Edge server visual task processing method and system

Abstract

The invention discloses a visual task processing method and system for an edge server, and belongs to the technical field of visual task processing, and the method comprises the steps: determining a local execution feasibility index of the local execution of a visual task by the edge server through multi-dimensional weighted evaluation, and combining the resource adaptation capability of a cloud server and the real-time demand of the visual task, quantizing a cloud uninstalling necessity index of cloud uninstalling the visual task, dynamically adjusting the original resolution of the visual task based on the direction of the cloud uninstalling necessity index in combination with the tolerance of the visual task to data quality and the residual resource space of the edge server to obtain an adjusted resolution, and adjusting the resolution of the visual task according to the adjusted resolution. According to the method, the local evaluation, cloud decision-making, self-optimization and local enhancement complete resource allocation closed loop is adopted, so that a set of feasible quantitative solution is provided for edge vision task processing.

Description

Technical Field

[0001] This invention relates to the field of visual task processing technology, and in particular to a visual task processing method and system for edge servers. Background Technology

[0002] With the popularization of scenarios such as industrial quality inspection, smart homes, and smart security, edge computing has gradually become the core architecture for visual task processing. It processes visual data on edge servers close to the acquisition end, which can reduce cloud bandwidth pressure and reduce latency.

[0003] Currently, most edge vision task processing technologies rely on "manually set thresholds," failing to consider the combined impact of hardware load, energy risks, and data adaptation. This leads to either idle edge resources (e.g., low CPU load but insufficient memory, yet still forcing local execution) or resource overload (e.g., CPU load below the threshold but temperature exceeding limits, causing device burnout). Furthermore, existing technologies often employ a simplistic "if the edge fails, then the cloud" logic, failing to quantify the collaborative relationship between edge gaps, cloud adaptation, and task timeliness. This results in a disconnect between cloud offloading priorities and actual needs, either "failing to offload when necessary" (forcing execution when the edge cannot), or "offloading when not necessary" (wasting cloud resources even when the edge can execute). Additionally, existing technologies often use "fixed resolution" or "random resolution downgrading," failing to balance offloading priorities, quality baselines, and resource remaining constraints. This causes resolution adjustments to become disconnected from task requirements, resulting in either "over-adjustment" (excessive accuracy loss) or "under-adjustment" (insufficient resource release). Consequently, edge servers face resource allocation and processing problems due to limited resources (CPU / memory / power constraints), diverse task requirements (different scenarios have significantly different requirements for accuracy and timeliness), and chaotic local-cloud collaboration. Summary of the Invention

[0004] The technical problem to be solved by this invention is that the existing technology has the disadvantages of "subjectivity" in local execution capability assessment, "blindness" in cloud offloading decision-making, and "arbitrariness" in resolution adjustment. To address this, we propose an edge server visual task processing method and system.

[0005] In a first aspect, one embodiment of the present invention provides a method for processing visual tasks on an edge server, the method comprising the following steps:

[0006] Before the edge server performs the vision task, hardware load data, energy status data and raw data characteristics of the vision task are collected. Through multi-dimensional weighted evaluation, the feasibility indicators of the local execution of the vision task on the edge server are determined.

[0007] Based on the gaps in the local execution feasibility indicators, and in combination with the resource adaptability of the cloud server and the real-time requirements of the visual task, the necessity indicators for cloud offloading of the visual task are quantified.

[0008] Based on the direction of the cloud offloading necessity index, and combined with the visual task's tolerance for data quality and the remaining resource space of the edge server, the original resolution of the visual task is dynamically adjusted to obtain the adjusted resolution.

[0009] Based on the adjusted resolution, the local execution capabilities of the edge server are reassessed, and the final execution plan for the vision task is output.

[0010] The final execution scheme includes local execution, cloud unloading, and edge-cloud collaboration.

[0011] Preferably, the method for determining the local execution feasibility indicators is as follows:

[0012] Before running a vision task on the edge server, collect the current CPU load rate, current memory load rate, real-time temperature, real-time power, and the resolution and complexity required by the vision task.

[0013] The overall hardware load is obtained based on the average of the current CPU load rate and the current memory load rate.

