Image multi-algorithm parallel scheduling method

By employing a parallel scheduling method for multiple image detection algorithms, and based on the execution characteristics of different image detection algorithms, batch aggregation and independent asynchronous scheduling channels are used to execute detection in parallel. Combined with confidence threshold filtering, this method solves the problems of low efficiency and poor scalability in multi-algorithm scenarios, and achieves efficient and reliable intelligent image detection.

CN123114702APending Publication Date: 2026-11-17GUANGDONG POWER GRID CO LTD DONGGUAN POWER SUPPLY BUREAU
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611524506.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-29
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, image intelligent detection systems suffer from low scheduling efficiency and poor system scalability in multi-algorithm scenarios. In particular, in scenarios such as power line inspection, industrial safety monitoring, park security, and daily office evidence collection, a single detection algorithm is difficult to meet the concurrent analysis needs of multiple targets and multiple business lines.

Method used

A parallel scheduling method for multiple image detection algorithms is adopted. By acquiring the image to be detected and the preset algorithm scheduling strategy, multiple image detection algorithms are scheduled to the first scheduling channel that supports batch aggregation scheduling and/or the second scheduling channel that supports independent asynchronous scheduling. Each image detection algorithm is executed in parallel, and the detection results are filtered according to the confidence threshold to obtain the effective detection results.

Benefits of technology

It improves the processing efficiency and resource utilization of multi-algorithm detection, reduces the interference of invalid detection results on subsequent processing, and enhances the scalability of the system and the effectiveness of the detection results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114702A_ABST
    Figure CN123114702A_ABST
Patent Text Reader

Abstract

This application relates to the fields of image recognition and intelligent monitoring, and particularly to a parallel scheduling method for multiple image algorithms. Addressing the problems of high algorithm interface coupling, low scheduling efficiency, insufficient result distribution, and weak correlation between alarm and review links in existing multi-algorithm image detection methods, this application proposes a parallel processing scheme based on a scheduling strategy: The scheme acquires the image to be detected and a preset algorithm scheduling strategy; multiple detection algorithms are assigned to different scheduling channels supporting batch aggregation processing and / or independent asynchronous processing according to their execution characteristics; after parallel execution, the detection results are filtered based on the confidence threshold corresponding to each algorithm to obtain valid detection results. This scheme can reduce the cost of algorithm access and replacement, improve concurrent detection efficiency, reduce the occupation of invalid results in subsequent processes, and enhance the accuracy, real-time performance, and traceability of detection results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of image recognition and intelligent monitoring, and in particular to a method for parallel scheduling of multiple image algorithms. Background Technology

[0002] In scenarios such as power line inspection, industrial safety monitoring, park security, and routine office evidence collection, computer vision-based image intelligent detection and alarm systems have become a key technological means to replace manual inspections and improve risk response efficiency. With the surge in the number of front-end monitoring devices and the increasing refinement of business scenarios, single detection algorithms are no longer sufficient to meet the concurrent analysis needs of multiple targets and business lines. Systems often need to integrate multiple AI models (such as helmet wearing detection, specific behavior violation identification, and equipment appearance fault analysis) to comprehensively cover various potential risks. Therefore, there is an urgent need for a comprehensive image intelligent processing solution with high concurrency scheduling capabilities, flexible auditing strategies, breakpoint resume mechanisms, and a strict process isolation architecture.

[0003] In existing technologies, image intelligent detection systems typically integrate one or more object detection algorithms. For multi-algorithm scenarios, common scheduling methods fall into two categories: one is the serial invocation method, where multiple algorithms are executed sequentially, and the total execution time equals the sum of the execution times of each algorithm; the other is the simple parallel method, where each algorithm is invoked independently and the results are aggregated. Regarding algorithm interface management, some systems provide unified algorithm invocation services externally through an API gateway model, shielding details such as algorithm addresses and authentication methods from the business layer.

[0004] However, existing technologies suffer from low scheduling efficiency and poor system scalability due to multiple algorithms. Summary of the Invention

[0005] This application provides a method for parallel scheduling of multiple algorithms for images, which solves the problems of low scheduling efficiency and poor system scalability in the prior art.

[0006] In a first aspect, embodiments of this application provide a method for parallel scheduling of multiple image algorithms, including:

[0007] The image to be detected and a preset algorithm scheduling strategy are obtained, wherein the algorithm scheduling strategy is configured with the execution strategy corresponding to each image detection algorithm;

[0008] According to the preset algorithm scheduling strategy, multiple image detection algorithms corresponding to the image to be detected are respectively scheduled to the corresponding target scheduling channels. The target scheduling channels include at least a first scheduling channel that supports batch aggregation scheduling and / or a second scheduling channel that supports independent asynchronous scheduling.

[0009] The multiple image detection algorithms are executed in parallel based on the target scheduling channel to output the detection results corresponding to each image detection algorithm;

[0010] The multiple detection results are filtered according to the confidence threshold corresponding to each image detection algorithm to obtain valid detection results.

[0011] In one possible implementation, acquiring the image to be detected includes at least one of the following methods:

[0012] In response to a screenshot trigger operation, a transparent overlay window is created on the target display and the target area is selected, and the image to be detected is cropped from the screen video stream; in response to a local image selection operation, the image to be detected is retrieved from the local storage area;

[0013] Based on the conditions for obtaining equipment monitoring data, historical monitoring images are obtained from the IoT device platform as images to be detected.

[0014] In one possible implementation, acquiring historical surveillance images from the IoT device platform as images to be detected includes:

[0015] The acquisition task is divided into a three-level task hierarchy, including batch tasks, device tasks, and download tasks;

[0016] Separately establish a metadata acquisition queue and an image download queue, and schedule the device tasks and the download tasks in parallel according to a preset concurrency level;

[0017] The execution status of the three-level tasks is persistently recorded, and the unfinished tasks are restored based on the persistent status after a task interruption is detected.

[0018] The recovered download task is reset to a pending download state and re-added to the image download queue.

[0019] In one possible implementation, scheduling multiple image detection algorithms corresponding to the image to be detected to their respective target scheduling channels according to a preset algorithm scheduling strategy includes:

[0020] Obtain the algorithm identifier corresponding to the currently enabled image detection algorithm, and query the preset algorithm scheduling strategy based on the algorithm identifier;

[0021] When the algorithm identifier has a corresponding dedicated execution strategy in the preset algorithm scheduling strategy, the image detection algorithm is assigned to the second scheduling channel, and an independent asynchronous execution task is established for the image detection algorithm;

[0022] When the algorithm identifier does not have a corresponding dedicated execution strategy, the image detection algorithm is assigned to the first scheduling channel, and the corresponding algorithm identifier is added to the general algorithm call set.

[0023] In one possible implementation, the parallel execution of the multiple image detection algorithms based on the target scheduling channel to output detection results corresponding to each image detection algorithm includes:

[0024] For the first scheduling channel, obtain the algorithm identifier of each general image detection algorithm, and combine multiple algorithm identifiers to form a general algorithm call parameter;

[0025] The general algorithm call parameters and the image to be detected are sent to the general algorithm service to obtain multiple detection results through a single batch call;

[0026] For the second scheduling channel, the independent asynchronous execution task is executed through the dedicated execution strategy to obtain multiple detection results.

[0027] In one possible implementation, filtering the multiple detection results based on the confidence threshold corresponding to each image detection algorithm to obtain valid detection results includes:

[0028] The multiple detection results are correlated according to the corresponding image detection algorithms, and the confidence threshold corresponding to each image detection algorithm is obtained;

[0029] For each detection result, the confidence level corresponding to the detection annotation data in the detection result is compared with the confidence level threshold corresponding to the image detection algorithm;

[0030] When the confidence level corresponding to the detection annotation data is greater than or equal to the confidence threshold, the detection annotation data is retained as valid detection annotation data;

[0031] If no valid detection label data is found in the detection results, the detection result is discarded.

[0032] In one possible implementation, the method further includes:

[0033] Based on the detection annotation data in the valid detection results, the corresponding detection area is determined in the image to be detected, and visual annotation information corresponding to the detection area is generated;

[0034] The image to be detected is annotated according to the visual annotation information to obtain an annotated image;

[0035] The labeled image is compressed, and the compressed labeled image is associated with the valid detection results to generate a corresponding alarm record;

[0036] The alarm records are persistently stored, and a preset review policy is used to determine whether the alarm records should enter the manual review or message push process.

[0037] In one possible implementation, the alarm record includes at least the recognition time, image detection algorithm identifier, alarm category, alarm confidence level, review status, push status, and storage path of the labeled image;

[0038] The review status is used to indicate whether the alarm record has been reviewed and the review result, and the push status is used to indicate whether the message corresponding to the alarm record has been sent.

[0039] In one possible implementation, determining whether the alarm record should enter the manual review or message push process according to a preset review strategy includes:

[0040] Obtain the manual review switch status corresponding to the image detection algorithm that generated the alarm record;

[0041] When the manual review switch is in the off state, the review status of the alarm record is set to reviewed, and a message push is triggered based on the alarm record;

[0042] When the manual review switch is on, the review status of the alarm record is set to pending review, and the alarm record is added to the manual review queue.

[0043] In response to the review operation of the alarm record in the manual review queue, the review status of the alarm record is updated according to the review result, and it is determined whether to trigger a message push based on the updated review status.

