Intelligent labeling system and method based on machine autonomous learning

Abstract

The invention relates to an intelligent labeling system and method based on machine autonomous learning, and the method comprises the following working steps: receiving and preprocessing original data, and dividing the original data into an initial data set and an intermediate data set; performing initial annotation on the initial data set to generate an initial annotation data set; training a machine learning model through the initial annotation data set to obtain a pre-training model; performing intelligent labeling on the intermediate data set through a pre-training model to generate an intermediate labeling data set; checking and correcting the intermediate annotation data set to form an intermediate correction data set, re-training the pre-training model through the intermediate correction data set, intelligently annotating the intermediate correction data set by using the re-trained pre-training model, and repeating the steps to form closed-loop feedback training; and outputting the finally labeled data. According to the method and the device, the efficiency, the accuracy and the adaptability of data annotation are improved, and an efficient and reliable solution is provided for various types of data annotation tasks.

Description

Technical Field

[0001] The present invention relates to the field of intelligent data processing technology, in particular to the field of data annotation processing technology, and specifically to an intelligent annotation system and method based on machine autonomous learning. Background Art

[0002] In the field of artificial intelligence and machine learning, data labeling is a crucial task. Whether it is a model used for image recognition, natural language processing or speech recognition, it relies on high-quality labeled data for training. It can be said that the quality of the labeled data directly affects the performance and accuracy of the model. Therefore, accurate labeling of raw data is regarded as the basis for building high-performance models. However, traditional raw data labeling usually relies on manual operations, and there are several major problems: Huge time and cost consumption: High-quality data annotation often requires a lot of manual labor, especially when facing large-scale data sets. The annotation process may take weeks or even months. In addition, the high cost of manual annotation is also an important factor limiting large-scale data annotation.

[0003] Subjectivity and inconsistency: Due to the different knowledge backgrounds, experience, and understanding abilities of the annotators, manual annotation is prone to subjective bias, leading to inconsistency in the annotation results. This inconsistency may affect the training effect of the model and reduce the generalization ability of the model.

[0004] Fatigue and error rate: Manual labeling work is usually monotonous, and labelers are prone to fatigue, which leads to an increase in labeling error rate, which not only affects the quality of the labeled data, but may also further affect the performance of the model.

[0005] In order to overcome the above-mentioned defects of manually annotating raw data, some semi-automatic and automatic annotation tools have been proposed. These tools usually perform preliminary annotation of data based on pre-trained models or rule engines, and then manually correct them. However, such methods still have some obvious limitations: Dependence on the performance of the initial model: Existing automatic labeling tools usually rely on the performance of pre-trained models. If the performance of the initial model is poor, the accuracy of automatic labeling will be difficult to guarantee, which in turn increases the workload of subsequent manual corrections.

[0006] Lack of continuous optimization capabilities: Traditional automatic annotation tools are usually unable to effectively use the results of manual corrections to continuously optimize the model. Even if the user corrects the annotation results, these correction information is difficult to feed back to the model for further learning and improvement.

[0007] Poor scalability: Existing automatic annotation tools are often designed for specific tasks or data types, and are difficult to adapt to the annotation needs of different fields or multiple types of data. For example, some tools may only be suitable for annotating text data, but perform poorly when processing image or audio data.

[0008] Therefore, how to overcome the defects of the existing intelligent labeling systems or methods is the technical problem that this application aims to solve. Summary of the invention

[0009] The purpose of the present invention is to overcome at least one of the above-mentioned technical defects and to provide an intelligent labeling system and method based on machine autonomous learning, which can quickly and accurately label data, improve the efficiency of data labeling while ensuring the accuracy of data labeling, and can be applied to various types of data with strong adaptability.

