Microscopic image real-time processing method based on edge calculation

Abstract

The invention discloses a microscopic image real-time processing method based on edge calculation. The microscopic image real-time processing method comprises the following steps: S1, acquiring a microscopic image frame sequence; s2, dividing the microscopic image frame sequence into image blocks and calculating local texture complexity; s3, classifying the image blocks into flat areas and high-detail areas; s4, carrying out mean filtering on the flat region, and carrying out denoising on the high-detail region by adopting a fast non-local mean algorithm; s5, fusing the two types of de-noised image blocks to generate a preliminary de-noised image frame; s6, an image pyramid structure is constructed, global denoising is carried out under low resolution, and detail enhancement is carried out under high resolution; s7, multi-scale results are fused, final de-noised image frames are generated, and inter-frame consistency detection is carried out; s8, dynamically adjusting the processing strategy according to the equipment resource state; and S9, carrying out structured coding on the de-noised image frame, and outputting the de-noised image frame to a feature recognition module. According to the method, region classification and resource scheduling are fused, and real-time de-noising and enhancement processing of the edge end of the microscopic image is realized.

Description

Technical Field

[0001] This invention relates to the field of image processing and edge computing technology, and in particular to a real-time microscopic image processing method based on edge computing. Background Technology

[0002] In applications such as medical imaging, biological research, and materials testing, microscopic images serve as crucial foundational data for microscopic structure analysis, and their processing quality and efficiency directly impact the accuracy of subsequent analysis and diagnosis. Current technologies typically employ a centralized architecture where images are acquired and then transmitted to a central server for processing. This involves transmitting large numbers of high-resolution images over a network to the cloud or servers for denoising, enhancement, and reconstruction. While this approach offers certain advantages in computing power, it suffers from significant limitations in edge computing environments.

[0003] First, existing solutions rely excessively on network transmission, leading to problems such as high latency, heavy bandwidth pressure, and data loss during image data upload, making it difficult to meet the demands of scenarios with high real-time requirements. Especially in resource-constrained or network-unstable environments, image processing is prone to interruption or lag, severely impacting processing efficiency and system stability. Second, traditional microscopic image denoising methods mostly employ a uniform algorithm flow, ignoring regional differences in image content and failing to differentiate processing based on the complexity of local image textures. This approach easily leads to loss of image details or waste of computational resources when dealing with microscopic images with significant structural changes, making it difficult to simultaneously achieve both clarity and computational efficiency. Furthermore, existing technologies generally lack the ability to perceive the resource status of edge devices, failing to dynamically adjust the processing flow, image block partitioning granularity, or algorithm complexity based on system load factors such as processor utilization and memory usage. This results in a significant decrease in system efficiency, or even processing failure, when edge device resources are limited.

[0004] To address the above problems, this invention proposes a real-time microscopic image processing method based on edge computing. By deploying an integrated acquisition and processing terminal at edge nodes, local image preprocessing, structure recognition, and intelligent scheduling are achieved. This method introduces a texture complexity analysis mechanism to select appropriate denoising strategies based on regional characteristics. Simultaneously, it constructs a multi-scale image pyramid for differentiated enhancement processing and combines it with a resource-aware task scheduling mechanism to achieve dynamic adaptation of image processing parameters and processes, effectively improving the stability, real-time performance, and processing accuracy of microscopic image processing in edge environments.

[0005] Therefore, how to provide a real-time microscopic image processing method based on edge computing is a challenge for those skilled in the art. The problems that staff urgently need to solve. Summary of the Invention

[0006] One objective of this invention is to propose a real-time microscopic image processing method based on edge computing. This invention integrates local texture analysis, regional differential denoising, multi-scale image enhancement, and edge resource awareness scheduling strategy to realize real-time acquisition, adaptive processing, and structured output of microscopic images at edge computing nodes. This effectively improves image clarity and system response speed, and solves problems such as high latency, severe loss of detail, and poor resource adaptation in traditional methods. It has good real-time performance and engineering adaptability.

[0007] A real-time microscopic image processing method based on edge computing according to an embodiment of the present invention includes the following steps: S1. Deploy a microscopic image acquisition and processing terminal at the edge computing node to acquire the sequence of microscopic image frames to be processed, and perform local caching and management of the image data at the edge. S2. Divide the acquired microscopic image frame sequence into multiple fixed-size image blocks, and calculate the local texture complexity index for each image block. S3. Classify the image blocks according to the local texture complexity index, dividing them into flat region image blocks and high-detail region image blocks. S4. For flat region image blocks, use a fixed-window-size mean filtering method for image denoising, outputting flat region denoised image blocks. For high-detail region image blocks, use a fast non-local mean algorithm based on neighborhood matching for image denoising, outputting high-detail region denoised image blocks. S5. Combine the flat region denoised image blocks with the high-detail region denoised image blocks. S6. Block-level fusion of denoised image blocks in detail regions to construct a complete preliminary denoised image frame; S7. Constructing a multi-scale image pyramid structure for the microscopic image based on the preliminary denoised image frame, performing global denoising at the low-resolution scale and local enhancement of high-frequency detail regions at the original resolution scale; S8. Fusing the multi-scale processing results to generate a complete denoised image frame, and performing inter-frame consistency detection on the image frame sequence; S9. Dynamically adjusting the image block partitioning granularity, denoising algorithm parameters, and execution order according to the current computing resource status of the edge device to achieve a resource-aware task scheduling strategy; S10. S2. Structure-encoding the denoised image frame and outputting it to the image feature recognition module at the edge node, while configuring a transmission strategy to upload the processing results to the remote diagnostic system.

[0008] Optionally, S1 specifically includes: S11. Deploy a microscopic image acquisition and processing terminal on the edge computing node. The microscopic image acquisition and processing terminal includes an image sensor, a data acquisition module, and an edge computing processing unit. S12. Acquire a continuous sequence of microscopic image frames through the image sensor, and the data acquisition module assigns the image frames numbers and timestamps. S13. Standardize the format of the microscopic image frames in the edge computing processing unit to unify the number of channels and resolution of the images. S14. Store the standardized microscopic image frames in the local cache area of ​​the edge node for temporary storage and read / write management. S15. Generate microscopic image frame cache index information and a timestamp sequence, and register them in the edge node task scheduling management module for subsequent image processing steps.