[0044] In one possible implementation, triggering the message push includes:

[0045] Retrieve the pre-configured alarm message receiver object;

[0046] An alarm message is generated based on the identification time and alarm category in the alarm record, and a compressed labeled image associated with the alarm record is obtained;

[0047] The alarm message and the compressed labeled image are sent to the alarm message receiving object;

[0048] The push status of the alarm record is updated based on the message sending result, wherein the push status includes at least sent, failed to send, and not sent.

[0049] Secondly, embodiments of this application provide an image multi-algorithm parallel scheduling device, comprising:

[0050] The acquisition module is used to acquire the image to be detected and a preset algorithm scheduling strategy, wherein the algorithm scheduling strategy is configured with the execution strategy corresponding to each image detection algorithm;

[0051] The scheduling module is used to schedule multiple image detection algorithms corresponding to the image to be detected to corresponding target scheduling channels according to a preset algorithm scheduling strategy. The target scheduling channels include at least a first scheduling channel that supports batch aggregation scheduling and / or a second scheduling channel that supports independent asynchronous scheduling.

[0052] The processing module is used to execute the multiple image detection algorithms in parallel based on the target scheduling channel, so as to output the detection results corresponding to each image detection algorithm;

[0053] The filtering module is used to filter the multiple detection results according to the confidence threshold corresponding to each image detection algorithm to obtain valid detection results.

[0054] In one possible implementation, the acquisition module is specifically used for:

[0055] In response to a screenshot trigger operation, a transparent overlay window is created on the target display and the target area is selected, and the image to be detected is cropped from the screen video stream; in response to a local image selection operation, the image to be detected is retrieved from the local storage area;

[0056] Based on the conditions for obtaining equipment monitoring data, historical monitoring images are obtained from the IoT device platform as images to be detected.

[0057] In one possible implementation, the acquisition module is specifically used for:

[0058] The acquisition task is divided into a three-level task hierarchy, including batch tasks, device tasks, and download tasks;

[0059] Separately establish a metadata acquisition queue and an image download queue, and schedule the device tasks and the download tasks in parallel according to a preset concurrency level;

[0060] The execution status of the three-level tasks is persistently recorded, and the unfinished tasks are restored based on the persistent status after a task interruption is detected.

[0061] The recovered download task is reset to a pending download state and re-added to the image download queue.

[0062] In one possible implementation, the scheduling module is specifically used for:

[0063] Obtain the algorithm identifier corresponding to the currently enabled image detection algorithm, and query the preset algorithm scheduling strategy based on the algorithm identifier;

[0064] When the algorithm identifier has a corresponding dedicated execution strategy in the preset algorithm scheduling strategy, the image detection algorithm is assigned to the second scheduling channel, and an independent asynchronous execution task is established for the image detection algorithm;

[0065] When the algorithm identifier does not have a corresponding dedicated execution strategy, the image detection algorithm is assigned to the first scheduling channel, and the corresponding algorithm identifier is added to the general algorithm call set.

[0066] In one possible implementation, the processing module is specifically used for:

[0067] For the first scheduling channel, obtain the algorithm identifier of each general image detection algorithm, and combine multiple algorithm identifiers to form a general algorithm call parameter;

[0068] The general algorithm call parameters and the image to be detected are sent to the general algorithm service to obtain multiple detection results through a single batch call;

[0069] For the second scheduling channel, the independent asynchronous execution task is executed through the dedicated execution strategy to obtain multiple detection results.

[0070] In one possible implementation, the filtering module is specifically used for:

[0071] The multiple detection results are correlated according to the corresponding image detection algorithms, and the confidence threshold corresponding to each image detection algorithm is obtained;

[0072] For each detection result, the confidence level corresponding to the detection annotation data in the detection result is compared with the confidence level threshold corresponding to the image detection algorithm;

[0073] When the confidence level corresponding to the detection annotation data is greater than or equal to the confidence threshold, the detection annotation data is retained as valid detection annotation data;

[0074] If no valid detection label data is found in the detection results, the detection result is discarded.

[0075] In one possible implementation, the device further includes an alarm module for:

[0076] Based on the detection annotation data in the valid detection results, the corresponding detection area is determined in the image to be detected, and visual annotation information corresponding to the detection area is generated;

[0077] The image to be detected is annotated according to the visual annotation information to obtain an annotated image;

[0078] The labeled image is compressed, and the compressed labeled image is associated with the valid detection results to generate a corresponding alarm record;

[0079] The alarm records are persistently stored, and a preset review policy is used to determine whether the alarm records should enter the manual review or message push process.

[0080] In one possible implementation, the alarm record includes at least the recognition time, image detection algorithm identifier, alarm category, alarm confidence level, review status, push status, and storage path of the labeled image;

[0081] The review status is used to indicate whether the alarm record has been reviewed and the review result, and the push status is used to indicate whether the message corresponding to the alarm record has been sent.

[0082] In one possible implementation, the alarm module is specifically used for:

[0083] Obtain the manual review switch status corresponding to the image detection algorithm that generated the alarm record;

[0084] When the manual review switch is in the off state, the review status of the alarm record is set to reviewed, and a message push is triggered based on the alarm record;

[0085] When the manual review switch is on, the review status of the alarm record is set to pending review, and the alarm record is added to the manual review queue.

[0086] In response to the review operation of the alarm record in the manual review queue, the review status of the alarm record is updated according to the review result, and it is determined whether to trigger a message push based on the updated review status.

[0087] In one possible implementation, the alarm module is specifically used for:

[0088] Retrieve the pre-configured alarm message receiver object;

[0089] An alarm message is generated based on the identification time and alarm category in the alarm record, and a compressed labeled image associated with the alarm record is obtained;

[0090] The alarm message and the compressed labeled image are sent to the alarm message receiving object;

[0091] The push status of the alarm record is updated based on the message sending result, wherein the push status includes at least sent, failed to send, and not sent.

[0092] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0093] The memory stores computer-executed instructions;

[0094] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0095] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0096] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0097] The image multi-algorithm parallel scheduling method provided in this application first acquires the image to be detected and a preset algorithm scheduling strategy to determine the corresponding scheduling method according to different image detection algorithms. After acquiring the image to be detected, tasks are allocated to multiple image detection algorithms corresponding to the image to be detected according to the algorithm scheduling strategy. Each image detection algorithm is scheduled to a target scheduling channel that matches its execution strategy. The target scheduling channel includes at least a first scheduling channel that supports batch aggregation scheduling and / or a second scheduling channel that supports independent asynchronous scheduling. For multiple image detection algorithms suitable for batch processing, the corresponding algorithm tasks can be aggregated through the first scheduling channel to reduce redundant task scheduling and resource initialization overhead. For image detection algorithms that need to be executed independently or have high requirements for execution latency, they are asynchronously scheduled through the second scheduling channel, so that different algorithm tasks can run in mutually independent execution contexts. Furthermore, multiple image detection algorithms are executed in parallel based on the target scheduling channels corresponding to each image detection algorithm, so that multiple algorithms can target the same image to be detected. Simultaneously, detection processing is performed, and corresponding detection results are output separately. This leverages the execution characteristics of different scheduling channels to improve the parallel processing capability of multiple algorithms and reduce the cumulative processing latency caused by the sequential execution of multiple algorithms. After obtaining the detection results corresponding to each image detection algorithm, the detection results are further filtered based on the confidence threshold corresponding to each image detection algorithm. The confidence of the detection results is compared with the corresponding confidence threshold, and the detection results that meet the corresponding confidence threshold are retained as valid detection results. This suppresses the output of low-confidence detection results while completing the parallel detection of multiple algorithms, thereby improving the effectiveness and reliability of the final detection results. Through the above method, different image detection algorithms are allocated to different scheduling channels according to their execution characteristics based on a preset algorithm scheduling strategy. Batch aggregation scheduling and independent asynchronous scheduling are used to achieve parallel execution of multiple algorithms. Then, the results are filtered by combining the confidence threshold matched to each algorithm. This can improve the efficiency of algorithm resource utilization and overall detection processing efficiency while ensuring the detection coverage capability of multiple algorithms and reducing the interference of invalid detection results on subsequent processing. Attached Figure Description

[0098] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0099] Figure 1 A flowchart illustrating the image multi-algorithm parallel scheduling method provided in this application embodiment. Figure 1 ;

[0100] Figure 2 A flowchart illustrating the image multi-algorithm parallel scheduling method provided in this application embodiment. Figure 2 ;

[0101] Figure 3 This is a schematic diagram of the structure of the image multi-algorithm parallel scheduling device provided in the embodiments of this application;

[0102] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0103] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0104] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0105] In scenarios such as power line inspection, industrial safety monitoring, park security, and routine office evidence collection, computer vision-based image intelligent detection and alarm systems have become a key technological means to replace manual inspections and improve risk response efficiency. With the surge in the number of front-end monitoring devices and the increasing refinement of business scenarios, single detection algorithms are no longer sufficient to meet the concurrent analysis needs of multiple targets and business lines. Systems often need to integrate multiple AI models (such as helmet wearing detection, specific behavior violation identification, and equipment appearance fault analysis) to comprehensively cover various potential risks. Therefore, there is an urgent need for a comprehensive image intelligent processing solution with high concurrency scheduling capabilities, flexible auditing strategies, breakpoint resume mechanisms, and a strict process isolation architecture.