[0010] In order to achieve the above object, the technical solution adopted by the present invention is: The present invention provides an intelligent labeling system based on machine autonomous learning, which includes: A data input module is used to receive and preprocess raw data, and preprocess the acquired raw data to form a raw data set; An initial annotation module, taking a part of the original data set as an initial data set and performing initial annotation on the initial data set to generate an initial annotated data set; An autonomous learning module trains a machine learning model using the initial annotated data set, and the model obtained after training is a pre-trained model; An intelligent labeling module, taking another part of the original data set as an intermediate data set, and intelligently labeling the intermediate data set through the pre-trained model to generate an intermediate labeled data set; The annotation correction module checks and corrects the intermediate annotation data set to form an intermediate correction data set, and the intermediate correction data set is fed back to the autonomous learning module. The autonomous learning module retrains the pre-trained model through the intermediate correction data set, uses the intermediate correction data set as the intermediate data set, and uses the re-trained pre-trained model to intelligently annotate the intermediate correction data set, thereby forming a closed-loop feedback training; The data output module outputs the final labeled data, i.e., the result data set, after multiple closed-loop feedback trainings.

[0011] For the above technical solution, the applicant has further specific optimized implementation methods.

[0012] Optionally, the initial annotation is performed using manual annotation or semi-automatic tools.

[0013] Optionally, the data output module performs statistical analysis based on the result data set, forms a labeling report, and outputs it simultaneously with the result data set.

[0014] In particular, the present invention also provides an intelligent labeling method based on machine autonomous learning, which includes the following working steps: Step S1: receiving and preprocessing raw data, and preprocessing the acquired raw data to form an original data set, and dividing the original data set into two parts, one of which is used as an initial data set, and the other is used as an intermediate data set; Step S2: performing initial annotation on the initial data set to generate an initial annotated data set; Step S3: training the machine learning model using the initial annotated data set, and the model obtained after training is a pre-trained model; Step S4: Intelligently labeling the intermediate data set through the pre-trained model to generate an intermediate labeled data set; Step S5: Check and correct the intermediate annotated data set to form an intermediate corrected data set, retrain the pre-trained model with the intermediate corrected data set, and use the re-trained pre-trained model to intelligently annotate the intermediate corrected data set, and repeat this process to form a closed-loop feedback training; Step S6: After multiple closed-loop feedback trainings, the training of the initial data set and the intermediate data set is completed, and the final labeled data, namely the result data set, is output.

[0015] For the above technical solution, the applicant has further specific optimized implementation methods.

[0016] Optionally, the initial annotation in step S2 is completed by manual annotation or semi-automatic tools to ensure the accuracy of the initial annotated data set after annotation.

[0017] Optionally, the checking and correction of the intermediate annotated data set in step S5 is performed by manual annotation or semi-automatic tools to ensure the accuracy of the intermediate corrected data set.

[0018] Optionally, the original data is at least one of image data, text data, and audio data.

[0019] Optionally, the machine learning model is trained using a deep learning algorithm based on the initial annotated data set, and the deep learning algorithm includes but is not limited to a convolutional neural network learning algorithm, a recurrent neural network learning algorithm, and a deep neural network learning algorithm.

[0020] Optionally, in step S5, the pre-trained model after closed-loop feedback training is verified and evaluated using the initial annotated data set. When the accuracy of the verification and evaluation exceeds a preset threshold, the initial data set and the intermediate data set are merged and outputted, otherwise the closed-loop feedback training is continued.

[0021] Optionally, the pre-trained model after step S5 is saved as a labeling prediction model, and the labeling prediction model is used to intelligently label data of the same type as the original data.

[0022] Compared with the prior art, the advantages of the intelligent labeling system and method based on machine autonomous learning in the patent application of this invention are: The intelligent labeling system and method based on machine autonomous learning of the present application can autonomously train the model based on the initial data labeling by introducing autonomous learning technology, and introduces a labeling correction and feedback mechanism. The user's correction results can be effectively fed back to the model training process, continuously improving the accuracy and consistency of the labeling, and forming a closed-loop system for continuously optimizing the model. The labeling results are continuously optimized through the closed-loop feedback mechanism, while minimizing human intervention and achieving efficient and accurate data labeling.

[0023] Furthermore, the present application can support labeling tasks for multiple data types (such as images, text, and audio). It only needs to adjust the labels during the initial labeling. The machine learning model with attention mechanism can flexibly adjust the labeling strategy according to different task requirements, and has strong scalability and adaptability.

