A smart medical computing power network resource dynamic allocation method and system

Abstract

The application discloses a kind of wisdom medical computing power network resource dynamic allocation method and system, it is related to resource allocation technical field, it includes according to the comparison result between the jolt intensity prediction result and preset intensity threshold value, determine a target computing power allocation strategy from at least two preset computing power allocation strategies;According to target computing power allocation strategy, collect first image data;First image data is sent to the first computing power resource node indicated by target computing power allocation strategy, so that the first computing power resource node executes the first image processing operation of pre-set after first image data, output second image data;The image quality check result of second image data is obtained;When image quality check result does not satisfy preset image quality condition, generate computing power scheduling request, send computing power scheduling request to second computing power resource node;Receive the third image data returned by second computing power resource node.The application has the effect of improving resource allocation efficiency.

Description

Technical Field

[0001] This application relates to the field of resource allocation technology, and in particular to a method and system for dynamic allocation of computing power network resources in smart healthcare. Background Technology

[0002] In remote mobile healthcare, especially in ambulance-based emergency consultations, real-time acquisition and high-quality transmission of medical image data are fundamental to remote diagnosis. Under current technology, ambulances are typically equipped with onboard cameras, data acquisition terminals, and wireless communication modules to transmit patient injury images back to remote hospitals in real time. However, the unique nature of mobile healthcare environments lies in the fact that the physical vehicle is constantly in motion, and changes in road conditions directly impact image acquisition quality. When the vehicle travels on bumpy roads, the onboard image acquisition unit is subjected to continuous mechanical vibration, leading to motion blur and inter-frame jitter in the acquired video stream, severely reducing image readability.

[0003] To address image quality issues caused by the aforementioned physical environment, existing technologies primarily tackle the problem from two dimensions. First, they introduce mechanical or electronic image stabilization devices at the image acquisition end, suppressing the impact of shaking through hardware-level attitude sensing and compensation. Second, they rely on the high computing power of cloud servers for video stabilization or image enhancement after image transmission. However, mechanical image stabilization has a limited compensation range and struggles to handle high-frequency, severe bumps such as those on gravel roads; electronic image stabilization often comes at the cost of sacrificing image quality or introducing latency. While relying entirely on cloud processing can achieve good restoration results, it requires continuous transmission of the complete video stream. On bumpy roads, not only does image blurring lead to a large amount of invalid data transmission, but network bandwidth fluctuations also cause keyframe packet loss, resulting in inconsistent image quality received from the cloud. More importantly, in existing solutions, image quality assessment is usually performed after cloud reception, making it impossible to verify image usability in real time at the acquisition site. By the time unusable images are discovered, the patient has already been transferred away from the bumpy road, losing the opportunity to acquire crucial diagnostic images.

[0004] Another overlooked issue is the allocation of computing resources. Existing telemedicine systems typically employ a static configuration approach, where vehicle-mounted terminals are fixed for initial encoding and compression, while cloud servers handle image analysis and diagnosis. This model is sustainable under stable road conditions but cannot dynamically adapt to fluctuations in computing power demands caused by changes in road surface. When encountering severe bumps, image blurriness increases dramatically, and the required computational complexity grows non-linearly. Static allocation schemes cannot provide sufficient real-time processing capabilities locally, nor can they handle sudden surges in repair requests in the cloud, leading to increased system response latency and even diagnostic blind spots. Furthermore, the combined effects of different vehicle models, driving habits, and road surface characteristics make the impact of bumps on image quality highly uncertain, and existing solutions lack the self-learning ability to adapt to this complex and ever-changing environment.

[0005] To address the aforementioned issues, there is an urgent need in this field for a method for dynamically allocating smart healthcare computing network resources that can dynamically predict the intensity of bumps based on road conditions ahead, hierarchically schedule edge-cloud computing resources, and establish a closed-loop quality verification and feedback repair mechanism during image transmission. This method would resolve the contradiction between image acquisition failures caused by changes in the physical environment and the surge in computing power demand in mobile healthcare. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides a method and system for dynamic allocation of computing power network resources in smart healthcare.

[0007] Firstly, this application provides a method for dynamically allocating computing power network resources for smart healthcare, applied to vehicle-mounted terminals, comprising the following steps: Acquire the target vehicle's current location data, attitude data, and image data of the road surface ahead; The positioning data, attitude data and image data are input into the road bump intensity prediction model to generate the bump intensity prediction result of the target vehicle in a preset future time period. Based on the comparison between the predicted turbulence intensity and the preset intensity threshold, a target computing power allocation strategy is determined from at least two preset computing power allocation strategies. According to the target computing power allocation strategy, the vehicle-mounted image acquisition unit is controlled to acquire raw image data and output the raw image data as the first image data. The first image data is sent to the first computing power resource node indicated by the target computing power allocation strategy, so that the first computing power resource node performs a preset first image processing operation on the first image data and outputs the second image data. Obtain the image quality verification result of the second image data; When the image quality verification result does not meet the preset image quality conditions, a computing power scheduling request carrying the original image data identifier is generated and sent to the second computing power resource node; the computing power capability of the second computing power resource node is greater than that of the first computing power resource node. The third image data returned by the second computing power resource node is generated by the second computing power resource node in response to the computing power scheduling request, after obtaining the corresponding original image data sequence according to the original image data identifier and performing a preset second image processing operation on the original image data sequence. The image quality of the third image data is higher than that of the second image data.

[0008] Preferably, the positioning data, attitude data, and image data are input into the road bump intensity prediction model to generate a bump intensity prediction result for the target vehicle within a preset future time period, including: The positioning data is compared with the road surface smoothness layer in the high-precision map database to generate the prior roughness index of the road segment ahead corresponding to the positioning data. Vibration feature extraction is performed on the attitude data to generate vibration energy spectrum data corresponding to the attitude data. Perform road surface semantic segmentation on the image data to generate data on the area ratio of road surface damage areas in the image data; The prior roughness index, vibration energy spectrum data, and area ratio data are input into the road surface bump intensity prediction model, and then the bump intensity prediction results are output.

[0009] Preferably, the preset at least two computing power allocation strategies include a first computing power allocation strategy, a second computing power allocation strategy, and a third computing power allocation strategy, wherein: The predicted bump intensity corresponding to the first computing power allocation strategy is less than the first preset threshold. The first computing power allocation strategy instructs the vehicle image acquisition unit to acquire video stream at the first frame rate as the first image data, instructs to send the first image data to the cloud computing power resource node as the first computing power resource node, and instructs the cloud computing power resource node to perform video diagnostic analysis operation as the first image processing operation. The bump intensity prediction result corresponding to the second computing power allocation strategy is between the first preset threshold and the second preset threshold. The second computing power allocation strategy instructs the vehicle image acquisition unit to acquire a continuous image sequence at a second frame rate higher than the first frame rate as the first image data, instructs the first image data to be sent to the edge computing power resource node as the first computing power resource node, and instructs the edge computing power resource node to perform inter-frame motion compensation and anti-shake processing operations as the first image processing operation. If the predicted bump intensity of the third computing power allocation strategy is greater than the second preset threshold, the third computing power allocation strategy instructs the vehicle-mounted image acquisition unit to acquire a continuous image sequence at the second frame rate as the first image data, and instructs the first image data to be sent to the edge computing power resource node as the first computing power resource node to perform the first image processing operation. While performing the first image processing operation, the continuous image sequence is pre-transmitted to the cloud computing power resource node as the original image data sequence for data backup in advance, so as to shorten the response delay when the computing power scheduling request is triggered later.

[0010] Preferably, before controlling the vehicle-mounted image acquisition unit to acquire and output the first image data, the method further includes: Obtain the network status data of the communication network currently accessed by the target vehicle and the load status data of the first computing power resource node; When the network status data meets the preset network transmission conditions and the load status data meets the preset load conditions, the vehicle-mounted image acquisition unit is controlled to acquire and output the first image data according to the target computing power allocation strategy. When the network status data does not meet the preset network transmission conditions or the load status data does not meet the preset load conditions, the vehicle-mounted image acquisition unit is controlled to acquire and output the fourth image data according to the preset backup computing power allocation strategy, and the fourth image data is sent to the preset backup computing power resource node.

[0011] Preferably, obtaining the image quality verification result of the second image data includes: The second image data is input into the medical image region segmentation model to generate region location data of at least one key diagnostic region in the second image data corresponding to the preset diagnostic interest area; Obtain a standard, clear medical anatomical image template corresponding to the preset diagnostic focus area as reference image data; Calculate the weighted structural similarity index between the key diagnostic regions in the second image data and the corresponding key diagnostic regions in the reference image data, and use it as the image quality verification result; wherein, the weight of the weighted structural similarity index is higher in the key diagnostic regions than in the non-key diagnostic regions.