[0009] Optionally, S2 specifically includes: S21. In the edge computing node, call the image preprocessing module to read the locally cached microscopic image frame data, and divide each frame image into multiple uniform, non-overlapping rectangular image blocks according to preset image block size parameters, and mark the row and column numbers of the image blocks and their position information in the whole frame image; S22. Perform edge-preserving filtering on each image block to enhance local texture features, and extract the pixel gray-level matrix from the filtered image block; S23. Calculate the local statistical features of the image block based on the pixel gray-level matrix, with basic parameters including gray-level mean, standard deviation, range, and gray-level histogram distribution density; S24. Construct the gray-level co-occurrence matrix of the image block and extract texture. The descriptive metrics include energy, contrast, entropy, and correlation, characterizing the detail complexity of image patches; S25, calculate the gradient magnitude map and gradient direction map in the image patch, and statistically analyze their mean, variance, and gradient direction concentration as auxiliary descriptive information for the image texture change trend; S26, fuse the local statistical features, texture descriptive metrics, and gradient features according to weights to generate a unified local texture complexity index for the image patch, and bind the unified local texture complexity index value after normalization with the spatial location index of the image patch to construct the texture complexity distribution map of the image frame; S27, output the texture complexity distribution map as the feature data structure of the image frame for subsequent image patch classification steps to call and process.

[0010] Optionally, S3 specifically includes: S31. Load the texture complexity distribution map corresponding to the microscopic image frame into the edge computing node, and sequentially traverse the texture complexity index values ​​of all image blocks; S32. According to the preset complexity classification threshold, perform interval discrimination on the texture complexity index of each image block. If the index value is less than the threshold, mark the image block as a flat region image block; if the index value is greater than or equal to the threshold, mark the image block as a high-detail region image block; S33. Add the classified image blocks to the flat region image block set and the high-detail region image block set respectively, and record the classification label, spatial location information and original image frame number of each image block; S34. Generate a region classification mapping map of the microscopic image frame. The region classification mapping map is a two-dimensional matrix structure with the same size as the microscopic image frame. The value corresponding to each position represents the classification label of the image block to which that position belongs; S35. Cache the flat region image block set, the high-detail region image block set and the region classification mapping map into the edge node image processing module for subsequent calling and scheduling of different region differential processing operations.

[0011] Optionally, S4 specifically includes: S41. Extract the index information of flat region image blocks from the region classification map and load the corresponding flat region image block data into the edge computing node; S42. For each flat region image block, set a sliding filter with a fixed window size, traverse each pixel in a row-column manner, extract a pixel set in its neighborhood, and calculate the mean of the gray values ​​of all pixels in the pixel set to generate the replacement value of the current pixel, completing the mean filtering operation of the entire image block; S43. Mark the processed image block as a flat region denoising image block and record its image frame number, image block index and processing timestamp to generate a flat region image block denoising dataset; S44. Extract the index information of high detail region image blocks from the region classification map and load the original data of the corresponding high detail region image blocks. According to the following steps: S45. For each high-detail region image block, determine its neighborhood range in the original image frame, slide the image block within the neighborhood according to a preset step size, and extract multiple neighborhood reference blocks; S46. Among all extracted reference blocks, use the fast nonlocal mean algorithm framework to match the pixel values ​​inside the image block with the pixels of the reference blocks in its neighborhood, and select the local optimal matching region according to the preset acceleration strategy; S47. After completing the neighborhood matching of all pixel positions, form the denoising result of the image block, generate the high-detail region denoised image block, and record its image frame number, index position and processing parameter configuration; S48. Output all flat region denoised image blocks and high-detail region denoised image blocks to the image block fusion module for unified fusion operation in subsequent image reconstruction processing steps.

[0012] Optionally, S5 specifically includes: S51. Load the datasets of flat region denoised image blocks and high detail region denoised image blocks into the image fusion module of the edge computing node, and establish a unified image frame fusion index table; S52. Based on the position information of each denoised image block in the original image frame, map the flat region denoised image blocks and high detail region denoised image blocks to their corresponding spatial regions respectively; S53. Construct a complete image block arrangement matrix in the two-dimensional coordinate system of the image frame to ensure that all image blocks are positioned and filled according to their original spatial order; S54. Perform edge smoothing processing on the boundary regions of the image blocks to eliminate possible grayscale abrupt changes or block effects at the splicing of adjacent image blocks; S55. Reconstruct all the positioned image block data to generate a preliminary denoised image frame with the corresponding image frame number, and cache the image frame in the local storage area of ​​the edge node; S56. Generate metadata information of the preliminary denoised image frame, including frame index number, construction time, image block fusion status and processing completion flag, and register it in the image frame processing record table.

[0013] Optionally, S6 specifically includes: S61. Load the initial denoised image frame into the edge computing node and call the multi-scale image processing module to perform scale decomposition to construct the image pyramid structure of the microscopic image; S62. Downsample the initial denoised image frame layer by layer according to the preset scale level to generate a multi-scale image set including the original resolution layer and multiple low-resolution layers; S63. Perform global denoising processing on the low-resolution layer image, using an image filtering method with unified parameter configuration to enhance the overall structural consistency and further eliminate large-scale background noise; S64. Load the texture complexity distribution map and region classification of the image frame into the original resolution layer. S65. For the high-frequency detail regions, edge-preserving filtering or local contrast enhancement methods are used to enhance the regions, highlighting the microstructural details and edge contours in the microscopic image. S66. The low-resolution layer processing results and the original resolution layer enhancement results are scale-aligned and interpolated to construct a complete multi-scale image fusion framework. S67. The fused multi-scale image frames are generated, and the processing parameters, image feature indicators, and processing time information of each scale image layer are recorded as the basic data structure for subsequent consistency detection and image output.

