Automatic data annotation method and device based on large model

Abstract

The invention belongs to the field of data processing, and particularly relates to an automatic data annotation method and device based on a large model, and the method comprises the steps: S1, obtaining original data which is image data; s2, features in the image data are recognized through a pre-trained deep learning model, first labeling is carried out, and a target labeling result output according to a standard format is formed; and S3, inputting the annotation result into a large-scale deep learning model, carrying out secondary standard by the large-scale deep learning model, and outputting an image annotation result of the image data. According to the method and the device, efficient, low-cost, high-quality and safe automatic data labeling is realized.

Description

Technical Field

[0001] The present application relates to the field of data processing, and in particular to a large model-based automatic data labeling method and device. Background Art

[0002] With the rapid development of artificial intelligence and machine learning technologies, data labeling, as a key link in model training, is particularly important. High-quality data labeling is the basis for training high-performance models. However, current data labeling mainly relies on manual work, which has many problems such as high cost, low efficiency, and uneven quality. Especially when facing large-scale data sets, manual labeling is particularly difficult and cannot meet the needs of rapid iteration and efficient training. In order to improve the efficiency and quality of data labeling, automated data labeling technology has gradually attracted attention.

[0003] Existing automated data labeling methods mainly include manual labeling, semi-automatic labeling, rule-based automatic labeling, and machine learning-based automatic labeling.

[0004] There are many manual annotation platforms on the market, such as Amazon Mechanical Turk and Figure Eight (formerly CrowdFlower). These platforms mainly rely on crowdsourcing to complete data annotation tasks through a large number of manual annotators. Manual annotation is inefficient and costly, and the quality of annotation depends on the experience and ability of the annotator, which is prone to inconsistency and subjective bias. Semi-automatic annotation tools such as Labelbox and SuperAnnotate improve the efficiency of annotation by combining manual and automated methods. The system first performs preliminary annotation, and then manually corrects and improves it. Although this method improves efficiency, it still requires a lot of manual participation, and the accuracy of automatic annotation is limited when facing complex tasks. Rule-based automatic annotation, such as NER (named entity recognition) tools, often use predefined dictionaries and rules. The disadvantage is that the flexibility of rules and templates is limited, and it is difficult to handle complex and changeable annotation tasks. The annotation effect depends on the completeness and accuracy of the rules. Some systems have begun to use traditional machine learning algorithms (such as SVM and random forest) for automatic annotation, and train classifiers to identify and annotate data. Traditional machine learning algorithms have limited performance when processing large-scale and complex data, and require a lot of feature engineering, and are not as effective as deep learning models. Summary of the invention

[0005] In order to solve the above problems, this application provides an automated data annotation method and device based on a large model, aiming to solve the problems of high cost, low efficiency, inconsistent annotation quality, poor adaptability, etc. in the prior art. By further optimizing and fine-tuning the pre-trained model, the system can efficiently and accurately handle large-scale and complex data annotation tasks, and improve the overall effect of data annotation.

[0006] The first aspect of the present application provides an automatic data annotation method based on a large model, which mainly includes:

[0007] Step S1, obtaining original data, wherein the original data is image data;

[0008] Step S2: Identify features in the image data through a pre-trained deep learning model, and perform a first annotation to form a target annotation result output in a standard format;

[0009] Step S3: input the annotation results into a large-scale deep learning model, and the large-scale deep learning model performs a second standardization to output image annotation results for the image data.

[0010] Preferably, step S1 further comprises:

[0011] Step S11, checking the correctness of the format and content of the original data;

[0012] Step S12, converting the original data into a unified data format;

[0013] Step S13, deleting duplicate data and filling in missing values ​​of original data;

[0014] Step S14: performing noise filtering on the original data.

[0015] Preferably, in step S13, the missing values ​​of the original data are filled by interpolation or filling with default values.

[0016] Preferably, in step S2, the standard format includes one or more annotation contents, each of which includes the target category, position and confidence level in the graph.

[0017] Preferably, step S3 further comprises:

[0018] Step S31: Encrypt the image annotation result.

[0019] The second aspect of the present application provides an automatic data annotation device based on a large model, which mainly includes:

[0020] A data acquisition module, used to acquire original data, wherein the original data is image data;

[0021] The image feature recognition module is used to identify features in image data through a pre-trained deep learning model and perform the first annotation to form a target annotation result output in a standard format;

[0022] The large model annotation module is used to input the annotation results into the large-scale deep learning model, and the large-scale deep learning model performs a second standard to output the image annotation results of the image data.

[0023] Preferably, the data acquisition module includes:

[0024] A content checking unit, used to check the correctness of the format and content of the original data;

[0025] A format conversion unit, used for converting the original data into a unified data format;

[0026] A data filling unit is used to delete duplicate data and fill in missing values ​​of original data;

[0027] The noise filtering unit is used to perform noise filtering on the original data.