[0012] Preferably, when the image quality verification result does not meet the preset image quality conditions, a computing power scheduling request carrying the original image data identifier is generated, including: When the weighted structural similarity index is less than the preset medical image quality standard threshold, the image quality verification result is determined to not meet the preset image quality conditions. Obtain the original burst image data corresponding to the second image data, and use it as the original image data pointed to by the original image data identifier; A computing power scheduling request is generated, which includes the original image data, the motion vector parameters corresponding to the second image data, and the image quality verification result.

[0013] Preferably, the second image processing operation includes a temporal restoration operation based on a generative adversarial network; the third image data is generated by the second computing power resource node in the following manner: Receive computing power scheduling request, parse and obtain the original image data sequence pointed to by the original image data identifier, the original image data sequence contains multiple consecutive frames of blurred image data; Obtain the real-time vibration frequency data of the target vehicle when the original image data sequence is acquired; The original image data sequence and the real-time vibration frequency data are input into a pre-trained generative adversarial network model. The generator in the generative adversarial network model generates repaired image data for at least one target frame in the original image data sequence. The repaired image data is combined with the image data of other frames in the original image data sequence, excluding the target frame, through temporal smoothing to generate a third image data containing the repaired image data.

[0014] Preferably, the generator in the generative adversarial network model generates repaired image data for at least one target frame in the original image data sequence, including: The attention mechanism is used to identify key regions in the original image data sequence that correspond to preset diagnostic areas of interest. When generating the repaired image data, higher computing resources are allocated to repair the critical regions than to non-critical regions.

[0015] Preferably, the weighted structural similarity index between the key diagnostic region in the second image data and the corresponding key diagnostic region in the reference image data is calculated, including: Obtain the diagnostic part identifier corresponding to the remote diagnostic task currently being performed by the target vehicle; Based on the diagnostic site identifier, retrieve the anatomical structure region data corresponding to the diagnostic site identifier from the preset medical knowledge base; The anatomical structure region data is registered with the second image data to generate the location data of the target key region in the second image data corresponding to the diagnostic site identifier; Based on the turbulence intensity prediction result, a first weight value corresponding to the target key region and a second weight value corresponding to the non-target key region in the second image data are determined, wherein the first weight value is higher than the second weight value. Based on the first weight value and the second weight value, calculate the weighted structural similarity index between the second image data and the reference image data.

[0016] Preferably, after generating repaired image data for at least one target frame in the original image data sequence through the generator in the generative adversarial network model, the method further includes: The repaired image data is input into a preset secondary quality verification model to generate a secondary quality verification result for the repaired image data. When the result of the secondary quality check is less than the preset secondary quality threshold, the adjacent frame data of the target frame in the original image data sequence are obtained; Based on the optical flow information between the adjacent frame data and the repaired image data, an iterative optimization operation is performed on the repaired image data to generate optimized repaired image data, until the secondary quality check result of the optimized repaired image data is not less than the preset secondary quality threshold or the number of iterations reaches the preset number threshold.

[0017] Secondly, this application provides a smart healthcare computing power network resource dynamic allocation system, comprising: The data acquisition module is used to acquire the target vehicle's current location data, attitude data, and image data of the road surface ahead; The bump intensity prediction module is used to input positioning data, attitude data and image data into the road bump intensity prediction model to generate the bump intensity prediction result of the target vehicle in a preset future time period. The allocation strategy module is used to determine a target computing power allocation strategy from at least two preset computing power allocation strategies based on the comparison between the turbulence intensity prediction result and the preset intensity threshold. The control acquisition module is used to control the vehicle-mounted image acquisition unit to acquire raw image data according to the target computing power allocation strategy, and output the raw image data as the first image data. The image processing module is used to send the first image data to the first computing power resource node indicated by the target computing power allocation strategy, so that the first computing power resource node performs a preset first image processing operation on the first image data and outputs the second image data. The verification result module is used to obtain the image quality verification result of the second image data. The scheduling module is used to generate a computing power scheduling request carrying the original image data identifier when the image quality verification result does not meet the preset image quality conditions, and send the computing power scheduling request to the second computing power resource node; the computing power capability of the second computing power resource node is greater than that of the first computing power resource node. The receiving module is used to receive the third image data returned by the second computing power resource node. The third image data is generated by the second computing power resource node in response to the computing power scheduling request, after obtaining the corresponding original image data sequence according to the original image data identifier, and performing a preset second image processing operation on the original image data sequence. The image quality of the third image data is higher than that of the second image data.

[0018] In summary, this application includes at least one of the following beneficial technical effects: 1. This application provides a method for dynamic allocation of computing power network resources in smart healthcare. By introducing a road bump intensity prediction model, based on the positioning, attitude, and road image data acquired in real time by the vehicle terminal, the method accurately predicts the bump intensity in the future time period. Based on the comparison between the prediction result and a preset intensity threshold, it dynamically selects an appropriate computing power allocation strategy. This mechanism breaks the limitations of traditional static allocation of computing power resources, enabling the system to optimize the scheduling of computing power resources before the vehicle enters a bumpy road section: low frame rate acquisition and lightweight cloud analysis are used on smooth road sections to save network and computing resources; edge computing is used on moderately bumpy road sections to improve frame rate and reduce jitter, balancing real-time performance and image quality; and parallel mode of edge processing and cloud data backup is initiated in advance on severely bumpy road sections, significantly reducing the risk of key frame loss due to sudden bumps. This effectively achieves precise matching between computing power resources and changes in the physical environment, avoids invalid data transmission, and improves the overall system's resource utilization efficiency and response speed. 2. This invention introduces a closed-loop quality verification and feedback repair mechanism into the image transmission chain. Upon obtaining the second image data after preliminary processing by the first computing power resource node, image quality verification is immediately performed. If the verification result does not meet preset conditions, a computing power scheduling request carrying the original image data identifier is automatically generated, seamlessly switching the processing task to the second computing power resource node with stronger computing power. This mechanism effectively solves the pain points of delayed image quality judgment and inability to provide on-site repair in traditional solutions. It ensures that even when images are blurred due to turbulence, the cloud's high computing power can be quickly triggered for deep repair. Furthermore, since the scheduling request carries the original data identifier, the second computing power resource node can directly retrieve the original image sequence for high-quality reconstruction, avoiding network delays and bandwidth occupation caused by repeated transmissions. This provides a continuous and reliable source of high-quality images for remote medical diagnosis. 3. This invention constructs a differentiated set of computing power allocation strategies, particularly a third computing power allocation strategy designed for scenarios with severe turbulence. While performing real-time edge computing processing, the original image data sequence is pre-synchronously backed up to cloud computing power resource nodes. This forward-looking data redundancy design, combined with subsequent precise scheduling based on image quality verification results, allows cloud nodes to utilize the backed-up original data sequence, along with parameters such as the vehicle's real-time vibration frequency, to perform temporal repair operations using advanced algorithms such as generative adversarial networks when image quality is substandard. This generates third-generation image data of higher quality than the initial processing. This collaborative working mode of processing, backing up, and repairing on demand not only significantly shortens the response delay from discovering quality problems to obtaining high-quality repair results, but also ensures the image clarity of the core anatomical structures on which remote diagnosis depends by allocating resources to the repair of key diagnostic areas through an attention mechanism. This fundamentally improves the accuracy and reliability of remote consultations in mobile medical environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the method for dynamically allocating computing power network resources for smart healthcare according to an embodiment of this application.

[0021] Figure 2 This is a schematic diagram of the system for dynamic allocation of smart healthcare computing power network resources according to an embodiment of this application. Detailed Implementation

[0022] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0023] Application Overview: In existing technologies, image acquisition and transmission in remote mobile healthcare, especially in emergency vehicle-mounted consultation scenarios, primarily rely on real-time video streams from vehicle-mounted cameras to cloud servers for analysis and diagnosis. However, existing systems generally employ a static configuration mode, where the initial encoding and compression are performed by the vehicle-mounted terminal, and image analysis and diagnosis are executed by the cloud server. This lacks the ability to perceive and adapt to changes in the physical environment. When an ambulance travels on bumpy roads, the vehicle-mounted image acquisition unit is subjected to continuous mechanical vibration interference, resulting in motion blur and inter-frame jitter in the acquired video stream, severely reducing the diagnostic value of the images. More critically, in existing solutions, image quality assessment is usually performed after cloud reception, making it impossible to verify image usability in real time at the acquisition site. By the time an image is found to be unusable, the patient has already been transferred away from the bumpy road, losing the opportunity to acquire crucial diagnostic images. Furthermore, the combined effects of different vehicle models, driving habits, and road surface characteristics make the impact of bumps on image quality highly uncertain. Existing solutions lack the self-learning ability to adapt to this complex and changing environment, leading to increased system response delays and even diagnostic blind spots when encountering high-intensity bumps.