[0106] In existing technologies, image intelligent detection systems typically integrate one or more object detection algorithms. For multi-algorithm scenarios, common scheduling methods fall into two categories: one is the serial invocation method, where multiple algorithms are executed sequentially, and the total execution time equals the sum of the execution times of each algorithm; the other is the simple parallel method, where each algorithm is invoked independently and the results are aggregated. Regarding algorithm interface management, some systems provide unified algorithm invocation services externally through an API gateway model, shielding details such as algorithm addresses and authentication methods from the business layer.

[0107] In the aforementioned existing technologies, multi-algorithm scheduling typically employs serial sequential calls or a simple parallel architecture without differentiation, resulting in a linear accumulation of the total time consumed by multi-model detection. Furthermore, the deep coupling between the interface protocols of each algorithm and the business code leads to problems such as low efficiency in multi-algorithm scheduling and poor system scalability.

[0108] Based on this, the inventors propose a parallel scheduling method for multiple image detection algorithms, arguing that different image detection algorithms differ in their computational resource requirements, processing methods, and execution latency, and that a matching scheduling method should be adopted according to the algorithm's own execution characteristics. In specific implementation, a corresponding execution strategy is pre-configured for each image detection algorithm, and different types of scheduling channels are established accordingly. Multiple image detection algorithms corresponding to the image to be detected are scheduled to a first scheduling channel supporting batch aggregation scheduling and / or a second scheduling channel supporting independent asynchronous scheduling, according to their respective execution strategies. This allows algorithms with similar execution characteristics to be batch aggregated, while algorithms requiring independent execution can run independently asynchronously. Subsequently, the corresponding multiple image detection algorithms are executed in parallel using each target scheduling channel, and the detection results output by each algorithm are uniformly filtered according to the corresponding confidence threshold. This decouples the algorithm scheduling logic from the specific business processing logic and achieves differentiated scheduling based on the execution characteristics of different algorithms. By adopting the above approach, on the one hand, the linear accumulation of detection time caused by serial calls of multiple image detection algorithms is avoided, and on the other hand, resource competition caused by indiscriminate parallelism of different algorithms is avoided. This allows computing resources to be rationally allocated according to the algorithm execution characteristics. At the same time, the need to modify business code when adding or replacing image detection algorithms is reduced through pre-configured algorithm scheduling strategies and general scheduling channels, thereby improving the processing efficiency, resource utilization, and scalability of multi-algorithm parallel detection.

[0109] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0110] It should be noted that the image multi-algorithm parallel scheduling method provided in this application can be applied to image intelligent recognition systems built on the Electron desktop application framework, as well as to other terminal or server systems with image acquisition, artificial intelligence algorithm invocation, detection result processing, and alarm management functions. Electron is a cross-platform desktop application development framework that supports building desktop applications using Web technologies; the main process is the only process in the Electron application, responsible for managing the application lifecycle, creating windows, and calling native operating system APIs, and has a complete Node.js environment; the rendering process is responsible for displaying the UI interface, with each window corresponding to an independent rendering process, which cannot directly access Node.js APIs by default; IPC is the inter-process communication mechanism between the main process and the rendering process, implemented through the ipcMain and ipcRenderer modules for asynchronous message passing; the preloaded script executes before the rendering process loads the webpage, and securely exposes restricted API interfaces to the rendering process through contextBridge, which is a key mechanism for achieving secure isolation.

[0111] The overall system architecture of this application includes a rendering process layer, a preloading layer, and a main process layer. The rendering process layer comprises a screenshot acquisition module, a manual image selection module, a batch download management module, and core modules for algorithm scheduling and decision-making, result aggregation and processing, alarm persistence, and message push. The preloading layer exposes a limited set of whitelisted IPC interfaces through a secure bridging mechanism. The main process layer manages system-level operations such as configuration, database, alarms, and downloads through a plug-in architecture. The underlying layer consists of a SQLite database and JSON configuration files, forming a data persistence layer. Data exchange between layers occurs via IPC mechanisms; the rendering process does not directly access operating system resources.

[0112] Figure 1 A flowchart illustrating the image multi-algorithm parallel scheduling method provided in this application embodiment. Figure 1 ;like Figure 1 As shown, the method includes:

[0113] S101. Obtain the image to be detected and the preset algorithm scheduling strategy.

[0114] The algorithm scheduling strategy is configured with the execution strategy corresponding to each image detection algorithm.

[0115] It should be understood that the images to be detected can be obtained from various data sources, and images from different sources, after acquisition, all enter the subsequent algorithm scheduling process through a unified data interface, thus avoiding the need to set up independent detection processes for different image sources. Simultaneously with acquiring the images to be detected, a preset algorithm scheduling strategy also needs to be acquired. This algorithm scheduling strategy can be stored in a configuration file and should include at least the following parameters: the current image detection algorithm's activation status, algorithm identifier, algorithm execution type, dedicated execution strategy, minimum confidence threshold, and manual review switch. To improve the flexibility of configuration adjustments, the algorithm configuration can be managed as a global configuration state. When a user modifies the algorithm's activation status, minimum confidence threshold, or manual review switch, the modified configuration is first updated to the current running state, allowing subsequent detection tasks to execute directly according to the latest configuration. Simultaneously, a debouncing mechanism is used to persist configuration changes. For example, during continuous modifications, a 500-millisecond delay is applied, and only the last configuration change is persisted to reduce frequent disk I / O. A forced write is then performed when the application exits, ensuring that configuration changes are not lost upon application exit.

[0116] For example, when a user sets the "Helmet Detection Algorithm" to enabled, the minimum confidence threshold to 0.80, and the manual review switch to off, and simultaneously sets the "Phishing Behavior Detection Algorithm" to enabled, the minimum confidence threshold to 0.70, and the manual review switch to on, the system can directly read the above runtime configuration in subsequent detection tasks without needing to restart the application.

[0117] It should be understood that by configuring runtime real-time updates and debouncing persistence mechanisms, parameters such as algorithm activation status, confidence threshold, and auditing strategy can take effect without restarting, improving the flexibility of algorithm scheduling strategy adjustments.

[0118] It should be noted that the specific process of acquiring the image to be detected in this embodiment will be described in detail later in this application, and will not be repeated in this embodiment.

[0119] S102. According to the preset algorithm scheduling strategy, the multiple image detection algorithms corresponding to the image to be detected are scheduled to the corresponding target scheduling channels respectively.

[0120] The target scheduling channel includes at least a first scheduling channel that supports batch aggregation scheduling and / or a second scheduling channel that supports independent asynchronous scheduling.

[0121] In one possible implementation, the algorithm identifier corresponding to the currently enabled image detection algorithm is first obtained, and the preset algorithm scheduling strategy is queried based on the algorithm identifier. Then, when the algorithm identifier has a corresponding dedicated execution strategy in the preset algorithm scheduling strategy, the image detection algorithm is assigned to the second scheduling channel, and an independent asynchronous execution task is established for the image detection algorithm. Then, when the algorithm identifier does not have a corresponding dedicated execution strategy, the image detection algorithm is assigned to the first scheduling channel, and the corresponding algorithm identifier is added to the general algorithm call set.

[0122] It should be noted that this embodiment adopts a hybrid parallel scheduling mechanism of "general aggregation + dedicated independence". Its basic principle is as follows: for multiple general image detection algorithms that can be called through a unified interface, instead of establishing separate independent requests, the algorithm identifiers of the multiple general algorithms are aggregated and then called uniformly through a batch request; for dedicated image detection algorithms with independent workflows, independent interfaces, or special calling procedures, independent asynchronous execution tasks are established respectively, allowing the dedicated algorithms to run independently according to their corresponding execution strategies. This forms a first scheduling channel and a second scheduling channel, enabling the two channels to execute simultaneously.

[0123] Specifically, the system first obtains the algorithm identifier corresponding to the currently enabled image detection algorithm, and then uses the algorithm identifier as the query key to query the preset dedicated strategy mapping table. If a dedicated execution strategy exists for a certain algorithm identifier, the image detection algorithm is identified as a dedicated algorithm and added to the second scheduling channel, while an independent asynchronous execution task is established for the algorithm; if a dedicated execution strategy does not exist for a certain algorithm identifier, it is identified as a general algorithm, and the algorithm identifier is added to the general algorithm call set corresponding to the first scheduling channel.

[0124] For example, suppose that the bird nest detection algorithm, the safety helmet detection algorithm, and the fishing behavior detection algorithm are currently enabled. The bird nest detection algorithm and the safety helmet detection algorithm can be called through a unified algorithm gateway, while the fishing behavior detection algorithm needs to be executed through a dedicated AI workflow. In this case, the system can add the bird nest detection algorithm and the safety helmet detection algorithm to the first scheduling channel and add the fishing behavior detection algorithm to the second scheduling channel.

[0125] In the specific implementation process, the algorithm component code can be used as the algorithm identifier, and a dedicated strategy mapping table can be used to determine whether a dedicated execution strategy has been registered. For example, after receiving an image to be detected, the strategy dispatcher StrategyRouter traverses the list of currently enabled algorithms. For each algorithm, the pre-defined dedicated strategy mapping table SpecificStrategyMap is checked, using the algorithm's component code as the key to find whether a dedicated execution strategy has been registered. If a dedicated strategy exists, such as the phishing behavior detection algorithm X000001 being executed through the Dify AI workflow platform, the algorithm is placed in a dedicated task queue, and each dedicated algorithm creates an independent asynchronous execution task; otherwise, the algorithm is assigned to the general algorithm queue, and its component code is collected into the general algorithm code array. General algorithms may include multiple target detection algorithms that are uniformly called through the AI ​​Gateway, while dedicated algorithms may include algorithms that require independent workflow execution.

