samples are random input samples of the first number. The decision control plane and the image transmission plane are trained using the training sample set. The decision control plane is responsible for processing image data and making decisions, while the image transmission plane is responsible for transmitting image data. The two planes may interact in some way (such as a network interface, a message queue, etc.). During the training process, a certain number of verification samples (i.e., random input samples) are randomly screened from the training sample set for verification during the training process to ensure the training effect of the model. The model is tested using the screened verification samples to obtain verification output. The verification output is the result of the model processing the verification samples, which is used to compare with the real output samples to evaluate the accuracy and performance of the model.

[0033] Step S233, based on the verification output and the mapped output sample, error compensation training is performed to obtain the constructed upper coordinator. The mapped output sample is a known sample with a real processing result or label, corresponding to the verification sample, and is used to evaluate the prediction accuracy of the model. Specifically, the error between the verification output and the mapping output sample is analyzed to identify the source and nature of the error, for example, whether it is a systematic error or a random error, a defect of the model itself or a problem of data preprocessing, etc. According to the results of the error analysis, the training strategy is adjusted, for example, including modifying the architecture of the model, adjusting the parameters of the model, improving the data preprocessing method, introducing regularization terms to reduce overfitting, etc. The adjusted model is retrained and tested with a new verification sample to evaluate the effect of the improvement, and multiple iterations are performed until the prediction error of the model reaches an acceptable range, and finally the constructed upper coordinator is obtained.

[0034] In a possible implementation, step S200 further includes step S240, traversing the target computing power network to determine distinctive network node features, wherein the network node features include computing power resource heterogeneity and network protocol differences. The entire target computing power network is comprehensively traversed to identify all nodes, links, and connection relationships between them in the network, and to obtain basic information such as the network topology and node distribution. During the traversal process, the characteristics of each network node are focused on, which may include the heterogeneity of computing power resources (i.e., different nodes may have computing power resources of different types and performances) and differences in network protocols (i.e., different nodes may follow different communication protocols).

[0035] Step S250, based on the network node characteristics, a unified perspective operation and maintenance conversion based on global image processing is performed to determine a unified management rule based on a third perspective. Using image processing technology, each node, link and other elements in the target computing power network are visualized to form a global network image, which clearly displays the network's topological structure, node distribution and connection relationship between nodes. Based on the network node characteristics, the global network image is further processed and analyzed to achieve a unified perspective operation and maintenance conversion, including abstracting and normalizing the differences between different nodes and different protocols, so as to uniformly manage and operate the entire network from a global perspective. On the basis of a unified perspective, a unified management rule based on a third perspective is formulated to cover all nodes and links in the entire network, ensuring consistency and efficiency in the management and operation process.

[0036] Step S260, based on the unified management rules, configure the upper coordinator. According to the determined unified management rules, the upper coordinator is configured accordingly, including setting management policies, configuring communication protocols, allocating computing resources, etc. Through configuration, it is ensured that the upper coordinator can manage and schedule the entire target computing network according to unified rules. After the configuration is completed, the upper coordinator needs to be tested and verified, including simulating various network scenarios and fault conditions to verify the performance and reliability of the upper coordinator.

[0037] Step S300, read the image processing task, perform task decomposition to determine multiple image processing nodes, couple and bind the multiple image processing nodes with the required computing power configuration, and determine the image processing target. The image processing task is read from the user request, the application call or the system internal task. The task contains information such as the image data to be processed, the processing requirements (such as denoising, enhancement, recognition, etc.) and the expected output results. After reading the task, the image processing task is decomposed and decomposed into multiple subtasks that can be executed in parallel or serially. Specifically, the image processing task is deeply analyzed to understand the specific requirements and goals of the task, including understanding the characteristics of the image data (such as resolution, color depth, etc.), the complexity of the processing requirements and the expected output results, etc., and the image processing task is divided into multiple subtasks. For example, a complex image processing task may include preprocessing (such as denoising, enhancement), feature extraction, classification and recognition, etc. Multiple image processing nodes are determined according to the results of task decomposition and analysis. Each node has specific processing capabilities and resources for executing different parts of the image processing task; then based on the image processing task The specific requirements of the image processing system and the processing capacity of the nodes are considered to evaluate the computing power configuration required for each node. According to the evaluation results, the corresponding computing power resources are allocated to each image processing node, including binding specific CPU, GPU and other resources to the corresponding nodes to ensure that the nodes can obtain sufficient computing power and resource support when performing tasks. Then, according to the characteristics of the task, the processing capacity of the nodes and the resource status, the different parts of the image processing task are mapped to the corresponding image processing nodes to ensure that the tasks can be efficiently and evenly allocated to each node. A reasonable task scheduling strategy (such as load balancing, priority sorting, and task dependency management) is formulated to optimize the utilization efficiency of the nodes and the execution speed of the tasks. Finally, the image processing goals are determined, including the expected output results (feature extraction, image classification, etc.), performance indicators (such as processing speed, accuracy, resource utilization, etc.), and constraints (such as time limit, resource limit, data privacy, etc.). By coupling and binding multiple image processing nodes with the required computing power configuration, it is ensured that the image processing task can be completed efficiently and accurately.