[0024] To address the aforementioned issues, the inventors discovered a predictable correlation between road bump intensity and image blur. By using onboard sensors to perceive vehicle attitude, positioning information, and images of the road ahead in real time, the intensity of bumps in the near future can be predicted, providing a basis for dynamic and proactive allocation of computing resources. Further research revealed significant differences in the importance of different image regions for diagnosis; the quality requirements for critical diagnostic areas such as wounds and pupils are far higher than for background areas. Therefore, image quality verification should focus on critical areas rather than the entire image. Simultaneously, for image degradation caused by varying degrees of bumps, a single level of computing power cannot simultaneously address real-time performance and restoration effectiveness. Therefore, a strategy of dynamically switching computing power allocation based on bump intensity thresholds was proposed: under stable road conditions, processing is done directly in the cloud; under moderate bumps, lightweight stabilization and verification are performed by edge nodes; and under high-intensity bumps, a high-order cloud model is triggered for deep restoration. Further experimental verification combined the bump intensity prediction results with image quality closed-loop feedback, introducing a dynamic adjustment mechanism for model parameters to form an adaptive computing power scheduling closed-loop system.

[0025] Specifically, the vehicle-mounted terminal first simultaneously acquires the target vehicle's positioning data, attitude data, and road surface image data. The positioning data is compared with a high-precision map road surface smoothness layer to generate a priori roughness index. Vibration energy spectra are extracted from the attitude data, and semantic segmentation is performed on the road surface image to obtain the area ratio of damaged regions. These multi-source features are input into a road bump intensity prediction model, outputting continuous prediction results of bump intensity over future time periods. Based on a comparison of the prediction results with preset thresholds, the system dynamically selects a computing power allocation strategy: when the frame rate is below the first threshold, video streams are acquired at a normal frame rate and directly uploaded to the cloud for diagnosis; when the frame rate is between the first and second thresholds, a high frame rate continuous shooting mode is switched, and the continuous images are sent to edge nodes for image stabilization; when the frame rate is above the second threshold, in addition to edge image stabilization, the continuous image sequence is pre-backed up to the cloud. The second image output after edge node processing extracts key diagnostic regions using a medical image segmentation model and calculates a weighted structural similarity index with a standard medical image template as a quality verification result. When the weighted structural similarity index is below the preset threshold, the system generates a computing power scheduling request carrying the original image identifier and sends it to the cloud. The cloud retrieves a pre-backed sequence of raw burst images based on the identifiers. Combined with the vehicle's real-time vibration frequency, it uses an attention mechanism within a generative adversarial network to focus on repairing key diagnostic areas, generating a high-quality third image that is returned to the in-vehicle terminal. The in-vehicle terminal then uploads this data as experience to the central server for periodic model updates, forming a closed-loop optimization process.

[0026] Compared to existing technologies, traditional methods rely on single video stream transmission and static computing power configuration, lacking the ability to perceive and adapt to changes in the physical environment. This leads to systemic image distortion and diagnostic delays when vehicles experience bumps. This solution innovatively integrates multimodal sensor data and prior road condition information, achieving proactive dynamic allocation of computing resources through bump intensity prediction, overcoming the lag in response to sudden bumps in traditional static allocation methods. Unlike existing post-processing image enhancement techniques, this solution establishes a closed-loop mechanism of "prediction-verification-repair-learning." During image transmission, a key-region-oriented weighted structural similarity assessment is introduced, ensuring that higher-order generative adversarial network repair is triggered only when automatic processing fails, balancing real-time performance and resource efficiency. The dynamic computing power level switching mechanism combines the low-latency anti-shake advantages of edge nodes with the high-order repair capabilities of cloud nodes. The pre-backup mechanism in the third strategy significantly shortens the repair response delay in high-intensity bump scenarios. The adaptive adjustment function of model parameters, through empirical data uploads and updates to the central server, ensures long-term detection accuracy and system self-evolution capabilities.

[0027] Through the above technical solutions, this application effectively overcomes the problems of image quality degradation and diagnostic blind spots caused by vehicle bumps, improving image clarity and usability in key diagnostic areas while ensuring real-time detection. The dynamic computing power allocation mechanism combines the advantages of efficient transmission under smooth road conditions with robust repair under bumpy road conditions, and the adaptive adjustment function of model parameters ensures that the system can adapt to the combined effects of different vehicle models, different driving habits, and different road surface characteristics.

[0028] After introducing the basic concept of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0029] Example 1: This application discloses a method for dynamic allocation of computing power network resources in smart healthcare.

[0030] Reference Figure 1 A method for dynamically allocating computing power network resources in smart healthcare, applied to vehicle-mounted terminals, includes the following steps: The system acquires the target vehicle's current location data, attitude data, and image data of the road surface ahead. The location data refers to the target vehicle's latitude and longitude coordinates acquired in real-time via a Global Navigation Satellite System (GNSS) receiver. Specifically, it can employ Differential Global Positioning System (DSBS) or Real-Time Dynamic Positioning (RTS) technology to achieve centimeter-level high-precision positioning, providing a precise spatial reference for subsequent road surface smoothness comparison. The attitude data refers to the raw signals of triaxial acceleration and triaxial angular velocity collected in real-time by the vehicle's inertial measurement unit, used to analyze the vibration characteristics of the vehicle caused by uneven road surfaces. The image data of the road surface ahead refers to a sequence of color or grayscale images collected in real-time by the vehicle's forward-facing camera, used for visually identifying damaged areas of the road surface and quantifying their distribution. The positioning data, attitude data and image data are input into the road bump intensity prediction model to generate the bump intensity prediction result of the target vehicle in a preset future time period. Based on the comparison between the predicted turbulence intensity and the preset intensity threshold, a target computing power allocation strategy is determined from at least two preset computing power allocation strategies; wherein, the computing power allocation strategy is used to indicate the acquisition method of the medical image data to be transmitted, the transmission path of the medical image data, and the computing power resource level for processing the medical image data; According to the target computing power allocation strategy, the vehicle-mounted image acquisition unit is controlled to acquire raw image data and output the raw image data as the first image data. The first image data is sent to the first computing power resource node indicated by the target computing power allocation strategy, so that the first computing power resource node performs a preset first image processing operation on the first image data and outputs the second image data. Obtain the image quality verification result of the second image data; When the image quality verification result does not meet the preset image quality conditions, a computing power scheduling request carrying the original image data identifier is generated and sent to the second computing power resource node; the computing power capability of the second computing power resource node is greater than that of the first computing power resource node. The third image data returned by the second computing power resource node is generated by the second computing power resource node in response to the computing power scheduling request, after obtaining the corresponding original image data sequence according to the original image data identifier and performing a preset second image processing operation on the original image data sequence. The image quality of the third image data is higher than that of the second image data.

[0031] The working process and principle of this application are as follows: First, the positioning data, attitude data, and image data of the road surface ahead of the target vehicle at the current moment are acquired. The positioning data is compared with the road surface smoothness layer of a high-precision map to generate a prior roughness index. Vibration energy spectrum is extracted from the attitude data. Semantic segmentation is performed on the image data to obtain the area ratio of the damaged area. The above multi-source features are input into the road bump intensity prediction model, and the continuous prediction results of bump intensity in the future time period are output. Based on the comparison between the prediction results and the preset intensity threshold, the target computing power allocation strategy is dynamically selected. According to the target strategy, the vehicle-mounted image acquisition unit is controlled to acquire raw image data and output it as the first image data. It is sent to the first computing power resource node to perform the first image processing operation to obtain the second image data. Then, the image quality verification result of the second image data is obtained. When the verification result does not meet the preset image quality conditions, a computing power scheduling request carrying the original image data identifier is generated and sent to the second computing power resource node. The second computing power resource node obtains the corresponding raw image data sequence according to the identifier, performs the second image processing operation based on the generative adversarial network, generates the third image data with higher image quality, and returns it to the vehicle terminal. In this way, dynamic optimization of the entire process, from road condition perception to computing power scheduling to quality closed loop, is achieved, ensuring that high-quality diagnostic images can be stably output even under complex road conditions.

[0032] Furthermore, the positioning data, attitude data, and image data are input into the road bump intensity prediction model to generate a bump intensity prediction result for the target vehicle within a preset future time period, including: The positioning data is compared with the road surface smoothness layer in the high-precision map database to generate the prior roughness index of the road segment ahead corresponding to the positioning data. Vibration feature extraction is performed on the attitude data to generate vibration energy spectrum data corresponding to the attitude data. Perform road surface semantic segmentation on the image data to generate data on the area ratio of road surface damage areas in the image data; The prior roughness index, vibration energy spectrum data, and area ratio data are input into the road surface bump intensity prediction model, and then the bump intensity prediction results are output.