[0014] Optionally, S7 specifically includes: S71. Load the fused image frame data output by the multi-scale image processing module into the edge computing node, and call the image reconstruction module to reconstruct and back-project the image results at each scale to generate a complete denoised image frame with uniform spatial resolution; S72. Perform pixel-level integrity verification on the reconstructed complete denoised image frame, check whether the image region filling is complete and whether the boundary stitching is continuous, and perform block edge smoothing to form a final image frame with complete structure and visual consistency; S73. Extract the frame number, timestamp, position information and processing parameters of the current final image frame, and register them in the image frame sequence index table for sequence management and consistency comparison; S74. Call the inter-frame consistency detection module, load the current image frame and its previous frame... For each frame of image data, pixel-level difference calculation, structural similarity comparison, and edge contour overlap analysis are performed on the corresponding regions. S75. Based on the inter-frame image change rate, texture distribution change range, and inter-frame difference distribution map, a consistency score index between the current frame and the previous frame is determined. S76. The consistency score is jointly judged with the image frame content stability parameter to output whether the current frame meets the inter-frame continuity standard. S77. For image frames with insufficient continuity, the key region content from the previous frame result is loaded using a local backoff strategy for compensation and reconstruction, and the availability status of the current image frame is marked. S78. Image frames that have completed inter-frame consistency detection are added to the effective frame buffer sequence for subsequent image analysis and structured coding processing steps.

[0015] Optionally, S8 specifically includes: S81. In the edge computing node, call the system resource monitoring module to collect real-time resource indicators of the current device, including processor utilization, memory usage, storage read / write speed, and network communication status. S82. Normalize the collected resource status information and compare it with the preset resource threshold configuration table to determine whether the current computing resources are in a high-availability, medium-load, or resource-scarce state. S83. When it is determined to be in a high-availability state, use a smaller image patch granularity, enable the full version of the image denoising algorithm processing path, and maintain the standard task execution order. S84. When it is determined to be in a medium-load state, use a medium-sized image patch granularity, call the standard configured image denoising algorithm parameters, and merge the processing tasks of non-critical areas according to task priority. S85. When a resource shortage is detected, a larger image block size is used for granular division, the denoising algorithm is adjusted to a simplified version, the enhancement processing flow for non-critical image areas is shut down, and the execution order in the current task queue is rearranged. S86. Based on the current resource status change trend, the parameter configuration table of the image processing module is dynamically updated, and the adjusted task scheduling strategy is synchronized to each processing sub-module in real time. S87. The time, parameters, and execution results of each resource status switch and scheduling strategy change are recorded to form a resource-aware historical log for subsequent system operation optimization and strategy evaluation. S88. The processing rhythm between the image acquisition module, image denoising module, and image fusion module is coordinated through the scheduling controller to ensure the continuity and stability of the image processing flow under resource change conditions.

[0016] Optionally, S9 specifically includes: S91. In the edge computing node, call the image encoding module to load the denoised image frame and its associated frame number, spatial index, timestamp, and processing parameter information; S92. Integrate the image frame data and metadata in a structured manner to generate a structured representation format for the image frame, and encode the image frame into an intermediate representation file that is easy to identify later according to a preset format standard; S93. Embed image processing path identifiers, region processing labels, and consistency detection results during the structured encoding process to ensure that the image frame has traceability and reconstructability; S94. Write the structured encoded image frame into the image processing output buffer of the edge node and call the feature recognition module interface to perform... Data push; S95, In the image feature recognition module, perform feature parsing tasks on the structured image frame, including key area localization, target edge extraction, and texture pattern analysis, and generate preliminary recognition results; S96, Call the edge transmission control module to generate a transmission control strategy based on the current network bandwidth status, task priority, and data importance level; S97, Encapsulate the preliminary recognition results and the original structured image frame together into a diagnostic data packet, upload it to the remote diagnostic system, and complete the remote synchronization of local processing results; S98, Record the timestamp, data packet number, and transmission status of image output and remote transmission process, generate a diagnostic data transmission log, and store it in the edge node system recording module.

[0017] The beneficial effects of this invention are: The real-time microscopic image processing method based on edge computing provided by this invention has significant advantages over existing technologies. Firstly, by deploying microscopic image acquisition and processing terminals on edge computing nodes, this method enables local acquisition, caching, and processing of image data. This avoids the bandwidth burden of uploading large numbers of high-resolution images to remote servers, significantly reducing transmission latency and network dependence, and improving the real-time performance and stability of the overall system. It is particularly suitable for applications with high timeliness requirements, such as on-site inspection and telemedicine.

[0018] Secondly, this invention introduces a region classification mechanism based on local texture complexity, which divides the image into flat regions and high-detail regions, and adopts differentiated denoising strategies, matching lightweight mean filtering algorithm and fast nonlocal mean algorithm respectively. This allows the image to effectively preserve structural detail information while ensuring processing efficiency, solving the problems of single processing strategy and severe detail loss in traditional methods, thereby significantly improving the quality and clarity of microscopic images.

[0019] Furthermore, this invention constructs a multi-scale image pyramid structure, performing global noise suppression at the low-resolution scale and detail enhancement processing on key regions at the original resolution scale. This achieves joint optimization at different scales, further enhancing the image's sense of hierarchy and recognition performance. Simultaneously, through a resource monitoring and scheduling mechanism, the system can perceive the computing status of edge devices in real time and dynamically adjust the image block partitioning granularity, algorithm parameters, and task execution order, ensuring stable operation even under limited computing power and improving the system's adaptability and resource utilization efficiency.