[0038] In a possible implementation, step S300 further includes step S310, if the image processing task is a multi-concurrent task, dissociate and determine a plurality of groups of image processing nodes. If the image processing task is a multi-concurrent task, the task is dissociated into a plurality of parts that can be processed in parallel according to the independence, data dependency and processing complexity of the task, for example, by image area, by processing steps, by task priority, etc. Each part can be assigned to one or more image processing nodes for processing. By dissociating the task into a plurality of subtasks and determining a plurality of groups of image processing nodes, the parallel processing capability of the distributed computing environment is fully utilized to improve the overall efficiency of image processing.

[0039] Step S320, integrating the multiple groups of image processing nodes based on computing power homology, performing computing power collision analysis based on time phase constraints, and determining the colliding image processing nodes. Computing power homology refers to the similarity of different computing nodes in terms of computing power, processing speed, resource utilization, etc., integrating multiple groups of image processing nodes based on computing power homology, specifically, performing performance evaluation on each image processing node, including key indicators such as computing speed, memory size, storage capacity, network bandwidth, etc., dividing nodes with similar computing power into different groups according to the performance evaluation results, and performing unified resource management on the nodes in each group, including task allocation, resource scheduling, load balancing, etc., by integrating similar nodes together, computing resources can be used more effectively, avoiding waste and fragmentation of resources, and similar task scheduling strategies can be adopted for nodes with similar computing power to reduce scheduling complexity and error rate, similar nodes can maintain high execution efficiency when processing the same or similar tasks, and reduce performance fluctuations caused by node differences; time phase constraints refer to the task being executed. The computing power collision analysis aims to identify computing power resource conflicts and competitions that may occur in the parallel processing process. When multiple tasks request the same or similar image processing nodes at the same time, computing power collisions may occur, resulting in task execution delays or failures. Specifically, by simulating the task scheduling process, it is predicted which nodes may become collision points, such as simulating different task allocation strategies and execution sequences, and monitoring the performance status of each node in real time, including key indicators such as CPU usage, memory occupancy, and network bandwidth. When it is found that the performance indicator of a node is close to or exceeds the threshold, it means that the node may soon become a collision point. A special conflict detection algorithm is used to identify potential computing power collisions. Through computing power collision analysis, it is determined which image processing nodes may become collision points. These nodes usually have high task loads and low resource availability.

[0040] Step S330, the target computing power network performs computing power resource perception, performs virtual-real interaction, traverses the collision image processing nodes for scheduling and avoidance, and determines the collision coordination strategy. Computing power resource perception means that the computing power network can obtain the computing power resource information of each node in the network in real time and accurately, including the performance, load, availability, etc. of the computing power nodes. Each computing power node actively reports its own computing power resource information to the computing power perception center. The computing power perception center aggregates the reported information of each node to build a global unified computing power resource view to ensure efficient use of computing power resources and avoid resource waste; virtual-real interaction refers to the interaction and mapping between the computing power resources of the real world and the task requirements of the virtual world. Specifically, it is to realize communication and data transmission between the virtual world and the real world through network protocols and interfaces, and use virtualization technology to map image processing tasks in the virtual world to the computing power resources in the real world for processing. The system can accurately match computing resources with task requirements and improve the efficiency and accuracy of task processing. It can check and analyze nodes that may collide with computing power due to the simultaneous processing of multiple image processing tasks one by one. According to the analysis results, it can reallocate tasks or adjust the execution order to avoid or reduce the impact of computing power collisions, ensure the smooth execution of image processing tasks, and improve the overall utilization of computing power resources and the stability of the system. It can then determine the collision coordination strategy. For example, when the system load is high, a more stringent scheduling and avoidance strategy can be adopted; when the system load is low, the strategy can be relaxed to improve resource utilization. It can combine load balancing, task priority adjustment, resource reservation and other strategies to formulate collision coordination strategies. It can reduce the impact of computing power collisions on system performance and task execution, optimize resource utilization and improve the overall performance of the system while solving the collision problem.

