Stream data image semantic segmentation model training method and system and readable storage medium
By dividing streaming data into multiple subsets and constructing a dynamic weight matrix ensemble model, the accuracy and stability issues of streaming data semantic segmentation models are solved without increasing resources, achieving efficient model updates and segmentation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Z-ONE TECH CO LTD
- Filing Date
- 2023-04-28
- Publication Date
- 2026-07-07
AI Technical Summary
When processing dynamically and continuously generated streaming data, existing technologies struggle to quickly and accurately update and iterate new knowledge for semantic segmentation models. Furthermore, without increasing data and computing power, the improvement in model performance is limited, especially when the distribution of streaming data is unstable, leading to a sharp decline in model recognition performance.
A streaming data partitioning strategy is adopted to divide the dataset into multiple subsets, train multiple image semantic segmentation models in each subset, and construct a total segmentation model based on the identifiable categories and dynamic voting weights of the models. The models are then integrated through a dynamic weight matrix to achieve semantic segmentation.
It improves the accuracy and stability of the model while using as little storage and training resources as possible, reduces computational pressure, ensures the stability of online model updates and iterations, and can handle streaming data and batch image data.
Smart Images
Figure CN116664830B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image segmentation technology, and in particular to a method, system, and readable storage medium for training a streaming data image semantic segmentation model. Background Technology
[0002] Image semantic segmentation is a crucial task in computer vision. Its goal is to classify each pixel in an image. Its applications include, but are not limited to, autonomous driving, image enhancement, and 3D reconstruction. If image segmentation could be performed quickly and accurately, many problems would be easily solved. However, practical applications deal with dynamically generated streaming data. Semantic segmentation models need to be updated and iteratively learned to acquire new knowledge, making training with the entire historical dataset uneconomical and impractical. In particular, the distribution of streaming data is dynamic and unstable, and it may contain knowledge that the model has not yet learned, leading to a sharp decline in model recognition performance. Currently, the main solutions include the following:
[0003] 1. Increase time and use as much data and computing power as possible to optimize the model, but the improvement is limited and it is not advisable to use it too often in practical applications due to time constraints.
[0004] 2. Using an integration strategy can improve the model performance to a certain extent without increasing data, computing power, or optimizing the model, but it cannot currently cover scenarios with streaming data and increased categories. Summary of the Invention
[0005] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0006] To improve the semantic segmentation performance of streaming image data with minimal storage and training resources, this invention provides a training method, system, and readable storage medium for a streaming image semantic segmentation model.
[0007] Firstly, the present invention provides a method for training a semantic segmentation model for streaming image data, which employs the following technical solution:
[0008] A method for training a semantic segmentation model for streaming image data includes:
[0009] The streaming dataset is divided into multiple data subsets based on the streaming data partitioning strategy;
[0010] Multiple image semantic segmentation models are trained based on the multiple data subsets respectively; wherein each data subset corresponds one-to-one with the image semantic segmentation model, and each image semantic segmentation model has an identifiable category for identifying the data category of the corresponding streaming data.
[0011] Based on the identifiable categories of the image semantic segmentation model and the dynamic voting weights of the image semantic segmentation model, the multiple image semantic segmentation models are integrated into a total segmentation model;
[0012] The overall segmentation model identifies the data category of the input streaming data and performs semantic segmentation based on the dynamic voting weights of the image semantic segmentation model.
[0013] Furthermore, the method of dividing the streaming dataset into multiple data subsets based on the streaming data partitioning strategy includes:
[0014] The streaming dataset is formed by acquiring dynamic streaming data of continuous images;
[0015] Based on the size of the streaming dataset, computing resources, and business iteration requirements, a preset streaming data partitioning strategy is determined, and the streaming dataset is divided into multiple data subsets.
[0016] Furthermore, in the above-mentioned method for training a streaming data image semantic segmentation model, the streaming data partitioning strategy includes at least one of the following: partitioning streaming data using a time window method, partitioning streaming data using a fixed size method, and partitioning streaming data using a scene clustering method.
[0017] Furthermore, in the above-mentioned method for training a streaming data image semantic segmentation model, the number of data subsets divided does not exceed ten.
[0018] Furthermore, in the above-mentioned method for training a streaming data image semantic segmentation model, the image semantic segmentation model can be selected from different semantic segmentation models.