[0024] Furthermore, the present application can provide detailed annotation reports and visual correction tools, so that users can conduct independent inspections and corrections, and at the same time can intuitively understand the annotation results and model performance, which facilitates subsequent data analysis and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Hereinafter, some specific embodiments of the present invention will be described in detail in an exemplary and non-limiting manner with reference to the accompanying drawings. The same reference numerals in the accompanying drawings indicate the same or similar components or parts. It should be understood by those skilled in the art that these drawings are not necessarily drawn to scale. In the accompanying drawings: Figure 1 It is a principle block diagram of intelligent labeling based on machine autonomous learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0026] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. 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.

[0027] In addition, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. Example

[0028] This embodiment describes an intelligent labeling system based on machine autonomous learning. Figure 1 As shown, it includes: The data input module is used to receive and preprocess the raw data, and preprocess the acquired raw data to form the raw data set. The raw data here includes different types of data formats such as images, text, audio, etc. The preprocessing steps may include data cleaning, format conversion, word segmentation (for text), image scaling and enhancement (for image data), etc., to ensure that the data has a unified standard format when entering the subsequent modules; The initial annotation module takes a part of the original data set as the initial data set and performs initial annotation on the initial data set to generate an initial annotated data set. The initial annotation is performed by manual annotation or semi-automatic tools. Manual annotation can be completed by experienced annotators to ensure the accuracy and consistency of the initial annotation data. Semi-automatic tools can assist annotators to improve efficiency and reduce manual operations for some simple tasks. The autonomous learning module trains the machine learning model through the initial labeled data set, and the model obtained after training is a pre-trained model. The machine learning model here uses a deep learning algorithm (such as convolutional neural network, recurrent neural network, Transformer, etc.) to learn the relationship between data features and labels in the initial labeled data set. Supervised learning, semi-supervised learning or unsupervised learning methods can be used in the model training process, and the specific choice depends on the nature of the data and the requirements of the task. The trained pre-trained model is used to preliminarily label another part of the data that is not labeled in the original data set; An intelligent labeling module, taking another part of the original data set as an intermediate data set, and intelligently labeling the intermediate data set through the pre-trained model to generate an intermediate labeled data set; The annotation correction module checks and corrects the intermediate annotation data set to form an intermediate correction data set, and the intermediate correction data set is fed back to the autonomous learning module. The autonomous learning module retrains the pre-trained model through the intermediate correction data set, uses the intermediate correction data set as the intermediate data set, and uses the re-trained pre-trained model to intelligently annotate the intermediate correction data set, thus forming a closed-loop feedback training. In other words, the pre-trained model of the previous module can be continuously optimized, and on the basis of the preliminary annotation to form the intermediate annotation data set, the closed-loop feedback training of the correction can be realized through subsequent manual inspection and correction, thereby gradually improving the annotation accuracy; The data output module outputs the final labeled data, i.e., the result data set, after multiple closed-loop feedback trainings. The data output module performs statistical analysis based on the result data set, forms a labeling report, and outputs it simultaneously with the result data set. The report content of the labeling report includes the accuracy statistics of the labeling, the learning curve of the model, the visualization of the labeling results, etc., to help users fully understand the quality of the labeling process and results.

[0029] The final result data set obtained can achieve better results than existing technologies, whether it is directly analyzed or used as the data basis for other subsequent functional modules (such as automatic processing model training, performance evaluation, etc.). Example

[0030] This embodiment also describes an intelligent labeling method based on machine autonomous learning. The architecture of this method is based on the system described in Example 1. The intelligent labeling method of this embodiment includes the following working steps: Step S1: receiving and preprocessing raw data, the raw data being at least one of image data, text data, and audio data. Preprocessing the acquired raw data to form a raw data set, and dividing the raw data set into two parts, one of which is used as an initial data set and the other as an intermediate data set. This step can import the raw data into the system, and the data input module preprocesses the data to ensure that the format and quality of the data meet the requirements of subsequent processing; Step S2: Initially annotate the initial data set to generate an initial annotated data set. The initial annotation is completed by manual annotation or semi-automatic tools to ensure the accuracy of the annotated initial annotated data set. Initially annotating part of the data by manual or semi-automatic methods can maximize the accuracy of the initial annotated data set, which will be used as a training set for initial training of the model; Step S3: Train the machine learning model using the initial annotated data set, and the model obtained after training is a pre-trained model. This training is the initial training of the machine learning model. This process can adjust the hyperparameters and architecture of the machine learning model, find the optimal model for the data features, and ensure that the model after initial training can at least accurately identify the patterns and features in the data; Step S4: intelligently annotating the intermediate data set through the pre-trained model to generate an intermediate annotated data set, so that the intermediate annotated data set can generate a preliminary annotation result accordingly; Step S5: Check and correct the intermediate labeled data set to form an intermediate corrected data set, retrain the pre-trained model through the intermediate corrected data set, and use the re-trained pre-trained model to intelligently annotate the intermediate corrected data set, and repeat this process to form a closed-loop feedback training and gradually optimize the pre-trained model; Step S6: After multiple closed-loop feedback trainings, the training of the initial data set and the intermediate data set is completed, and the final labeled data, namely the result data set, is output.