[0033] In a specific embodiment, the construction and execution process of the road surface bump intensity prediction model aims to transform multi-source heterogeneous sensing data into a quantitative prediction of future bump intensity. Its core lies in revealing the coupling relationship between vehicle dynamic response and road surface geometry through precise physical modeling and feature extraction. Regarding the processing of positioning data, the system acquires high-precision latitude and longitude positioning data output by the vehicle-mounted global navigation satellite system receiver in real time. This positioning data is then spatially indexed and matched with the road surface smoothness layer in the high-precision map database pre-stored in the vehicle's edge computing unit. The road surface smoothness layer in the high-precision map database is generated based on laser point cloud data collected from multiple previous surveys. Each road segment is associated with an international roughness index, representing the cumulative degree of longitudinal road surface undulation. The specific implementation of spatial matching is as follows: a spatial buffer is established with the current positioning point as the center and a preset radius of 100 meters. The international roughness index sequence corresponding to all road segments within the buffer coverage area is extracted, and this sequence is averaged by distance to generate a prior roughness index for the road segment ahead corresponding to the current positioning point. The calculation formula for this prior roughness index is: in, The prior roughness index of the road section ahead; The number of the road segment within the buffer zone, with a value ranging from 1 to... ; For the first The higher the value of the International Roughness Index for each road segment, the rougher the road surface is. From the current location point to the th The Euclidean distance between the center points of each road segment is expressed in meters. The calculation formula uses an inverse distance-squared weighted average, aiming to assign higher weights to road segments closer to the current vehicle, as the roughness of closer road segments has a more direct impact on the driving experience in the near future. Prior roughness index. The value range is typically between 0.5 and 20 meters per kilometer. When the value is less than 2 meters per kilometer, the road surface can be considered extremely smooth, such as a highway surface; when... When the value is greater than 8 meters per kilometer, it indicates that you are about to enter a bumpy section of road, such as a gravel road or a severely damaged rural road.

[0034] Secondly, regarding attitude data processing, the system collects three-axis acceleration and three-axis angular velocity data in real time from the onboard inertial measurement unit at a sampling frequency of 100 Hz, which together constitute the original six-degree-of-freedom attitude data. Since the original data includes motion components caused by non-road excitations such as vehicle acceleration / deceleration and steering, decoupling processing is required. Specifically, a sliding window filtering algorithm is used, with a window length set to 1 second, containing 100 sampling points. The acceleration data within the window is detrended by subtracting the average value within the window, thereby eliminating low-frequency components caused by vehicle acceleration / deceleration and retaining high-frequency vibration components caused by road surface irregularities. Subsequently, a fast Fourier transform is performed on the processed acceleration data to convert the time-domain signal into a frequency-domain signal, generating vibration energy spectrum data. The core indicators of the vibration energy spectrum data are the total vibration energy and the dominant frequency distribution. Total vibration energy... The calculation formula is: in, The total vibrational energy is expressed in meters per square second squared, representing the intensity of the vibration. For frequency indexing, from 1 Hz to hertz, The value of is set to 50 Hz according to the Nyquist sampling theorem, which is half of the sampling frequency of 100 Hz. For frequency The corresponding power spectral density is obtained by dividing the square of the modulus of the Fourier transform by the frequency resolution. Total vibrational energy. The higher the value, the more severe the vibration the vehicle is currently experiencing. Simultaneously, the system extracts the frequency corresponding to the peak power in the vibration energy spectrum as the dominant frequency. The unit is Hertz. When a vehicle travels on a gravel road, the main frequency... The frequency is usually concentrated in the high-frequency range of 10 to 20 Hz; when driving on undulating roads, the frequency is concentrated in the low-frequency range of 1 to 5 Hz.

[0035] Next, regarding the processing of the road surface image data, the vehicle's forward-facing camera acquires RGB images at a rate of 30 frames per second. This image data is input into a lightweight deep learning semantic segmentation model. This model uses an improved MobileNetV3 as the backbone network, combined with the DeepLabV3+ decoder structure to achieve real-time inference on the vehicle's edge computing unit. In the model's output segmentation results, each pixel is assigned a category label, including background, lane lines, gravel areas, pothole areas, and crack areas. Based on the segmentation results, the system calculates the area proportion of the damaged road surface. The formula for calculating this ratio is: in, The value represents the proportion of the area occupied by the damaged road surface, and is a dimensionless coefficient with a value between 0 and 1. The total number of pixels in the image that are segmented into damage categories such as gravel, potholes, or cracks; This represents the total number of pixels in the image. This ratio directly quantifies the severity of road surface damage in the foreground view. For example, when... When the value is close to 0, it indicates that the road surface is basically in good condition; when... A value greater than 0.3 indicates that more than 30% of the road surface ahead is damaged, and the vehicle will soon experience severe bumps.

[0036] Finally, the prior roughness index calculated above is used... Total vibrational energy in vibrational energy spectrum data and clock speed and the proportion of road surface damage area By concatenating the features, a four-dimensional feature vector is formed. The feature vector is input into the road bump intensity prediction model, which outputs the predicted bump intensity of the target vehicle within a 5-second time window. The road bump intensity prediction model adopts a temporal convolutional network architecture, which consists of three one-dimensional convolutional layers, two gated recurrent unit layers, and one fully connected output layer. The kernel sizes of the three one-dimensional convolutional layers are 3, 5, and 3, respectively, with a stride of 1. The padding mode is set to "same" to maintain the sequence length, and the activation function is a linear rectified function. The hidden layer dimension of the gated recurrent unit layer is set to 64 to capture the temporal dependencies between features. The fully connected output layer contains one neuron, and the activation function is the sigmoid function, which compresses the output value to between 0 and 1, serving as the normalized bump intensity prediction result. The model is trained using supervised learning, with training data derived from historical driving records. Each training sample contains a feature vector sequence from the past second and a corresponding actual bump intensity label for the next 5 seconds. The actual bump intensity label is calculated using the root mean square value of vertical acceleration collected by the onboard accelerometer and normalized to the 0-1 range. The loss function is mean squared error, and the optimizer uses the Adam algorithm. The initial learning rate is set to 0.001, the batch size to 32, and the model is trained for 100 epochs. To prevent overfitting, a random deactivation layer with a dropout rate of 0.3 is added after the gated recurrent unit layer. After training, the model learns a non-linear mapping relationship from current perceived features to future bump intensity. The output is the predicted bump intensity. As the core basis for subsequent computing power allocation strategy decisions, when When the value is less than 0.3, the road condition is considered stable; when... A value between 0.3 and 0.7 indicates a moderately bumpy road condition; when... When the value is greater than 0.7, it is judged as a high-intensity bumpy road condition. This prediction mechanism ensures that the system can complete the dynamic scheduling preparation of computing resources about 5 seconds before the vehicle actually enters the bumpy road section, effectively overcoming the lag problem of traditional passive response schemes.

[0037] Furthermore, the preset at least two computing power allocation strategies include a first computing power allocation strategy, a second computing power allocation strategy, and a third computing power allocation strategy, wherein: The predicted bump intensity corresponding to the first computing power allocation strategy is less than the first preset threshold. The first computing power allocation strategy instructs the vehicle image acquisition unit to acquire video stream at the first frame rate as the first image data, instructs to send the first image data to the cloud computing power resource node as the first computing power resource node, and instructs the cloud computing power resource node to perform video diagnostic analysis operation as the first image processing operation. The bump intensity prediction result corresponding to the second computing power allocation strategy is between the first preset threshold and the second preset threshold. The second computing power allocation strategy instructs the vehicle image acquisition unit to acquire a continuous image sequence at a second frame rate higher than the first frame rate as the first image data, instructs the first image data to be sent to the edge computing power resource node as the first computing power resource node, and instructs the edge computing power resource node to perform inter-frame motion compensation and anti-shake processing operations as the first image processing operation. If the predicted bump intensity of the third computing power allocation strategy is greater than the second preset threshold, the third computing power allocation strategy instructs the vehicle-mounted image acquisition unit to acquire a continuous image sequence at the second frame rate as the first image data, and instructs the first image data to be sent to the edge computing power resource node as the first computing power resource node to perform the first image processing operation. While performing the first image processing operation, the continuous image sequence is pre-transmitted to the cloud computing power resource node as the original image data sequence for data backup in advance, so as to shorten the response delay when the computing power scheduling request is triggered later.