[0019] Furthermore, in the above-mentioned method for training a streaming data image semantic segmentation model, the integration of multiple image semantic segmentation models into a total segmentation model based on the identifiable categories of the image semantic segmentation models and the dynamic voting weights of the image semantic segmentation models includes:
[0020] The dynamic voting weight is determined based on the identifiable category;
[0021] Construct a dynamic weighted voting matrix based on the aforementioned dynamic voting weights;
[0022] The overall segmentation model is determined based on the dynamic weight voting matrix and each of the image semantic segmentation models.
[0023] Furthermore, in the above-mentioned method for training a streaming image semantic segmentation model, the dynamic voting weight... ,in,
[0024] ,
[0025] The dynamic weighted voting matrix ,
[0026] The total segmentation model .
[0027] Secondly, this invention provides a method for semantic segmentation of streaming data images, employing the following technical solution:
[0028] A method for semantic segmentation of streaming data images, comprising:
[0029] The overall segmentation model is obtained based on the streaming data image semantic segmentation model training method described in any one of the first aspects of the present invention;
[0030] The segmentation results are obtained by semantic segmentation of the streaming data image based on the overall segmentation model.
[0031] Thirdly, this invention provides a training system for a streaming data image semantic segmentation model, employing the following technical solution:
[0032] A streaming data image semantic segmentation model training system includes:
[0033] The data partitioning module is used to divide a streaming dataset into multiple data subsets based on a streaming data partitioning strategy.
[0034] The model training module is used to train multiple image semantic segmentation models based on the multiple data subsets respectively; wherein each data subset corresponds one-to-one with the image semantic segmentation model, and each image semantic segmentation model has an identifiable category for identifying the data category of the corresponding streaming data.
[0035] An ensemble learning module is used to integrate multiple image semantic segmentation models into a total segmentation model based on the identifiable categories of the image semantic segmentation models and the dynamic voting weights of the image semantic segmentation models; the total segmentation model identifies the data categories of the input streaming data and performs semantic segmentation according to the dynamic voting weights of the image semantic segmentation models.
[0036] Fourthly, the present invention provides an electronic device that adopts the following technical solution:
[0037] An electronic device, comprising:
[0038] processor;
[0039] A memory communicatively connected to the processor stores instructions executable by the processor, which are executed by the processor to enable the processor to perform a streaming data image semantic segmentation model training method according to any one of the first aspects of the present invention.
[0040] Fifthly, the present invention provides a computer-readable storage medium, which adopts the following technical solution:
[0041] A computer-readable storage medium storing computer instructions that, when executed by a processor, implement a streaming data image semantic segmentation model training method as described in any one of the first aspects of the present invention.
[0042] The streaming image semantic segmentation model proposed in this invention utilizes dynamic weights to weight the recognition categories of multiple image semantic segmentation models within the model. This addresses the difficulties of training image semantic segmentation models using streaming data and scenarios with increasing categories, improving model performance while minimizing storage and training resources. It possesses at least one of the following advantages: 1. High accuracy: the segmentation model learns new knowledge without forgetting old knowledge; 2. Low computational pressure: the model does not require the use of the entire streaming data at once during training, significantly reducing data storage and computational pressure; 3. Wide range of engineering applications: it can process both streaming and batch image data, ensuring the stability of online model updates and iterations. Attached Figure Description
[0043] Figure 1 This is a flowchart of an embodiment of a streaming data image semantic segmentation model training method of the present invention.
[0044] Figure 2 This is a flowchart of another embodiment of the streaming data image semantic segmentation model training method of the present invention.
[0045] Figure 3 This is a flowchart illustrating another embodiment of the streaming data image semantic segmentation model training method of the present invention.
[0046] Figure 4 This is a flowchart of another embodiment of the streaming data image semantic segmentation model training method of the present invention.
[0047] Figure 5 This is a flowchart illustrating another embodiment of the streaming data image semantic segmentation model training method of the present invention.
[0048] Figure 6 This is a schematic diagram of an embodiment of the streaming data image semantic segmentation model training system of the present invention.
[0049] Figure labeling: 1. Data partitioning module; 2. Model training module; 3. Ensemble learning module. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0051] The method steps described in this embodiment of the invention can be executed in the order described in the specific implementation, or the execution order of each step can be adjusted according to actual needs, provided that the technical problem can be solved. These are not listed one by one here.
[0052] The following is in conjunction with the appendix Figure 1-6 The present invention will be described in further detail below.