[0126] Furthermore, to decouple algorithm scheduling from specific algorithm interfaces, an algorithm gateway can be used as a unified entry point for general algorithms. The algorithm gateway aggregates multiple image detection algorithms and provides a unified calling interface to the business side, thereby avoiding the business side directly handling the interface addresses, parameter formats, and authentication methods of different algorithms. When a new image detection algorithm is added, it can be added primarily by configuring the algorithm identifier, algorithm type, and corresponding strategy, without changing the core scheduling logic.

[0127] Understandably, by allocating multiple image detection algorithms to batch aggregation channels and independent asynchronous channels based on their execution characteristics, the efficiency of batch calls for general algorithms and the independent execution requirements of specialized algorithms can be balanced simultaneously, avoiding the problem of different algorithms waiting for each other due to a single serial call method. At the same time, since the algorithm scheduling logic is decoupled from the specific algorithm interface, when adding or replacing image detection algorithms, the algorithm can be integrated by adding corresponding algorithm configurations and dedicated strategies, thereby improving the scalability of the multi-algorithm detection system.

[0128] S103. Execute multiple image detection algorithms in parallel based on the target scheduling channel to output the detection results corresponding to each image detection algorithm.

[0129] In one possible implementation, for the first scheduling channel, the algorithm identifier of each general image detection algorithm is first obtained, and multiple algorithm identifiers are combined to form a general algorithm call parameter; then the general algorithm call parameter and the image to be detected are sent to the general algorithm service to obtain multiple detection results through a single batch call; for the second scheduling channel, independent asynchronous execution tasks are executed through a dedicated execution strategy to obtain multiple detection results.

[0130] In this embodiment, after the algorithm distribution is completed, the system simultaneously starts the first scheduling channel and the second scheduling channel. The first scheduling channel is responsible for the batch aggregation and invocation of general algorithms, while the second scheduling channel is responsible for the independent asynchronous invocation of special algorithms, thereby forming a parallel execution relationship among multiple algorithms.

[0131] For the first scheduling channel, all algorithm identifiers are concatenated with commas to form an algorithm component code set. This set, along with the image to be detected, is sent to the batch prediction interface of the AI ​​algorithm gateway in a single batch request via multipart / form-data format. If interface authentication is required, dynamic authentication information can be generated before sending the batch request, such as POST / ai-gateway / predict. This request uses the HMAC-SHA256 dynamic signature authentication mechanism. Each request calculates the signature in real-time based on the current GMT time and key. The request header format can be hmacusername="...", algorithm="hmac-sha256", headers="x-date", signature="...", allowing different requests to have different authentication credentials. This method reduces the security risks associated with the long-term reuse of fixed authentication information. Subsequently, the algorithm gateway calls the corresponding detection model based on the multiple algorithm identifiers and returns the detection results corresponding to each algorithm.

[0132] For the second scheduling channel, independent asynchronous tasks are established based on the dedicated execution strategy corresponding to each dedicated image detection algorithm. Taking the phishing behavior detection algorithm, which needs to be executed through the Dify workflow, as an example, the image to be detected is first uploaded to the Dify file service to obtain the file identifier. Then, a blocking workflow execution request is initiated from the main process through the IPC proxy, passing in the file identifier, image width, and height parameters. All algorithm calls are executed in full parallelism through Promise.all(), and the failure of a single algorithm does not affect the normal operation of other algorithms. The detection results are uniformly aggregated by the ResultAggregator. The detection results returned by each channel are expanded into a one-dimensional result array. For each result, the minimum confidence threshold minConfidence configured for the algorithm is associated, and confidence filtering is performed: only detection annotation data with a confidence greater than or equal to the threshold are retained; if the valid detection result of an algorithm is empty, the matching item is discarded.

[0133] Furthermore, an asynchronous parallel task mechanism can be used to execute multiple general-purpose algorithms and multiple specialized algorithms simultaneously. For example, a batch request corresponding to the first scheduling channel and multiple independent tasks corresponding to the second scheduling channel can be launched simultaneously using a parallel task mechanism such as Promise.all. When one algorithm fails, the corresponding algorithm can be marked as failed, while the normal detection results of other algorithms are retained, so that the failure of a single algorithm does not cause the entire image detection task to fail.

[0134] For example, for a surveillance image that simultaneously contains people, safety helmets, and water activities, common algorithms such as safety helmet detection and personnel detection can be submitted to the algorithm gateway for batch detection. Meanwhile, the phishing behavior detection algorithm, which requires a dedicated workflow, can be submitted independently to its corresponding workflow. These multiple algorithms execute within the same time window, ultimately yielding results for safety helmet detection, personnel detection, and phishing behavior detection, respectively.

[0135] Understandably, aggregating general algorithms into a single batch request reduces the overhead of connection establishment, authentication, and data transmission caused by multiple algorithms making separate network requests. Creating independent asynchronous tasks for specialized algorithms ensures they have independent execution space and call flows. Furthermore, parallel execution eliminates the need for algorithms to wait sequentially in a fixed order, thus making the overall detection response time primarily dependent on the execution path of the longest-running algorithm, rather than the simple sum of all algorithm execution times. Simultaneously, a single-algorithm exception isolation mechanism prevents the failure of a single algorithm call from interrupting the entire multi-algorithm detection process, improving its stability.

[0136] S104. Filter multiple detection results according to the confidence threshold corresponding to each image detection algorithm to obtain valid detection results.

[0137] In one possible implementation, multiple detection results are first associated according to their corresponding image detection algorithms, and the confidence thresholds for each image detection algorithm are obtained. Then, for each detection result, the confidence level of the detection annotation data in the detection result is compared with the confidence threshold of the image detection algorithm. When the confidence level of the detection annotation data is greater than or equal to the confidence threshold, the detection annotation data is retained as valid detection annotation data. When there is no valid detection annotation data in the detection result, the detection result is discarded.

[0138] It should be understood that, since different image detection algorithms may have different recognition capabilities and output confidence distributions, this embodiment does not use a uniform confidence threshold. Instead, it configures a corresponding minimum confidence threshold for each image detection algorithm. After the detection results are returned, each detection result is first associated with the image detection algorithm that generated it based on the algorithm identifier, and the confidence threshold corresponding to that image detection algorithm is read. Subsequently, a confidence judgment is performed on each detection annotation data in the detection results, comparing the confidence in the detection annotation data with the minimum confidence threshold of the corresponding algorithm. When the confidence corresponding to the detection annotation data is greater than or equal to the confidence threshold, the detection annotation data is determined to be valid detection annotation data. When all detection annotation data corresponding to a certain detection result does not reach the corresponding threshold, the detection result is discarded and will not proceed to the subsequent alarm processing process.

[0139] For example, if the minimum confidence threshold for a safety helmet detection algorithm is 0.80, and its detection results include three labeled data points with confidence levels of 0.91, 0.76, and 0.84, the system will retain the two labeled data points with confidence levels of 0.91 and 0.84, and discard the labeled data point with a confidence level of 0.76. If all the labeled data points output by a certain algorithm are below the corresponding threshold, the detection results corresponding to that algorithm will not be included in the subsequent alarm generation process.

[0140] In one implementation, after the detection results are filtered, they can be structurally associated according to the correspondence between the basic information of the alarm record and the detection results. This allows subsequent alarm records to not only save the alarm conclusion, but also to save information such as the algorithm identifier that generated the alarm, the detection category, the detection box position, and the corresponding confidence level, thereby forming a complete chain of detection evidence.

[0141] Understandably, by configuring independent confidence thresholds for different image detection algorithms, the detection result filtering process can be adapted to the recognition features of different algorithms, avoiding the situation where some algorithms filter too strictly or not enough due to the use of a uniform threshold. At the same time, directly removing detection annotation data that does not meet the confidence requirements during the filtering stage can reduce invalid detection results from entering the subsequent annotation, alarm storage, and message push processes, thereby reducing the subsequent processing load and reducing the occurrence of invalid alarms due to low-confidence detection results.

[0142] The image multi-algorithm parallel scheduling method provided in this application first acquires the image to be detected and a preset algorithm scheduling strategy to determine the corresponding scheduling method according to different image detection algorithms. After acquiring the image to be detected, tasks are allocated to multiple image detection algorithms corresponding to the image to be detected according to the algorithm scheduling strategy. Each image detection algorithm is scheduled to a target scheduling channel that matches its execution strategy. The target scheduling channel includes at least a first scheduling channel that supports batch aggregation scheduling and / or a second scheduling channel that supports independent asynchronous scheduling. For multiple image detection algorithms suitable for batch processing, the corresponding algorithm tasks can be aggregated through the first scheduling channel to reduce redundant task scheduling and resource initialization overhead. For image detection algorithms that need to be executed independently or have high requirements for execution latency, they are asynchronously scheduled through the second scheduling channel, so that different algorithm tasks can run in mutually independent execution contexts. Furthermore, multiple image detection algorithms are executed in parallel based on the target scheduling channels corresponding to each image detection algorithm, so that multiple algorithms can target the same image to be detected. Simultaneously, detection processing is performed, and corresponding detection results are output separately. This leverages the execution characteristics of different scheduling channels to improve the parallel processing capability of multiple algorithms and reduce the cumulative processing latency caused by the sequential execution of multiple algorithms. After obtaining the detection results corresponding to each image detection algorithm, the detection results are further filtered based on the confidence threshold corresponding to each image detection algorithm. The confidence of the detection results is compared with the corresponding confidence threshold, and the detection results that meet the corresponding confidence threshold are retained as valid detection results. This suppresses the output of low-confidence detection results while completing the parallel detection of multiple algorithms, thereby improving the effectiveness and reliability of the final detection results. Through the above method, different image detection algorithms are allocated to different scheduling channels according to their execution characteristics based on a preset algorithm scheduling strategy. Batch aggregation scheduling and independent asynchronous scheduling are used to achieve parallel execution of multiple algorithms. Then, the results are filtered by combining the confidence threshold matched to each algorithm. This can improve the efficiency of algorithm resource utilization and overall detection processing efficiency while ensuring the detection coverage capability of multiple algorithms and reducing the interference of invalid detection results on subsequent processing.