[0020] In summary, this invention not only achieves efficient real-time processing of microscopic images through technical means, but also constructs a deployable, schedulable, and adaptable image processing solution under an edge computing architecture. It has good engineering practicality and promotional value, and is significantly superior to existing centralized microscopic image processing systems. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0022] Figure 1 This is an overall flowchart of a real-time microscopic image processing method based on edge computing proposed in this invention. Figure 2 This is a flowchart of image block classification and differential denoising for flat and high-detail regions in a real-time microscopic image processing method based on edge computing proposed in this invention. Figure 3 This is a schematic diagram of the edge resource perception and dynamic scheduling mechanism of image processing parameters in a real-time microscopic image processing method based on edge computing proposed in this invention. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0024] refer to Figure 1-3 A real-time microscopic image processing method based on edge computing includes the following steps: S1. Deploy a microscopic image acquisition and processing terminal at the edge computing node to acquire the sequence of microscopic image frames to be processed, and perform local caching and management of the image data at the edge. S2. Divide the acquired microscopic image frame sequence into multiple fixed-size image blocks, and calculate the local texture complexity index for each image block. S3. Classify the image blocks according to the local texture complexity index, dividing them into flat region image blocks and high-detail region image blocks. S4. For flat region image blocks, use a fixed-window-size mean filtering method for image denoising, outputting flat region denoised image blocks. For high-detail region image blocks, use a fast non-local mean algorithm based on neighborhood matching for image denoising, outputting high-detail region denoised image blocks. S5. Combine the flat region denoised image blocks with the high-detail region denoised image blocks. S6. Block-level fusion of denoised image blocks in detail regions to construct a complete preliminary denoised image frame; S7. Constructing a multi-scale image pyramid structure for the microscopic image based on the preliminary denoised image frame, performing global denoising at the low-resolution scale and local enhancement of high-frequency detail regions at the original resolution scale; S8. Fusing the multi-scale processing results to generate a complete denoised image frame, and performing inter-frame consistency detection on the image frame sequence; S9. Dynamically adjusting the image block partitioning granularity, denoising algorithm parameters, and execution order according to the current computing resource status of the edge device to achieve a resource-aware task scheduling strategy; S10. S2. Structure-encoding the denoised image frame and outputting it to the image feature recognition module at the edge node, while configuring a transmission strategy to upload the processing results to the remote diagnostic system.

[0025] This invention constructs a complete edge computing-based microscopic image processing chain, enabling the entire process of image acquisition, preprocessing, region classification, differential denoising, multi-scale enhancement, and structured output to be completed locally in a closed loop at the edge node. This significantly reduces the network latency and computational bottleneck problems caused by the reliance on cloud processing in traditional systems, and improves the response speed, edge autonomy, and system stability of microscopic image processing. It is particularly suitable for deployment scenarios with high time requirements and complex network conditions.

[0026] In this embodiment, S1 specifically includes: S11. Deploy a microscopic image acquisition and processing terminal on the edge computing node. The microscopic image acquisition and processing terminal includes an image sensor, a data acquisition module, and an edge computing processing unit. S12. Acquire a continuous sequence of microscopic image frames through the image sensor, and the data acquisition module assigns the image frames numbers and timestamps. S13. Standardize the format of the microscopic image frames in the edge computing processing unit to unify the number of channels and resolution of the images. S14. Store the standardized microscopic image frames in the local cache area of ​​the edge node for temporary storage and read / write management. S15. Generate microscopic image frame cache index information and a timestamp sequence, and register them in the edge node task scheduling management module for subsequent image processing steps.

[0027] By building an integrated module for image acquisition, numbering, format standardization, and index cache management at edge nodes, the system can complete data format unification and access structure optimization at the source stage, providing efficient and continuous input support for subsequent image block partitioning and texture analysis. This reduces preprocessing redundancy caused by format inconsistencies between traditional acquisition devices and processing terminals, and improves the efficiency of image data processing and flow.

[0028] In this embodiment, S2 specifically includes: S21. In the edge computing node, call the image preprocessing module to read the locally cached microscopic image frame data, and divide each frame image into multiple uniform, non-overlapping rectangular image blocks according to preset image block size parameters, and mark the row and column numbers of the image blocks and their position information in the whole frame image; S22. Perform edge-preserving filtering on each image block to enhance local texture features, and extract the pixel gray-level matrix from the filtered image block; S23. Calculate the local statistical features of the image block based on the pixel gray-level matrix, with basic parameters including gray-level mean, standard deviation, range, and gray-level histogram distribution density; S24. Construct the gray-level co-occurrence matrix of the image block and extract texture. The descriptive metrics include energy, contrast, entropy, and correlation, characterizing the detail complexity of image patches; S25, calculate the gradient magnitude map and gradient direction map in the image patch, and statistically analyze their mean, variance, and gradient direction concentration as auxiliary descriptive information for the image texture change trend; S26, fuse the local statistical features, texture descriptive metrics, and gradient features according to weights to generate a unified local texture complexity index for the image patch, and bind the unified local texture complexity index value after normalization with the spatial location index of the image patch to construct the texture complexity distribution map of the image frame; S27, output the texture complexity distribution map as the feature data structure of the image frame for subsequent image patch classification steps to call and process.

[0029] This paper introduces three complementary image descriptor factors: local statistical features, texture co-occurrence matrix index, and gradient direction characteristics. By normalizing and fusing these factors, a unified texture complexity index is generated, enabling fine modeling of local structural changes in images. This effectively distinguishes between information-dense and texture-single regions in images, providing a high-precision decision-making basis for subsequent differential processing and solving the problem of insufficient image content representation in existing methods.

[0030] In this embodiment, S3 specifically includes: S31. Load the texture complexity distribution map corresponding to the microscopic image frame into the edge computing node, and sequentially traverse the texture complexity index values ​​of all image blocks; S32. According to the preset complexity classification threshold, perform interval discrimination on the texture complexity index of each image block. If the index value is less than the threshold, mark the image block as a flat region image block; if the index value is greater than or equal to the threshold, mark the image block as a high-detail region image block; S33. Add the classified image blocks to the flat region image block set and the high-detail region image block set respectively, and record the classification label, spatial location information and original image frame number of each image block; S34. Generate a region classification mapping map of the microscopic image frame. The region classification mapping map is a two-dimensional matrix structure with the same size as the microscopic image frame. The value corresponding to each position represents the classification label of the image block to which that position belongs; S35. Cache the flat region image block set, the high-detail region image block set and the region classification mapping map into the edge node image processing module for subsequent calling and scheduling of different region differential processing operations.

[0031] Automated classification is achieved by setting a threshold based on the complexity index of image patches, and a region classification mapping map is constructed for image patch spatial mapping and structure scheduling. This not only improves the accuracy of image patch classification, but also supports subsequent modules to load different processing paths and algorithm parameters according to region type, thus optimizing the overall system performance in terms of computing resource allocation and time overhead.