[0014] The temperature risk coefficient is obtained by proportionally relating the real-time temperature to the first difference and the second difference between the preset safe temperature and the temperature threshold. The energy consumption risk coefficient is obtained by proportionally relating the real-time power to the preset energy consumption threshold. The temperature risk coefficient and the energy consumption risk coefficient are then weighted and summed to obtain the energy risk coefficient.

[0015] The resolution matching coefficient is obtained by the proportional relationship between the visual data resolution and the third and fourth differences between the preset maximum edge support resolution and the preset base resolution, respectively. The complexity matching coefficient is obtained by the proportional relationship between the target feature complexity and the preset maximum edge support complexity. The resolution matching coefficient and the complexity matching coefficient are then weighted and summed to obtain the data fit.

[0016] Multiply the complement of the overall hardware load, the energy risk coefficient, and the data fit by the product to obtain the local execution feasibility index, which is recorded as the local execution confidence level.

[0017] Preferably, based on the local execution confidence level, the local execution capability is evaluated as follows:

[0018] If the local execution confidence is greater than the preset maximum local execution confidence, the edge server is evaluated to have high reliability in local processing, and the necessity index for cloud offloading of visual tasks is quantified.

[0019] If the local execution confidence is less than the preset minimum local execution confidence, the edge server is deemed to lack local processing capabilities and will be offloaded to the cloud.

[0020] Preferably, the method for determining the cloud uninstallation necessity index is as follows:

[0021] The gap in edge capabilities is defined by supplementing the local feasibility indicators.

[0022] Collect bandwidth and load data of the cloud server and resolution requirements of the visual task to determine the data compatibility with the cloud capabilities;

[0023] By combining the maximum allowable latency of the vision task with the edge processing latency of the edge server, the data timeliness factor for the real-time requirements of the task is calculated.

[0024] The edge capability gap based on the local execution confidence level is multiplied by the data adaptability, and then added to the data timeliness factor to obtain the cloud uninstallation necessity index, which is recorded as the cloud uninstallation priority.

[0025] Preferably, based on the cloud uninstallation priority and the determination of the necessity of cloud uninstallation, the following is performed:

[0026] If the cloud uninstallation priority is less than 1, then the necessity of cloud uninstallation is low, and local processing is selected first, and the original resolution of the visual task is dynamically adjusted.

[0027] If the cloud uninstallation priority is ≥1, then the necessity of cloud uninstallation is high, and cloud processing is selected as the priority.

[0028] Preferably, the method for dynamically adjusting the resolution parameter is as follows:

[0029] The uninstallation priority item is determined based on the linear transformation result of the cloud uninstallation priority.

[0030] The data quality coefficient is determined by comparing the accuracy loss requirements of the visual task with the accuracy loss data after adjusting the historical resolution.

[0031] The edge remaining resource coefficient is determined based on the ratio of the fifth difference between the maximum resources of the edge server and the currently used resources to the resources required by the task.

[0032] Multiply the unloading priority item by the data quality coefficient and add it to the edge remaining resource coefficient to obtain the resolution adjustment coefficient;

[0033] Based on the absolute value of the resolution adjustment coefficient, the original resolution of the visual task is reduced proportionally to obtain the adjusted resolution.

[0034] Preferably, the specific formula for calculating the adjusted resolution is as follows:

[0035] ;

[0036] In the formula: K res Res is the resolution adjustment factor. daj For the adjusted resolution, Res original This is the original resolution.

[0037] Secondly, embodiments of the present invention also provide an edge server vision task processing system, including a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that the processor executes the computer program to implement the steps of the above-described method.

[0038] Preferably, the system includes:

[0039] Perception layer module: includes the resource display panel of the edge server, temperature sensor, power meter, bandwidth tester and visual data acquisition device, used to collect hardware load, energy status, network status and visual raw data;

[0040] Decision-making layer module: includes local evaluation unit, cloud decision-making unit and resolution adjustment unit, used to execute multi-dimensional evaluation, quantitative decision and dynamic adjustment logic;

[0041] Execution layer module: includes edge server processing unit and cloud interaction unit, used to complete the local execution or cloud unloading of visual tasks according to the execution plan output by the decision layer;

[0042] Storage module: Used to store the historical execution logs of the edge server, the requirement parameters of visual tasks, and the system configuration thresholds.