[0143] In one possible implementation, acquiring the image to be detected includes at least one of the following methods:

[0144] First, in response to a screenshot trigger operation, a transparent overlay window is created on the target display and the target area is selected, and the image to be detected is cropped from the screen video stream. Second, in response to a local image selection operation, the image to be detected is retrieved from the local storage area. Third, based on the device monitoring data acquisition conditions, historical monitoring images are retrieved from the IoT device platform as the image to be detected.

[0145] In the specific implementation process, screenshot acquisition entry, local image selection entry, and device monitoring image acquisition entry can be set so that the image to be detected can be obtained through at least the following methods.

[0146] The first method is desktop screenshot. When a user needs to inspect the image content on the current monitor, they can trigger a screenshot operation using a preset shortcut key, such as F1. On the desktop, the user can use the F1 shortcut key to bring up the multi-monitor full-screen transparent overlay screenshot tool. The window management module calls the operating system's screen enumeration interface to obtain information about all connected monitors, including boundary coordinates, scaling ratio, display ID, etc., and creates an ElectronBrowserWindow instance for each monitor, setting the window properties to borderless, transparent background, window on top, and full-screen overlay of the corresponding monitor's workspace. All screenshot windows are created in parallel. When the user clicks on a screenshot window on any monitor, the screenshot windows on the other monitors automatically close, leaving only the screenshot window on the currently active monitor.

[0147] Furthermore, the video stream capture module obtains the screen video source of the currently active display through Electron's desktopCapturer API, calls the browser's standard getDisplayMedia interface to obtain the screen video stream, and assigns it to the srcObject property of the hidden video element. The screenshot device's interactive state is managed by a four-state finite state machine: idle, selected, created, dragged, and adjusted. When the user presses the left mouse button and drags, the current mouse position is updated in real time, the selection rectangle is dynamically calculated, and the clip-path of the cutout mask is updated. After releasing the mouse, if both the width and height of the selection exceed the minimum threshold of 10 pixels, the selection is confirmed to be created, and the dashed border of the selection, the 8-way zoom handle, and the operation bar are displayed. The cutout mask uses CSS's clip-path polygon function to "cut out" a hole in the selection on a fixed-position, semi-transparent black mask layer, making the inside of the selection transparent and allowing the user to see the original screen image below. The image cropping engine creates a full-size off-screen canvas, draws the current video frame onto the full-size canvas, calculates the scaling ratio from the display coordinates to the original coordinates (scaleX = videoWidth / window width, scaleY = videoHeight / window height), multiplies the display coordinates of the user-selected area by the scaling ratio to obtain the selection coordinates at the original resolution, then creates a cropCanvas, uses drawImage to crop the corresponding pixel area from the selection position on the full-size canvas, outputs a PNG format cropped image as the image to be detected, encapsulates it as a File, and directly passes it to the algorithm scheduling and decision module. The screenshot tool also supports a timed screenshot mode, which can be automatically triggered by setInterval or requestAnimationFrame at preset time intervals of 3 seconds, 5 seconds, 10 seconds, 15 seconds, or a user-defined value ≥3 seconds to execute the image cropping and detection process; when the timed mode is started, it automatically switches to pass-through mode so that the timed screenshot operation does not interfere with the user's operation of the underlying application. In pass-through mode, calling Electron's setIgnoreMouseEvents(true,{forward:true}) causes mouse events to pass through the screenshot window, and the control interface can be brought back via a floating button.

[0148] In another implementation, the images to be detected can be obtained using local image selection. Users can trigger the local image selection operation through the file selection interface or the F2 shortcut key, selecting one or more images from local storage as the images to be detected. A maximum of 50 local images can be selected at a time, and the format is limited to jpg, jpeg, and png. For example, during a manual inspection process, a user can select multiple historical images at once, allowing the selected images to sequentially enter the subsequent algorithm detection process.

[0149] In another implementation, the image to be detected can be acquired using a device monitoring image acquisition method. When a user needs to detect historical monitoring images stored in the IoT device platform, they can request the corresponding historical monitoring images from the IoT device platform according to preset acquisition conditions. Acquisition conditions may include a target time range, a device list, and preset bit parameters, etc.

[0150] In one feasible approach, to accommodate the batch acquisition needs of a large number of devices and a large number of historical images, the image acquisition task can be divided into a three-level task hierarchy: batch task, device task, and download task. The batch task is the top-level container, describing the overall information of a download operation, including the time range, preset bit filter list, overall status, storage directory path, cumulative active runtime, and a set of subordinate device tasks. The device task is the middle layer, associated with specific IoT devices, including device ledger information such as line name, pole number, installation direction, metadata acquisition status, device-level download status, skip reason statistics dictionary, and a set of subordinate download tasks. The download task is the smallest execution unit, corresponding to a single image file download operation, including the IoT storage path fileFolder / fileName, preset bit number, shooting time, download status, download progress percentage, local storage path, and error details.

[0151] Furthermore, separate metadata retrieval queues and image download queues can be established, with corresponding concurrency levels set for each. For example, the default concurrency level for both queues can be set to 4, and can be adjusted within the range of 1 to 10 based on the number of devices, network bandwidth, and terminal performance. When the number of active metadata requests is less than the upper limit and there are devices waiting to be processed, the metadata retrieval queue retrieves the task from the head of the queue, calls startDeviceMetadata to confirm the start, sends an HTTP POST request to the IoT platform to query the list of image files for that device within the batch time range, parses the returned CONTENT JSON field, extracts the fileFolder, fileName, and presetPosition for each image, performs deduplication and filtering by comparing it with the preset position filter list configured in the batch and existing local files, and calls finishDeviceMetadata to create a corresponding DownloadTask subtask for that device and add it to the download queue. Metadata requests are bound using an independent AbortController, supporting timeout and user-initiated cancellation. The download execution queue maintains a global active download counter. When the active count is less than the download concurrency limit and there are queued tasks in the queue, the DownloadTask at the head of the queue is retrieved, its status is set to downloading, and a GET request is initiated to download the file using the connection pool configured with keepAlive:true and maxSockets:10 in Node.js's HTTPS module. The file is then written to the local disk in a streaming manner using fs.createWriteStream. After the download is complete, the status is updated to completed, the active count is decremented by one, and the next round of scheduling is triggered. If an error occurs, the error information is recorded and the status is set to failed.

[0152] To prevent the loss of completed tasks during batch downloads due to application crashes, network interruptions, or other anomalies, the execution status of batch tasks, device tasks, and download tasks can be persisted. For example, manifest.json can serve as the master index file, recording the current batch IDs and status summaries; each batch can be stored independently as a batches / batchID.json file, containing the complete three-level task status for that batch. The saveState function is triggered debouncing every second, iterating through all batches, devices, and download tasks, serializing the current status into JSON and writing it to the corresponding file; before the application exits normally, the before-quit event forces a call to the flush function to ensure the last status change is completely written to disk. When the application starts, the restoreState function is called. First, it reads manifest.json to obtain a list of batch IDs, and then reads each batch file and deserializes it into a batch object. For all batches whose status is not completed or canceled, the recovery logic is executed: batch-level running / pending is set to pending, while partial / failed / incomplete remains unchanged; device metadata status is reset from pending or requesting to pending to re-enqueue metadata retrieval, and incomplete retains completed download records; download task status is reset from downloading to queued to re-enqueue download, while queued / failed / incomplete retains its original state; after recovery, the dual-queue scheduling loop is restarted, and unfinished tasks are automatically resumed from the breakpoint. Cancellation operations support three levels: single task, single device, and entire batch. Each active HTTP / HTTPS request is bound to an independent AbortController instance. Single task-level cancellation calls the abort() method of the corresponding request for that task. Device-level cancellation iterates through all download tasks under that device, calls abort() one by one, and cancels metadata requests. Batch-level cancellation iterates through all devices in the batch, recursively executes device-level cancellation, and cleans up pending queue items. Cancelled tasks support retrying. Tasks in the failed / canceled / incomplete state are reset to queued or pending and then re-enter the scheduling queue.

