Intelligent ai image real-time processing method and system for edge computing scenarios

Abstract

This application relates to an intelligent AI image real-time processing method and system for edge computing scenarios. The method, applied to edge computing devices, includes: during system construction, establishing a linear processing pipeline and preparing a multi-level inference model; during runtime, sequentially executing image input, adaptive preprocessing based on feature analysis, multi-scale region confidence filtering, dynamically calling the matched inference model for target recognition based on the filtered confidence level, result fusion and annotation, and output along the pipeline. This method, through a linear pipeline + hardware / software decoupling architecture combined with a collaborative mechanism of preprocessing confidence-guided dynamic inference, simplifies the processing flow, reduces latency, and allows for flexible algorithm updates, meeting the stringent requirements of real-time image processing in edge scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent AI image processing technology, and in particular to an intelligent AI real-time image processing method and system for edge computing scenarios. Background Technology

[0002] Intelligent AI image target detection and recognition technology uses intelligent methods such as machine learning and deep learning to automatically locate and classify specific targets in images, and has been widely used in fields such as security monitoring, autonomous driving, and industrial inspection. Currently, mainstream implementation solutions typically rely on the powerful computing resources deployed in the cloud or dedicated computing centers to run complex detection models.

[0003] However, such solutions have significant drawbacks: First, the intelligent processing flow is complex, usually involving multi-stage and multi-model collaboration and frequent data exchange, resulting in high system latency and poor real-time performance; second, they heavily rely on centralized computing power, making the system expensive and difficult to adapt to edge computing scenarios with limited network conditions, limited hardware and software resources, or extremely high real-time requirements; finally, the algorithm model is tightly coupled with the hardware system, making model updates and iterations difficult and the system lacks flexibility.

[0004] Existing technologies, such as Chinese patents with publication numbers CN119516207A, CN120294825A, and CN120070906A, while each achieving AI image processing, dynamic adjustment of acquisition frequency, or multi-agent collaborative recognition, share a common drawback: either the process is complex and reliant on centralized computing, or it employs multi-level module coupling. None of these solutions address the fundamental contradiction of achieving high real-time performance and high-precision target detection and recognition on resource-constrained edge computing devices. Therefore, there is an urgent need for an intelligent AI image target detection and recognition solution with a streamlined, lightweight architecture, capable of integrated end-to-end real-time processing, and with flexibly updatable algorithms. Summary of the Invention

[0005] Therefore, it is necessary to provide a method and system for real-time intelligent AI image processing in edge computing scenarios to address the aforementioned technical problems.

[0006] A real-time intelligent AI image processing method for edge computing scenarios, the method being executed by a system deployed at the edge, the system having a linear processing pipeline and dynamically loadable intelligent algorithm modules, the method comprising:

[0007] Step S100, System Construction and Initialization:

[0008] In response to system startup, a linear processing pipeline from image input to image output is established and initialized through the control module, and at least two inference models with different computing power scales in the intelligent algorithm module are loaded and prepared; wherein, the invocation of the inference model is associated with the processing results and status of the image preprocessing module and the image inference module in the linear processing pipeline;

[0009] Step S200: Streamlined real-time image processing:

[0010] S210. Obtain raw image data through the image input module in the linear processing pipeline;

[0011] S220. The image preprocessing module performs adaptive preprocessing based on image feature analysis on the original image data, outputs a preprocessed image, and sends the preprocessed image to the image preprocessing module and the image fusion module in parallel.

[0012] S230. The image preprocessing module performs multi-scale target candidate region extraction and confidence-based screening on the preprocessed image, and outputs the preprocessed image sequence and the screening confidence of each region.

[0013] S240. Through the image reasoning module, based on the screening confidence of each region output by the image preprocessing module, the corresponding level of reasoning model provided by the intelligent algorithm module is dynamically matched and called to process the preprocessed image sequence and generate target detection and recognition results.

[0014] S250. The target detection and recognition results are fused and labeled onto the preprocessed image from the image preprocessing module through the image fusion module to generate a result image.