[0038] In one specific embodiment, the dynamic selection mechanism of the computing power allocation strategy is based on the bump intensity prediction results output by the road bump intensity prediction model. The comparison result with two preset thresholds triggers differentiated data acquisition, transmission, and processing links. The first preset threshold... With the second preset threshold The value is based on statistical analysis of a large amount of real vehicle road test data: when the root mean square value of vertical acceleration is less than The image quality is generally acceptable, given... to Lightweight image stabilization is required at times, greater than Deep repair must be initiated at this time. The above physical baseline is mapped to the normalized turbulence intensity prediction results through calibration experiments. ,set up , 7.

[0039] when The first computing power allocation strategy is executed: the vehicle-mounted image acquisition unit acquires video streams at a first frame rate of 30 frames per second, compresses them using H.265 encoding, and sends them directly to the cloud computing resource nodes via the 5G network. The cloud nodes deploy a real-time diagnostic analysis model based on the YOLOv8 architecture, identify and annotate key diagnostic areas in the video frames, and generate second image data with annotation information, which is then sent back to the vehicle terminal or synchronized to the consultation center.

[0040] when The second computing power allocation strategy is then implemented: the vehicle-mounted image acquisition unit switches to a high frame rate continuous shooting mode of 60 frames per second, sending the continuous image sequence to the edge computing resource nodes in real time. The edge nodes use an optical flow method based on the Farneback algorithm for inter-frame motion estimation, calculating the dense optical flow field and then performing an affine transformation on the parameter matrix. The current frame is subjected to inverse transformation compensation to generate a second image data after image stabilization and forward it to the cloud. At the same time, the image quality verification result is fed back to the vehicle terminal.

[0041] when A third computing power allocation strategy is implemented: the vehicle-mounted image acquisition unit also acquires a continuous image sequence at 60 frames per second, but activates two parallel links. The first link sends the images to the edge node in real time for image stabilization; the second link lightly compresses the same image sequence using JPEG-XL encoding and pre-transmits it to the cloud node as raw image data for backup in a circular buffer, with a storage depth set to 10 seconds. When subsequent image quality checks fail and trigger a computing power scheduling request, the cloud node directly reads the backed-up raw image sequence from the local buffer to initiate deep repair, reducing the original raw data acquisition latency of over 800 milliseconds to less than 15 milliseconds, ensuring end-to-end processing timeliness under high-intensity bumpy scenarios.

[0042] Furthermore, before the vehicle-mounted image acquisition unit acquires and outputs the first image data, the following steps are also included: Obtain the network status data of the communication network currently accessed by the target vehicle and the load status data of the first computing power resource node; When the network status data meets the preset network transmission conditions and the load status data meets the preset load conditions, the vehicle-mounted image acquisition unit is controlled to acquire and output the first image data according to the target computing power allocation strategy. When the network status data does not meet the preset network transmission conditions or the load status data does not meet the preset load conditions, the vehicle-mounted image acquisition unit is controlled to acquire and output the fourth image data according to the preset backup computing power allocation strategy, and the fourth image data is sent to the preset backup computing power resource node.

[0043] In one specific embodiment, to ensure the feasibility of the computing power allocation strategy under actual dynamic network environments and node load conditions, the system first quantitatively evaluates the communication network status and the load status of the first computing power resource node before executing the target computing power allocation strategy. Based on the evaluation results, it decides whether to execute the original strategy or switch to the backup strategy. The core of this pre-verification mechanism lies in constructing quantifiable network transmission conditions and load conditions. Its judgment is based on the quantitative decomposition of the real-time transmission requirements of remote medical care and the real-time monitoring of the processing capabilities of edge / cloud nodes.

[0044] First, regarding the acquisition and evaluation of communication network status data, the vehicle-mounted terminal sends three Internet Control Message Protocol (ICMP) echo request messages (ping packets) to the IP address of the first computing resource node (which may be an edge node or a cloud node) at a frequency of once per second, with each transmission interval being 200 milliseconds. Three round-trip delay sample values ​​are obtained by measuring the round-trip time from sending the request to receiving the response. , , The unit is milliseconds. Simultaneously, the standard deviation of these three sample values ​​is calculated as a network jitter metric. The formula for calculating network jitter is: in, This is the arithmetic mean of three round-trip delay samples, also in milliseconds. Network jitter metric. This reflects the stability of network latency; a higher value indicates a more unstable network. For real-time video transmission, The value should be as small as possible. Subsequently, the system calculates the round-trip delay based on the average of the most recent 10 measurements. and average jitter As the core representation of the current network state data, the preset network transmission conditions are defined as simultaneously satisfying two inequalities: and .in, This is the round-trip latency threshold, and its value is based on the end-to-end latency requirements of remote real-time diagnostics: to ensure the real-time performance of video interaction, the end-to-end latency should be less than 300 milliseconds. Considering the latency allocation in video encoding, transmission, decoding, and processing, the latency budget allocated to network transmission is 100 milliseconds, therefore, it is set as follows: millisecond. The jitter threshold is set to 20 milliseconds. This value ensures that the video playback buffer is not frequently interrupted or lost due to excessive jitter.

[0045] Secondly, regarding the acquisition and evaluation of the load status data of the first computing resource node, the vehicle terminal obtains the current CPU utilization of the node by sending a Hypertext Transfer Protocol (HTTP) GET request to the load monitoring API interface of the first computing resource node. Memory usage and the current active task queue length Among them, CPU utilization This is a percentage value, ranging from 0 to 100; memory usage. Also a percentage value; Task queue length This is a dimensionless integer representing the number of tasks currently queued and waiting to be processed by the node. To comprehensively evaluate node load, the system calculates a comprehensive load index. The calculation formula is: in, , , These are preset weighting coefficients, representing the contribution of CPU, memory, and task queue to the overall load, respectively. Based on typical bottleneck analysis of edge computing nodes, the CPU is often the primary bottleneck for image processing tasks; therefore, they are set as follows: Memory is secondary; settings The task queue has a relatively small impact; setting... . This is the maximum task queue length that a node can process in parallel, set according to the number of CPU cores configured on the node. For an edge node with an 8-core CPU, The value is 16, meaning each core handles an average of 2 tasks. Overall Load Index This is a dimensionless numerical value, ranging from 0 to 1, with values ​​closer to 1 indicating higher node load. The preset load condition is defined as follows: ,in The load threshold is set to 0.8. The logic behind this threshold is as follows: when the overall load exceeds 80%, the remaining processing capacity of the node is limited, and newly added tasks may face significant queuing delays, making it impossible to guarantee the real-time requirements of debouncing or diagnostic analysis. Therefore, the load condition is deemed not met.

[0046] When both of the above conditions are met, that is, when the network status data meets the preset network transmission conditions... milliseconds and And the load status data meets the preset load conditions. When the system determines that the current environment is suitable for executing the original target computing power allocation strategy, it controls the vehicle-mounted image acquisition unit to acquire and output the first image data according to the strategy.

[0047] Conversely, when network status data fails to meet preset network transmission conditions or load status data fails to meet preset load conditions, the system determines that the original strategy cannot be reliably executed and immediately switches to the preset backup computing power allocation strategy. The design principle of the backup computing power allocation strategy is "degradation protection," meaning that in situations of limited resources or network degradation, priority is given to ensuring the basic acquisition and local storage of critical diagnostic images, avoiding data loss due to the pursuit of real-time transmission. Specifically, the backup strategy instructs the vehicle-mounted image acquisition unit to operate at a lower third frame rate. Acquire fourth image data. Third frame rate. The setting is based on maximizing the acquisition time and minimizing the data volume under extreme conditions. In this embodiment, it is preferably set to 5 frames per second, only one-sixth of the normal frame rate. The acquired fourth image data is not transmitted in real time. Instead, it is compressed using the JPEG baseline format and stored in the vehicle terminal's large-capacity solid-state drive in a cyclic overwrite manner, with a storage depth set to 30 minutes. Simultaneously, a backup strategy sends the fourth image data to a preset backup computing resource node. This backup node is typically the embedded graphics processor (GPU) on the vehicle terminal itself, rather than relying on network edge or cloud nodes. The local GPU performs simplified image enhancement operations, such as histogram equalization and adaptive noise reduction filtering, generating images with slightly improved quality but still not meeting diagnostic standards. These images are then uploaded in batches after the network recovers. This backup mechanism ensures that the original image data can still be preserved under the worst communication conditions, providing a data foundation for subsequent supplementary acquisition or delayed diagnosis, and avoiding diagnostic blind spots caused by network or node failures.