[0153] It should be understood that by setting up multiple unified image acquisition entry points, real-time desktop screenshots, local historical images, and batch historical monitoring images from IoT device platforms can be uniformly converted into images to be detected and enter the same algorithm scheduling process, avoiding the need to repeatedly build detection and processing links for different data sources. Through three-level task management, dual-queue concurrency, and state persistence mechanisms, the concurrent processing capability during the acquisition of a large number of device monitoring images can be improved, and unfinished tasks can be recovered after the application exits abnormally, reducing duplicate downloads and data omissions.

[0154] Figure 2 A flowchart illustrating the image multi-algorithm parallel scheduling method provided in this application embodiment. Figure 2 ;like Figure 2 As shown, in this embodiment... Figure 1 Based on the embodiments, the method is described in detail below, including:

[0155] S201. Based on the detection annotation data in the valid detection results, determine the corresponding detection area in the image to be detected, and generate visual annotation information corresponding to the detection area.

[0156] After obtaining valid detection results, this embodiment converts information such as target category, detection box coordinates, and confidence level from the detection annotation data into visual annotation information that can be overlaid onto the image to be detected. Specifically, an off-screen canvas with the same pixel size as the original image to be detected can be created, the image to be detected can be drawn onto the canvas, and then each detection annotation data in the valid detection results can be traversed. The detection area is determined according to the target category and detection box coordinates corresponding to the detection annotation data, and a rectangular detection box is generated around the detection area. At the same time, category labels and confidence level information are generated above the detection box.

[0157] To help users distinguish different detection categories within the same image, a corresponding color value can be generated based on the category name using a stable hash algorithm. This ensures that the label color for the same category remains consistent across different images, while different categories have different label colors. Subsequently, a border of the corresponding color is drawn on the detection bounding box, and a label background, target category name, and confidence score are drawn above the detection bounding box.

[0158] For example, based on the labeled category name, a hue value is generated using a stable hash function, with a hue range of 200°–340°. After HSL-to-RGB conversion, a unique and stable color value is obtained for each category, with similar categories using the same color and dissimilar categories using different colors. This color is used as the outline color to draw detection bounding boxes on the canvas. Above the bounding boxes, a background color block for the label is first drawn, followed by the category name and confidence value in white text, formatted as "title confidence%". After annotation, the canvas's `toBlob` method outputs a PNG formatted Blob image. Thus, the structured information in the detection results, such as the bounding box coordinates, category, and confidence score, is converted into visual annotation information, corresponding to the original image.

[0159] Understandably, by converting structured detection results into visual annotations on images, abstract detection box coordinates, target categories, and confidence levels can be transformed into intuitive visual evidence, allowing users to determine the target location and recognition result corresponding to the alarm without having to view the algorithm's original data separately. At the same time, using stable category color mapping can improve the recognizability of multi-category detection results in the same image, providing an intuitive basis for subsequent manual review and alarm tracing.

[0160] S202. Label the image to be detected based on the visual annotation information to obtain the labeled image.

[0161] In this embodiment, the image to be detected can be used as a base image, and the visual annotation information generated in S201 can be superimposed on the image to be detected to generate an annotated image corresponding to the original image. The annotated image not only contains the content of the original image, but also superimposes the detection bounding box, category label, and confidence score. The annotation process can be completed in an off-screen canvas. After all valid detection annotation data has been drawn, the canvas content can be converted into an image data object, such as a PNG format Blob object, through the canvas's image output interface, and the Blob object can be used as input for subsequent image compression and storage.

[0162] In one implementation, the system can also determine whether to save the unannotated original image based on the original screenshot save switch in the system configuration. When the original screenshot save switch is turned on, the original image can be saved to the user-preset screenshot storage directory before annotation is performed. When the switch is turned off, only the subsequently generated annotated image is retained to reduce the local storage space usage.

[0163] Understandably, by directly overlaying the detection results onto the original image to generate an annotated image, an intuitive correspondence can be established between the detection algorithm output and the actual image content, forming detection evidence that can be independently saved, previewed, and transmitted. At the same time, annotating via an off-screen canvas avoids directly modifying the original image data, allowing the original image and the annotated image to be saved separately according to the configuration, thus balancing the needs of preserving detection evidence and protecting the original data.

[0164] S203. Compress the labeled image and associate the compressed labeled image with the valid detection results to generate corresponding alarm records.

[0165] In one possible implementation, the alarm record includes at least the recognition time, image detection algorithm identifier, alarm category, alarm confidence level, review status, push status, and storage path of the labeled image; wherein, the review status is used to indicate whether the alarm record has been reviewed and the review result, and the push status is used to indicate whether the message corresponding to the alarm record has been sent.

[0166] To reduce the amount of data in local storage and subsequent message transmission of labeled images, compression can be performed on the labeled images after annotation. Specifically, the labeled image blob output from the canvas can be passed to a WebWorker thread, where a binary search algorithm is used to automatically find the optimal JPEG compression quality parameter `quality`. The initial search interval is [0.1, 1.0]. In each iteration, the midpoint of the interval is used as the `quality` value for JPEG compression encoding, and the search interval is narrowed based on the comparison between the output file size and the target upper limit of 2MB. After 6 iterations with a tolerance of 0.005, the final compressed blob is output. Compression is performed in a separate Worker thread without blocking the main UI thread. The labeled images and detection results are structurally associated and stored in the database, and a separate image preview window is available for full-size viewing.

[0167] After compression, the compressed labeled images are saved to local storage, and the corresponding storage path is obtained. Then, the storage path is structurally associated with the valid detection results to generate alarm records. Alarm records may include fields such as alarm identifier, recognition time, algorithm identifier, algorithm name, alarm category, detection labeled data, detection confidence level, labeled image path, review status, push status, and review remarks. The detection labeled data may further include information such as detection area coordinates, target category, and confidence level.

[0168] In one implementation, a matching array can be formed according to the correspondence between alarm records and detection algorithms, with each matching item corresponding to the detection result of an image detection algorithm. For the same alarm record, the detection results of multiple algorithms can be saved simultaneously, allowing reviewers to view the detection basis generated by different algorithms for the same image to be detected.

[0169] For example, if an image to be detected produces valid results after being processed by a safety helmet detection algorithm and a personnel detection algorithm, the detection bounding box, category, and confidence score of the safety helmet detection algorithm, as well as the corresponding detection data of the personnel detection algorithm, can be associated with the path of the finally generated labeled image in the same alarm record. In this way, when reviewers view the alarm later, they can simultaneously view the basic alarm information, the algorithm detection results, and the visualized labeled image.

[0170] Understandably, by simultaneously storing detection algorithm information, detection annotation data, confidence level, review status, push status, and annotation image path in the alarm log, the algorithm detection results, visual evidence, and alarm processing process can be structurally linked to form a complete alarm evidence chain. By placing the image compression operation in a separate thread, the impact of the image processing process on the main thread and user interface can be reduced. At the same time, by setting image size constraints, the amount of data in subsequent storage and message transmission processes can be reduced.

[0171] S204. Persistently store alarm records and determine whether alarm records should enter the manual review or message push process according to the preset review policy.

[0172] In one possible implementation, the manual review switch status corresponding to the image detection algorithm that generates the alarm record is first obtained; when the manual review switch is in the off state, the review status of the alarm record is set to reviewed, and a message push is triggered based on the alarm record; when the manual review switch is in the on state, the review status of the alarm record is set to pending review, and the alarm record is added to the manual review queue; then, in response to the review operation of the alarm record in the manual review queue, the review status of the alarm record is updated according to the review result, and it is determined whether to trigger a message push based on the updated review status.

[0173] In this embodiment, after generating an alarm record, the alarm record is persistently stored through a database service provided by the main process. For example, alarm data can be sent to the main process via an inter-process communication interface, and the main process can call the SQLite database service to create an alarm record, writing information such as the labeled image path, algorithm matching result, detection label data, review status, and review remarks into the database. By associating the labeled images in the file system with the structured alarm records in the database, a correspondence of "alarm record - detection result - labeled image" can be formed.

[0174] Furthermore, the initial review status of an alarm record can be determined based on the manual review switch corresponding to the image detection algorithm that generates the alarm record. When the manual review switch is off, the detection result corresponding to the algorithm can be directly marked as reviewed and the review result can be set to pass; when the manual review switch is on, the review status of the detection result is set to pending review, and the corresponding alarm record is added to the manual review queue.

[0175] For example, after long-term operation and verification, the manual review switch of the safety helmet detection algorithm can be configured to be turned off. At this time, when the safety helmet detection algorithm generates a detection result that meets the confidence threshold, the corresponding result can be directly set as reviewed and enter the message push process. For phishing behavior detection algorithms that require further manual confirmation, the manual review switch can be set to on, so that its detection result enters the manual review queue and is judged by the reviewers to determine whether the detection result is true.

[0176] When the manual review switch is enabled, reviewers can select alarm records to be reviewed from the manual review queue, view the corresponding labeled images, detection categories, detection areas, and detection confidence levels, and perform review operations. The system updates the review status and results of the corresponding alarm records based on the review operations, and determines whether to trigger a message push based on the updated review status.

[0177] In one implementation, the review status and push status are managed independently. The review status may include at least the following states: pending review, approved, and rejected; the push status may include at least the following states: not sent, sent, and sent failed. When an alarm has been approved but the message fails to be sent, there is no need to perform manual review again; simply re-push the message based on the failed message status.

[0178] After the alarm records are persisted, they can be displayed, filtered, manually reviewed, and exported in the alarm center. For example, alarm records can be retrieved based on criteria such as identification time, alarm category, algorithm identifier, review status, and push status, so as to query and trace historical alarms.