[0032] In this embodiment, S4 specifically includes: S41. Extract the index information of flat region image blocks from the region classification map and load the corresponding flat region image block data into the edge computing node; S42. For each flat region image block, set a sliding filter with a fixed window size, traverse each pixel in a row-column manner, extract a pixel set in its neighborhood, and calculate the mean of the gray values ​​of all pixels in the pixel set to generate the replacement value of the current pixel, completing the mean filtering operation of the entire image block; S43. Mark the processed image block as a flat region denoising image block and record its image frame number, image block index and processing timestamp to generate a flat region image block denoising dataset; S44. Extract the index information of high detail region image blocks from the region classification map and load the original data of the corresponding high detail region image blocks. According to the following steps: S45. For each high-detail region image block, determine its neighborhood range in the original image frame, slide the image block within the neighborhood according to a preset step size, and extract multiple neighborhood reference blocks; S46. Among all extracted reference blocks, use the fast nonlocal mean algorithm framework to match the pixel values ​​inside the image block with the pixels of the reference blocks in its neighborhood, and select the local optimal matching region according to the preset acceleration strategy; S47. After completing the neighborhood matching of all pixel positions, form the denoising result of the image block, generate the high-detail region denoised image block, and record its image frame number, index position and processing parameter configuration; S48. Output all flat region denoised image blocks and high-detail region denoised image blocks to the image block fusion module for unified fusion operation in subsequent image reconstruction processing steps.

[0033] A dual-channel differentiated image denoising workflow is adopted. For flat areas, a low-computational-complexity mean filtering strategy is used to ensure system processing efficiency. For high-detail areas, a fast nonlocal mean algorithm is introduced, enhancing structure preservation while reducing the computation time of traditional nonlocal mean algorithms, achieving a balance between quality and efficiency. This structure significantly overcomes the problem of existing unified filtering strategies failing to balance performance and effectiveness.

[0034] In this embodiment, S5 specifically includes: S51. Load the datasets of flat region denoised image blocks and high detail region denoised image blocks into the image fusion module of the edge computing node, and establish a unified image frame fusion index table; S52. Based on the position information of each denoised image block in the original image frame, map the flat region denoised image blocks and high detail region denoised image blocks to their corresponding spatial regions respectively; S53. Construct a complete image block arrangement matrix in the two-dimensional coordinate system of the image frame to ensure that all image blocks are positioned and filled according to their original spatial order; S54. Perform edge smoothing processing on the boundary regions of the image blocks to eliminate possible grayscale abrupt changes or block effects at the splicing of adjacent image blocks; S55. Reconstruct all the positioned image block data to generate a preliminary denoised image frame with the corresponding image frame number, and cache the image frame in the local storage area of ​​the edge node; S56. Generate metadata information of the preliminary denoised image frame, including frame index number, construction time, image block fusion status and processing completion flag, and register it in the image frame processing record table.

[0035] By integrating denoised image blocks and reconstructing image frames, combined with spatial positioning information and edge smoothing mechanisms, the processing strategy effectively alleviates grayscale jumps and structural discontinuities during image block stitching, improves the overall continuity and structural fidelity of the reconstructed image, and solves the visual inconsistencies and texture distortion defects that are prone to occur in traditional block-level reconstruction.

[0036] In this embodiment, S6 specifically includes: S61. Load the initial denoised image frame into the edge computing node and call the multi-scale image processing module to perform scale decomposition to construct the image pyramid structure of the microscopic image; S62. Downsample the initial denoised image frame layer by layer according to the preset scale level to generate a multi-scale image set including the original resolution layer and multiple low-resolution layers; S63. Perform global denoising processing on the low-resolution layer image, using an image filtering method with unified parameter configuration to enhance the overall structural consistency and further eliminate large-scale background noise; S64. Load the texture complexity distribution map and region classification of the image frame into the original resolution layer. S65. For the high-frequency detail regions, edge-preserving filtering or local contrast enhancement methods are used to enhance the regions, highlighting the microstructural details and edge contours in the microscopic image. S66. The low-resolution layer processing results and the original resolution layer enhancement results are scale-aligned and interpolated to construct a complete multi-scale image fusion framework. S67. The fused multi-scale image frames are generated, and the processing parameters, image feature indicators, and processing time information of each scale image layer are recorded as the basic data structure for subsequent consistency detection and image output.

[0037] By constructing a multi-scale pyramid structure, background noise suppression is performed across the entire image at the low-resolution layer, while detail enhancement is performed by focusing on key areas at the original resolution layer. This enables collaborative processing and optimized output of images at different scales, improving both the overall smoothness of the image and the clarity of key structures, effectively enhancing the accuracy of feature recognition and adapting it to subsequent AI recognition and visual analysis tasks.

[0038] In this embodiment, S7 specifically includes: S71. Load the fused image frame data output by the multi-scale image processing module into the edge computing node, and call the image reconstruction module to reconstruct and back-project the image results at each scale to generate a complete denoised image frame with uniform spatial resolution; S72. Perform pixel-level integrity verification on the reconstructed complete denoised image frame, check whether the image region filling is complete and whether the boundary stitching is continuous, and perform block edge smoothing to form a final image frame with complete structure and visual consistency; S73. Extract the frame number, timestamp, position information and processing parameters of the current final image frame, and register them in the image frame sequence index table for sequence management and consistency comparison; S74. Call the inter-frame consistency detection module, load the current image frame and its previous frame... For each frame of image data, pixel-level difference calculation, structural similarity comparison, and edge contour overlap analysis are performed on the corresponding regions. S75. Based on the inter-frame image change rate, texture distribution change range, and inter-frame difference distribution map, a consistency score index between the current frame and the previous frame is determined. S76. The consistency score is jointly judged with the image frame content stability parameter to output whether the current frame meets the inter-frame continuity standard. S77. For image frames with insufficient continuity, the key region content from the previous frame result is loaded using a local backoff strategy for compensation and reconstruction, and the availability status of the current image frame is marked. S78. Image frames that have completed inter-frame consistency detection are added to the effective frame buffer sequence for subsequent image analysis and structured coding processing steps.

[0039] By introducing an inter-frame consistency detection mechanism and integrating pixel difference, structural similarity and texture coherence evaluation models, we can achieve automatic analysis and back-down repair of the stability between consecutive image frames. This solves the problem of inter-frame jumps and discrimination errors that occur when traditional image processing methods process video streams or image sequences, and improves the processing reliability of image sequences in dynamic scenes.