[0041] Step S400, based on the upper coordinator, perceive the network computing power status and perform decision analysis based on the image processing target to determine the virtual layer strategy. The upper coordinator obtains the computing power resource status in the network in real time through its built-in monitoring mechanism or the interface with the external monitoring system, including the utilization rate, idle rate, load status, etc. of computing units such as CPU, GPU, FPGA, etc., processes and analyzes the collected computing power status data to extract useful information and trends, and evaluates the network computing power status based on the results of data analysis to determine whether the current computing power resources meet the needs of the image processing task, and whether there are problems such as resource bottlenecks or waste, and then performs decision analysis based on the image processing target. Specifically, the upper coordinator deeply understands the specific requirements and goals of the image processing task, including indicators such as processing speed, accuracy, and resource utilization, matches the needs of the image processing task with the current network computing power status, analyzes whether there is a shortage or surplus of computing power resources, and makes a decision based on the task requirements and Based on the analysis results of the computing power status, appropriate decision plans are formulated, such as adjusting the allocation strategy of computing power resources, optimizing the execution plan of image processing tasks, starting or stopping certain computing nodes, etc., and then determining the virtual layer strategy. According to the results of the decision analysis, the upper coordinator will determine the specific strategy of the virtual layer, which may include resource virtualization, isolation and sharing, dynamic scheduling and fault recovery. Resource virtualization refers to the virtualization of physical computing power resources into logical resource pools so that they can be dynamically allocated according to task requirements; isolation and sharing refers to the use of virtualization technology to achieve resource isolation and sharing between different tasks to ensure the security and efficiency of tasks; dynamic scheduling refers to the dynamic adjustment of the allocation and scheduling strategy of virtual resources according to the changes in the network computing power status and the demand for image processing tasks; fault recovery refers to the formulation of a fault recovery strategy to ensure that task execution can be quickly restored when computing power resources fail.

[0042] Step S500, based on the virtual-real mapping relationship, the virtual layer strategy is converted based on the target computing power network to determine the physical layer strategy, and the physical layer strategy includes a network transmission strategy and an image processing strategy. The virtual-to-real mapping relationship enables the strategies and resource configurations in the virtual layer to be mapped to the physical resources and network topology of the physical layer. The conversion of the virtual layer strategy based on the target computing network means that the abstract strategies and resource management schemes formulated in the virtual layer are adapted, optimized and adjusted according to the actual situation and characteristics of the target computing network to form strategies that can be effectively executed in the physical network. Specifically, a comprehensive understanding of the target computing network is conducted, including its network architecture, node distribution, resource capacity, performance parameters, etc. The strategies formulated in the virtual layer are deeply analyzed to clarify their goals, constraints and expected effects. According to the resource situation of the target computing network, the resource allocation strategy in the virtual layer is adapted. For example, if some nodes in the target network have abundant resources, more tasks will be scheduled to these nodes. Otherwise, task allocation needs to be optimized to avoid resource overload. According to the performance characteristics of the target computing network, the task scheduling scheme in the virtual layer is optimized. For example, low-latency links in the network can be used to accelerate data transmission. Or use load balancing technology to balance the load of each node and improve the overall performance; according to the security requirements of the target computing power network, strengthen the security policies in the virtual layer, for example, deploy firewalls, intrusion detection systems and other security devices to prevent network attacks and data leakage; refine the policies in the virtual layer into specific operational steps that can be executed in the physical layer, for example, refine the resource allocation policy in the virtual layer into specific resource allocation instructions and parameter settings in the physical layer; after policy conversion, a set of policies that can be executed in the physical layer is obtained, namely the physical layer policy, which clearly specifies the specific operations and rules on how to configure resources, schedule tasks, and ensure security in the physical network. Among them, the physical layer policy mainly includes network transmission policy and image processing policy. The network transmission policy mainly involves how to effectively transmit image data in the physical network, including routing selection, flow control, data compression and security policy; the image processing policy refers to the efficient and accurate processing of image data at the receiving end, including preprocessing, processing algorithm selection, parallel processing and resource allocation.

