Distributed neural network model division method for edge video analysis

Abstract

The invention discloses a distributed neural network model division method for edge video analysis, which relates to the technical field of edge computing, and comprises the following steps: extracting visual features of video frames, and classifying the video frames; dynamically determining a division mode of a neural network model according to a classification result of the video frames, and dynamically adjusting a task allocation proportion between different devices; monitoring the resource use condition of each edge device in real time, and carrying out quantitative evaluation on the resource availability of the device to obtain a device resource score; according to the resource scores of the edge devices, dynamically adjusting the division of the distributed neural network model, and carrying out fusion processing on the visual features of the edge devices; and the model division effect is evaluated in combination with the equipment resource score. According to the method, the problems of low accuracy and poor real-time performance in the edge video analysis process are solved, the calculation load of different edge devices is effectively balanced, and the model division strategy is flexibly adjusted according to the dynamic change of the video data.

Description

Technical Field

[0001] The present invention relates to the field of edge computing technology, and more specifically, to a distributed neural network model partitioning method for edge video analysis. Background Art

[0002] Edge video analysis refers to the collection, processing, analysis and decision-making of video data directly on the edge of the network, such as cameras, edge servers and other devices close to the video data source, rather than transmitting all video data to the cloud or data center for processing. It uses the computing power and storage resources of edge devices to analyze the content in the video locally in real time and extract valuable information. In many current application scenarios involving video analysis, the demand for real-time processing and analysis of video data is growing. Edge computing has gradually become a key technical means to solve the real-time problem of video analysis because it can process data close to the data source, effectively reduce data transmission delays, and improve system response speed.

[0003] Deficiencies of existing technologies: Edge devices usually face the challenge of limited computing resources, storage resources, and energy resources, and it is difficult to independently process complex neural network models to achieve efficient video analysis. Traditional neural network model partitioning methods often do not fully consider the special needs and complex situations in edge video analysis scenarios. For example, in some industrial production scenarios, video data not only contains regular equipment operation pictures, but also complex and uncommon pictures such as abnormal light and shadow flickering and smoke caused by equipment failure. The characteristics of these abnormal pictures are very different from normal pictures, and their distribution in time and space is random. When dealing with such scenes, the existing model partitioning methods cannot effectively balance the computing load of different edge devices, and it is difficult to flexibly adjust the model partitioning strategy according to the dynamic changes of video data, resulting in low analysis accuracy and poor real-time performance.

[0004] In view of the above problems, the present invention proposes a solution. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a distributed neural network model partitioning method for edge video analysis, which solves the problems raised in the above-mentioned background technology by providing a distributed neural network model partitioning method for edge video analysis.

[0006] To achieve the above object, the present invention provides the following technical solutions: A distributed neural network model partitioning method for edge video analysis includes the following steps: Preprocess the input video data, extract the visual features of the video frames according to the local binary pattern (LBP) algorithm, and classify the video frames using the support vector machine classification model; According to the classification results of the video frames, the division method of the neural network model is dynamically determined, and the task allocation ratio between different devices is dynamically adjusted; Monitor the resource usage of each edge device in real time, establish a resource evaluation model, quantitatively evaluate the resource availability of the device, and obtain the device resource score; According to the resource scores of edge devices, the division of the distributed neural network model is dynamically adjusted, and the visual features of each edge device are fused. The distributed neural network model is optimized according to the results of fusion processing, and the model partitioning effect is evaluated in combination with the device resource score.

[0007] In a preferred embodiment, the visual features include texture features and color features; the texture feature extraction process is as follows: Preprocess the input video data; The texture features of the video frame are extracted by the local binary pattern (LBP) algorithm through the lightweight feature extractor on the edge device. The calculation formula is as follows: ; In the formula, is the center pixel coordinate, is the number of neighborhood pixels, where is a sign function, when hour, ;when hour, , It is Device texture features.

[0008] In a preferred embodiment, the color feature extraction process is as follows: The color representation of the video frame is in RGB space (red, green, blue); For each frame, histograms are calculated for the R (red), G (green), and B (blue) channels; The histogram of each video frame is normalized by dividing each frequency value of the histogram by the total number of pixels so that the histogram has a uniform scale; The color features of the video frame are obtained by analyzing and processing the color histogram of each video frame .