[0043] The technical effects and advantages of this invention are as follows:

[0044] This invention achieves "multi-dimensional comprehensive evaluation" through hardware load averaging, energy risk weighting, and data adaptation matching logic. The average of the current CPU load rate and the current memory load rate avoids the bias of a single indicator. For example, if the CPU load is 60% but the memory load is 80%, the overall load is 70%, which reduces local execution confidence and prevents memory overload. Weighted calculation of temperature and power further provides early warning of equipment overheating. If the temperature exceeds the limit, the energy risk coefficient decreases, and the local execution confidence decreases accordingly, thus preventing equipment burnout. Finally, the "matching degree between data resolution and edge capabilities" is evaluated. For example, high-resolution medical images have low adaptability to low-configuration edge servers, reducing local execution confidence and guiding tasks to the cloud. This improves the utilization rate of edge resources and reduces equipment failure rate.

[0045] This invention achieves a balance between necessity and feasibility by using a reverse-driven edge gap, a positive-support cloud adaptation, and a timeliness factor constraint logic. Specifically, the complement of local execution confidence can quantify edge capability gaps; for example, an edge gap of 0.7 indicates that 70% of tasks require cloud support, ensuring the necessity of offloading. Secondly, the "matching degree between data and cloud capabilities" is evaluated; for example, low-resolution images have high adaptability to high-bandwidth cloud environments, increasing the priority of cloud offloading and ensuring its feasibility. Finally, the difference between the maximum allowable latency of a task and the edge latency quantifies real-time requirements; for example, urgent tasks have a high timeliness factor, increasing the priority of cloud offloading and ensuring tasks do not time out. This improves the success rate of cloud offloading and reduces the average task latency.

[0046] This invention achieves "no degradation in accuracy + full release of resources" by anchoring the direction of unloading priority, constraining the magnitude of the quality coefficient, and limiting the upper limit of remaining resources. Specifically, the adjustment direction is anchored by the linear transformation of the cloud unloading priority; for example, if the cloud priority is high, the adjustment coefficient is negative, guiding the reduction in resolution to release resources. Secondly, the adjustment magnitude is constrained by the task's tolerance to resolution. Finally, the upper limit of adjustment is limited by the sufficiency of remaining resources at the edge; if there is sufficient remaining memory at the edge, the adjustment coefficient is positive, avoiding unnecessary resolution reduction. This ensures that the task accuracy loss is controlled within the required minimum and improves the efficiency of edge resource release. Attached Figure Description

[0047] Figure 1 This is a flowchart outlining the method steps for processing visual tasks on this edge server.

[0048] Figure 2 This is a schematic diagram illustrating the evaluation of local execution capability based on local execution feasibility indicators in this invention;

[0049] Figure 3This is a schematic diagram illustrating the determination of the necessity of cloud uninstallation based on necessary indicators for cloud uninstallation in this invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings and preferred embodiments.

[0051] Reference Figures 1 to 3 As shown, the present invention provides a technical solution: a method for processing visual tasks on an edge server, the method comprising the following steps:

[0052] Step S001: Before the edge server performs the vision task, collect the hardware load data, energy status data and raw data characteristics of the vision task of the edge server, and determine the local execution feasibility indicators of the vision task on the edge server through multi-dimensional weighted evaluation.

[0053] The method for determining the feasibility indicators for local implementation is as follows:

[0054] Before running a vision task on the edge server, collect the current CPU load rate, current memory load rate, real-time temperature, real-time power, and the resolution and complexity required by the vision task.

[0055] The temperature risk coefficient is obtained by calculating the ratios of the real-time temperature to the first and second preset safe temperature and temperature threshold, respectively. The energy consumption risk coefficient is obtained by calculating the ratio of the real-time power to the preset energy consumption threshold. Finally, the temperature risk coefficient and the energy consumption risk coefficient are weighted and summed to obtain the energy risk coefficient. ;

[0056] In the formula: NF is the energy risk coefficient, which is a quantitative indicator of the edge server's hardware safety threshold and reflects the risk of hardware damage due to overload. The value of the energy risk coefficient NF ranges from 0 to 1 (0 represents completely unusable, and 1 represents absolutely safe). W is the real-time temperature, i.e., the real-time temperature of the edge server's CPU / GPU, which can be read through hardware sensors. safe For safe temperature, that is, the "upper limit of normal operating temperature" for edge servers, W max The temperature threshold is the maximum temperature that the edge server's hardware can withstand, where W ≤ W safe hour, The value is 1; when W ≥ W max hour, The value is 0, where P represents real-time energy consumption, i.e., the real-time power consumption of the edge server, which can be read through the power management chip. max This represents the energy consumption threshold, i.e., the maximum allowable power of the edge server. w1 and w2 are both weighting coefficients, and w1 + w2 = 1. For example, setting w1 = 0.7 and w2 = 0.3 is worth noting. The core logic is "temperature safety first, energy safety second";