[0043] Step S600, based on the physical layer strategy, in response to the target computing power network node of the target computing power network, execute the image processing task. In the physical layer of the computing power network, according to the pre-established physical layer strategy, the image processing task is assigned to a specific node in the target computing power network for processing, including task scheduling, resource allocation, algorithm execution, etc. Specifically, the target computing power network node receives the image processing task, the node parses the received image processing task, and according to the task scheduling rules in the physical layer strategy, the image processing task is assigned to the appropriate node for processing; sufficient computing resources (such as CPU, GPU, etc.) are allocated to the target computing power network node to ensure that the image processing task can be executed smoothly, and necessary storage resources are allocated to the task to store input data, intermediate results and final output, to ensure smooth communication between nodes and between nodes and external networks, so as to facilitate data transmission and result return; according to the specific requirements of the image processing task, a suitable image processing algorithm is selected, and the selected algorithm is configured, including parameter setting, optimization strategy, etc., to ensure that the algorithm can be executed efficiently and accurately, and the image processing algorithm is executed on the target computing power network node to process the input data and generate the expected output result, and the result of the image processing task is verified to ensure the accuracy and completeness of the result.

[0044] In a possible implementation, the image processing method based on the computing power network further includes step S710, reading the image processing records in a predetermined time zone, and mining the image processing rules based on the time phase cycle. Reading the image processing records in the predetermined time zone refers to retrieving the image processing records in a specific time zone from a database or log file, which may contain information such as the timestamp, processing type, and processing results of the image processing, and analyzing these records through statistical methods (such as time series analysis, cluster analysis, etc.) to identify the distribution rules and periodic characteristics of image processing activities in different time periods (such as daily, weekly, monthly, etc.), that is, external factors that affect image processing activities (such as user behavior, system load, etc.). Image processing rules are used to optimize the scheduling and management of image processing tasks and improve the overall performance of the system and user experience.

[0045] Step S720, based on the image processing rules, identify the computing power resource code and the node computing power ratio, and generate a deterministic network. According to the mined image processing rules, evaluate the demand for computing power resources in different time periods, identify the computing power code (which may be a unique identifier) ​​of each image processing node (or computing resource) and their computing power ratio in different time periods (that is, the proportion of the processing tasks they undertake to the total task volume), which helps to understand the distribution and utilization of the system's computing power resources, and provide basic data for subsequent deterministic network generation, and then generate a deterministic network. A deterministic network is a network architecture that can ensure that data transmission and processing are deterministic and stable. In the field of image processing, a deterministic network can be used to ensure that image processing tasks can be executed according to a predetermined schedule and performance requirements. Specifically, according to the characteristics and requirements of the image processing tasks, a network architecture with determinism and stability is designed. According to the image processing rules and the node computing power ratio, sufficient computing power resources are reserved for image processing tasks in different time periods. At the same time, an efficient resource scheduling algorithm is designed to ensure that tasks can be allocated to appropriate computing power nodes for execution as needed.

[0046] Step S730: configuring a resource reservation mechanism based on image processing for the deterministic network. In a deterministic network architecture, according to the characteristics and requirements of image processing tasks, sufficient resources (such as computing resources, network resources, etc.) are reserved for these tasks in advance to ensure that the tasks can be executed according to the predetermined requirements and achieve the expected performance indicators, including resource evaluation, reservation strategy formulation, reservation implementation and other links. Specifically, the image processing tasks are evaluated for resources to understand their requirements for computing resources, storage resources and network resources, such as analyzing the complexity of the task, processing speed requirements, data throughput, etc., and determining the minimum amount of resources required for task execution; based on the results of resource evaluation, a resource reservation strategy is formulated to ensure the rationality and effectiveness of resource reservation, such as reserving more resources for high-priority or high-complexity image processing tasks to ensure their smooth execution; at the same time, the overall utilization of resources must also be considered to avoid excessive reservation leading to resource waste; after determining the resource reservation strategy, the reservation implementation is carried out, including configuring the corresponding resource reservation mechanism in the deterministic network architecture, such as setting a resource reservation table, updating the routing strategy, adjusting the resource allocation algorithm, etc., to ensure that sufficient resources are reserved for the image processing tasks during task execution and to prevent other tasks from preempting these resources.