[0009] In a preferred embodiment, the process of dynamically determining the division mode of the neural network model is as follows: When multiple consecutive frames are detected as normal frames, the part of the model with relatively small computational workload and low reliance on global features is allocated to edge devices with relatively weak computing resources for processing; Assign the parts with large computational workload and need to integrate global information to edge devices or the cloud with strong computing power for processing; When an abnormal frame is detected, due to the complexity and importance of the features of the abnormal frame, more computing tasks are concentrated on devices with stronger computing power to ensure that the abnormal features can be accurately extracted.

[0010] In a preferred embodiment, the process of dynamically adjusting the task allocation ratio between different devices is as follows: According to the frequency of abnormal frames and distribution, dynamically adjust the task allocation ratio between different devices; If the frequency of abnormal frames is greater than the set frequency threshold, abnormal frames appear frequently, and the task load of devices with stronger computing power needs to be increased; If the frequency of abnormal frames is less than the set frequency threshold, the abnormal frames are relatively sparse. Under the premise of ensuring accuracy, tasks should be reasonably allocated to improve overall efficiency.

[0011] In a preferred embodiment, the process of obtaining the device resource score is as follows: The resource usage includes CPU usage, remaining storage space and device battery power; By establishing a resource evaluation model, the resource availability of each device is quantitatively evaluated to obtain the device resource score; the specific calculation formula is as follows: ;In the formula It is Equipment resource rating, It is CPU usage of each device, It is CPU usage of each device, It is The remaining storage space of the device It is Total storage space of devices, It is Device storage space weighting factor, It is Device battery level, It is The device battery power weighting factor.

[0012] In a preferred embodiment, according to the resource score of the edge device, the partitioning process of the distributed neural network model is dynamically adjusted as follows: When the resource score of a device is less than the set score threshold, reduce the computing tasks assigned to the device and transfer the tasks to devices with higher resource scores; When the resource score of a device is greater than the set score threshold, there is no need to reduce the computing tasks assigned to the device.

[0013] In a preferred embodiment, the process of fusing the visual features of each edge device is as follows: After the distributed neural network model is divided, the visual features obtained by different devices when processing video data are fused. The specific calculation formula is as follows: ;In the formula It is the fusion of visual features. It is Texture features of the device, It is The texture feature weights of each device, It is The color characteristics of each device, It is The color feature weights of each device.

[0014] In a preferred embodiment, it is characterized in that the distributed neural network model is optimized according to the result of the fusion processing as follows: By calculating the error between the analysis result and the true label, the error information is fed back to the sub-model on each device using the back-propagation algorithm; Adopt a dynamic weight adjustment strategy to adjust the weight of the sub-model in the model optimization process according to its performance in different scenarios; And the fused visual feature weights of each device are optimized according to the adjusted weights in the model optimization process.

[0015] In a preferred embodiment, the steps of evaluating the model partitioning effect are as follows: The total equipment resource score and fused visual features are defined as input variables, and they are divided into different fuzzy sets respectively; The model partitioning effect is defined as the output variable, which is divided into fuzzy sets; Formulate fuzzy rules to describe the impact of total equipment resource score and fused visual features on model partitioning effect; Perform fuzzy reasoning based on fuzzy rules and evaluate the model partitioning effect.

[0016] Technical effects and advantages of a distributed neural network model partitioning method for edge video analysis of the present invention: 1. The present invention can share the computing load by dividing the neural network model into multiple sub-models and distributing them to different edge devices, so that each device only needs to process part of the computing tasks, avoiding the performance bottleneck of a single device. The edge device can process the tasks of the sub-model in parallel, thereby improving the overall computing efficiency and response speed. Through model partitioning, the communication requirements between the edge device and the cloud or central server can be reduced. Only part of the data and model parameters need to be transmitted, which greatly reduces the pressure on bandwidth. This not only saves bandwidth resources, but also reduces the performance degradation caused by network delay and improves real-time performance. Edge video analysis often requires low latency, especially in real-time analysis of video streams. By performing local reasoning and processing on the edge device, the amount of data that needs to be transmitted and processed can be reduced, improving real-time performance. Even in an environment with poor network conditions, distributed model partitioning can still ensure that each edge device can run efficiently and meet the needs of real-time analysis. The computing power of the edge device is limited compared to the cloud server. By dividing the model into multiple smaller parts, it can adapt to the computing power of different edge devices and ensure that each device can run efficiently. The partitioning method can be adjusted according to the performance characteristics of different devices to further improve the processing power of the edge device.