[0057] The resolution matching coefficient is obtained by calculating the ratios of the third and fourth differences between the visual data resolution and the preset maximum edge support resolution and the preset base resolution, respectively. The complexity matching coefficient is obtained by calculating the ratio between the target feature complexity and the preset maximum edge support complexity. Finally, the resolution matching coefficient and the complexity matching coefficient are weighted and summed to obtain the data fit. ;

[0058] In the formula: BS represents the local data fit, which is a matching index between the edge server's computing power and the characteristics of the visual task data, reflecting whether the edge can efficiently process the current visual data. The value range of the data fit BS is 0-1 (0 represents a complete mismatch, 1 represents a perfect match). Res represents the visual data resolution, i.e., the resolution of the image / video frame, which can be read through image metadata. base The base resolution, which is the lowest resolution that edge servers can handle "without stress," is Res max The maximum supported resolution at the edge is the highest resolution that the edge server can stably process, where Res ≤ Res base hour, The value is 1; when Res≥Res max hour, The value is 0, C obj The feature complexity of the target, i.e., the number and complexity of targets in a visual task, can be automatically counted using image analysis tools. C max The maximum supported complexity at the edge is defined as the highest target complexity that the edge server can stably handle. w3 and w4 are weighting coefficients, and w3 + w4 = 1. For example, if w3 = 0.6 and w4 = 0.4, it's worth noting that... The core logic is "resolution matching takes precedence, target complexity matching takes secondary role";

[0059] The overall hardware load is calculated by averaging the current CPU load and memory load. The complement of the overall hardware load, the energy risk coefficient, and the data fit are then multiplied to obtain a local execution feasibility index, which is denoted as the local execution confidence level. ;

[0060] In the formula: BZX is the local execution confidence level, F1 is the current CPU load rate, that is, the real-time CPU usage percentage of the edge server (e.g., 0.7 means 70% load), F2 is the current memory load rate, that is, the real-time memory usage percentage of the edge server, NF is the energy risk coefficient, and BS is the local data adaptability.

[0061] Preferably, visual tasks (such as image recognition and video analysis) are computationally intensive tasks, and CPU / memory load directly reflects the real-time resource availability of the edge server. When calculating the overall hardware load, the average value of CPU and memory load is taken. This can avoid misjudgments caused by fluctuations in a single indicator, through... The reverse correlation of hardware load ensures that "the more idle the hardware, the higher the basic feasibility of local execution" (e.g., when the average load is 70%, this item is 0.3, representing 30% hardware support). Edge servers are often deployed in non-data center scenarios (such as outdoor monitoring and industrial equipment), relying on batteries or limited power supplies. Excessive energy consumption can lead to overheating, shortened battery life, or hardware damage (e.g., temperature exceeding a threshold triggers frequency throttling). The energy risk factor (NF) combines temperature and power consumption risks. When the temperature approaches or exceeds the threshold, the NF value decreases significantly, thus incorporating device safety considerations into the decision-making process. The data characteristics of visual tasks (such as resolution and target complexity) are strongly correlated with edge computing capabilities (low-resolution images can be processed quickly at the edge, while high-resolution / complex target images require more resources. If the data and edge capabilities do not match, processing may fail even if the hardware is idle). Therefore, the data fit factor (BS) can assess the degree of matching between task requirements and edge hardware processing capabilities, and when the requirements exceed the hardware limits, the BS will decrease. The above three dimensions are the core constraints for edge servers to process visual tasks locally. They directly determine whether the task can be executed "stable, efficient and accurate". None of them can be missing. Moreover, these three dimensions cover the three core issues of "whether it can be processed (hardware), whether it dares to process (energy consumption) and whether it is suitable to process (data)", which is a complete evaluation framework for edge local execution capabilities.