[0179] Understandably, by automatically linking alarm generation, alarm storage, manual review, and message push, a complete closed loop from image detection to alarm processing can be formed. By configuring manual review switches according to different algorithms, detection results that do not require manual review can directly enter subsequent processing, while detection results that require manual confirmation enter the review queue, thus avoiding the use of the same review process for all algorithms. At the same time, maintaining the review status and push status separately can prevent alarms from being re-reviewed due to message sending failures, thereby improving the traceability and manageability of the alarm processing process.

[0180] In S204 above, when it is determined from the review status that the message push conditions are met, the message push operation can be further executed.

[0181] In one possible implementation, triggering a push notification includes:

[0182] First, obtain the pre-configured alarm message receiving object; then, generate an alarm message based on the identification time and alarm category in the alarm record, and obtain the compressed labeled image associated with the alarm record; then, send the alarm message and the compressed labeled image to the alarm message receiving object; finally, update the push status of the alarm record based on the message sending result, where the push status includes at least sent, sent failed, and not sent.

[0183] First, the pre-configured alarm message receiving objects are obtained, such as one or more pre-configured receiving numbers. Then, an alarm message is generated based on the identification time, alarm category, and approved detection results in the alarm record, and the compressed labeled image associated with the alarm record is obtained. Finally, the alarm message and labeled image are sent to the corresponding alarm message receiving objects.

[0184] In one specific implementation, the labeled image can be uploaded to a messaging platform to obtain the corresponding image file identifier. Then, an image message is constructed based on this identifier, and simultaneously, a text message containing summary information about the detection time and detection category is also constructed. For example, the text message can adopt the format "[Identification Time] Category A and Category B have been identified; please pay attention promptly." For multiple alarm message recipients, separate message sending tasks can be established to send image messages and text messages in parallel. Specifically, a parallel sending mechanism can be adopted, and the `allSettled` method can be used to ensure that the failure of a single recipient to send messages will not affect the sending of other recipients.

[0185] After transmission is complete, the push status in the alarm record is updated based on the transmission results of each receiving object. When at least one receiving object successfully transmits, the push status can be set to "sent"; when all receiving objects fail to transmit, the push status is set to "transmission failed"; when the alarm record has not yet met the push conditions, the "not sent" status is maintained.

[0186] For example, if an alarm record needs to send messages to three alarm recipients, and two of them successfully receive the message while one fails to receive it, the push status of the alarm record can be set to "sent" while retaining the specific sending result so that it can be re-pushed to the failed recipient later without having to re-execute image detection and manual review.

[0187] Understandably, by sending alarm messages and labeled images together, recipients can directly obtain corresponding visual detection evidence upon receiving alarm notifications, reducing the information gaps caused by relying solely on textual descriptions. By establishing separate sending tasks for multiple recipients, the impact of an abnormal sending result on a single recipient can be reduced. Furthermore, by persistently recording the sending results, a basis can be provided for subsequent failure re-pushing and alarm tracing.

[0188] In summary, the overall workflow of this embodiment can be summarized as follows: First, the image to be detected is obtained through at least one of the following methods: multi-monitor screenshot, local image selection, or batch download from an IoT device platform. Then, the currently enabled image detection algorithm and its corresponding execution strategy are read, and algorithms that can be called through a unified interface are assigned to the first scheduling channel, while algorithms with dedicated execution strategies are assigned to the second scheduling channel. Subsequently, multiple general algorithms are batch aggregated and called through the first scheduling channel, while multiple dedicated algorithms are independently and asynchronously called through the second scheduling channel, and the detection results generated by multiple channels are uniformly aggregated. Then, the detection annotation data is filtered according to the minimum confidence threshold corresponding to each image detection algorithm to obtain valid detection results. Next, the detection area is determined in the original image based on the valid detection results, and detection boxes, category labels, and confidence information are generated through Canvas to obtain an annotated image. The annotated image is compressed, and the compressed annotated image is structurally associated with the detection results to generate an alarm record. Finally, the alarm record is determined to enter the manual review process or directly enter the message push process according to the algorithm-level manual review strategy, and the review status and push status are persistently managed respectively.

[0189] For example, in a power equipment inspection scenario, users can use the F1 shortcut to capture the equipment monitoring screen on a monitor, or use F2 to select historical inspection images, or download historical monitoring images in batches from an IoT device platform according to a device list and time range. For a single monitoring image, the system can simultaneously call general algorithms such as safety helmet detection, personnel detection, and equipment anomaly detection, and independently call behavior recognition algorithms that require dedicated workflows. Assuming the confidence level of the safety helmet detection result is 0.92 and the confidence level of the personnel detection result is 0.95, both exceeding their respective thresholds, the system generates corresponding detection boxes and category labels on the original image, compresses the labeled image, and saves it along with the detection results as an alarm record. If the safety helmet detection algorithm is configured to not require manual review, the result can directly enter the push process; if the behavior recognition algorithm is configured to require manual review, the corresponding result enters the manual review queue, where reviewers check the labeled image and detection information to confirm whether to push. In this way, a complete closed loop is formed from image acquisition, algorithm scheduling, parallel detection, result filtering, visual labeling, alarm storage, manual review, to message push.

[0190] This embodiment unifies the organization of multi-source image acquisition, multi-algorithm hybrid parallel scheduling, algorithm-level confidence filtering, detection result visualization, alarm structured storage, and differentiated review and message push, enabling images from different sources to reuse the same detection and alarm processing chain. The scheduling mechanism of "batch aggregation of general algorithms + independent asynchronous execution of dedicated algorithms" reduces redundant calls and enables parallel execution between algorithms; the three-level task management and breakpoint recovery mechanism improves the continuity of batch historical image acquisition; the algorithm-level review strategy allows different algorithms to use different alarm processing paths; and the structured association between detection results and labeled images preserves a complete alarm evidence chain. Therefore, this embodiment not only achieves multi-algorithm image detection but also forms an integrated technical process from data acquisition to closed-loop alarm processing.

[0191] Figure 3 This is a schematic diagram of the structure of the image multi-algorithm parallel scheduling device provided in the embodiments of this application; as shown below. Figure 3 As shown, the device includes:

[0192] The acquisition module 301 is used to acquire the image to be detected and the preset algorithm scheduling strategy. The algorithm scheduling strategy is configured with the execution strategy corresponding to each image detection algorithm.

[0193] The scheduling module 302 is used to schedule multiple image detection algorithms corresponding to the image to be detected to the corresponding target scheduling channels according to a preset algorithm scheduling strategy. The target scheduling channels include at least a first scheduling channel that supports batch aggregation scheduling and / or a second scheduling channel that supports independent asynchronous scheduling.

[0194] The processing module 303 is used to execute multiple image detection algorithms in parallel based on the target scheduling channel, so as to output the detection results corresponding to each image detection algorithm;

[0195] The filtering module 304 is used to filter multiple detection results according to the confidence threshold corresponding to each image detection algorithm to obtain effective detection results.

[0196] In one possible implementation, the acquisition module 301 is specifically used for:

[0197] In response to a screenshot trigger operation, a transparent overlay window is created on the target display and the target area is selected, and the image to be detected is cropped from the screen video stream; in response to a local image selection operation, the image to be detected is retrieved from the local storage area.

[0198] Based on the conditions for obtaining equipment monitoring data, historical monitoring images are obtained from the IoT device platform as images to be detected.

[0199] In one possible implementation, the acquisition module 301 is specifically used for:

[0200] The acquisition task is divided into a three-level task hierarchy, including batch tasks, device tasks, and download tasks;

[0201] Separate metadata acquisition queues and image download queues are established, and device tasks and download tasks are scheduled in parallel according to preset concurrency levels;

[0202] The execution status of level 3 tasks is persistently recorded, and unfinished tasks are restored based on the persistent status after task interruption is detected.

[0203] Among them, the restored download tasks were reset to the pending download status and re-added to the image download queue.

[0204] In one possible implementation, the scheduling module 302 is specifically used for:

[0205] Obtain the algorithm identifier corresponding to the currently enabled image detection algorithm, and query the preset algorithm scheduling strategy based on the algorithm identifier;

[0206] When the algorithm identifier has a corresponding dedicated execution strategy in the preset algorithm scheduling strategy, the image detection algorithm is assigned to the second scheduling channel, and an independent asynchronous execution task is established for the image detection algorithm;

[0207] When there is no corresponding dedicated execution strategy for the algorithm identifier, the image detection algorithm is assigned to the first scheduling channel, and the corresponding algorithm identifier is added to the general algorithm call set.

[0208] In one possible implementation, the processing module 303 is specifically used for:

[0209] For the first scheduling channel, obtain the algorithm identifier of each general image detection algorithm, and combine multiple algorithm identifiers to form a general algorithm call parameter;

[0210] Send the general algorithm call parameters and the image to be detected to the general algorithm service to obtain multiple detection results through a single batch call;

[0211] For the second scheduling channel, independent asynchronous execution tasks are executed through a dedicated execution strategy to obtain multiple detection results.

[0212] In one possible implementation, the filtering module 304 is specifically used for:

[0213] Multiple detection results are correlated according to their corresponding image detection algorithms, and the confidence thresholds for each image detection algorithm are obtained.