[0048] Furthermore, the image quality verification result of the second image data is obtained, including: The second image data is input into the medical image region segmentation model to generate region location data of at least one key diagnostic region in the second image data corresponding to the preset diagnostic interest area; Obtain a standard, clear medical anatomical image template corresponding to the preset diagnostic focus area as reference image data; Calculate the weighted structural similarity index between the key diagnostic regions in the second image data and the corresponding key diagnostic regions in the reference image data, and use it as the image quality verification result; wherein, the weight of the weighted structural similarity index is higher in the key diagnostic regions than in the non-key diagnostic regions.

[0049] Furthermore, a weighted structural similarity index is calculated between the key diagnostic regions in the second image data and the corresponding key diagnostic regions in the reference image data, including: Obtain the diagnostic part identifier corresponding to the remote diagnostic task currently being performed by the target vehicle; Based on the diagnostic site identifier, retrieve the anatomical structure region data corresponding to the diagnostic site identifier from the preset medical knowledge base; The anatomical structure region data is registered with the second image data to generate the location data of the target key region in the second image data corresponding to the diagnostic site identifier; Based on the turbulence intensity prediction result, a first weight value corresponding to the target key region and a second weight value corresponding to the non-target key region in the second image data are determined, wherein the first weight value is higher than the second weight value. Based on the first weight value and the second weight value, calculate the weighted structural similarity index between the second image data and the reference image data.

[0050] In one specific embodiment, the generation of image quality verification results and the calculation of the weighted structural similarity index are the core links in realizing the closed-loop feedback repair mechanism. The purpose is to transform the subjective concept of "image clarity" into a quantifiable and comparable objective mathematical indicator, and to ensure that the indicator focuses on the anatomical areas that are truly important for remote diagnosis.

[0051] First, for the extraction of key diagnostic regions from the second image data, the system inputs the image data after image stabilization into a pre-trained medical image region segmentation model. This model uses a U-Net architecture and is trained on a training set containing 5000 sets of emergency trauma images. Each set of data has been annotated by experienced radiologists with key diagnostic sites such as open wounds, pupil areas, and fracture ends. When the second image data is input into the model, it outputs a probability map of the same size as the input image. The value of each pixel in the map represents the probability that the pixel belongs to a predefined diagnostic area of ​​interest. The system binarizes this probability map by setting a segmentation threshold of 0.5 and extracts the contours of the binarized connected components to generate region location data for at least one key diagnostic region. This data is stored in the form of a list of polygon vertex coordinates.

[0052] Secondly, regarding the acquisition of reference image data, the system retrieves standard, clear anatomical image templates corresponding to the preset diagnostic focus areas from the local storage of the vehicle terminal or the cloud-based medical knowledge base, based on the nature of the current remote diagnostic task. These templates are not fixed but dynamically matched according to the diagnostic site identifier. For example, if the current task is identified as "trauma assessment," the retrieved standard template is a high-resolution, distortion-free, and uniformly lit typical wound image; if the task is identified as "pupil light reflex examination," the retrieved standard template is a close-up image of the eye with clear and sharp pupil boundaries and visible iris texture. These standard template images have all been reviewed by medical experts to ensure that their anatomical structures are typical and the image quality meets diagnostic requirements. Their resolution is uniformly standardized to the same 1920×1080 pixels as the acquired images to eliminate the impact of scale differences on subsequent similarity calculations.

[0053] Next, for the calculation of the weighted structural similarity index, the system first obtains the diagnostic site identifier corresponding to the remote diagnostic task currently being performed by the target vehicle, such as "pupil examination". Based on this identifier, the system retrieves the anatomical structure region data corresponding to the diagnostic site identifier from the medical knowledge base. Subsequently, the system uses a registration algorithm based on mutual information to find the spatial transformation matrix through iterative optimization, spatially aligning the second image data with the standard template image, thereby generating positional data on the standard template that precisely corresponds to the target key region in the second image.

[0054] After completing region registration, the system calculates the weighted structural similarity index (SSIM). The traditional SSIM calculation is based on the brightness, contrast, and structural similarity of local image windows. In this embodiment, a spatial variable weighting mechanism is introduced to focus the evaluation on key diagnostic regions. The system determines a first weight value corresponding to the target key region and a second weight value corresponding to non-target key regions based on the turbulence intensity prediction results and diagnostic site identification. The logic for setting the weight values ​​is: the more severe the turbulence, the higher the weight of the key region should be. The specific calculation formula is as follows: ,in As key area weights, The base weight value is determined based on the diagnostic site identification; for example, the base weight for the pupil area is set to 3.0. The bump impact factor is preset to 1.5; This is the prediction result for turbulence intensity. Weighting of non-critical areas. Fixed at 1.

[0055] After obtaining the weights, the system divides the entire image into multiple non-overlapping 8×8 pixel windows, calculates the SSIM value for each window, and assigns different weights based on whether the window center falls within the critical diagnostic region. The formula for calculating the weighted SSIM is as follows: ,in, The total number of windows, For the first The weight of each window, For the first The SSIM value calculated for each window and the corresponding window in the standard template. This is the final image quality verification result, with a value ranging from 0 to 1. The closer to 1, the higher the image quality.

[0056] Furthermore, when the image quality verification result does not meet the preset image quality conditions, a computing power scheduling request carrying the original image data identifier is generated, including: When the weighted structural similarity index is less than the preset medical image quality standard threshold, the image quality verification result is determined to not meet the preset image quality conditions. Obtain the original burst image data corresponding to the second image data, and use it as the original image data pointed to by the original image data identifier; A computing power scheduling request is generated, which includes the original image data, the motion vector parameters corresponding to the second image data, and the image quality verification result.

[0057] Furthermore, the second image processing operation includes a temporal repair operation based on a generative adversarial network; the third image data is generated by the second computing power resource node in the following manner: Receive computing power scheduling request, parse and obtain the original image data sequence pointed to by the original image data identifier, the original image data sequence contains multiple consecutive frames of blurred image data; Obtain the real-time vibration frequency data of the target vehicle when the original image data sequence is acquired; The original image data sequence and the real-time vibration frequency data are input into a pre-trained generative adversarial network model. The generator in the generative adversarial network model generates repaired image data for at least one target frame in the original image data sequence. The repaired image data is combined with the image data of other frames in the original image data sequence, excluding the target frame, through temporal smoothing to generate a third image data containing the repaired image data.

[0058] In one specific embodiment, temporal restoration based on generative adversarial networks is the core technology for image quality restoration under high-intensity bumpy road conditions. After receiving a computing power scheduling request, the cloud computing resource node parses and obtains the original image data identifier. Based on this identifier, it reads the corresponding original image data sequence from a pre-set circular buffer. According to the third computing power allocation strategy, this sequence consists of 300 frames (5 seconds of continuous images pre-backed up before the vehicle enters the high-intensity bumpy road section), each frame with a resolution of 1920×1080 pixels, lightly compressed using JPEG-XL encoding with a compression ratio of 4:1. The decoded image sequence constitutes a five-dimensional tensor of shape [300, 1080, 1920, 3].

[0059] Meanwhile, the cloud node parses the real-time vibration frequency data corresponding to the image acquisition time from the computing power scheduling request. This data originates from the vehicle-mounted inertial measurement unit and is recorded at a sampling rate of 100 Hz. Since the image frame rate of 60 frames per second is inconsistent with the vibration sampling rate, time alignment is required: taking the acquisition timestamp of each frame as the center, vibration sampling points within a 5-millisecond window before and after are taken, and the vibration dominant frequency characteristic value associated with that frame is calculated using an inverse distance-weighted average. This yields vibration feature vectors corresponding one-to-one with 300 frames of images. The unit for each element is Hertz, with values ​​ranging from 1 to 50 Hertz.

[0060] Subsequently, the original image data sequence and vibration feature vectors are input into a pre-trained generative adversarial network (GAN) model. This model employs an improved Temporal-GAN architecture, with the generator consisting of a spatiotemporal encoder and decoder. The encoder compresses the input sequence into a spatiotemporal feature tensor through a 3D convolutional layer. The vibration feature vector, after being mapped by a fully connected layer, undergoes an affine transformation on the encoder output through a feature linear modulation layer, thus guiding the generation process based on vibration conditions. The decoder gradually restores the image resolution and outputs the repaired image sequence. The model training data comes from 100,000 pairs of clear-blurred image sequences collected from a simulated bumpy platform. The loss function consists of a weighted sum of adversarial loss, content loss, and perceptual loss, with weights set to 1, 100, and 10, respectively. During generator inference, an attention mechanism automatically identifies the keyframe with the most severe motion blur as the target frame, allocating more computational resources for repair.