[0062] Understandably, the local execution confidence score (BZX) is a value between 0 and 1. The closer BZX is to 1, the better the edge server's local execution conditions are in terms of resources, energy security, and task adaptability. Conversely, the closer BZX is to 0, the greater the challenges to local execution. This metric is achieved by multiplying load complementation, energy risk, and data adaptability. Poor performance in any of these dimensions will lower the overall confidence score, which aligns with the "weakest link principle" in resource allocation decisions.

[0063] Furthermore, based on the local execution confidence level, the local execution capability is evaluated as follows:

[0064] If the local execution confidence is greater than the preset maximum local execution confidence, the edge server is evaluated to have high reliability in local processing, and the necessity of cloud offloading of visual tasks is quantified.

[0065] If the local execution confidence level is lower than the preset minimum local execution confidence level, it is determined that the edge server lacks local processing capabilities and will be offloaded to the cloud.

[0066] Understandably, the preset local minimum execution confidence (e.g., 0.3) and local maximum execution confidence (e.g., 0.7) are used as system configuration thresholds to transform continuous confidence values ​​into discrete preliminary decisions. This design allows the system to quickly decide to unload when it is "clearly infeasible," while introducing a more refined follow-up evaluation in the "potentially feasible" gray area.

[0067] Thus, by collecting multi-dimensional data and performing weighted calculations, a quantitative indicator of local execution feasibility—the local execution confidence score (BZX)—was obtained, providing a basis for subsequent unloading decisions.

[0068] Step S002: Based on the gap in the feasibility indicators for local execution, and combined with the resource adaptability of the cloud server and the real-time requirements of the visual task, quantify the necessity indicators for cloud offloading of the visual task.

[0069] The method for determining the necessity indicators for cloud uninstallation is as follows:

[0070] Collect bandwidth, load data, and resolution requirements of the visual task from the cloud server to determine the data compatibility with cloud capabilities: ;

[0071] In the formula: YS represents the data adaptation degree in the cloud, and Res represents the data adaptation degree in the cloud. cloud-max The maximum supported resolution in the cloud, and obtained from the cloud service documentation, is B. min B represents the minimum bandwidth required in the cloud and is obtained from the cloud service documentation. B is the network bandwidth, estimated using "file transfer speed". L is... cloud The cloud resource load rate is estimated through the status page of the cloud service. w5 and w6 are both weighting coefficients, and w5+w6=1. For example, w5=0.7 and w6=0.3.

[0072] By combining the maximum allowable latency of the vision task with the edge processing latency of the edge server, the data timeliness factor for the task's real-time requirements is calculated: ;

[0073] In the formula: SS is the data timeliness factor, D max For the maximum allowable delay, D edge For edge processing latency;

[0074] The complement of the local execution feasibility indicators is used as the edge capability gap. The edge capability gap based on the local execution confidence is multiplied by the data adaptability, and then added to the data timeliness factor to obtain the cloud offloading necessity indicator, which is recorded as the cloud offloading priority. ;

[0075] In the formula: YX is the cloud unloading priority, YS is the cloud data adaptation degree, that is, the matching degree between visual data and cloud resources (e.g., the adaptation degree is 0.8 when the network bandwidth is sufficient and 0.2 when the bandwidth is insufficient), which can be measured by network speed test tools, and SS is the data timeliness factor, that is, the time sensitivity of visual tasks (e.g., the timeliness factor of real-time video monitoring is 0.9 and that of static image recognition is 0.3), which can be preset by task type or dynamically detected (e.g., video frame rate).

[0076] Preferably, edge capability gap The data adaptability (YS) directly reflects the insufficiency of local resources. The cloud data adaptability (YS) evaluates the smoothness of task execution in the cloud from three perspectives: resolution requirements, network bandwidth, and cloud load. The higher the value, the more suitable the cloud is for handling the task. The data timeliness factor (SS) quantifies the urgency of the task. When the allowable delay time for the task is very tight, the data timeliness factor (SS) approaches 1, which will significantly increase the priority of cloud offloading, because the cloud can usually provide more stable or faster processing speeds.

[0077] Understandably, the cloud uninstallation priority YX is a comprehensive evaluation value, which is composed of... The computational component represents a combination of "necessity" and "feasibility," and the more insufficient its local capabilities (gap), the more significant the gap. The larger the value of this component (YS), the better the cloud adaptation (YS). The data timeliness factor SS is added, so that tasks with high real-time requirements may be given a higher cloud offloading priority even if the local capabilities are adequate, due to timeliness pressure. This indicator aims to balance resources, performance and timeliness.