[0053] Streaming imagery refers to a continuous stream of image data that is generated in real-time and requires processing. This image data is typically generated in real-time by devices such as cameras, surveillance systems, and sensors, and is continuously produced at a certain rate. Unlike batch image data, streaming imagery is characterized by its real-time nature, high noise levels, multimodal nature, and unstructured characteristics. Processing streaming imagery requires efficient real-time processing capabilities and computer vision algorithms, such as stream processing engines, image recognition, object detection, and object tracking. Streaming imagery has applications in many fields, such as video surveillance, intelligent transportation, and robot vision.
[0054] This invention discloses a method for training a semantic segmentation model for streaming image data, referring to... Figure 1 A method for training a semantic segmentation model for streaming image data, comprising:
[0055] S1, the streaming dataset is divided into multiple data subsets based on the streaming data partitioning strategy;
[0056] S2, train multiple image semantic segmentation models based on the multiple data subsets respectively; wherein each data subset corresponds one-to-one with the image semantic segmentation model, and each image semantic segmentation model has an identifiable category for identifying the data category of the corresponding streaming data.
[0057] S3, based on the identifiable categories of the image semantic segmentation model and the dynamic voting weights of the image semantic segmentation model, the multiple image semantic segmentation models are integrated into a total segmentation model; the total segmentation model identifies the data category of the input streaming data and performs semantic segmentation according to the dynamic voting weights of the image semantic segmentation model.
[0058] When training a semantic segmentation model for streaming image data, the streaming dataset is first divided into multiple subsets according to a pre-defined streaming data partitioning strategy, and a corresponding image semantic segmentation model is trained using each subset. Next, the identifiable categories of each image semantic segmentation model are determined. Then, based on the identifiable categories, a dynamic voting weight is determined for each image semantic segmentation model, and all image semantic segmentation models are merged and integrated into a single overall segmentation model based on this dynamic voting weight.
[0059] The streaming data image semantic segmentation model proposed in this invention has the following advantages: 1. High accuracy: Since the overall segmentation model is obtained by integrating image semantic segmentation models trained on multiple data subsets, the overall segmentation model can learn new knowledge without forgetting old knowledge; 2. Low computational pressure: During model training, only different computing resources are needed to train the segmentation model corresponding to each data subset, without using the full streaming data at once, greatly reducing the data storage and computational pressure; 3. Wide range of engineering applications: It can process streaming data and batch image data, ensuring the stability of online model updates and iterations.
[0060] Furthermore, as a specific embodiment of the present invention, refer to... Figure 2 or Figure 3 Step S1, dividing the streaming dataset into multiple data subsets includes:
[0061] S11, acquire dynamic streaming data of continuous images to form the streaming dataset;
[0062] S12, determine a preset streaming data partitioning strategy based on the size of the streaming dataset, computing resources and business iteration requirements, and divide the streaming dataset into multiple data subsets.
[0063] Specifically, in step S11, continuous image data is first acquired using an in-vehicle camera system and generated as dynamic streaming data. The acquired dynamic streaming data is then preprocessed, including denoising, augmentation, and format conversion, to form the streaming dataset, thereby improving data quality and usability.
[0064] In step S12, the data is partitioned based on the streaming dataset: continuous image data is divided into multiple data subsets according to a preset streaming data partitioning strategy. Simultaneously, the partitioned data subsets can be stored in a database or file system for subsequent data access and processing, or directly input into an image semantic segmentation model.
[0065] Furthermore, in one specific embodiment of the present invention, the number of data subsets does not exceed ten. In actual computation, each data subset requires data processing and computation; too many data subsets increase computational burden and storage overhead. Simultaneously, too many data subsets may contain similar or duplicate data, leading to data redundancy and redundant computation. Too many data subsets may require different processing algorithms or parameters, increasing algorithm complexity and potentially increasing error rate and inaccuracies.
[0066] Furthermore, as a specific embodiment of the present invention, the streaming data partitioning strategy is as follows: The streaming data is partitioned using a time window method. Specifically, the streaming dataset is divided into multiple consecutive time windows according to time, with each time window containing data within a certain time range. For example, the dataset can be divided into time windows per second, per minute, or even per week, with each time window containing data from a period preceding the current time.
[0067] Furthermore, as a specific embodiment of the present invention, the streaming data partitioning strategy can also be: using a fixed-size method to partition the streaming data. Specifically, the streaming dataset is divided into multiple fixed-size subsets according to the amount of data, with each subset containing a certain number of data points. The specific size of the data subsets is determined based on the size of the streaming dataset, computing resources, and business iteration requirements. For example, the dataset can be divided into subsets of 1000 data points each, with each subset containing 1000 consecutive data points.