[0017] 2. The present invention distributes the neural network model to multiple edge devices. Even if a device fails, the system can still process it through other devices, ensuring the fault tolerance of the system.

[0018] This also improves the robustness of the system, especially in complex edge computing scenarios, ensuring the stable operation of video analysis tasks. This method allows the dynamic addition or reduction of edge devices as needed, adapting to edge computing networks of different sizes. Model partitioning can be flexibly adjusted according to the addition or exit of devices. When dealing with large-scale video stream analysis, it can be efficiently expanded to support large-scale distributed edge computing. By distributing computing tasks on multiple edge devices, device resources can be used more effectively and excessive load on a single device can be avoided. This saves computing resources and reduces the energy consumption of the device.

[0019] This distributed neural network model partitioning method provides an efficient, scalable, and low-latency solution for edge video analysis. By reasonably dividing the computing tasks, it can not only fully utilize the computing power of edge devices, but also reduce the burden of network bandwidth and improve the real-time and robustness of the system. These advantages make this technology widely used in scenarios that require edge computing support, such as real-time video surveillance, intelligent transportation, and security. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural schematic diagram of a distributed neural network model partitioning method for edge video analysis of the present invention. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] Embodiment 1, Figure 1 The present invention provides a distributed neural network model partitioning method for edge video analysis.

[0023] S10, preprocessing the input video data, extracting visual features of the video frames according to a local binary pattern (LBP) algorithm, and classifying the video frames using a support vector machine classification model; The visual features include texture features and color features; The input video data is preprocessed, and the texture features of the video frame are extracted using the local binary pattern (LBP) algorithm through the lightweight feature extractor on the edge device. The calculation formula is as follows: ; In the formula, is the center pixel coordinate, is the number of neighborhood pixels, where is a sign function, when hour, ;when hour, .

[0024] The process of counting the color features of a video frame through a color histogram is as follows: The color representation of the video frame is in RGB space (red, green, blue); For each frame, the histogram can be calculated for its R (red), G (green), and B (blue) channels respectively; Normalize the histogram by dividing each frequency value by the total number of pixels so that the histogram of each video frame has a uniform scale; Get the color features of the video frame through the color histogram .

[0025] Use the pre-trained support vector machine (SVM) classification model to classify the video frames into normal frames and abnormal frames; For normal frames, their features are relatively stable and regular; Abnormal frames have unique abnormal features, such as abnormal light and shadow changes and irregular object shapes.

[0026] S20, dynamically determining the division method of the neural network model according to the classification result of the video frame, and dynamically adjusting the task allocation ratio between different devices; According to the classification results of the video frames, the division method of the neural network model is dynamically determined. When multiple consecutive frames are detected as normal frames, the part of the model with relatively small computational workload and low dependence on global features (such as shallow convolutional layers) is allocated to edge devices with relatively weak computing resources for processing, while the part with large computational workload and requiring comprehensive global information (such as deep fully connected layers) is allocated to edge devices with strong computing power or the cloud for processing; When an abnormal frame is detected, considering the feature complexity and importance of the abnormal frame, more computing tasks are concentrated on devices with stronger computing power to ensure that the abnormal features can be accurately extracted.

[0027] The process of dynamically adjusting the task allocation ratio between different devices is as follows: According to the frequency of abnormal frames and distribution, dynamically adjust the task allocation ratio between different devices; If the frequency of abnormal frames is greater than the set frequency threshold, abnormal frames appear frequently, and the task load of devices with stronger computing power needs to be increased; If the frequency of abnormal frames is less than the set frequency threshold, the abnormal frames are relatively sparse. Under the premise of ensuring accuracy, tasks should be reasonably allocated to improve overall efficiency.