[0028] Preferably, in the data filling unit, missing values ​​of the original data are filled by interpolation or filling with default values.

[0029] Preferably, in the image feature recognition module, the standard format includes one or more annotation contents, and each annotation content includes the target category, position and confidence level in the graphic.

[0030] Preferably, the large model annotation module includes:

[0031] The data encryption unit is used to encrypt the image annotation results.

[0032] This application achieves efficient, low-cost, high-quality and secure automated data labeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flowchart of a preferred embodiment of the automatic data annotation method based on a large model of the present application. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical scheme and advantages of the implementation of this application clearer, the technical scheme in the implementation of this application will be described in more detail in combination with the drawings in the implementation of this application. In the drawings, the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions. The described implementation is a part of the implementation of this application, not all of the implementations. The implementation described below with reference to the drawings is exemplary and is intended to be used to explain this application, and cannot be understood as a limitation on this application. Based on the implementation in this application, all other implementations obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. The implementation of this application is described in detail below in combination with the drawings.

[0035] This application provides an automated data annotation method based on a large model, such as Figure 1 As shown, it mainly includes:

[0036] Step S1, obtaining original data, wherein the original data is image data.

[0037] In some optional implementations, step S1 further includes:

[0038] Step S11, checking the correctness of the format and content of the original data;

[0039] Step S12, converting the original data into a unified data format;

[0040] Step S13, deleting duplicate data and filling in missing values ​​of original data;

[0041] Step S14: performing noise filtering on the original data.

[0042] In this embodiment, after obtaining the original data from various data sources (such as databases, file systems, API interfaces, etc.), the original data needs to be preprocessed. Step S11 is a preliminary check, which performs a preliminary check on the obtained data to ensure the basic correctness of the data format and content. Step S12 is format conversion: converting the data into a unified format for subsequent processing and labeling. Step S13 is deduplication and missing value filling: detecting and deleting duplicate data, ensuring the uniqueness of the data set, and processing missing values ​​in the data. In some optional implementations, in step S13, the missing values ​​of the original data are filled by interpolation and filling default values, thereby improving data integrity. Step S14 is noise filtering: removing noise and irrelevant information in the data to improve data quality.

[0043] The large-scale data set is then divided into multiple batches for easier system processing and resource management. The batched data is then imported into the system to prepare for subsequent large model processing.

[0044] Step S2: Identify features in the image data through a pre-trained deep learning model, and perform the first annotation to form a target annotation result output in a standard format.

[0045] This step outputs image features through image target detection. First, the input image and video data are preprocessed by format conversion, size adjustment, denoising, etc. to ensure data quality and model input requirements. Then, deep learning algorithms such as convolutional neural networks (CNN) are used to extract features from image and video data. Common models include ResNet, VGG, etc. Then, pre-trained visual models (such as YOLO, Faster R-CNN, etc.) are used for preliminary recognition and annotation, and the results of model recognition are converted into standardized annotation formats, and annotation files (such as JSON, XML, etc.) are generated. In some optional implementations, in step S2, the standard format includes one or more annotation contents, and each annotation content includes the target category, position and confidence in the graphic.

[0046] It should also be noted that the pre-trained model can be fine-tuned according to specific tasks and data sets to improve recognition accuracy and labeling quality in specific application scenarios. On the other hand, the preliminary labeling results are optimized, including removing redundant labels, merging overlapping labels, and improving labeling consistency to ensure high quality of the labeling results.

[0047] Step S3: input the annotation results into a large-scale deep learning model, and the large-scale deep learning model performs a second standardization to output image annotation results for the image data.

[0048] It is understandable that large-scale pre-trained models based on deep learning (such as BERT, GPT) are usually called big models. They are ready-made models that show powerful capabilities in natural language processing tasks. These models can capture rich semantic information by pre-training on large-scale data sets and are suitable for a variety of downstream tasks. For example, the big model is used in step S3 of this application to process the text content output by step S2 to perform image annotation on the original image data.

[0049] It should also be noted that the large model can be fine-tuned according to specific tasks and data sets to ensure that the model can accurately understand and annotate the data.

[0050] Finally, the results of large model annotation are generated into a standardized output format (such as JSON, XML, etc.), and the annotation results are stored in a database or file system for subsequent query and use. Before that, the annotation results can also be checked for consistency and accuracy to ensure the correctness and reliability of the annotation; or optimization algorithms can be applied to improve the annotation quality, and manual review can be performed if necessary to further ensure accuracy.

[0051] In some optional implementations, step S3 further includes:

[0052] Step S31: Encrypt the image annotation result.

[0053] In this embodiment, encryption and other security technologies are used in the labeling process to protect data privacy and prevent data leakage and abuse.