[0040] In this embodiment, S8 specifically includes: S81. In the edge computing node, call the system resource monitoring module to collect real-time resource indicators of the current device, including processor utilization, memory usage, storage read / write speed, and network communication status. S82. Normalize the collected resource status information and compare it with the preset resource threshold configuration table to determine whether the current computing resources are in a high-availability, medium-load, or resource-scarce state. S83. When it is determined to be in a high-availability state, use a smaller image patch granularity, enable the full version of the image denoising algorithm processing path, and maintain the standard task execution order. S84. When it is determined to be in a medium-load state, use a medium-sized image patch granularity, call the standard configured image denoising algorithm parameters, and merge the processing tasks of non-critical areas according to task priority. S85. When a resource shortage is detected, a larger image block size is used for granular division, the denoising algorithm is adjusted to a simplified version, the enhancement processing flow for non-critical image areas is shut down, and the execution order in the current task queue is rearranged. S86. Based on the current resource status change trend, the parameter configuration table of the image processing module is dynamically updated, and the adjusted task scheduling strategy is synchronized to each processing sub-module in real time. S87. The time, parameters, and execution results of each resource status switch and scheduling strategy change are recorded to form a resource-aware historical log for subsequent system operation optimization and strategy evaluation. S88. The processing rhythm between the image acquisition module, image denoising module, and image fusion module is coordinated through the scheduling controller to ensure the continuity and stability of the image processing flow under resource change conditions.

[0041] By dynamically adjusting the image block partitioning granularity and algorithm path based on the current resource status of edge devices, a resource-aware scheduling mechanism is constructed to ensure that the most adaptive image processing scheme can be selected under different load conditions. This ensures that the system can still run stably under resource constraints, effectively avoids the risks of processing interruption and large fluctuations in image quality, and improves the practicality and robustness of the edge computing system.

[0042] In this embodiment, S9 specifically includes: S91. In the edge computing node, call the image encoding module to load the denoised image frame and its associated frame number, spatial index, timestamp, and processing parameter information; S92. Integrate the image frame data and metadata in a structured manner to generate a structured representation format for the image frame, and encode the image frame into an intermediate representation file that is easy to identify later according to a preset format standard; S93. Embed image processing path identifiers, region processing labels, and consistency detection results during the structured encoding process to ensure that the image frame has traceability and reconstructability; S94. Write the structured encoded image frame into the image processing output buffer of the edge node and call the feature recognition module interface to perform... Data push; S95, In the image feature recognition module, perform feature parsing tasks on the structured image frame, including key area localization, target edge extraction, and texture pattern analysis, and generate preliminary recognition results; S96, Call the edge transmission control module to generate a transmission control strategy based on the current network bandwidth status, task priority, and data importance level; S97, Encapsulate the preliminary recognition results and the original structured image frame together into a diagnostic data packet, upload it to the remote diagnostic system, and complete the remote synchronization of local processing results; S98, Record the timestamp, data packet number, and transmission status of image output and remote transmission process, generate a diagnostic data transmission log, and store it in the edge node system recording module.

[0043] By performing structured encoding and content annotation on the processed image frames, completing preliminary feature parsing and recognition tasks at edge nodes, and then uploading them to the remote diagnostic system through an adaptive bandwidth scheduling mechanism, a fusion architecture of edge intelligent recognition and remote expert collaborative judgment is realized, taking into account both local response capabilities and global diagnostic capabilities, and constructing an efficient and reliable image analysis closed loop.

[0044] Example 1: To verify the feasibility of this invention in practice, it was applied to a real-world scenario in a microscopic imaging laboratory. The research team needed to perform continuous, dynamic, and high-definition imaging analysis on biological samples to observe the transient response of cell membrane structures to different stimuli. The experiment used a high-resolution digital microscope for image acquisition, generating 20 frames per second, each with a resolution of 4096×3072, resulting in a massive amount of raw data. Since the experiment could not be interrupted and preliminary image analysis results were required within 3 seconds of image acquisition for subsequent automatic annotation and rapid response judgment, extremely high demands were placed on the real-time performance and stability of image processing. However, the existing system used a traditional central server model for image processing, uploading image data to a remote server for denoising and enhancement. The average processing delay exceeded 8 seconds, and the upload process was limited by local area network bandwidth, frequently resulting in data backlog and image frame drops, severely impacting experimental progress and data integrity.

[0045] To address the aforementioned issues, the project team deployed an image processing terminal with an integrated edge computing module next to the experimental platform. This terminal features a built-in quad-core ARM processor and a low-power GPU coprocessor, supporting TensorRT model acceleration and local caching optimization mechanisms. Supported by the method proposed in this invention, the system completes the entire process of image acquisition, region classification, differential denoising, multi-scale enhancement, and structured encoding at the edge nodes, eliminating the need to upload the original images to the cloud. Only key results and metadata are retained for remote backup and synchronization. The method first automatically divides each frame into 128×128 pixel image blocks and calculates texture complexity indices to achieve accurate identification of flat and high-detail regions. Fast mean filtering is used in flat regions to improve processing speed, while an optimized fast nonlocal mean algorithm is applied in high-detail regions for high-fidelity denoising, preserving cell membrane boundaries and microstructural features. Subsequently, the system uses a multi-scale image pyramid structure to perform full-image denoising at low resolution and high-frequency texture enhancement at the original resolution, effectively improving the overall image quality.

[0046] In actual operation, the average latency for edge image frame processing was 2.3 seconds, lower than the project's proposed upper limit of 3 seconds. Acquisition and standardization took approximately 0.6 seconds, classification and denoising approximately 1.1 seconds, and multi-scale enhancement and structured output approximately 0.6 seconds. The average inter-frame consistency score was 0.94, and the average image sharpness PSNR was 34.8 dB, an improvement of approximately 2.6 dB compared to the original system. The system operated continuously for 72 hours without any processing interruptions or buffer overflows, achieving an image frame integrity rate of 99.8%. Furthermore, the experiment tested the device's resource adaptability under different loads: when CPU utilization exceeded 85%, the system automatically adjusted the image block granularity to 256×256 and switched to a simplified model, ensuring latency remained within 3.2 seconds and maintaining stable system availability.