[0015] S260. Output the resulting image through the image output module.

[0016] In one embodiment, the image preprocessing module further includes: analyzing at least one feature of the input image, such as sharpness, contrast, and noise, in real time, and dynamically selecting and combining corresponding image enhancement, geometric correction, or denoising algorithms for processing based on the analysis results.

[0017] In one embodiment, the method further includes: scanning the preprocessed image using sliding windows of different sizes to generate candidate regions at multiple scales;

[0018] Calculate the initial confidence that each candidate region contains the target;

[0019] High-confidence candidate regions are selected based on dynamic thresholds, and non-maximum suppression is applied to remove redundancy.

[0020] Each filtered region is scaled to a predetermined size to form the preprocessed image sequence.

[0021] In one embodiment, the method further includes: a preset mapping relationship between screening confidence and inference model level, wherein the inference model level includes at least a high-precision level, a balanced level, and a fast level;

[0022] Based on the interval of the screening confidence level corresponding to each region in the preprocessed image sequence, the corresponding inference model level in the mapping relationship is selected.

[0023] The control module schedules the loading of the corresponding inference model instance at the inference model level in the intelligent algorithm module into the computing unit of the image inference module for inference calculation.

[0024] In one embodiment, the method further includes: real-time monitoring of the computational load of the image inference module;

[0025] The number of preprocessed image sequences simultaneously fed into the image inference module is dynamically adjusted based on the computational load to balance processing latency and throughput.

[0026] In one embodiment, the intelligent algorithm module is independent of the hardware platform of the image inference module, and the inference model is stored in the form of an instruction set or intermediate representation supported by the hardware platform, which is loaded, replaced or updated by the control module as needed.

[0027] An edge computing system for performing the above method includes:

[0028] Control processor;

[0029] A linear processing pipeline managed and scheduled by the control processor; the linear processing pipeline comprises components sequentially connected via data channels:

[0030] Image input interface unit, used to receive image signals;

[0031] Image preprocessing hardware unit for performing adaptive preprocessing based on image feature analysis;

[0032] Image preprocessing hardware unit for performing multi-scale target candidate region extraction and confidence-based filtering;

[0033] An image inference hardware unit is configured to receive and execute inference models of different levels dynamically specified by the control processor, based on the region filtering confidence level output by the image preprocessing hardware unit.

[0034] An image fusion hardware unit is used to fuse the recognition results with the image;

[0035] Image output interface unit, used to output processed image signals;

[0036] The memory stores the system program and an independent intelligent algorithm library, which contains at least two different levels of inference models that can be selected by the control processor and loaded into the image inference hardware unit.

[0037] The control processor is configured to: initialize the linear processing pipeline, and select a matching inference model from the intelligent algorithm library and load it into the image inference hardware unit based on the screening confidence level generated by the image preprocessing hardware unit.

[0038] In one embodiment, the image preprocessing hardware unit, the image preprocessing hardware unit, and the image fusion hardware unit are implemented by an FPGA or an ASIC; the image inference hardware unit is a processor or programmable logic device containing a dedicated AI computing core.

[0039] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0040] Step S100, System Construction and Initialization:

[0041] In response to system startup, a linear processing pipeline from image input to image output is established and initialized through the control module, and at least two inference models with different computing power scales in the intelligent algorithm module are loaded and prepared; wherein, the invocation of the inference model is associated with the processing results and status of the image preprocessing module and the image inference module in the linear processing pipeline;

[0042] Step S200: Streamlined real-time image processing:

[0043] S210. Obtain raw image data through the image input module in the linear processing pipeline;

[0044] S220. The image preprocessing module performs adaptive preprocessing based on image feature analysis on the original image data, outputs a preprocessed image, and sends the preprocessed image to the image preprocessing module and the image fusion module in parallel.

[0045] S230. The image preprocessing module performs multi-scale target candidate region extraction and confidence-based screening on the preprocessed image, and outputs the preprocessed image sequence and the screening confidence of each region.