[0054] This application uses the deep learning capabilities of a large model to handle complex annotation tasks, including semantic understanding and context association, to improve the comprehensiveness and accuracy of annotations, with the following specific advantages:

[0055] By replacing most manual operations with automated systems, labor costs can be significantly reduced. Especially in the annotation of large-scale data sets, the cost advantage is more obvious.

[0056] By leveraging the high-speed computing power of large models, large amounts of data can be quickly processed, significantly improving labeling efficiency and meeting the needs of rapid iteration and efficient training.

[0057] Through the consistency and high precision of large models, the stability and consistency of annotation quality are ensured, eliminating the subjective differences of manual annotation.

[0058] Large models have strong transfer learning capabilities and can quickly adapt to data labeling needs in different fields and types without the need for frequent adjustments and retraining.

[0059] Reduce the chance of manual access to data, improve data privacy and security through built-in security measures and encryption technology, and reduce the risk of data leakage.

[0060] The deep learning capabilities based on large models can handle complex labeling tasks, improve labeling accuracy and effectiveness, and reduce the difficulty and error rate of manual labeling.

[0061] The second aspect of the present application provides an automatic data annotation device based on a large model corresponding to the above method, which mainly includes:

[0062] A data acquisition module, used to acquire original data, wherein the original data is image data;

[0063] The image feature recognition module is used to identify features in image data through a pre-trained deep learning model and perform the first annotation to form a target annotation result output in a standard format;

[0064] The large model annotation module is used to input the annotation results into the large-scale deep learning model, and the large-scale deep learning model performs a second standard to output the image annotation results of the image data.

[0065] In some optional implementations, the data acquisition module includes:

[0066] A content checking unit, used to check the correctness of the format and content of the original data;

[0067] A format conversion unit, used for converting the original data into a unified data format;

[0068] A data filling unit is used to delete duplicate data and fill in missing values ​​of original data;

[0069] The noise filtering unit is used to perform noise filtering on the original data.

[0070] In some optional implementations, in the data filling unit, missing values ​​of the original data are filled by interpolation or filling with default values.

[0071] In some optional implementations, in the image feature recognition module, the standard format includes one or more annotation contents, and each annotation content includes a target category, a position, and a confidence level in the graphic.

[0072] In some optional implementations, the large model annotation module includes: a data encryption unit, used to encrypt the image annotation result.

[0073] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. An automatic data annotation method based on a large model, characterized in that: include: Step S1, obtaining original data, wherein the original data is image data; Step S2: Identify features in the image data through a pre-trained deep learning model, and perform a first annotation to form a target annotation result output in a standard format; Step S3: input the annotation results into a large-scale deep learning model, and the large-scale deep learning model performs a second standardization to output image annotation results for the image data.

2. The large model-based automatic data annotation method according to claim 1, characterized in that: Step S1 further comprises: Step S11, checking the correctness of the format and content of the original data; Step S12, converting the original data into a unified data format; Step S13, deleting duplicate data and filling in missing values ​​of original data; Step S14: performing noise filtering on the original data.

3. The large model-based automatic data annotation method according to claim 2, characterized in that: In step S13, the missing values ​​of the original data are filled by interpolation or filling with default values.

4. The large model-based automatic data annotation method according to claim 1, characterized in that: In step S2, the standard format includes one or more annotation contents, each of which includes the target category, position and confidence level in the graph.

5. The large model-based automatic data annotation method according to claim 1, characterized in that: Step S3 further comprises: Step S31: Encrypt the image annotation result.

6. An automatic data annotation device based on a large model, characterized in that: include: A data acquisition module, used to acquire original data, wherein the original data is image data; The image feature recognition module is used to identify features in image data through a pre-trained deep learning model and perform the first annotation to form a target annotation result output in a standard format; The large model annotation module is used to input the annotation results into the large-scale deep learning model, and the large-scale deep learning model performs a second standard to output the image annotation results of the image data.

7. The large model-based automatic data annotation device according to claim 6, characterized in that: The data acquisition module comprises: A content checking unit, used to check the correctness of the format and content of the original data; A format conversion unit, used for converting the original data into a unified data format; A data filling unit is used to delete duplicate data and fill in missing values ​​of original data; The noise filtering unit is used to perform noise filtering on the original data.

8. The large model-based automatic data annotation device according to claim 7, characterized in that: In the data filling unit, missing values ​​of the original data are filled by interpolation and filling with default values.

9. The large model-based automatic data annotation device according to claim 6, characterized in that: In the image feature recognition module, the standard format includes one or more annotation contents, and each annotation content includes the target category, position and confidence level in the image.

10. The large model-based automatic data annotation device according to claim 6, characterized in that: The large model annotation module includes: The data encryption unit is used to encrypt the image annotation results.

Images