[0028] S30, monitor the resource usage of each edge device in real time, establish a resource evaluation model, quantitatively evaluate the resource availability of the device, and obtain a device resource score; The resource usage includes CPU usage, remaining storage space and device battery power; By establishing a resource evaluation model, the resource availability of each device is quantitatively evaluated to obtain the device resource score; the specific calculation formula is as follows: ;In the formula It is Equipment resource rating, It is CPU usage of each device, It is CPU usage of each device, It is The remaining storage space of the device It is Total storage space of devices, It is Device storage space weighting factor, It is Device battery level, It is The device battery power weight factor.

[0029] In Edge Video Analytics, the quantitative evaluation of device resource availability and the role of device resource scoring are crucial. By effectively evaluating the device resources, it is possible to ensure that video analysis tasks are efficiently executed on edge devices and achieve the expected performance and results. The following are several key roles of device resource scoring for edge video analytics: 1. Optimize resource allocation and load balancing Edge devices usually have limited resources, and video analysis requires large amounts of computing, storage, and network resources. By quantitatively evaluating the resource availability of each device, tasks can be dynamically assigned based on the device's resource score. Devices with higher resource scores can take on more computing tasks, while devices with lower resources are assigned lighter tasks. This can achieve load balancing and avoid performance degradation caused by overloading certain devices.

[0030] 2. Improve real-time performance and response speed Edge video analysis tasks often require real-time processing of large amounts of video data, especially in the fields of surveillance, traffic management, and smart manufacturing. If the computing power or network bandwidth of the device is weak, data delays or processing failures may occur. Quantitatively evaluating the resource availability of the device and assigning resource scores can effectively predict whether the device can process video data on time, thereby ensuring real-time performance and response speed.

[0031] 3. Decision support and dynamic scheduling The device resource score can be used as an input to the scheduling algorithm to help the system dynamically select the device that is most suitable for performing a specific video analysis task. For example, in scenarios that require high computing resources, the system can prioritize devices with higher scores to handle complex video analysis tasks, while in situations where resources are limited, tasks can be assigned to multiple devices for parallel processing. Dynamic scheduling can not only improve efficiency, but also optimize energy use and reduce resource waste.

[0032] 4. Enhance system robustness and fault tolerance In edge computing, devices may experience resource depletion due to environmental factors, hardware failures, or other problems. By monitoring the resource scores of devices in real time, the system can promptly identify which devices have insufficient resources and reschedule or migrate tasks, thereby maintaining the overall robustness and fault tolerance of the system and preventing single point failures from causing service interruptions.

[0033] 5. Support scalability and elasticity As edge computing devices increase, the system needs to be able to flexibly expand and handle the increasing number of video analysis tasks. Device resource scoring can provide scalability support for the system and help determine whether new devices need to be added or existing devices need to be upgraded. By quantitatively evaluating device resources, the system can better adapt to the needs of different scales and loads.

[0034] 6. Reduce costs and energy consumption Edge devices have limited computing resources and are usually powered by batteries, so efficient use of resources is critical to cost control and extending the life of the device. By quantitatively evaluating the resource status of the device, the system can effectively control the computing requirements of video analysis, avoid unnecessary calculations and transmissions, and thus reduce energy consumption and operation and maintenance costs.

[0035] 7. Improve analysis accuracy Running video analysis algorithms on devices with higher resource scores can use more complex models and algorithms to improve the accuracy of analysis results. Devices with lower resources can choose simplified models for basic processing, ensuring that edge devices perform appropriate analysis tasks based on their own capabilities.

[0036] Device resource scoring plays a critical role in edge video analysis. It helps achieve better task scheduling, load balancing, real-time performance, system robustness, and scalability through real-time evaluation and quantification of device resources. Optimizing the use of device resources can not only improve the performance of video analysis, but also effectively reduce costs and energy consumption, and improve the efficiency and intelligence level of the overall system.