[0068] Furthermore, as a specific embodiment of the present invention, the streaming data partitioning strategy can also be as follows: Streaming data can be partitioned using a scenario clustering method, where scenarios can be divided into event scenarios and environment scenarios. Partitioning streaming data by event scenario clustering involves dividing the streaming dataset into multiple subsets based on the occurrence of a specific event, with each subset containing data related to that event. For example, the dataset can be divided into subsets where a certain abnormal event occurs, with each subset containing data related to that abnormal event. Partitioning streaming data by environment scenario clusters similar environments within the streaming dataset together, forming multiple data clusters, and treating each data cluster as a data subset. For example, data about highways can be divided into a data cluster, forming a data subset with highways as the primary environment.
[0069] Furthermore, as a specific embodiment of the present invention, step S2, training corresponding image semantic segmentation models based on each of the data subsets and obtaining the identifiable categories of each image semantic segmentation model, includes:
[0070] Training the model: Using each data subset and its annotation information as the training set, the selected image semantic segmentation model is trained to obtain the corresponding image semantic segmentation model.
[0071] Model evaluation: Each image semantic segmentation model is evaluated, including its accuracy, robustness, and generalization ability, and the image semantic segmentation model is tuned.
[0072] Category recognition: Test each image semantic segmentation model, obtain the category information that the model can recognize, and analyze and statistically analyze the recognition results to obtain the identifiable categories.
[0073] Furthermore, as a specific embodiment of the present invention, when training an image semantic segmentation model using data subsets, different data subsets can be trained using the same or different image semantic segmentation models. The image semantic segmentation models used include, but are not limited to, FCN, U-Net, DeepLab, Mask R-CNN, and PSPNet. When selecting an image semantic segmentation model, factors such as the model's training difficulty, computational complexity, accuracy, and generalization ability, as well as the characteristics and size of the dataset, can be considered. Simultaneously, a suitable model can be selected by comprehensively considering the specific application scenario and task requirements.
[0074] Furthermore, as a specific embodiment of the present invention, refer to... Figure 4 Step S3, determining the dynamic voting weights of each image semantic segmentation model based on the identifiable category, and then integrating the image semantic segmentation models to obtain the overall segmentation model, includes:
[0075] S31, determine the dynamic voting weight based on the identifiable category;
[0076] S32, Construct a dynamic weighted voting matrix based on the dynamic voting weights;
[0077] S33, the overall segmentation model is determined based on the dynamic weight voting matrix and each of the image semantic segmentation models.
[0078] In this invention, integration based on dynamic voting weights is based on the identifiable categories of the models. In principle, if model M1 has not learned about category C2 during training, then during the inference phase, M1's ability to identify category C2 is considered zero, meaning model M1's dynamic voting weight for category C2 is zero. Similarly, if the entire model set M = {M1, M2, ..., M...}... n In the model, only M1 has learned about category C2 during training, so the dynamic voting weight of model M1 in category C2 is 1.
[0079] Specifically, firstly, after testing each image semantic segmentation model to obtain identifiable categories, the dynamic voting weights of each image semantic segmentation model are calculated, including:
[0080] Let W ij For image semantic segmentation model M i In category C j Dynamic voting weights , then; among which,
[0081] ,
[0082] Next, based on the dynamic voting weight W ij Construct a dynamic weighted voting matrix W:
[0083] ,
[0084] Finally, based on the dynamic weight voting matrix W and the image semantic segmentation models M1, M2, ..., M... n Determine the total segmentation model P:
[0085] .
[0086] By following the steps above, the overall segmentation model that can be used for semantic segmentation of streaming data images can be determined.
[0087] Furthermore, the model used in this invention is not limited to a specific self-supervised learning and semantic segmentation model; it is applicable to any model that meets the requirements of street scene semantic segmentation scenarios, and the calling interface can be adapted for compatibility.
[0088] Furthermore, the category-based dynamic weighted voting integration used in this invention is not limited to a specific model or scenario, and is applicable to both streaming data and dynamic category scenarios.
[0089] Reference Figure 5 The implementation principle of one embodiment of the streaming data image semantic segmentation model training method of the present invention is as follows:
[0090] To improve the semantic segmentation performance of streaming images with minimal storage and training resources, this invention first divides the pre-collected streaming dataset into multiple data subsets according to a preset partitioning strategy, and then trains corresponding image semantic segmentation models M1, M2, ..., M using each data subset. n Next, the identifiable categories of each image semantic segmentation model are determined. Then, based on the identifiable categories, the dynamic voting weight W of each image semantic segmentation model is determined. ij Based on this dynamic voting weight W ij The dynamic weight voting matrix W is further determined, and based on this dynamic weight voting matrix W, all image semantic segmentation models are integrated to generate the overall segmentation model P.