[0031] Optionally, the checking and correction of the intermediate annotated data set in step S5 is performed by manual annotation or semi-automatic tools to ensure the accuracy of the intermediate corrected data set.

[0032] Optionally, the machine learning model is trained using a deep learning algorithm based on the initial annotated data set, and the deep learning algorithm includes but is not limited to a convolutional neural network learning algorithm, a recurrent neural network learning algorithm, and a deep neural network learning algorithm.

[0033] Optionally, in step S5, the pre-trained model after closed-loop feedback training is verified and evaluated using the initial annotated data set. When the accuracy of the verification and evaluation exceeds a preset threshold, the initial data set and the intermediate data set are merged and outputted, otherwise the closed-loop feedback training is continued.

[0034] Optionally, the pre-trained model after step S5 is saved as a labeling prediction model, and the labeling prediction model is used to intelligently label data of the same type as the original data.

[0035] Compared with the shortcomings of the prior art, the corresponding invention purposes and advantages of the intelligent labeling system and method based on machine autonomous learning in the patent application of this invention are: 1. Existing technologies have the problem of high time and high cost consumption Existing Disadvantages: Traditional manual labeling methods require a lot of manpower and time, especially when dealing with large-scale data sets, the labeling process often takes weeks or months. This high time and cost consumption limits the efficiency of data labeling and also increases the overall cost of the project.

[0036] Purpose and advantages of the invention: The present invention significantly reduces the workload of manual labeling by introducing autonomous learning and intelligent labeling technology. Through machine autonomous learning, the system can quickly and automatically label large-scale data, thereby significantly shortening the labeling time, reducing the labeling cost, and improving the overall work efficiency.

[0037] 2. Existing technologies are subjective and inconsistent Existing shortcomings: Since manual labeling depends on the knowledge, experience and understanding of the labeler, different labelers may label the same data differently, resulting in subjectivity and inconsistency in the labeling results. This inconsistency will affect the training effect of the model and reduce the accuracy and generalization ability of the model.

[0038] Purpose and advantages of the invention: The invention reduces the interference of human subjective factors and ensures the consistency of the annotation results through the standardized annotation process of the autonomous learning model. The system makes the annotation process more unified and standardized through continuous model optimization, thereby improving the accuracy of annotation and the performance of the model.

[0039] 3. Existing automatic annotation tools have a dependency on the initial model Existing shortcomings: Existing automatic labeling tools usually rely on pre-trained models. If the initial model performance is poor, the accuracy of automatic labeling will be difficult to guarantee, which will increase the workload of manual correction and limit the actual application effect of automatic labeling tools.

[0040] Purpose and advantages of the invention: The autonomous learning module of the present invention can autonomously train the model based on the initial annotation data and continuously optimize it during the annotation process. Even if the initial model performance is limited, through continuous feedback and iteration, the system can gradually improve the accuracy of annotation and reduce the burden of manual correction.

[0041] 4. Existing technologies lack the ability to continuously optimize Current shortcomings: Traditional automatic annotation tools usually cannot effectively use the results of manual corrections for continuous optimization. The correction information during the annotation process is not fed back to the model, resulting in the model being unable to learn and improve from it, ultimately affecting the quality of the annotation results.

[0042] Purpose and advantages of the invention: The present invention designs a closed-loop feedback mechanism, and the user's correction results can be fed back to the model in real time, forming a continuously optimized closed-loop system. Through this mechanism, the system can continuously learn and improve, improve the annotation ability of the model, and make the quality of the annotation results continue to improve over time.