[0046] S240. Through the image reasoning module, based on the screening confidence of each region output by the image preprocessing module, the corresponding level of reasoning model provided by the intelligent algorithm module is dynamically matched and called to process the preprocessed image sequence and generate target detection and recognition results.

[0047] S250. The target detection and recognition results are fused and labeled onto the preprocessed image from the image preprocessing module through the image fusion module to generate a result image.

[0048] S260. Output the resulting image through the image output module.

[0049] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0050] Step S100, System Construction and Initialization:

[0051] In response to system startup, a linear processing pipeline from image input to image output is established and initialized through the control module, and at least two inference models with different computing power scales in the intelligent algorithm module are loaded and prepared; wherein, the invocation of the inference model is associated with the processing results and status of the image preprocessing module and the image inference module in the linear processing pipeline;

[0052] Step S200: Streamlined real-time image processing:

[0053] S210. Obtain raw image data through the image input module in the linear processing pipeline;

[0054] S220. The image preprocessing module performs adaptive preprocessing based on image feature analysis on the original image data, outputs a preprocessed image, and sends the preprocessed image to the image preprocessing module and the image fusion module in parallel.

[0055] S230. The image preprocessing module performs multi-scale target candidate region extraction and confidence-based screening on the preprocessed image, and outputs the preprocessed image sequence and the screening confidence of each region.

[0056] S240. Through the image reasoning module, based on the screening confidence of each region output by the image preprocessing module, the corresponding level of reasoning model provided by the intelligent algorithm module is dynamically matched and called to process the preprocessed image sequence and generate target detection and recognition results.

[0057] S250. The target detection and recognition results are fused and labeled onto the preprocessed image from the image preprocessing module through the image fusion module to generate a result image.

[0058] S260. Output the resulting image through the image output module.

[0059] The aforementioned intelligent AI image real-time processing method and system for edge computing scenarios lays the foundation for low-latency, deterministic processing data pathways by constructing a highly integrated linear pipeline hardware architecture for edge computing. On this architecture, a creative decoupling design between intelligent algorithms and the hardware inference platform enables independent management and dynamic deployment of algorithm models. Furthermore, a dynamic multi-level inference scheduling intelligent collaborative mechanism based on pre-processing confidence screening is embedded within the linear pipeline, allowing the initial judgment in the pre-processing stage to guide the allocation of computing resources in subsequent inference stages in real time and with high accuracy. These three technical features are progressively layered and deeply coupled: the linear pipeline provides a stable and efficient operating platform for dynamic scheduling; the hardware-software decoupling provides flexible and variable algorithm resources for dynamic scheduling; and the intelligent scheduling mechanism fully utilizes the former two, achieving an optimal balance between processing real-time performance and recognition accuracy under the limited computing power constraints of edge devices. This system-level, integrated collaborative innovation fundamentally solves the core pain points of existing technologies, such as complex processes, poor real-time performance, high costs, and difficulty in adapting to rapid iteration needs. It provides efficient, flexible, and reliable intelligent image processing solutions for application scenarios with stringent real-time requirements, such as edge computing and embodied intelligence. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a real-time intelligent AI image processing method for edge computing scenarios in one embodiment.

[0061] Figure 2 This is a structural block diagram of an intelligent AI image real-time processing system for edge computing scenarios in one embodiment;

[0062] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0064] In one embodiment, such as Figure 1 As shown, an intelligent AI image target detection and recognition method is provided, which includes two main stages:

[0065] System construction and initialization phase

[0066] S100, System Construction and Initialization.

[0067] In response to system startup, a linear processing pipeline from image input to image output is established and initialized through the control module, and at least two inference models with different computing power scales in the intelligent algorithm module are loaded and prepared; wherein, the invocation of the inference model is associated with the processing results and status of the image preprocessing module and the image inference module in the linear processing pipeline.