[0047] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. An image processing method based on a computing network, characterized in that: The method comprises: Read the distributed target computing power network, perform multi-granularity reconstruction deployment and computing power resource encoding based on the computing power type of image processing, and determine the computing power virtual architecture, where a virtual-real mapping is established between the computing power virtual architecture and the target computing power network; Based on the computing power virtual architecture, the data plane and the control plane are separated, and an upper coordinator is established in combination with data-driven training. The upper coordinator is configured with unified management rules based on the third perspective; Read the image processing task, perform task decomposition to determine multiple image processing nodes, couple and bind the multiple image processing nodes with the required computing power configuration, and determine the image processing target; Based on the upper coordinator, the network computing power status is sensed and decision analysis based on the image processing target is performed to determine the virtual layer strategy; Based on the virtual-real mapping relationship, the virtual layer strategy is converted based on the target computing power network to determine the physical layer strategy, where the physical layer strategy includes a network transmission strategy and an image processing strategy; Based on the physical layer strategy, the target computing network node of the target computing network executes the image processing task; The step of establishing a higher-level coordinator includes: Guided by decision control based on image processing, a decision control plane is constructed based on the computing power virtual architecture; Taking network communication transmission as the demand, an image transmission plane is constructed based on the computing power virtual architecture; Combining the decision control plane with the image transmission plane, calling sample data supervised training until convergence, and generating the upper coordinator; The calling of sample data supervised training until convergence to generate the upper coordinator includes: Retrieving and calling historical image processing records as training samples, wherein the training samples include input samples and output samples divided by mapping; Performing supervised training on the decision control plane and the image transmission plane based on the training samples, and screening verification samples to obtain verification outputs, wherein the verification samples are random input samples of a first number; Based on the verification output and the mapped output sample, error compensation training is performed to obtain the constructed upper coordinator; The upper coordinator is configured with unified management rules based on the third perspective, including: Traversing the target computing power network to determine distinctive network node characteristics, wherein the network node characteristics include computing power resource heterogeneity and network protocol differences; Based on the network node characteristics, unified perspective operation and maintenance conversion based on global image processing is performed to determine unified management rules based on the third perspective; Based on the unified management rule, the upper coordinator is configured.

2. The image processing method based on computing network according to claim 1, characterized in that: The determining of the computing power virtual architecture includes: Based on the computing power type, the distributed target computing power network is clustered and segmented to determine the first computing power Based on the distributed location and computing power type, perform hierarchical distributed network deployment on the first computing power, and determine to initialize and reconstruct the network; Based on the image processing elements, the initialized reconstruction network is traversed to perform multi-level granularity division of each reconstructed grid, and multiple groups of computing power grids are determined, wherein each computing power network node corresponds to a group of computing power grids; Based on the initialized reconstruction network and the multiple groups of computing power grids, the computing power virtual architecture is constructed.

3. The image processing method based on computing network according to claim 2, characterized in that: The target computing power network has unique distributed coding, and the coding sequence is scalable; If there are physical changes in the target computing power network, the computing power virtual architecture is updated synchronously.

4. The image processing method based on computing network according to claim 1, characterized in that: The method further comprises: If the image processing task is a multi-concurrent task, dissociate and determine multiple groups of image processing nodes; Integrate the multiple groups of image processing nodes based on computing power homology, perform computing power collision analysis based on time phase constraints, and determine the collision image processing nodes; The target computing power network perceives computing power resources, performs virtual-real interaction, traverses the collision image processing nodes for scheduling and avoidance, and determines a collision coordination strategy.

5. The image processing method based on computing network according to claim 1, characterized in that: The method further comprises: Read the image processing records of the predetermined time zone and mine the image processing rules based on the time phase cycle; Based on the image processing rules, identify the computing resource coding and node computing power ratio, and generate a deterministic network; An image processing based resource reservation mechanism is configured for the deterministic network.