[0043] 5. Existing technologies have poor scalability Existing shortcomings: Existing automatic annotation tools are often designed for specific tasks or data types, and are difficult to adapt to the needs of data annotation in different fields or multiple types. Such tools with poor scalability often perform less than ideal when faced with diverse data.

[0044] Purpose and advantages of the invention: The present invention supports annotation tasks of various data types (such as images, texts, and audios), and the system has strong scalability and adaptability. Whether facing annotation requirements in different fields or processing data in different formats, the system can flexibly adjust the annotation strategy to ensure high quality of the annotation results.

[0045] 6. The existing technology lacks a closed-loop mechanism for annotation correction and feedback Existing shortcomings: In traditional annotation tools, users’ corrections to annotation results are often only partial changes, and the correction information cannot be effectively fed back to the system, making it impossible for the annotation model to learn and optimize from it.

[0046] Purpose and advantages of the invention: The present invention introduces a closed-loop feedback mechanism, and the user's corrected data will be fed back to the model to further optimize the annotation model. This closed-loop feedback mechanism ensures that the system can make timely adjustments and improvements after receiving the corrected data, so that the accuracy and consistency of the annotation are continuously improved.

[0047] At present, the intelligent annotation system and method based on machine autonomous learning of the present application have been widely used in the work of standardizing documents and image processing. The following is an explanation with several examples for easy understanding.

[0048] Application scenario 1: Intelligent labeling system and method for rules and regulations analysis 1. Application scenarios In the field of natural language processing (NLP), regulatory analysis is a common task used to identify and classify regulations (such as laws, regulations, departmental regulations, etc.) expressed in text. Traditional regulatory annotation usually requires manual annotation of large amounts of text, which is not only time-consuming and labor-intensive, but also susceptible to subjective bias. Therefore, the introduction of intelligent annotation systems and methods based on machine autonomous learning can greatly improve the efficiency and accuracy of regulatory annotation.

[0049] 2. Implementation process 2-1. Data preparation and input Collect a large amount of text data, such as system specifications, operating procedures, technical reports, etc. The data input module pre-processes the text data, including noise removal, stop word filtering, word segmentation and other steps to ensure the uniformity of data format.

[0050] 2-2. Initial marking Manual annotators first annotate a small portion of text with regulations, labeling the regulation category (such as "laws", "regulations" or "departmental regulations"). These initial annotated data will be used as training sets for the initial training of the model.

[0051] 2-3. Model training The self-learning module uses the initial annotated data to train the regulation analysis model. The model uses deep learning algorithms (such as the Transformer-based BERT model) to learn the relationship between text features and regulation labels. During the model training process, the model parameters are optimized to improve the accuracy of the model's regulation classification.

[0052] 2-4. Intelligent labeling The unlabeled text data is intelligently labeled using the trained regulatory analysis model. The intelligent labeling module automatically analyzes the regulatory tendency of each text and assigns it a corresponding regulatory label. During the labeling process, the system generates preliminary regulatory labeling data based on the model's prediction results.

[0053] 2-5. Annotation correction and feedback The user reviews the intelligent annotation results and corrects any possible incorrect annotations. The annotation correction module feeds these corrected data back to the autonomous learning module, and the model uses these corrected data for retraining and optimization, thereby improving the model's annotation capabilities.

[0054] 2-6. Result Output Finally, the system outputs the annotated regulation data and generates an annotation report, which includes the accuracy of regulation classification, the model’s learning curve, and the distribution of regulation categories, helping users understand the quality of the annotation results and the performance of the model.

[0055] 3. Effects and advantages By introducing intelligent annotation systems and methods, the efficiency of regulations annotation has been greatly improved, manual intervention has been reduced, and the annotation results have become more consistent and accurate. As the model is continuously optimized, the system can gradually learn and adapt to new expressions of regulations, ensuring that the annotation results remain high quality when processing large amounts of text data from different sources.