[0047] As can be seen from the above practical application examples, the real-time microscopic image processing method based on edge computing proposed in this invention effectively solves the problems of image upload delay, single processing path and poor adaptability to device resources in traditional methods. It realizes efficient local processing and intelligent scheduling of high-resolution images, and provides a practical and stable technical path for microscopic image scenarios with high-frequency sampling and high timeliness requirements.

[0048] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An edge-computing-based microscopic image real-time processing method, characterized in that, Comprise the following steps: S1, deploy a microscopic image acquisition and processing terminal at the edge computing node, acquire a sequence of microscopic image frames to be processed, and locally cache and manage the image data on the edge side; S2, divide the acquired sequence of microscopic image frames into a plurality of fixed-size image blocks, and calculate a local texture complexity index for each image block; S3, classify the image blocks according to the local texture complexity index, into flat region image blocks and high detail region image blocks; S4, for flat region image blocks, use a fixed window size mean filter method for image denoising processing, output flat region denoising image blocks, for high detail region image blocks, use a fast non-local mean algorithm based on neighborhood matching for image denoising processing, output high detail region denoising image blocks; S5, fuse the flat region denoising image blocks and the high detail region denoising image blocks at the block level to construct a complete preliminary denoising image frame; S6, based on the preliminary denoising image frame, construct a multi-scale image pyramid structure of the microscopic image, perform global denoising processing at a low resolution scale, and perform local enhancement processing of high frequency detail regions at an original resolution scale; S7, fuse the multi-scale processing results to generate a complete denoising image frame, and perform inter-frame consistency detection on the image frame sequence; S8, dynamically adjust the image block division granularity, denoising algorithm parameters and execution order according to the current computing resource state of the edge device; S9, structure the denoising processed image frame, output to the image feature recognition module in the edge node, and configure a transmission strategy to upload the processing result to a remote diagnosis system. 2.The microscopic image real-time processing method based on edge computing according to claim 1, wherein, The S1 specifically comprises: S11, deploy a microscopic image acquisition and processing terminal at the edge computing node, the microscopic image acquisition and processing terminal comprising an image sensor, a data acquisition module and an edge computing processing unit; S12, acquire a continuous sequence of microscopic image frames through the image sensor, and complete the numbering and time labeling of the microscopic image frames by the data acquisition module; S13, perform format standardization processing on the microscopic image frames in the edge computing processing unit, unify the channel number and resolution of the image; S14, store the format standardized microscopic image frames into the local cache area of the edge node for temporary storage and read-write management; S15, generate microscopic image frame cache index information and time stamp sequence, and register it in the edge node task scheduling management module. 3.The microscopic image real-time processing method based on edge computing according to claim 1, wherein, The S2 specifically comprises: S21, calling an image preprocessing module in the edge computing node, reading locally cached microscopic image frame data, and regularly dividing each image according to preset image block size parameters, dividing into multiple rectangular image blocks of consistent size and non-overlapping, and marking the row and column numbers of the image blocks and the position information in the whole frame image; S22, performing an edge-preserving filtering operation on each image block, and extracting a pixel gray matrix in the filtered image block; S23, calculating the local statistical features of the image block based on the pixel gray matrix, the basic parameters including the average value, the standard deviation, the range and the gray histogram distribution density; S24, constructing a gray co-occurrence matrix of the image block, extracting texture description indicators including energy, contrast, entropy and correlation, representing the detail complexity of the image block; S25, calculating the gradient amplitude graph and the gradient direction graph in the image block, and calculating the mean value, variance and gradient direction concentration; S26, fusing the local statistical features, texture description indicators and gradient features according to the weight, generating a unified local texture complexity index of the image block, and binding the normalized texture complexity index value with the image block space position index to construct a texture complexity distribution graph of the image frame; S27, outputting the texture complexity distribution graph as the feature data structure of the image frame. 4.The method of claim 1, wherein, The S3 specifically comprises: S31, loading the texture complexity distribution graph corresponding to the microscopic image frame in the edge computing node, and sequentially traversing the texture complexity index values of all image blocks; S32, according to the preset complexity classification threshold, interval discriminating the texture complexity index of each image block, if the index value is less than the threshold, the image block is marked as a flat area image block; if the index value is greater than or equal to the threshold, the image block is marked as a high detail area image block; S33, adding the classified image blocks to the flat area image block set and the high detail area image block set respectively, and recording the classification label, the spatial position information and the original image frame number of each image block; S34, generating a region classification mapping graph of the microscopic image frame, the region classification mapping graph is a two-dimensional matrix structure consistent with the size of the microscopic image frame, and the value corresponding to each position represents the classification label of the image block to which the position belongs; S35, caching the flat area image block set, the high detail area image block set and the region classification mapping graph to the edge node image processing module.

5. The microscopic image real-time processing method based on edge computing according to claim 1, characterized in that, The S4 specifically comprises: S41, index information of the flat region image block is extracted from the region classification mapping diagram, and corresponding flat region image block data is loaded in the edge computing node; S42, for each flat region image block, a sliding filter with a fixed window size is set, each pixel point is traversed in a row-column manner, a pixel set is extracted in the neighborhood, mean value calculation is performed on all pixel gray values in the pixel set, a substitute value of the current pixel is generated, and mean value filtering operation of the entire image block is completed; S43, the processed image block is marked as a flat region denoising image block, and its image frame number, image block index and processing timestamp are recorded to generate a flat region image block denoising dataset; S44, index information of the high-detail region image block is extracted from the region classification mapping diagram, and corresponding high-detail region image block original data is loaded; S45, for each high-detail region image block, the neighborhood range thereof in the original image frame is determined, image block sliding is performed in the neighborhood according to a preset step size, and a plurality of neighborhood reference blocks are extracted; S46, in all extracted reference blocks, a fast non-local mean algorithm framework is used to perform matching operation on the pixel values in the image block and the reference blocks in the neighborhood, and a locally optimal matching region is selected according to a preset acceleration strategy; S47, after the neighborhood matching of all pixel positions is completed, a denoising result of the image block is formed, a high-detail region denoising image block is generated, and its image frame number, index position and processing parameter configuration are recorded; S48, all flat region denoising image blocks and high-detail region denoising image blocks are uniformly output to an image block fusion module. 6.The microscopic image real-time processing method based on edge computing according to claim 1, wherein, The S5 specifically includes: S51, the dataset of the flat region denoising image block and the high-detail region denoising image block is loaded in the image fusion module of the edge computing node, and a unified image frame fusion index table is established; S52, according to the position information of each denoising image block in the original image frame, the flat region denoising image block and the high-detail region denoising image block are respectively mapped to the corresponding spatial region; S53, a complete image block arrangement matrix is constructed under the image frame two-dimensional coordinate system; S54, edge smoothing processing is performed on the image block boundary region to eliminate possible gray value mutation or block effect at the splicing position of adjacent image blocks; S55, all positioned image block data is reconstructed to generate a preliminary denoising image frame corresponding to the image frame number, and the image frame is cached to the local storage area of the edge node; S56, metadata information of the preliminary denoising image frame is generated, including frame index number, construction time, image block fusion state and processing completion flag, and is registered to an image frame processing record table.