[0091] This invention also discloses a method for semantic segmentation of streaming data images.
[0092] A method for semantic segmentation of streaming data images, comprising:
[0093] Step 1: Based on the above-mentioned streaming data image semantic segmentation model training method of the present invention, the overall segmentation model is obtained.
[0094] For example, a streaming dataset is divided into multiple data subsets based on a streaming data partitioning strategy; multiple image semantic segmentation models are trained based on the multiple data subsets respectively; wherein each data subset corresponds one-to-one with an image semantic segmentation model, and each image semantic segmentation model has a identifiable category for identifying the data category of the corresponding streaming data; the multiple image semantic segmentation models are integrated into a total segmentation model based on the identifiable categories of the image semantic segmentation models and the dynamic voting weights of the image semantic segmentation models; the total segmentation model identifies the data category of the input streaming data and performs semantic segmentation according to the dynamic voting weights of the image semantic segmentation models.
[0095] For details, please refer to the content on training methods for streaming data image semantic segmentation models.
[0096] Step 2: Perform semantic segmentation on the streaming data image based on the overall segmentation model to obtain the segmentation result.
[0097] Furthermore, as a specific embodiment of the present invention, the second step: performing semantic segmentation on the streaming data image based on the overall segmentation model to obtain the segmentation result specifically includes:
[0098] First, image preprocessing: preprocessing the streaming data images acquired by the vehicle, including image resizing, color space conversion, normalization, and other operations, to adapt to the input requirements of the overall segmentation model;
[0099] Secondly, model inference: using the trained overall segmentation model, inference is performed on the preprocessed streaming data image to obtain the image segmentation result;
[0100] Third, post-processing: Post-processing of the segmentation results includes noise removal, hole filling, and connection of broken links to obtain more accurate segmentation results;
[0101] Fourth, visualization: The segmentation results can be visualized using methods such as color mapping, border annotation, and visual masks, making the segmentation results clearer and easier to understand.
[0102] Based on the above-mentioned streaming data image semantic segmentation model training method, this invention also proposes a streaming data image semantic segmentation model training system.
[0103] Reference Figure 6 A streaming data image semantic segmentation model training system includes a data partitioning module 1, a model training module 2, and an ensemble learning module 3.
[0104] Data partitioning module 1 is used to divide a streaming dataset into multiple data subsets based on a streaming data partitioning strategy;
[0105] Model training module 2 is used to train multiple image semantic segmentation models based on the multiple data subsets respectively; wherein each data subset corresponds one-to-one with the image semantic segmentation model, and each image semantic segmentation model has an identifiable category for identifying the data category of the corresponding streaming data;
[0106] The integrated learning module 3 is used to integrate the multiple image semantic segmentation models into a total segmentation model based on the identifiable categories of the image semantic segmentation model and the dynamic voting weights of the image semantic segmentation model; the total segmentation model identifies the data category of the input streaming data and performs semantic segmentation according to the dynamic voting weights of the image semantic segmentation model.
[0107] Optionally, in one specific embodiment of the present invention, the data partitioning module 1 includes a stream database, a partitioner, and a data subset library. The stream database is used to store the collected stream dataset. The partitioner stores the logical language corresponding to a preset stream data partitioning strategy, which is used to partition the stream data in the stream database based on the preset stream data partitioning strategy when it is necessary to partition the stream data. The data subset library is used to store the data subsets after partitioning by the partitioner.
[0108] Optionally, as a specific embodiment of the present invention, the model training module 2 includes multiple trainers. Each trainer can load different image semantic segmentation models as needed. The training samples consisting of each data subset are used to train the image semantic segmentation models in each trainer, resulting in multiple image semantic segmentation models. At the same time, each trained image semantic segmentation model can also be tested within the trainer to obtain the recognizable types that each image semantic segmentation model can recognize.
[0109] Optionally, as a specific embodiment of the present invention, the ensemble learning module 3 includes a dynamic weight voter and an integrator. The dynamic weight voter is used to determine the dynamic voting weights of each image semantic segmentation model based on the identifiable categories of each model. The integrator is used to determine the final overall segmentation model based on the obtained dynamic voting weights and the corresponding image semantic segmentation models.