[0061] While the restored image sequence output by the generator significantly improves the quality of individual frames, brightness jitter may exist between adjacent frames. To address this, the system performs temporal smoothing: it calculates the bidirectional optical flow between each restored image frame and the preceding and following frames, maps the pixels of the preceding and following frames to the coordinate system of the current frame based on the optical flow information, and generates the final output third image data through a weighted average, with a smoothing coefficient set to 0.3. This processing ensures temporal continuity while maintaining the sharpness of individual frames.

[0062] Through actual measurement and verification, the peak signal-to-noise ratio (PSNR) of a target frame in the original blurred image sequence was 18.5 dB, and the structural similarity index was 0.62. After generative adversarial network (GAN) repair, the two indicators were improved to 31.2 dB and 0.93, respectively. After temporal smoothing, the PSNR slightly decreased to 30.8 dB, but the structural similarity index remained unchanged at 0.93, and the standard deviation of the PSNR difference between adjacent frames decreased from 2.5 dB to 0.4 dB. This proves that the image sequence has been significantly optimized in temporal stability, meeting the dual requirements of remote diagnosis for image quality and viewing continuity.

[0063] Furthermore, the generator in the generative adversarial network model generates repaired image data for at least one target frame in the original image data sequence, including: The attention mechanism is used to identify key regions in the original image data sequence that correspond to preset diagnostic areas of interest. When generating the repaired image data, higher computing resources are allocated to repair the critical regions than to non-critical regions.

[0064] In one specific embodiment, the introduction of an attention mechanism in the generative adversarial network model aims to prioritize the allocation of limited computational resources to areas critical to remote diagnosis, such as wounds, pupils, or fracture ends, thereby maximizing diagnostic value under overall resource constraints.

[0065] First, regarding input feature extraction for the attention mechanism, the encoder part of the generator transforms the input blurred image sequence into multi-scale spatiotemporal feature maps through multiple layers of 3D convolution operations. Taking a certain intermediate layer in the generator as an example, the feature map output by this layer has a dimension of... These correspond to the number of time frames, the number of channels, the height, and the width, respectively. This feature map contains abstract spatiotemporal information about the input sequence and forms the basis for subsequent attention calculations.

[0066] Secondly, regarding the generation of attention weights, the system inputs the feature map into a lightweight attention sub-network. This network consists of two stacked 1×1×1 convolutional layers, compressing the number of channels in the feature map first to 64, then to 1, outputting a single-channel attention score map. A sigmoid activation function is then applied to this score map, mapping the value of each element to between 0 and 1, generating a self-learning attention weight map. The closer the value is to 1, the more important the spatiotemporal location is.

[0067] However, relying solely on self-learning attention may deviate from medical priors. Therefore, the system introduces a guidance mechanism: projecting the key diagnostic region location data generated earlier by the medical image segmentation model onto the current feature map scale. Since the original image resolution is 1920×1080, and the current feature map resolution is one-quarter of the original (480×270), a binary prior mask is generated through coordinate mapping. The value is assigned as 1 at the spatiotemporal location corresponding to the key area, and 0 for the rest.

[0068] Subsequently, the self-learning attention weight map will be used. With prior mask We perform weighted fusion to obtain the attention-guided weight map. : The fusion coefficient of 0.6 was set based on experimental optimization, aiming to balance the model's ability to automatically discover potentially important regions with the guidance of prior medical knowledge.

[0069] After obtaining the guided attention weight map, the system applies it to subsequent calculations in the generator. Specifically, a feature recalibration method is used: the guided attention weight map is multiplied element-wise with the original feature map to obtain a weighted feature map. Through this operation, feature values ​​corresponding to key regions are preserved or even amplified, while feature values ​​of non-key regions are suppressed. The weighted feature map is then fed into the decoder, and in the final generated restored image, the restoration quality of key regions is significantly better than that of non-key regions.

[0070] Furthermore, after generating repaired image data for at least one target frame in the original image data sequence through the generator in the generative adversarial network model, the method further includes: The repaired image data is input into a preset secondary quality verification model to generate a secondary quality verification result for the repaired image data. When the result of the secondary quality check is less than the preset secondary quality threshold, the adjacent frame data of the target frame in the original image data sequence are obtained; Based on the optical flow information between the adjacent frame data and the repaired image data, an iterative optimization operation is performed on the repaired image data to generate optimized repaired image data, until the secondary quality check result of the optimized repaired image data is not less than the preset secondary quality threshold or the number of iterations reaches the preset number threshold.

[0071] In one specific embodiment, the secondary quality verification and iterative optimization mechanism is a closed-loop control link that further ensures that the image quality meets the diagnostic requirements after the initial repair by the generative adversarial network.

[0072] First, for secondary quality verification, the system inputs the restored image data output by the generator into a pre-defined secondary quality verification model. This model employs a BRISQUE improved architecture based on deep regression, trained on a database containing 20,000 medical images, and outputs a quality score ranging from 0 to 100. A higher score indicates better image quality. The preset secondary quality threshold... The value was set at 75, which comes from clinical testing: when the score is below 75, the accuracy of physicians in identifying key lesions decreases significantly.

[0073] when When the system determines that the repaired image meets the diagnostic requirements, it is directly output as the third image data. When this happens, the iterative optimization process is triggered.

[0074] The core of iterative optimization is to use redundant information from temporally adjacent frames to correct the current frame. The system acquires the target frame. adjacent frames before and after and The Farneback algorithm was used to calculate the optical flow field. and The unit is pixels. Based on optical flow information, pixels from adjacent frames are inversely mapped to the current frame coordinate system, generating two prediction frames. and The current frame and the predicted frame are then weighted and fused to generate the optimized restored image. : The fusion weight is 0.3, which was determined based on experimental optimization to achieve a balance between suppressing noise and preserving details.

[0075] Will Re-enter the quality assessment model to obtain a new quality score. .like The iteration terminates and outputs. ;like And if the number of iterations is less than the preset threshold of 5, then... Repeat the above process as the new current frame until the quality meets the standard or the maximum number of iterations is reached.

[0076] Example 2: This application also discloses a smart medical computing power network resource dynamic allocation system.

[0077] Reference Figure 2 A smart healthcare computing power network resource dynamic allocation system, applied to vehicle-mounted terminals, includes: The data acquisition module is used to acquire the target vehicle's current location data, attitude data, and image data of the road surface ahead; The bump intensity prediction module is used to input positioning data, attitude data and image data into the road bump intensity prediction model to generate the bump intensity prediction result of the target vehicle in a preset future time period. The allocation strategy module is used to determine a target computing power allocation strategy from at least two preset computing power allocation strategies based on the comparison between the turbulence intensity prediction result and the preset intensity threshold. The control acquisition module is used to control the vehicle-mounted image acquisition unit to acquire raw image data according to the target computing power allocation strategy, and output the raw image data as the first image data. The image processing module is used to send the first image data to the first computing power resource node indicated by the target computing power allocation strategy, so that the first computing power resource node performs a preset first image processing operation on the first image data and outputs the second image data. The verification result module is used to obtain the image quality verification result of the second image data. The scheduling module is used to generate a computing power scheduling request carrying the original image data identifier when the image quality verification result does not meet the preset image quality conditions, and send the computing power scheduling request to the second computing power resource node; the computing power capability of the second computing power resource node is greater than that of the first computing power resource node. The receiving module is used to receive the third image data returned by the second computing power resource node. The third image data is generated by the second computing power resource node in response to the computing power scheduling request, after obtaining the corresponding original image data sequence according to the original image data identifier, and performing a preset second image processing operation on the original image data sequence. The image quality of the third image data is higher than that of the second image data.

[0078] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0079] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0080] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A method for dynamically allocating computing power network resources in smart healthcare, characterized in that, Applied to vehicle-mounted terminals, the method includes the following steps: Acquire the target vehicle's current location data, attitude data, and image data of the road surface ahead; The positioning data, attitude data and image data are input into the road bump intensity prediction model to generate the bump intensity prediction result of the target vehicle in a preset future time period. Based on the comparison between the predicted turbulence intensity and the preset intensity threshold, a target computing power allocation strategy is determined from at least two preset computing power allocation strategies. According to the target computing power allocation strategy, the vehicle-mounted image acquisition unit is controlled to acquire raw image data and output the raw image data as the first image data. The first image data is sent to the first computing power resource node indicated by the target computing power allocation strategy, so that the first computing power resource node performs a preset first image processing operation on the first image data and outputs the second image data. Obtain the image quality verification result of the second image data; When the image quality verification result does not meet the preset image quality conditions, a computing power scheduling request carrying the original image data identifier is generated and sent to the second computing power resource node; the computing power capability of the second computing power resource node is greater than that of the first computing power resource node. The third image data returned by the second computing power resource node is generated by the second computing power resource node in response to the computing power scheduling request, after obtaining the corresponding original image data sequence according to the original image data identifier and performing a preset second image processing operation on the original image data sequence. The image quality of the third image data is higher than that of the second image data.