[0068] After the system powers on, the control module performs initialization, which is divided into three stages: First, the control module and interface module initialize their hardware and load their software. Second, the control module sequentially initializes the hardware and software of each module: image input, preprocessing, pre-processing, inference, fusion, and output. Third, the control module establishes and verifies the data links between these modules, forming a stable linear processing pipeline. Simultaneously, the control module loads pre-stored inference models of different computing power scales from the intelligent algorithm module into a preparatory state, preparing for subsequent dynamic calls. This stage establishes the physical foundation of the linear pipeline and the resource state for algorithm decoupling preparation.

[0069] Pipeline image real-time processing stage

[0070] S200, pipelined real-time image processing. After initialization, the system enters the real-time processing loop.

[0071] S210 obtains raw image data through the image input module in the linear processing pipeline.

[0072] The image input module decodes the image signal from the image signal source, obtains the raw image data, and sends it into the pipeline.

[0073] S220 performs adaptive preprocessing based on image feature analysis on the original image data through the image preprocessing module, outputs the preprocessed image, and sends the preprocessed image to the image preprocessing module and the image fusion module in parallel.

[0074] S230 performs multi-scale target candidate region extraction and confidence-based screening on the preprocessed image through the image preprocessing module, and outputs the preprocessed image sequence and the screening confidence of each region.

[0075] S240, through the image inference module, dynamically matches and calls the corresponding level of inference model provided by the intelligent algorithm module based on the screening confidence of each region output by the image preprocessing module, processes the preprocessed image sequence, and generates target detection and recognition results.

[0076] S250 uses an image fusion module to fuse and label the target detection and recognition results onto a preprocessed image from an image preprocessing module, generating a result image.

[0077] S260 outputs the result image through the image output module.

[0078] The aforementioned intelligent AI image real-time processing method for edge computing scenarios lays the foundation for low-latency, deterministic processing data pathways by constructing a highly integrated linear pipeline hardware architecture for edge computing. On this architecture, a creative decoupling design between intelligent algorithms and the hardware inference platform enables independent management and dynamic deployment of algorithm models. Furthermore, a dynamic multi-level inference scheduling intelligent collaborative mechanism based on pre-processing confidence screening is embedded within the linear pipeline, allowing the initial judgment in the pre-processing stage to guide the allocation of computing resources in subsequent inference stages in real time and with high accuracy. These three technical features are progressively layered and deeply coupled: the linear pipeline provides a stable and efficient operating platform for dynamic scheduling; the hardware-software decoupling provides flexible and variable algorithm resources for dynamic scheduling; and the intelligent scheduling mechanism fully utilizes the former two, achieving an optimal balance between processing real-time performance and recognition accuracy under the limited computing power constraints of edge devices. This system-level, integrated collaborative innovation fundamentally solves the core pain points of existing technologies, such as complex processes, poor real-time performance, high costs, and difficulty in adapting to rapid iteration needs. It provides efficient, flexible, and reliable intelligent image processing solutions for application scenarios with stringent real-time requirements, such as edge computing and embodied intelligence.

[0079] In one embodiment, the image preprocessing module analyzes at least one feature of the input image, such as sharpness, contrast, and noise, in real time, and dynamically selects and combines corresponding image enhancement, geometric correction, or denoising algorithms for processing based on the analysis results.

[0080] Specifically, the adaptive preprocessing based on image feature analysis in step S220 includes four sub-stages:

[0081] Phase 1: Image Feature Analysis. The image preprocessing module receives the raw image data and extracts its multi-dimensional feature parameters in real time, including but not limited to sharpness, brightness distribution, contrast, noise index, color features, and texture complexity, and normalizes these parameter values.

[0082] Phase Two: Adaptive Preprocessing Algorithm Selection. The preprocessing module performs a weighted scoring of multi-dimensional feature parameter values ​​based on the feature analysis results. The scoring formula is as follows:

[0083]

[0084] Where S is the weighted score, It is the first i The weights of each feature ( =1), It is the first iThere are n normalized feature values, where n is the total number of features. Based on the different intervals in which the score value S falls, the optimal combination of algorithms is selected from the preprocessing algorithm library, including algorithms such as grayscale transformation, brightness enhancement, contrast enhancement, smoothing filtering, sharpening filtering, geometric transformation, and color correction.