[0214] For each detection result, the confidence level corresponding to the detection annotation data in the detection result is compared with the confidence level threshold corresponding to the image detection algorithm;

[0215] When the confidence level of the detected labeled data is greater than or equal to the confidence level threshold, the detected labeled data will be retained as valid detected labeled data.

[0216] If no valid detection labeling data is found in the detection results, the detection result will be discarded.

[0217] In one possible implementation, the device further includes an alarm module for:

[0218] Based on the detection annotation data in the valid detection results, the corresponding detection area is determined in the image to be detected, and the visual annotation information corresponding to the detection area is generated.

[0219] The image to be detected is annotated based on the visual annotation information to obtain an annotated image;

[0220] The labeled images are compressed, and the compressed labeled images are associated with the valid detection results to generate corresponding alarm records;

[0221] Alarm records are persistently stored, and the system determines whether to proceed to manual review or message push based on a preset review policy.

[0222] In one possible implementation, the alarm record includes at least the recognition time, image detection algorithm identifier, alarm category, alarm confidence level, review status, push status, and storage path of the labeled image;

[0223] The review status indicates whether the alarm record has been reviewed and the review result, while the push status indicates whether the message corresponding to the alarm record has been sent.

[0224] In one possible implementation, the alarm module is specifically used for:

[0225] Obtain the manual review switch status corresponding to the image detection algorithm that generates alarm records;

[0226] When the manual review switch is off, the review status of the alarm record is set to reviewed, and a message push is triggered based on the alarm record;

[0227] When the manual review switch is turned on, the review status of the alarm record is set to pending review, and the alarm record is added to the manual review queue.

[0228] In response to the review operation of alarm records in the manual review queue, the review status of the alarm records is updated according to the review results, and a message push is triggered based on the updated review status.

[0229] In one possible implementation, the alarm module is specifically used for:

[0230] Retrieve the pre-configured alarm message receiver object;

[0231] An alarm message is generated based on the identification time and alarm category in the alarm record, and a compressed labeled image associated with the alarm record is obtained;

[0232] Send the alarm message and the compressed labeled image to the alarm message receiving object;

[0233] The push status of the alarm record is updated based on the message sending result. The push status includes at least three states: sent, failed to send, and not sent.

[0234] The image multi-algorithm parallel scheduling device provided in this application embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0235] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 40 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the device 40 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0236] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.

[0237] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0238] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0239] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0240] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0241] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0242] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0243] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0244] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0245] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0246] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0247] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0248] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0249] Based on the above disclosure, those skilled in the art will understand that the computer-readable storage medium and its related implementation methods provided in the embodiments of the present invention are not limited to the specific hardware or logical partitioning forms described above. In actual product or system deployments, computer execution instructions can be encapsulated in different software distribution packages or firmware modules and distributed via networks, optical discs, or mobile storage devices. When computer execution instructions are loaded and executed by one or more processors, in addition to implementing the aforementioned method steps, hardware resource configurations can be automatically identified according to specific application scenarios, and thread priorities or memory allocation strategies can be dynamically adjusted to optimize execution efficiency without manual intervention. Furthermore, the computer-readable storage medium can also be integrated into IoT terminals, edge computing nodes, or cloud servers to form a distributed instruction storage and collaborative execution system to meet differentiated needs such as high concurrency, low latency, or data privacy protection.

[0250] Furthermore, it should be emphasized that the logical functions corresponding to the methods, units, modules, and storage media in this application, without departing from the core ideas of this invention, allow developers to make equivalent reconstructions, renamings, or re-divisions based on the characteristics of specific programming languages ​​or frameworks. For example, the functions of multiple "units" can be merged into a single service process, or message queues or pipe communication can be used instead of direct function calls. Even if these implementation details are not exactly the same as the foregoing embodiments, as long as they ultimately achieve the same technical effect and no creative effort is required, they should all be considered to fall within the protection scope of this invention. Similarly, any intermediate data processing forms, exception handling mechanisms, or logging methods embodied in the instructions in the computer-readable storage medium when executed are also derivative technologies of this invention and should not be construed as substantial limitations on the original solution.

[0251] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0252] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A parallel scheduling method for multiple image algorithms, characterized in that, include: The image to be detected and a preset algorithm scheduling strategy are obtained, wherein the algorithm scheduling strategy is configured with the execution strategy corresponding to each image detection algorithm; According to the preset algorithm scheduling strategy, multiple image detection algorithms corresponding to the image to be detected are respectively scheduled to the corresponding target scheduling channels. The target scheduling channels include at least a first scheduling channel that supports batch aggregation scheduling and / or a second scheduling channel that supports independent asynchronous scheduling. The multiple image detection algorithms are executed in parallel based on the target scheduling channel to output the detection results corresponding to each image detection algorithm; The multiple detection results are filtered according to the confidence threshold corresponding to each image detection algorithm to obtain valid detection results.

2. The method according to claim 1, characterized in that, Acquiring the image to be detected includes at least one of the following methods: In response to a screenshot trigger operation, a transparent overlay window is created on the target display and the target area is selected, and the image to be detected is cropped from the screen video stream; in response to a local image selection operation, the image to be detected is retrieved from the local storage area; Based on the conditions for obtaining equipment monitoring data, historical monitoring images are obtained from the IoT device platform as images to be detected.

3. The method according to claim 2, characterized in that, The step of acquiring historical surveillance images from the IoT device platform as images to be detected includes: The acquisition task is divided into a three-level task hierarchy, including batch tasks, device tasks, and download tasks; Separately establish a metadata acquisition queue and an image download queue, and schedule the device tasks and the download tasks in parallel according to a preset concurrency level; The execution status of the three-level tasks is persistently recorded, and the unfinished tasks are restored based on the persistent status after a task interruption is detected. The recovered download task is reset to a pending download state and re-added to the image download queue.

4. The method according to claim 1, characterized in that, The step of scheduling multiple image detection algorithms corresponding to the image to be detected to their respective target scheduling channels according to a preset algorithm scheduling strategy includes: Obtain the algorithm identifier corresponding to the currently enabled image detection algorithm, and query the preset algorithm scheduling strategy based on the algorithm identifier; When the algorithm identifier has a corresponding dedicated execution strategy in the preset algorithm scheduling strategy, the image detection algorithm is assigned to the second scheduling channel, and an independent asynchronous execution task is established for the image detection algorithm; When the algorithm identifier does not have a corresponding dedicated execution strategy, the image detection algorithm is assigned to the first scheduling channel, and the corresponding algorithm identifier is added to the general algorithm call set.

5. The method according to claim 4, characterized in that, The parallel execution of the multiple image detection algorithms based on the target scheduling channel, to output the detection results corresponding to each image detection algorithm, includes: For the first scheduling channel, obtain the algorithm identifier of each general image detection algorithm, and combine multiple algorithm identifiers to form a general algorithm call parameter; The general algorithm call parameters and the image to be detected are sent to the general algorithm service to obtain multiple detection results through a single batch call; For the second scheduling channel, the independent asynchronous execution task is executed through the dedicated execution strategy to obtain multiple detection results.

6. The method according to claim 1, characterized in that, The step of filtering the multiple detection results according to the confidence threshold corresponding to each image detection algorithm to obtain valid detection results includes: The multiple detection results are correlated according to the corresponding image detection algorithms, and the confidence threshold corresponding to each image detection algorithm is obtained; For each detection result, the confidence level corresponding to the detection annotation data in the detection result is compared with the confidence level threshold corresponding to the image detection algorithm; When the confidence level corresponding to the detection annotation data is greater than or equal to the confidence threshold, the detection annotation data is retained as valid detection annotation data; If no valid detection label data is found in the detection results, the detection result is discarded.

7. The method according to claim 6, characterized in that, The method further includes: Based on the detection annotation data in the valid detection results, the corresponding detection area is determined in the image to be detected, and visual annotation information corresponding to the detection area is generated; The image to be detected is annotated according to the visual annotation information to obtain an annotated image; The labeled image is compressed, and the compressed labeled image is associated with the valid detection results to generate a corresponding alarm record; The alarm records are persistently stored, and a preset review policy is used to determine whether the alarm records should enter the manual review or message push process.

8. The method according to claim 7, characterized in that, The alarm record includes at least the recognition time, image detection algorithm identifier, alarm category, alarm confidence level, review status, push status, and storage path of the labeled image; The review status is used to indicate whether the alarm record has been reviewed and the review result, and the push status is used to indicate whether the message corresponding to the alarm record has been sent.

9. The method according to claim 8, characterized in that, The step of determining whether the alarm record should enter the manual review or message push process according to the preset review strategy includes: Obtain the manual review switch status corresponding to the image detection algorithm that generated the alarm record; When the manual review switch is in the off state, the review status of the alarm record is set to reviewed, and a message push is triggered based on the alarm record; When the manual review switch is on, the review status of the alarm record is set to pending review, and the alarm record is added to the manual review queue. In response to the review operation of the alarm record in the manual review queue, the review status of the alarm record is updated according to the review result, and it is determined whether to trigger a message push based on the updated review status.

10. The method according to claim 9, characterized in that, The triggered message push includes: Retrieve the pre-configured alarm message receiver object; An alarm message is generated based on the identification time and alarm category in the alarm record, and a compressed labeled image associated with the alarm record is obtained; The alarm message and the compressed labeled image are sent to the alarm message receiving object; The push status of the alarm record is updated based on the message sending result, wherein the push status includes at least sent, failed to send, and not sent.