[0056] Application scenario 2: Intelligent system for image classification and annotation 1. Application scenarios In the field of computer vision, image classification is another common task. For example, in scenarios such as automatic inspection, drone image analysis, and power transmission line detection, image data annotation is a key step in model training. Traditional image classification annotation usually relies on a lot of manual operations, with low annotation efficiency and prone to errors. Intelligent annotation systems and methods based on machine autonomous learning can significantly improve the annotation efficiency and accuracy of image classification.

[0057] 2. Implementation process 2-1. Data preparation and input Collect and organize large amounts of image data, such as automatic inspection, drone image analysis, transmission line detection, etc. The data input module pre-processes the image, including resizing, standardization, enhancement (such as rotation, flipping, cropping), etc., to ensure the quality and consistency of the image data.

[0058] 2-2. Initial marking Manual labelers perform initial classification and labeling on a portion of the images. For example, they label the automatic inspection images with categories (such as "tree obstacle", "line foreign objects", "insulator contamination and damage", etc.). These labeled data will be used to train the image classification model.

[0059] 2-3. Model training The self-learning module uses the initial annotated data to train the image classification model. The model can use the convolutional neural network (CNN) architecture to automatically learn the mapping relationship between image features and classification labels. During the training process, the system will improve the accuracy of the model's image classification through optimization algorithms.

[0060] 2-4. Intelligent labeling The trained image classification model is used to intelligently annotate unlabeled images. The intelligent annotation module automatically assigns corresponding classification labels to each image through model prediction and generates preliminary annotation results.

[0061] 2-5. Annotation correction and feedback The user checks the results of the intelligent labeling and corrects possible classification errors. The labeling correction module feeds back these correction information to the autonomous learning module to further optimize the model. As the feedback data increases, the classification accuracy of the model will gradually improve, reducing the workload of subsequent corrections.

[0062] 2-6. Result Output Finally, the system outputs the annotated image data and generates an annotation report, which includes classification accuracy, error analysis, category distribution, etc., so that users can evaluate the quality of the annotation results and the performance of the model.

[0063] 3. Effects and advantages Through the application of intelligent annotation systems and methods, the efficiency of image classification annotation has been greatly improved, reducing the workload of manual annotation. The closed-loop feedback mechanism of the system ensures continuous optimization of the model, and the accuracy and consistency of classification results have been significantly improved. Especially when processing massive image data, the system can quickly generate high-quality annotation results, providing a solid data foundation for downstream applications (such as model training and performance evaluation). Application scenario 3: Intelligent annotation system and method for drone images 1. Application scenarios Drone image annotation is an important application in the field of drone AI. For example, during drone inspections, maintenance engineers need to annotate a large number of inspection photos. For example, during the training of the transmission line inspection model, the photos taken by the drone are annotated to help train the model for transmission line inspection. Due to the high professional requirements for drone image annotation and the huge workload, the traditional manual annotation method is inefficient and prone to errors. By introducing intelligent annotation systems and methods, the accuracy and efficiency of annotation can be improved.

[0064] 2. Implementation process 2-1. Data preparation and input Collect a large amount of drone image data, such as aerial work, infrared images, etc. The data input module pre-processes the image data, including standardization, denoising, contrast enhancement and other operations to ensure the quality of the image.

[0065] 2-2. Initial marking Maintenance engineers initially annotate some images. For example, in the case of power line inspection (which will also be used as an example below), the annotations may include information such as the location, size, and type of the power line and its fault area. These annotated data are used to train the drone image analysis model.

[0066] 2-3. Model training The self-learning module uses the initial labeled data to train the drone image analysis model. The model uses a deep learning architecture (such as a segmentation model based on U-Net) to learn the relationship between image features and transmission lines and their fault areas. During the model training process, parameters are continuously optimized to improve the accuracy of detection and segmentation.

[0067] 2-4. Intelligent labeling The trained model is used to intelligently annotate the unlabeled image data, and the system automatically identifies and annotates the transmission lines and their fault areas. The intelligent annotation module generates preliminary image annotation results, including the boundaries, size, type, and other information of the transmission lines and their fault areas.

[0068] 2-5. Annotation correction and feedback Maintenance engineers check the intelligent labeling results and correct the wrong labels. The labeling correction module feeds the corrected data back to the autonomous learning module, and the model uses the corrected data for retraining and optimization, thereby improving the model's labeling ability.