[0110] This invention also proposes an electronic device, comprising:
[0111] A processor, and a memory communicatively connected to the processor; the memory stores instructions executable by the processor, which are executed by the processor to enable the processor to perform a streaming data image semantic segmentation model training method according to the above embodiments.
[0112] This invention also discloses a computer-readable storage medium.
[0113] A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a streaming data image semantic segmentation model training method as described in any of the above embodiments. The computer-readable storage medium may include any entity or device capable of carrying a computer program, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), and a software distribution medium, etc. The computer program includes computer program code. The computer program code may be in the form of source code, object code, an executable file, or some intermediate form, etc. The computer-readable storage medium may include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), and a software distribution medium, etc.
[0114] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.
[0115] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a system including a processing module or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).
[0116] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for training a semantic segmentation model for streaming image data, characterized in that, include: The streaming dataset is divided into multiple data subsets based on the streaming data partitioning strategy; Multiple image semantic segmentation models are trained based on the aforementioned multiple data subsets; The data subsets therein correspond one-to-one with the image semantic segmentation models, and each image semantic segmentation model has an identifiable category for identifying the data category of the corresponding streaming data; Based on the identifiable categories of the image semantic segmentation model, dynamic voting weights are determined, a dynamic weight voting matrix is constructed based on the dynamic voting weights, and the multiple image semantic segmentation models are integrated into a total segmentation model based on the dynamic weight voting matrix. The dynamic voting weight The dynamic weighted voting matrix The total segmentation model Among them, W ij It is an image semantic segmentation model M i In category C j The dynamic voting weights are W, which is an n×m matrix, where n is the number of image semantic segmentation models and m is the number of categories. The overall segmentation model identifies the data category of the input streaming data and performs semantic segmentation based on the dynamic voting weights of the image semantic segmentation model.
2. The method for training a streaming data image semantic segmentation model according to claim 1, characterized in that, The streaming data partitioning strategy divides the streaming dataset into multiple data subsets, including: The streaming dataset is formed by acquiring dynamic streaming data of continuous images; Based on the size of the streaming dataset, computing resources, and business iteration requirements, a preset streaming data partitioning strategy is determined, and the streaming dataset is divided into multiple data subsets.
3. The method for training a streaming data image semantic segmentation model according to claim 2, characterized in that, The streaming data partitioning strategy includes at least one of the following: partitioning streaming data using a time window method, partitioning streaming data using a fixed size method, and partitioning streaming data using a scenario clustering method.
4. The method for training a streaming data image semantic segmentation model according to claim 2, characterized in that, The number of data subsets that are divided does not exceed ten.
5. The method for training a streaming data image semantic segmentation model according to claim 1, characterized in that, The image semantic segmentation model can select different semantic segmentation models.
6. A method for semantic segmentation of streaming data images, characterized in that, include: The overall segmentation model is obtained based on the streaming data image semantic segmentation model training method as described in any one of claims 1-5; The segmentation results are obtained by semantic segmentation of the streaming data image based on the overall segmentation model.
7. A training system for a streaming data image semantic segmentation model, characterized in that, include: The data partitioning module is used to divide a streaming dataset into multiple data subsets based on a streaming data partitioning strategy. The model training module is used to train multiple image semantic segmentation models based on the multiple data subsets respectively; wherein each data subset corresponds one-to-one with the image semantic segmentation model, and each image semantic segmentation model has an identifiable category for identifying the data category of the corresponding streaming data. An integrated learning module is used to determine dynamic voting weights based on the identifiable categories of the image semantic segmentation model, construct a dynamic weight voting matrix based on the dynamic voting weights, and integrate the multiple image semantic segmentation models into a total segmentation model based on the dynamic weight voting matrix. The dynamic voting weight The dynamic weighted voting matrix The total segmentation model Among them, W ij It is an image semantic segmentation model M i In category C j The dynamic voting weights are W, which is an n×m matrix, where n is the number of image semantic segmentation models and m is the number of categories. The overall segmentation model identifies the data category of the input streaming data and performs semantic segmentation based on the dynamic voting weights of the image semantic segmentation model.
8. An electronic device, characterized in that, include: processor; A memory communicatively connected to the processor, the memory storing instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform a streaming data image semantic segmentation model training method according to any one of claims 1-5.
9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, which, when executed by a processor, implement a streaming data image semantic segmentation model training method as described in any one of claims 1-5.