2. The method for dynamic allocation of intelligent medical computing power network resources according to claim 1, characterized in that, The positioning data, attitude data, and image data are input into the road bump intensity prediction model to generate a bump intensity prediction result for the target vehicle within a preset future time period, including: The positioning data is compared with the road surface smoothness layer in the high-precision map database to generate the prior roughness index of the road segment ahead corresponding to the positioning data. Vibration feature extraction is performed on the attitude data to generate vibration energy spectrum data corresponding to the attitude data. Perform road surface semantic segmentation on the image data to generate data on the area ratio of road surface damage areas in the image data; The prior roughness index, vibration energy spectrum data, and area ratio data are input into the road surface bump intensity prediction model, and then the bump intensity prediction results are output.

3. The method for dynamic allocation of intelligent medical computing power network resources according to claim 1, characterized in that, The preset at least two computing power allocation strategies include a first computing power allocation strategy, a second computing power allocation strategy, and a third computing power allocation strategy, wherein: The predicted bump intensity corresponding to the first computing power allocation strategy is less than the first preset threshold. The first computing power allocation strategy instructs the vehicle image acquisition unit to acquire video stream at the first frame rate as the first image data, instructs to send the first image data to the cloud computing power resource node as the first computing power resource node, and instructs the cloud computing power resource node to perform video diagnostic analysis operation as the first image processing operation. The bump intensity prediction result corresponding to the second computing power allocation strategy is between the first preset threshold and the second preset threshold. The second computing power allocation strategy instructs the vehicle image acquisition unit to acquire a continuous image sequence at a second frame rate higher than the first frame rate as the first image data, instructs the first image data to be sent to the edge computing power resource node as the first computing power resource node, and instructs the edge computing power resource node to perform inter-frame motion compensation and anti-shake processing operations as the first image processing operation. If the predicted bump intensity of the third computing power allocation strategy is greater than the second preset threshold, the third computing power allocation strategy instructs the vehicle-mounted image acquisition unit to acquire a continuous image sequence at the second frame rate as the first image data, and instructs the first image data to be sent to the edge computing power resource node as the first computing power resource node to perform the first image processing operation. While performing the first image processing operation, the continuous image sequence is pre-transmitted to the cloud computing power resource node as the original image data sequence for data backup in advance, so as to shorten the response delay when the computing power scheduling request is triggered later.

4. The method for dynamic allocation of intelligent medical computing power network resources according to claim 1, characterized in that, Before the vehicle-mounted image acquisition unit acquires and outputs the first image data, the following steps are also included: Obtain the network status data of the communication network currently accessed by the target vehicle and the load status data of the first computing power resource node; When the network status data meets the preset network transmission conditions and the load status data meets the preset load conditions, the vehicle-mounted image acquisition unit is controlled to acquire and output the first image data according to the target computing power allocation strategy. When the network status data does not meet the preset network transmission conditions or the load status data does not meet the preset load conditions, the vehicle-mounted image acquisition unit is controlled to acquire and output the fourth image data according to the preset backup computing power allocation strategy, and the fourth image data is sent to the preset backup computing power resource node.

5. The method for dynamic allocation of intelligent medical computing power network resources according to claim 1, characterized in that, Obtain the image quality verification results of the second image data, including: The second image data is input into the medical image region segmentation model to generate region location data of at least one key diagnostic region in the second image data corresponding to the preset diagnostic interest area; Obtain a standard, clear medical anatomical image template corresponding to the preset diagnostic focus area as reference image data; Calculate the weighted structural similarity index between the key diagnostic regions in the second image data and the corresponding key diagnostic regions in the reference image data, and use it as the image quality verification result; wherein, the weight of the weighted structural similarity index is higher in the key diagnostic regions than in the non-key diagnostic regions.

6. The method for dynamic allocation of intelligent medical computing power network resources according to claim 1, characterized in that, When the image quality verification result does not meet the preset image quality conditions, a computing power scheduling request carrying the original image data identifier is generated, including: When the weighted structural similarity index is less than the preset medical image quality standard threshold, the image quality verification result is determined to not meet the preset image quality conditions. Obtain the original burst image data corresponding to the second image data, and use it as the original image data pointed to by the original image data identifier; A computing power scheduling request is generated, which includes the original image data, the motion vector parameters corresponding to the second image data, and the image quality verification result.

7. The method for dynamic allocation of intelligent medical computing power network resources according to claim 1, characterized in that, The second image processing operation includes a temporal repair operation based on a generative adversarial network; the third image data is generated by the second computing power resource node in the following way: Receive computing power scheduling request, parse and obtain the original image data sequence pointed to by the original image data identifier, the original image data sequence contains multiple consecutive frames of blurred image data; Obtain the real-time vibration frequency data of the target vehicle when the original image data sequence is acquired; The original image data sequence and the real-time vibration frequency data are input into a pre-trained generative adversarial network model. The generator in the generative adversarial network model generates repaired image data for at least one target frame in the original image data sequence. The repaired image data is combined with the image data of other frames in the original image data sequence, excluding the target frame, through temporal smoothing to generate a third image data containing the repaired image data.

8. The method for dynamic allocation of intelligent medical computing power network resources according to claim 7, characterized in that, The generator in the generative adversarial network model generates repaired image data for at least one target frame in the original image data sequence, including: The attention mechanism is used to identify key regions in the original image data sequence that correspond to preset diagnostic areas of interest. When generating the repaired image data, higher computing resources are allocated to repair the critical regions than to non-critical regions.

9. A method for dynamic allocation of intelligent medical computing power network resources according to claim 5, characterized in that, Calculate the weighted structural similarity index between the key diagnostic regions in the second image data and the corresponding key diagnostic regions in the reference image data, including: Obtain the diagnostic part identifier corresponding to the remote diagnostic task currently being performed by the target vehicle; Based on the diagnostic site identifier, retrieve the anatomical structure region data corresponding to the diagnostic site identifier from the preset medical knowledge base; The anatomical structure region data is registered with the second image data to generate the location data of the target key region in the second image data corresponding to the diagnostic site identifier; Based on the turbulence intensity prediction result, a first weight value corresponding to the target key region and a second weight value corresponding to the non-target key region in the second image data are determined, wherein the first weight value is higher than the second weight value. Based on the first weight value and the second weight value, calculate the weighted structural similarity index between the second image data and the reference image data.

10. A method for dynamic allocation of intelligent medical computing power network resources according to claim 8, characterized in that, After generating repaired image data for at least one target frame in the original image data sequence using the generator in the generative adversarial network model, the method further includes: The repaired image data is input into a preset secondary quality verification model to generate a secondary quality verification result for the repaired image data. When the result of the secondary quality check is less than the preset secondary quality threshold, the adjacent frame data of the target frame in the original image data sequence are obtained; Based on the optical flow information between the adjacent frame data and the repaired image data, an iterative optimization operation is performed on the repaired image data to generate optimized repaired image data, until the secondary quality check result of the optimized repaired image data is not less than the preset secondary quality threshold or the number of iterations reaches the preset number threshold.

11. A smart healthcare computing power network resource dynamic allocation system, applied to the smart healthcare computing power network resource dynamic allocation method described in any one of claims 1-10, characterized in that, include: The data acquisition module is used to acquire the target vehicle's current location data, attitude data, and image data of the road surface ahead; The bump intensity prediction module is used to input positioning data, attitude data and image data into the road bump intensity prediction model to generate the bump intensity prediction result of the target vehicle in a preset future time period. The allocation strategy module is used to determine a target computing power allocation strategy from at least two preset computing power allocation strategies based on the comparison between the turbulence intensity prediction result and the preset intensity threshold. The control acquisition module is used to control the vehicle-mounted image acquisition unit to acquire raw image data according to the target computing power allocation strategy, and output the raw image data as the first image data. The image processing module is used to send the first image data to the first computing power resource node indicated by the target computing power allocation strategy, so that the first computing power resource node performs a preset first image processing operation on the first image data and outputs the second image data. The verification result module is used to obtain the image quality verification result of the second image data. The scheduling module is used to generate a computing power scheduling request carrying the original image data identifier when the image quality verification result does not meet the preset image quality conditions, and send the computing power scheduling request to the second computing power resource node. The computing power of the second computing resource node is greater than that of the first computing resource node; The receiving module is used to receive the third image data returned by the second computing power resource node. The third image data is generated by the second computing power resource node in response to the computing power scheduling request, after obtaining the corresponding original image data sequence according to the original image data identifier, and performing a preset second image processing operation on the original image data sequence. The image quality of the third image data is higher than that of the second image data.

Images