[0069] 2-6. Result Output The system outputs the labeled drone image data and generates a detailed labeling report. The report includes labeling accuracy, labeling area statistics, model performance indicators, etc., to help maintenance engineers and researchers evaluate the labeling quality and model performance.

[0070] 3. Effects and advantages The efficiency and accuracy of drone image annotation have been greatly improved through the intelligent annotation system and method. The system reduces the annotation burden of maintenance engineers, allowing them to focus on more complex tasks. At the same time, the system's feedback mechanism ensures that the annotation results are continuously optimized, providing high-quality data support for the training of inspection photo analysis models, which helps to improve the reliability and practicality of drone AI applications.

[0071] In summary, the intelligent labeling system and method based on machine autonomous learning of the present application, by introducing autonomous learning technology, can autonomously train the model based on the initial data labeling, and introduces a labeling correction and feedback mechanism. The user's correction results can be effectively fed back to the model training process, continuously improving the accuracy and consistency of the labeling, and forming a closed-loop system that continuously optimizes the model. The labeling results are continuously optimized through the closed-loop feedback mechanism, while minimizing human intervention, achieving efficient and accurate data labeling.

[0072] Furthermore, the present application can support labeling tasks for multiple data types (such as images, text, and audio). It only needs to adjust the labels during the initial labeling. The machine learning model with attention mechanism can flexibly adjust the labeling strategy according to different task requirements, and has strong scalability and adaptability.

[0073] Furthermore, the present application can provide detailed annotation reports and visual correction tools, so that users can conduct independent inspections and corrections, and at the same time can intuitively understand the annotation results and model performance, which facilitates subsequent data analysis and decision-making.

[0074] Specifically, the beneficial effects of this application are summarized as follows: 1. Improve annotation efficiency and reduce labor costs Effect: The present invention significantly reduces the workload of manual labeling by introducing machine autonomous learning technology. The system can autonomously learn and perform intelligent labeling based on the initial labeling, which significantly shortens the labeling time, especially when processing large-scale data. This not only improves the labeling efficiency, but also greatly reduces the cost of manual labeling.

[0075] 2. Improve annotation accuracy and consistency Effect: Traditional manual labeling is easily affected by subjective factors of the labelers, resulting in inconsistency in the labeling results. This invention ensures the standardization and consistency of the labeling results through machine autonomous learning and closed-loop feedback mechanism, reducing the errors caused by human subjective factors. With the continuous optimization of the model, the accuracy of the labeling results has also been continuously improved.

[0076] 3. Support annotation of multi-task and multi-type data Effect: The present invention is flexible in design and can adapt to the labeling tasks of various data types (such as images, texts, audio, etc.). This multi-tasking capability enables the system to play a good role in applications in different fields, whether it is natural language processing, computer vision, or medical image analysis, the system can respond efficiently.

[0077] 4. Closed-loop feedback mechanism to continuously optimize model performance Effect: Through a closed-loop feedback mechanism, the system feeds back the user's annotation correction results to the model for further training and optimization. This mechanism enables the model to continuously learn and improve its own performance, ensuring that the accuracy and effect of annotations are continuously optimized over time to adapt to the needs of new data and new scenarios.

[0078] 5. Reduce manual intervention and improve automation level Effect: The intelligent labeling system of the present invention can automatically complete most labeling tasks with less human intervention. This greatly reduces the reliance on manual labeling, makes the data labeling process more automated and intelligent, and frees up the workload of labelers, allowing them to focus on more complex tasks.

[0079] 6. High scalability and adaptability Effect: The system adopts a modular design, each module has clear functions, can be flexibly adjusted according to different annotation tasks, and has good scalability. Users can customize the functions of the system according to actual needs, adapt to the annotation needs of different fields and application scenarios, and ensure that the system can run stably and efficiently in various environments.

[0080] 7. Transparency and traceability of the labeling process Effect: The system provides detailed annotation reports and visualization tools, allowing users to intuitively understand the annotation process and results. Each step of the annotation is traceable, making it easier for users to conduct quality assessments and locate problems. This transparency and traceability enhances the credibility of the annotation work and makes it easier for users to review and optimize the annotation results.