[0037] S40, dynamically adjust the division of the distributed neural network model according to the resource score of the edge device, and perform fusion processing on the visual features of each edge device; Dynamically adjust the partitioning of the distributed neural network model based on the resource scores of edge devices; When the resource score of a device is less than the set score threshold, reduce the computing tasks assigned to the device and transfer the tasks to the device with a higher resource score; When the resource score of a device is greater than the set score threshold, there is no need to reduce the computing tasks assigned to the device; For example, if the CPU usage of an edge device is close to 100% and the remaining battery power is low, it means that the device is short of resources. At this time, some of the convolutional layer computing tasks originally assigned to it are transferred to other devices with sufficient computing resources and energy resources; At the same time, during the task transfer process, by optimizing the data transmission protocol and caching mechanism, it ensures that data can be transmitted to the target device quickly and accurately, reducing the additional delay caused by task transfer.

[0038] After the distributed neural network model is divided, the visual features obtained by different devices when processing video data are fused. The specific calculation formula is as follows: ;In the formula It is the fusion of visual features. It is Texture features of the device, It is Texture feature weights of each device, It is The color characteristics of each device, It is The color feature weights of each device.

[0039] S50, optimizing the distributed neural network model according to the result of the fusion processing, and evaluating the model division effect in combination with the device resource score.

[0040] According to the fusion results, the distributed neural network model is optimized; By calculating the error between the analysis result and the true label, the error information is fed back to the sub-model on each device using the back-propagation algorithm; A dynamic weight adjustment strategy is adopted to adjust the weight of the sub-model in the model optimization process according to its performance in different scenarios; And optimize the fused visual feature weights of each device according to the adjusted weights in the model optimization process; The weighted sum of the device resource scores is used to obtain the total device resource score. The steps for evaluating the model partitioning effect by combining the fusion visual features are as follows: In step C1, the total device resource score and the fused visual features are defined as input variables, and they are divided into different fuzzy sets.

[0041] For example, "Low", "Medium", "High" for the total device resource score, "Low", "Medium", "High" for the fused vision features.

[0042] Step C2, define the model partitioning effect as the output variable and divide it into fuzzy sets, for example, "High", "Low", for the model partitioning effect.

[0043] Step C3, formulate a set of fuzzy rules to describe the impact of different input variables on output variables. The definition of rules can be based on professional knowledge or obtained through data analysis and experiments. For example: The total device resource score is marked as R, the fused visual features are marked as F, and the model partition effect is marked as P, which can be defined Rule 1: IF (R is Low) AND (F is Low) THEN (P is Low ) Rule 2: IF (R is High) AND (F is High) THEN (P is High ) ... Step C4, perform fuzzy reasoning according to fuzzy rules and evaluate the model division effect.

[0044] It should be noted that the division of fuzzy sets can be adjusted according to actual conditions. For example, although this embodiment takes three fuzzy sets as examples, in fact, the total device resource score, fusion visual features, and model division effects can be divided into more than three sets to facilitate better and more accurate identification.

[0045] Furthermore, for the total device resource score and the judgment of the high, medium and low of the fused visual features, the threshold can be set according to the actual situation; when the total device resource score is higher than 80, it is marked as "High", when the fused visual features are higher than 100, it is marked as "High", and so on. I will not go into details here.

[0046] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0047] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0048] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0049] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0050] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0051] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A distributed neural network model partitioning method for edge video analysis, characterized in that: The following steps are involved: Preprocess the input video data, extract the visual features of the video frames according to the local binary pattern (LBP) algorithm, and classify the video frames using the support vector machine classification model; According to the classification results of the video frames, the division method of the neural network model is dynamically determined, and the task allocation ratio between different devices is dynamically adjusted; Monitor the resource usage of each edge device in real time, establish a resource evaluation model, quantitatively evaluate the resource availability of the device, and obtain the device resource score; According to the resource scores of edge devices, the division of the distributed neural network model is dynamically adjusted, and the visual features of each edge device are fused. The distributed neural network model is optimized according to the results of fusion processing, and the model partitioning effect is evaluated in combination with the device resource score.

2. A distributed neural network model partitioning method for edge video analysis according to claim 1, characterized in that: The visual features include texture features and color features; the texture feature extraction process is as follows: Preprocess the input video data; The texture features of the video frame are extracted by the local binary pattern (LBP) algorithm through the lightweight feature extractor on the edge device. The calculation formula is as follows: ; In the formula, is the center pixel coordinate, is the number of neighborhood pixels, where is a sign function, when hour, ;when hour, , It is Device texture features.