[0078] In addition, the following judgments are made based on cloud uninstallation priority and the necessity of cloud uninstallation:

[0079] If the cloud uninstallation priority is less than 1, then the necessity of cloud uninstallation is low, and local processing is preferred, and the original resolution of the visual task is dynamically adjusted.

[0080] If the priority of cloud uninstallation is ≥1, then the necessity of cloud uninstallation is high, and cloud processing is selected as the priority.

[0081] Understandably, the threshold "1" serves as a balance point. When YX < 1, it means that, based on comprehensive evaluation, local processing or adjusted local processing is a better or more feasible choice. When YX ≥ 1, it is strongly recommended to offload the task to the cloud to avoid potential performance issues or timeouts at the edge. This judgment logic transforms continuous priority values ​​into clear binary decision paths.

[0082] Thus, through quantitative calculation, the cloud uninstallation priority YX was obtained, providing a core decision-making basis for determining whether cloud uninstallation is necessary.

[0083] Step S003: Based on the direction of the cloud offloading necessity index, and combined with the visual task's tolerance for data quality and the remaining resource space of the edge server, dynamically adjust the original resolution of the visual task to obtain the adjusted resolution.

[0084] The method for dynamically adjusting the resolution parameter is as follows:

[0085] The data quality coefficient is determined by comparing the accuracy loss requirements of the visual task with the accuracy loss data after adjusting for historical resolution. ;

[0086] In the formula: ZS is the data quality coefficient, L req For the accuracy loss requirement, i.e. the maximum allowable accuracy loss for the task, L res This refers to the loss of accuracy due to reduced resolution.

[0087] The remaining resource coefficient at the edge is determined based on the ratio of the fifth difference between the maximum resources and currently used resources of the edge server to the resources required by the task. ;

[0088] In the formula: BY is the marginal residual resource coefficient, R max R is the maximum resource for edge servers. used R is the currently used resource, estimated through the device's resource display panel (such as memory / CPU usage on some industrial devices) or "Device Status". task The resources required for the task are estimated based on the historical performance of the visual task.

[0089] Based on the linear transformation result of the cloud offloading priority, the offloading priority term is determined. This offloading priority term is then multiplied by the data quality coefficient and added to the edge remaining resource coefficient to obtain the resolution adjustment coefficient. ;

[0090] In the formula: K res ZS is the resolution adjustment factor, ZS is the data quality factor, and BY is the edge remaining resource factor.

[0091] Based on the absolute value of the resolution adjustment factor, the original resolution of the visual task is reduced proportionally to obtain the adjusted resolution. The specific formula for calculating the adjusted resolution is as follows:

[0092] ;

[0093] In the formula: K res Res is the resolution adjustment factor. daj For the adjusted resolution, Res original This is the original resolution.

[0094] Preferred, uninstallation priority item This maps YX to a range centered around 0.5. When the cloud uninstallation priority YX is low (low uninstallation necessity), this item is positive, guiding the resolution adjustment coefficient K. res If positive, combining the marginal residual resource coefficient BY may result in The smaller the resolution, the less adjustment is needed, or even no adjustment is needed at all, prioritizing quality. The data quality coefficient ZS constrains the upper limit of adjustment. The more sensitive the task is to accuracy (accuracy loss requirement L), the better. req (Small), the larger the ZS data quality coefficient, the better. The more limited the magnification effect, the more the edge residual resource coefficient BY reflects the sufficiency of local resources. The more abundant the resources, the larger the edge residual resource coefficient BY, which may increase the resolution adjustment coefficient K. res A positive value is used to suppress unnecessary resolution reduction.

[0095] Understandably, the resolution adjustment factor K res It can be positive or negative, and its absolute value The resolution reduction ratio is determined, and the goal of the entire adjustment strategy is to release edge resources by appropriately reducing the resolution, under the premise that the cloud unloading priority is not high (YX<1), thereby improving the feasibility of local execution. However, it must be subject to the dual constraints of task quality baseline (ZS) and local resource margin (BY) to prevent excessive reduction of resolution from affecting task performance.