[0081] 8. Improve data annotation quality and help downstream applications Effect: Efficient and accurate annotation not only improves the quality of data, but also provides a reliable data foundation for downstream machine learning model training. High-quality annotated data helps improve the training effect and actual application performance of the model, thereby promoting the development of the entire data-driven application.

[0082] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with this technology to understand the contents of the present invention and implement them accordingly. They cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. An intelligent labeling system based on machine autonomous learning, characterized in that: include: A data input module is used to receive and preprocess raw data, and preprocess the acquired raw data to form a raw data set; An initial annotation module, taking a part of the original data set as an initial data set and performing initial annotation on the initial data set to generate an initial annotated data set; An autonomous learning module trains a machine learning model using the initial annotated data set, and the model obtained after training is a pre-trained model; An intelligent labeling module, taking another part of the original data set as an intermediate data set, and intelligently labeling the intermediate data set through the pre-trained model to generate an intermediate labeled data set; The annotation correction module checks and corrects the intermediate annotation data set to form an intermediate correction data set, and the intermediate correction data set is fed back to the autonomous learning module. The autonomous learning module retrains the pre-trained model through the intermediate correction data set, uses the intermediate correction data set as the intermediate data set, and uses the re-trained pre-trained model to intelligently annotate the intermediate correction data set, thereby forming a closed-loop feedback training; The data output module outputs the final labeled data, i.e., the result data set, after multiple closed-loop feedback trainings.

2. The intelligent labeling system based on machine autonomous learning according to claim 1 is characterized in that: The initial annotation is performed by manual annotation or semi-automatic tools.

3. The intelligent labeling system based on machine autonomous learning according to claim 1 is characterized in that: The data output module performs statistical analysis based on the result data set, forms a labeling report and outputs it simultaneously with the result data set.

4. An intelligent labeling method based on machine autonomous learning, characterized in that: The following steps are included: Step S1: receiving and preprocessing raw data, and preprocessing the acquired raw data to form an original data set, and dividing the original data set into two parts, one of which is used as an initial data set, and the other is used as an intermediate data set; Step S2: Initially annotating the initial data set to generate an initial annotated data set; Step S3: training the machine learning model using the initial annotated data set, and the model obtained after training is a pre-trained model; Step S4: Intelligently annotating the intermediate data set by using the pre-trained model to generate an intermediate annotated data set; Step S5: Check and correct the intermediate annotated data set to form an intermediate corrected data set, retrain the pre-trained model with the intermediate corrected data set, and use the re-trained pre-trained model to intelligently annotate the intermediate corrected data set, and repeat this process to form a closed-loop feedback training; Step S6: After multiple closed-loop feedback trainings, the training of the initial data set and the intermediate data set is completed, and the final labeled data, namely the result data set, is output.

5. The intelligent labeling method based on machine autonomous learning according to claim 4 is characterized in that: The initial annotation in step S2 is completed by manual annotation or semi-automatic tools to ensure the accuracy of the initial annotation data set after annotation.

6. The intelligent labeling method based on machine autonomous learning according to claim 4 is characterized in that: In step S5, the checking and correction of the intermediate annotated data set is performed by manual annotation or semi-automatic tools to ensure the accuracy of the intermediate corrected data set.

7. The intelligent labeling system based on machine autonomous learning according to claim 4 is characterized in that: The original data is at least one of image data, text data, and audio data.

8. The intelligent labeling method based on machine autonomous learning according to any one of claims 4 to 7, characterized in that: The machine learning model is trained according to the initial labeled data set through a deep learning algorithm, and the deep learning algorithm includes but is not limited to a convolutional neural network learning algorithm, a recurrent neural network learning algorithm, and a deep neural network learning algorithm.

9. The intelligent labeling method based on machine autonomous learning according to any one of claims 4 to 7, characterized in that: In step S5, the pre-trained model after closed-loop feedback training is verified and evaluated using the initial annotated data set. When the accuracy of the verification and evaluation exceeds a preset threshold, the initial data set and the intermediate data set are merged and outputted, otherwise the closed-loop feedback training is continued.

10. The intelligent labeling method based on machine autonomous learning according to claim 9, characterized in that: The pre-trained model after step S5 is saved as a labeling prediction model, and the labeling prediction model is used to intelligently label data of the same type as the original data.

Images