3. A distributed neural network model partitioning method for edge video analysis according to claim 2, characterized in that: The color feature extraction process is as follows: The color representation of the video frame is RGB space (red, green, blue); For each frame, histograms are calculated for the R (red), G (green), and B (blue) channels; The histogram of each video frame is normalized by dividing each frequency value of the histogram by the total number of pixels so that the histogram has a uniform scale; The color features of the video frame are obtained by analyzing and processing the color histogram of each video frame .

4. The distributed neural network model partitioning method for edge video analysis according to claim 3 is characterized in that: The process of dynamically determining the partitioning method of the neural network model is as follows: When multiple consecutive frames are detected as normal frames, the part of the model with relatively small computational workload and low reliance on global features is allocated to edge devices with relatively weak computing resources for processing; Assign the parts with large computational workload and need to integrate global information to edge devices or the cloud with strong computing power for processing; When an abnormal frame is detected, due to the complexity and importance of the features of the abnormal frame, more computing tasks are concentrated on devices with stronger computing power to ensure that the abnormal features can be accurately extracted.

5. A distributed neural network model partitioning method for edge video analysis according to claim 4, characterized in that: The process of dynamically adjusting the task allocation ratio between different devices is as follows: Dynamically adjust the task allocation ratio between different devices according to the occurrence frequency f and distribution of abnormal frames; If the frequency of abnormal frames is greater than the set frequency threshold, abnormal frames appear frequently, and the task load of devices with stronger computing power needs to be increased; If the frequency of abnormal frames is less than the set frequency threshold, the abnormal frames are relatively sparse. Under the premise of ensuring accuracy, tasks should be reasonably allocated to improve overall efficiency.

6. A distributed neural network model partitioning method for edge video analysis according to claim 5, characterized in that: The process of obtaining device resource scores is as follows: The resource usage includes CPU usage, remaining storage space and device battery power; By establishing a resource evaluation model, the resource availability of each device is quantitatively evaluated to obtain the device resource score; the specific calculation formula is as follows: ;In the formula It is Equipment resource rating, It is CPU usage of each device, It is CPU usage of each device, It is The remaining storage space of the device It is Total storage space of devices, It is Device storage space weighting factor, It is Device battery level, It is The device battery power weight factor.

7. A distributed neural network model partitioning method for edge video analysis according to claim 6, characterized in that: According to the resource scores of edge devices, the partitioning process of the distributed neural network model is dynamically adjusted as follows: When the resource score of a device is less than the set score threshold, reduce the computing tasks assigned to the device and transfer the tasks to devices with higher resource scores; When the resource score of a device is greater than the set score threshold, there is no need to reduce the computing tasks assigned to the device.

8. A distributed neural network model partitioning method for edge video analysis according to claim 7, characterized in that: The process of fusing the visual features of each edge device is as follows: After the distributed neural network model is divided, the visual features obtained by different devices when processing video data are fused. The specific calculation formula is as follows: ;In the formula It is the fusion of visual features. It is Texture features of the device, It is The texture feature weights of each device, It is The color characteristics of each device, It is The color feature weights of each device.

9. The distributed neural network model partitioning method for edge video analysis according to claim 7, characterized in that: The optimization process of the distributed neural network model based on the results of fusion processing is as follows: By calculating the error between the analysis result and the true label, the error information is fed back to the sub-model on each device using the back-propagation algorithm; Adopt a dynamic weight adjustment strategy to adjust the weight of the sub-model in the model optimization process according to its performance in different scenarios; And the fused visual feature weights of each device are optimized according to the adjusted weights in the model optimization process.

10. The distributed neural network model partitioning method for edge video analysis according to claim 7, characterized in that: The steps to evaluate the model partitioning effect are as follows: The total equipment resource score and fused visual features are defined as input variables, and they are divided into different fuzzy sets respectively; The model partitioning effect is defined as the output variable, which is divided into fuzzy sets; Formulate fuzzy rules to describe the impact of total equipment resource score and fused visual features on model partitioning effect; Perform fuzzy reasoning based on fuzzy rules and evaluate the model partitioning effect.

Images