[0085] The third stage: Adaptive preprocessing. The selected algorithm combination is run to perform geometric correction, denoising, and enhancement on the original image, achieving adaptive distortion correction, noise removal, and improved sharpness and contrast.

[0086] Fourth stage: Preprocessing result output. The processed image is output in parallel to the image preprocessing module and the image fusion module.

[0087] This embodiment achieves intelligent and adaptive preprocessing strategies through real-time image feature analysis and weighted scoring mechanisms. It can automatically take the most effective enhancement measures for input images of different qualities, providing a high-quality and standardized input data base for all subsequent processing stages, thereby improving the robustness and recognition accuracy of the entire system in complex environments from the source.

[0088] In one embodiment, the preprocessed image is scanned using sliding windows of different sizes to generate candidate regions at multiple scales; the initial confidence that each candidate region contains the target is calculated; high-confidence candidate regions are selected based on a dynamic threshold, and non-maximum suppression is applied to remove redundancy; each selected region is scaled to a predetermined size to form a preprocessed image sequence.

[0089] Specifically, step S230, which involves multi-scale target candidate region extraction and confidence-based filtering, includes four sub-stages:

[0090] Phase 1: Multi-scale candidate target region selection. The image preprocessing module receives the preprocessed image and scans it using a multi-scale sliding window approach to extract multiple candidate regions that can cover targets of different sizes. For each candidate region, a preliminary target region selection confidence score is calculated based on its texture, edge, color, and other features.

[0091] The second stage: High-probability region selection. An adaptive confidence threshold is set, and candidate regions with confidence levels higher than this threshold are selected as high-probability regions. Subsequently, a non-maximum suppression algorithm is applied to eliminate redundant regions with high overlap among these high-probability regions, generating a refined list of candidate regions, which are then sorted in descending order of confidence level.

[0092] The third stage: Dynamic scaling and region optimization. For each candidate region in the list, an optimal scaling ratio is dynamically calculated based on its own size and the input size required by the image inference module. Based on this scaling ratio, each region is precisely cropped and scaled to ensure that all output regions have a uniform image size, meeting the input requirements of the inference module.

[0093] Phase 4: Output of Preprocessing Results. The processed region image sequence is output to the image inference module, and metadata such as the coordinates of each region in the original image, the filtering confidence level, and the cropping and scaling parameters used are output to the control module.

[0094] This embodiment is crucial for ensuring the system's real-time performance. Multi-scale scanning ensures that targets of varying sizes are not overlooked, while confidence-based filtering and non-maximum suppression algorithms significantly eliminate background interference and repetitive regions, greatly reducing the amount of data required for complex AI inference. Dynamic scaling further optimizes the data format, reducing unnecessary computational overhead. This strategy of coarse screening followed by fine processing significantly reduces the load on the core inference process while ensuring no loss of critical information, laying the foundation for high-frame-rate processing on edge devices.

[0095] In one embodiment, a mapping relationship between screening confidence and inference model level is preset, and the inference model level includes at least high-precision level, balanced level and fast level; according to the interval of the screening confidence corresponding to each region in the preprocessed image sequence, the corresponding inference model level in the mapping relationship is selected; the control module schedules the loading of the inference model instance of the corresponding inference model level in the intelligent algorithm module into the computing unit of the image inference module for inference calculation.

[0096] Specifically, the process of dynamically matching and calling the corresponding level of inference model described in step S240, combined with the output filtering confidence, includes multi-level inference task scheduling:

[0097] The image inference module receives the pre-processed image sequence and its corresponding screening confidence level. The system pre-defines the mapping rules between confidence intervals and model levels. For example, high-confidence regions (confidence ≥ 0.8) are mapped to the high-precision level, calling the full model in the intelligent algorithm module for inference to pursue the highest recognition accuracy; medium-confidence regions (0.5 ≤ confidence < 0.8) are mapped to the balanced level, calling a model of medium complexity to balance accuracy and speed; and low-confidence regions (confidence < 0.5) are mapped to the fast level, calling a lightweight basic model to ensure processing speed.