7. The microscopic image real-time processing method based on edge computing according to claim 1, characterized in that, The S6 specifically includes: S61, load the preliminary denoising image frame in the edge computing node, and call the multi-scale image processing module to perform scale decomposition on the preliminary denoising image frame, and construct the image pyramid structure of the microscopic image; S62, downsample the preliminary denoising image frame layer by layer according to a preset scale level, and generate a multi-scale image set including an original resolution layer and a plurality of low resolution layers; S63, perform a global denoising processing operation on the low resolution layer image, and use an image filtering method with a unified parameter configuration to enhance the overall structural consistency and further eliminate large-scale background noise; S64, load the texture complexity distribution map and the region classification mapping map of the image frame in the original resolution layer, and extract the image position index corresponding to the high-frequency detail region; S65, for the high-frequency detail region, use an edge-preserving filtering method or a local contrast enhancement method to perform region enhancement processing, and highlight the microstructure details and edge contours in the microscopic image; S66, perform scale alignment and interpolation mapping on the low resolution layer processing result and the original resolution layer enhancement result, and construct a complete multi-scale image fusion framework; S67, generate a fused multi-scale image frame, and record the processing parameters, image feature indicators and processing time information of each scale image layer. 8.The method of claim 1, wherein, The S7 specifically includes: S71, load the fused image frame data output by the multi-scale image processing module in the edge computing node, call the image reconstruction module to recombine and perform reverse projection processing on the scale image results, and generate a complete denoising image frame with a unified spatial resolution; S72, perform pixel-level integrity checking on the complete denoising image frame after reconstruction, detect whether the image region filling is complete and the boundary splicing is continuous, and perform block edge smoothing processing to form a final image frame with complete structure and consistent vision; S73, extract the frame number, timestamp, position information and processing parameters of the current final image frame, and register them into an image frame sequence index table; S74, call the inter-frame consistency detection module, load the image data of the current image frame and the previous frame, and perform pixel-level difference calculation, structure similarity comparison and edge contour coincidence analysis on the corresponding regions; S75, determine the consistency score index between the current frame and the previous frame according to the inter-frame image change rate, texture distribution change range and difference distribution map between adjacent frames; S76, jointly judge the consistency score and the image frame content stability parameter, and output whether the current frame meets the inter-frame continuity standard; S77, for the image frame with insufficient continuity, load the key region content in the previous frame result for compensation reconstruction through a local rollback strategy, and mark the availability state of the current image frame; S78, add the image frame after the inter-frame consistency detection to the valid frame buffer sequence. 9.The microscopic image real-time processing method based on edge computing according to claim 1, wherein, The S8 specifically includes: S81, call the system resource monitoring module in the edge computing node, real-time collection of current device resource indicators including processor usage, memory occupancy, storage read-write rate and network communication state; S82, normalize the collected resource state information, and compare with the preset resource threshold configuration table to determine whether the current computing resource is in a high availability, medium load or resource shortage state; S83, when it is determined to be in a high availability state, a smaller size image block division granularity is used, a complete version of the image denoising algorithm processing path is enabled, and the standard task execution order is maintained; S84, when it is determined to be in a medium load state, a medium size image block division granularity is used, the standard configured image denoising algorithm parameters are called, and the processing tasks of non-critical regions are combined and executed according to the task priority; S85, when it is determined to be in a resource shortage state, a larger size image block division granularity is used, the denoising algorithm is adjusted to a simplified version, the enhancement processing flow of non-critical image regions is closed, and the execution order in the current task queue is rearranged; S86, according to the current resource state change trend, dynamically update the parameter configuration table of the image processing module, and real-time synchronize the adjusted task scheduling strategy to each processing submodule; S87, record the time, parameters and execution results of each resource state switching and scheduling strategy change to form a resource perception history log; S88, coordinate the processing rhythm between the image acquisition module, the image denoising module and the image fusion module through the scheduling controller.

10. The microscopic image real-time processing method based on edge computing according to claim 1, characterized in that, The S9 specifically comprises: S91, call the image encoding module in the edge computing node, load the image frame completed denoising processing and the frame number, spatial index, timestamp and processing parameter information associated with the image frame; S92, structure the image frame data and metadata information, and generate an image frame structured representation format; S93, embed the image processing path identifier, region processing label and consistency detection result in the structured encoding process; S94, write the structured encoded image frame into the image processing output buffer of the edge node, and call the feature recognition module interface for data pushing; S95, perform feature analysis tasks including key region positioning, target edge extraction and texture pattern analysis on the structured image frame in the image feature recognition module to generate preliminary recognition results; S96, call the edge transmission control module, generate a transmission control strategy according to the current network bandwidth state, task priority and data importance level; S97, package the preliminary recognition results and the original structured image frame into a diagnostic data packet together, and upload it to the remote diagnosis system to complete the remote synchronization of local processing results; S98, record the timestamp, data packet number and transmission state of the image output and remote transmission process, generate a diagnostic data transmission log, and store it in the edge node system record module.

Images