[0096] Thus, by using the quantitative formula that considers unloading tendency, quality tolerance, and resource status, the resolution adjustment coefficient K for the current task and system state has been dynamically calculated. res and utilize The adjusted resolution Res was obtained daj This prepares us for a reassessment of local execution capabilities.

[0097] Step S004: Based on the adjusted resolution, re-evaluate the local execution capabilities of the edge server and output the final execution plan for the vision task.

[0098] The specific method for outputting the final execution plan is as follows:

[0099] The adjusted resolution Res obtained in step S003 daj Substitute the original original resolution Res into the calculation formula of data fit BS in step S001, and recalculate the data fit BS.

[0100] Next, the recalculated BS and the current hardware load are used. And the energy risk coefficient NF, according to the formula Recalculate the local execution confidence level BZX;

[0101] Finally, the final execution plan is determined by comparing the reassessed local execution confidence level BZX with the preset local execution confidence level threshold.

[0102] If the local execution confidence BZX is greater than the local maximum execution confidence (0.7), then the final execution plan will be "local execution", that is, the visual task processing will be completed on the edge server using the adjusted resolution.

[0103] If the local execution confidence BZX is less than the local minimum execution confidence (0.3), the final execution plan will be "cloud unloading", which means uploading the visual task (which can use the original resolution or the adjusted resolution) to the cloud server for processing.

[0104] In addition, as an optional collaborative solution, if the system supports it and the task is divisible, when the local execution confidence BZX is close to the threshold, an "edge-cloud collaboration" solution can be output. For example, some preprocessing or low-complexity subtasks can be executed at the edge, while the core complex analysis can be offloaded to the cloud. Specifically, if 0.3 ≤ local execution confidence BZX ≤ 0.7 and 0.5 ≤ cloud offloading priority YX ≤ 1, the edge server handles "lightweight tasks" (such as preliminary screening of object detection), and the cloud handles "heavyweight tasks" (such as accurate classification of object recognition) to achieve resource complementarity.

[0105] Preferably, the re-evaluation step is a key quality control and decision-making loop. It verifies whether the resolution adjustment truly makes the task suitable for local execution. The maximum local execution confidence (0.7) and the minimum local execution confidence (0.3) are set by the system based on the expected success rate of the task and are used to make the final binary decision.

[0106] Understandably, the final execution plan is a comprehensive decision made after initial assessment, determination of cloud necessity, and self-optimization (resolution adjustment). This process forms a closed loop of "assessment-decision-adjustment-reassessment," aiming to maximize the utilization of edge resources while ensuring that the task's requirements in terms of quality and timeliness are met. Furthermore, this plan clearly defines three exit points: "local execution," "cloud offloading," and the optional "edge-cloud collaboration," covering the main edge computing scenarios.

[0107] Thus, through the complete processing flow, the system can dynamically and quantitatively output the optimal visual task execution plan based on real-time hardware status, network conditions, task characteristics, and cloud capabilities, achieving efficient and intelligent allocation of edge server resources.

[0108] Based on the same inventive concept as the above method, this embodiment of the invention also provides an edge server vision task processing system, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the above-described edge server vision task processing method.

[0109] It should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should also be within the scope of protection of this invention.

Claims

1. A method for processing visual tasks on an edge server, characterized in that, The method includes the following steps: Before the edge server performs the vision task, hardware load data, energy status data and raw data characteristics of the vision task are collected. Through multi-dimensional weighted evaluation, the feasibility indicators of the local execution of the vision task on the edge server are determined. Based on the gaps in the local execution feasibility indicators, and in combination with the resource adaptability of the cloud server and the real-time requirements of the visual task, the necessity indicators for cloud offloading of the visual task are quantified. Based on the direction of the cloud offloading necessity index, and combined with the visual task's tolerance for data quality and the remaining resource space of the edge server, the original resolution of the visual task is dynamically adjusted to obtain the adjusted resolution. Based on the adjusted resolution, the local execution capabilities of the edge server are reassessed, and the final execution plan for the vision task is output. The final execution scheme includes local execution, cloud unloading, and edge-cloud collaboration.