[0098] Based on this mapping rule, the control module schedules the intelligent algorithm module to load the corresponding model onto the computing core (such as NPU or DSP core) of the image inference module for inference computation. For video streams, the system also establishes a cross-frame target association mechanism, creating tracking chains for consecutively appearing similar targets. For targets in a stable tracking state, subsequent frames can prioritize the use of fast-level models for verification, further saving computing resources.

[0099] This embodiment innovatively achieves on-demand allocation of computing power. It breaks away from the traditional approach of indiscriminately using a single complex model for the entire map or all regions, instead allocating computing resources differently based on the probability that a region contains an objective. Regions with a high probability of objective are allocated more computing power for precision, while regions with a low probability are allocated less computing power for speed. This refined resource scheduling strategy, linked to the results of pre-processing, ensures both high accuracy and high real-time performance in edge environments with limited computing power, demonstrating the depth of system-level optimization.

[0100] In one embodiment, the computational load of the image inference module is monitored in real time; based on the computational load, the number of preprocessed image sequences simultaneously sent to the image inference module for processing is dynamically adjusted to balance processing latency and throughput.

[0101] Step S240 also includes a load balancing mechanism: the image inference module monitors its own computational load in real time, such as CPU / NPU utilization. When the load exceeds a preset first threshold, it indicates that the system is under heavy processing pressure, so the number of preprocessed image sequence batches sent to the inference module at the same time is dynamically reduced to prioritize the processing latency of a single frame and ensure real-time performance. When the load is below a second threshold, it indicates that there is surplus computing power, so the number of batches is appropriately increased to improve the overall throughput of the system.

[0102] This embodiment enables the system to adapt to changes in the scenario, maintaining latency when resources are scarce and improving performance when resources are idle, thereby enhancing the system's dynamic adaptability and stability.

[0103] In one embodiment, the intelligent algorithm module is independent of the hardware platform of the image inference module. The inference model is stored in an instruction set or intermediate representation form supported by the hardware platform, and is loaded, replaced, or updated by the control module as needed. The intelligent algorithm module is completely independent of the hardware platform of the image inference module. It can be stored in the system's non-volatile memory or external storage devices. The inference model contained therein can be an optimized version compiled for different hardware platforms, such as ARM NPU, FPGA, and DSP. The control module can select and load the matching model file from the intelligent algorithm module according to the actual deployed hardware. Users can import new model files through the interface module to replace or update the content in the intelligent algorithm module without changing the hardware system or other software. Its advantages are: it achieves complete hardware and software decoupling. The training, optimization, and updating of the algorithm model are completely independent of the lifecycle of the hardware system. This greatly improves the system's flexibility and maintainability, enables rapid adaptation to the iterative upgrades of AI algorithms, and allows the same hardware system to be flexibly applied to different detection and recognition tasks. For example, it can be used for face recognition today and for vehicle detection tomorrow by loading a new model, significantly expanding the system's application scope and reducing long-term usage costs.

[0104] In one embodiment, an edge computing system for performing the above-described method is provided. It includes:

[0105] Control processor;

[0106] A linear processing pipeline managed and scheduled by the control processor; the linear processing pipeline comprises components sequentially connected via data channels:

[0107] Image input interface unit, used to receive image signals;

[0108] Image preprocessing hardware unit for performing adaptive preprocessing based on image feature analysis;

[0109] Image preprocessing hardware unit for performing multi-scale target candidate region extraction and confidence-based filtering;

[0110] An image inference hardware unit is configured to receive and execute inference models of different levels dynamically specified by the control processor, based on the region filtering confidence level output by the image preprocessing hardware unit.

[0111] An image fusion hardware unit is used to fuse the recognition results with the image;

[0112] Image output interface unit, used to output processed image signals;

[0113] The memory stores the system program and an independent intelligent algorithm library, which contains at least two different levels of inference models that can be selected by the control processor and loaded into the image inference hardware unit.

[0114] The control processor is configured to: initialize the linear processing pipeline, and select a matching inference model from the intelligent algorithm library and load it into the image inference hardware unit based on the screening confidence level generated by the image preprocessing hardware unit.

[0115] Specifically, such as Figure 2 As shown, the system includes: a control module 170 as the control core, which can be implemented by a microprocessor such as ARM / RISC-V; a linear processing pipeline managed by the control module 170, which is essentially a hardware processing chain consisting of an image input module 110, an image preprocessing module 120, an image preprocessing module 130, an image inference module 140, an image fusion module 150, and an image output module 160 connected sequentially via a high-speed data bus or a dedicated data channel; an independent intelligent algorithm module 190, which can reside in a specific partition of memory; and an interface module 180 providing human-computer interaction. The responsibilities of the control module 170 include pipeline initialization, process control, and the most critical function—intelligently selecting the corresponding level of model from the intelligent algorithm module 190 based on the screening confidence level reported in real time by the image preprocessing module 130, and loading it into the image inference module 140.

[0116] In one embodiment, the image preprocessing hardware unit, the image preprocessing hardware unit, and the image fusion hardware unit are implemented by an FPGA or an ASIC; the image inference hardware unit is a processor or programmable logic device containing a dedicated AI computing core.

[0117] Specifically, the image preprocessing module 120, image preprocessing module 130, and image fusion module 150 in the system can be implemented using FPGA (Field-Programmable Gate Array) or ASIC (Application-Specific Integrated Circuit) chips. These devices excel at parallel, pipelined fixed logic operations, enabling them to efficiently perform deterministic operations such as image format conversion, scaling, filtering, and overlay with extremely low latency and power consumption. The image inference module 140 can be implemented using a SoC (such as the HiSilicon series or Fudan Microelectronics Qinglong series) with an integrated dedicated AI computing core (NPU), a DSP with parallel computing capabilities, or even an FPGA with a built-in AI IP core, to efficiently execute neural network inference. The advantages are: through a heterogeneous computing architecture, different computing tasks are allocated to the hardware units best suited to them. Fixed pipelined operations are handled by the FPGA / ASIC, freeing up the main processor load; flexible and computationally intensive AI inference is accelerated by dedicated AI hardware. This hardware-software co-design is the optimal technical path to meet the stringent performance, power consumption, and cost requirements of edge computing scenarios.

[0118] Those skilled in the art will understand that Figure 1 The steps in the flowchart are not necessarily executed in the strict order indicated by the arrows; some steps may be executed in parallel or alternately. At least some of these steps may include multiple sub-steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and their execution order may also be different from the order shown in the diagram.

[0119] In one embodiment, a computer device is provided, which may be an edge server or an embedded device on which the system of this application is deployed, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, input device, and display device connected via a system bus. The processor provides computing and control capabilities, equivalent to the aforementioned control module 170, and may integrate or connect to other processing units to act as an image inference module, etc. The memory of the computer device includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs (implementing control logic), and intelligent algorithm modules 190, etc. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface of the computer device is used for network communication with external devices (such as cameras, central servers). When the computer program is executed by the processor, it implements the aforementioned intelligent AI image target detection and recognition method.

[0120] In one embodiment, the steps implemented by the processor when executing a computer program include the method steps described in the above embodiments.