2. The edge server visual task processing method according to claim 1, characterized in that, The method for determining the feasibility indicators for local implementation is as follows: Before running a vision task on the edge server, collect the current CPU load rate, current memory load rate, real-time temperature, real-time power, and the resolution and complexity required by the vision task. The overall hardware load is obtained based on the average of the current CPU load rate and the current memory load rate. The temperature risk coefficient is obtained by proportionally relating the real-time temperature to the first difference and the second difference between the preset safe temperature and the temperature threshold. The energy consumption risk coefficient is obtained by proportionally relating the real-time power to the preset energy consumption threshold. The temperature risk coefficient and the energy consumption risk coefficient are then weighted and summed to obtain the energy risk coefficient. The resolution matching coefficient is obtained by the proportional relationship between the visual data resolution and the third and fourth differences between the preset maximum edge support resolution and the preset base resolution, respectively. The complexity matching coefficient is obtained by the proportional relationship between the target feature complexity and the preset maximum edge support complexity. The resolution matching coefficient and the complexity matching coefficient are then weighted and summed to obtain the data fit. Multiply the complement of the overall hardware load, the energy risk coefficient, and the data fit by the product to obtain the local execution feasibility index, which is recorded as the local execution confidence level.

3. The edge server visual task processing method according to claim 2, characterized in that, Based on the aforementioned local execution confidence level, the local execution capability is evaluated as follows: If the local execution confidence is greater than the preset maximum local execution confidence, the edge server is evaluated to have high reliability in local processing, and the necessity index for cloud offloading of visual tasks is quantified. If the local execution confidence is less than the preset minimum local execution confidence, the edge server is deemed to lack local processing capabilities and will be offloaded to the cloud.

4. The edge server visual task processing method according to claim 3, characterized in that, The method for determining the necessity index for cloud-based uninstallation is as follows: The gap in edge capabilities is defined by supplementing the local feasibility indicators. Collect bandwidth and load data of the cloud server and resolution requirements of the visual task to determine the data compatibility with the cloud capabilities; By combining the maximum allowable latency of the vision task with the edge processing latency of the edge server, the data timeliness factor for the real-time requirements of the task is calculated. The edge capability gap based on the local execution confidence level is multiplied by the data adaptability, and then added to the data timeliness factor to obtain the cloud uninstallation necessity index, which is recorded as the cloud uninstallation priority.

5. The edge server visual task processing method according to claim 4, characterized in that, Based on the cloud uninstallation priority and the determination of the necessity of cloud uninstallation, the following is made: If the cloud uninstallation priority is less than 1, then the necessity of cloud uninstallation is low, and local processing is selected first, and the original resolution of the visual task is dynamically adjusted. If the cloud uninstallation priority is ≥1, then the necessity of cloud uninstallation is high, and cloud processing is selected as the priority.

6. The edge server visual task processing method according to claim 5, characterized in that, The method for dynamically adjusting the resolution parameter is as follows: The uninstallation priority item is determined based on the linear transformation result of the cloud uninstallation priority. The data quality coefficient is determined by comparing the accuracy loss requirements of the visual task with the accuracy loss data after adjusting the historical resolution. The edge remaining resource coefficient is determined based on the ratio of the fifth difference between the maximum resources of the edge server and the currently used resources to the resources required by the task. Multiply the unloading priority item by the data quality coefficient and add it to the edge remaining resource coefficient to obtain the resolution adjustment coefficient; Based on the absolute value of the resolution adjustment coefficient, the original resolution of the visual task is reduced proportionally to obtain the adjusted resolution.

7. The edge server visual task processing method according to claim 6, characterized in that, The specific formula for calculating the adjusted resolution is as follows: ； In the formula: K res Res is the resolution adjustment factor. daj For the adjusted resolution, Res original This is the original resolution.

8. An edge server vision task processing system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

9. The edge server visual task processing system according to claim 8, characterized in that, The system includes: Perception layer module: includes the resource display panel of the edge server, temperature sensor, power meter, bandwidth tester and visual data acquisition device, used to collect hardware load, energy status, network status and visual raw data; Decision-making layer module: includes local evaluation unit, cloud decision-making unit and resolution adjustment unit, used to execute multi-dimensional evaluation, quantitative decision and dynamic adjustment logic; Execution layer module: includes edge server processing unit and cloud interaction unit, used to complete the local execution or cloud unloading of visual tasks according to the execution plan output by the decision layer; Storage module: Used to store the historical execution logs of the edge server, the requirement parameters of visual tasks, and the system configuration thresholds.

Images