[0121] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0122] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0123] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0124] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A real-time intelligent AI image processing method for edge computing scenarios, characterized in that, The method is executed by a system deployed at the edge, the system having a linear processing pipeline and dynamically loadable intelligent algorithm modules, the method comprising: Step S100, System Construction and Initialization: In response to system startup, a linear processing pipeline from image input to image output is established and initialized through the control module, and at least two inference models with different computing power scales in the intelligent algorithm module are loaded and prepared; wherein, the invocation of the inference model is associated with the processing results and status of the image preprocessing module and the image inference module in the linear processing pipeline; Step S200: Streamlined real-time image processing: S210. Obtain raw image data through the image input module in the linear processing pipeline; S220. The image preprocessing module performs adaptive preprocessing based on image feature analysis on the original image data, outputs a preprocessed image, and sends the preprocessed image to the image preprocessing module and the image fusion module in parallel. S230. The image preprocessing module performs multi-scale target candidate region extraction and confidence-based screening on the preprocessed image, and outputs the preprocessed image sequence and the screening confidence of each region. S240. Through the image reasoning module, based on the screening confidence of each region output by the image preprocessing module, the corresponding level of reasoning model provided by the intelligent algorithm module is dynamically matched and called to process the preprocessed image sequence and generate target detection and recognition results. S250. The target detection and recognition results are fused and labeled onto the preprocessed image from the image preprocessing module through the image fusion module to generate a result image. S260. Output the resulting image through the image output module.

2. The method according to claim 1, characterized in that, The adaptive preprocessing based on image feature analysis in step S220 specifically includes: The image preprocessing module analyzes at least one feature of the input image, such as sharpness, contrast, and noise, in real time, and dynamically selects and combines corresponding image enhancement, geometric correction, or denoising algorithms based on the analysis results.

3. The method according to claim 1, characterized in that, The multi-scale target candidate region extraction and confidence-based screening in step S230 include: The preprocessed image is scanned using sliding windows of different sizes to generate candidate regions at multiple scales; Calculate the initial confidence that each candidate region contains the target; High-confidence candidate regions are selected based on dynamic thresholds, and non-maximum suppression is applied to remove redundancy. Each filtered region is scaled to a predetermined size to form the preprocessed image sequence.

4. The method according to claim 1, characterized in that, The dynamic matching and invocation of the corresponding level of inference model in step S240 includes: A preset mapping relationship is established between the screening confidence level and the inference model level, wherein the inference model level includes at least high-precision level, balanced level and fast level; Based on the interval of the screening confidence level corresponding to each region in the preprocessed image sequence, the corresponding inference model level in the mapping relationship is selected. The control module schedules the loading of the corresponding inference model instance at the inference model level in the intelligent algorithm module into the computing unit of the image inference module for inference calculation.

5. The method according to claim 4, characterized in that, Step S240 further includes: Real-time monitoring of the computational load of the image inference module; The number of preprocessed image sequences simultaneously fed into the image inference module is dynamically adjusted based on the computational load to balance processing latency and throughput.

6. The method according to claim 1, characterized in that, The intelligent algorithm module is independent of the hardware platform of the image inference module. The inference model is stored in the form of an instruction set or intermediate representation supported by the hardware platform, and is loaded, replaced or updated by the control module as needed.

7. An edge computing system for performing the method according to any one of claims 1 to 6, characterized in that, include: Control processor; A linear processing pipeline managed and scheduled by the control processor; the linear processing pipeline comprises components sequentially connected via data channels: Image input interface unit, used to receive image signals; Image preprocessing hardware unit for performing adaptive preprocessing based on image feature analysis; Image preprocessing hardware unit for performing multi-scale target candidate region extraction and confidence-based filtering; An image inference hardware unit is configured to receive and execute inference models of different levels dynamically specified by the control processor, based on the region filtering confidence level output by the image preprocessing hardware unit. An image fusion hardware unit is used to fuse the recognition results with the image; Image output interface unit, used to output processed image signals; The memory stores the system program and an independent intelligent algorithm library, which contains at least two different levels of inference models that can be selected by the control processor and loaded into the image inference hardware unit. The control processor is configured to: initialize the linear processing pipeline, and select a matching inference model from the intelligent algorithm library and load it into the image inference hardware unit based on the screening confidence level generated by the image preprocessing hardware unit.

8. The edge computing system according to claim 7, characterized in that, The image preprocessing hardware unit, the image preprocessing hardware unit, and the image fusion hardware unit are implemented by FPGA or ASIC; the image inference hardware unit is a processor or programmable logic device containing a dedicated AI computing core.

Images