A semi-supervised automatic data annotation method based on deep learning, computer equipment and storage media

Abstract

This invention provides a semi-supervised automatic data annotation method, computer device, and storage medium based on deep learning. The method includes: acquiring a frame-aligned multimodal dataset to be annotated; manually annotating a portion of the dataset to be annotated, dividing the dataset into an annotated dataset and an unannotated dataset; automatically annotating the unannotated dataset using the BEV deep learning algorithm, obtaining an annotated dataset and a new unannotated dataset; determining whether the new unannotated dataset is empty; if the new unannotated dataset is not empty, evaluating the new annotated dataset and the new unannotated dataset using a binary classification deep learning algorithm. The method designed in this invention uses a semi-supervised approach to automatically annotate the dataset, while simultaneously employing a binary classification deep learning algorithm to evaluate and optimize the annotation results, which can significantly improve the accuracy of the annotation results output by the deep learning algorithm for automatic annotation.

Description

Technical Field

[0001] This invention relates to the field of data annotation technology, specifically to a semi-supervised automatic data annotation method, computer equipment, and storage medium based on deep learning. Background Technology

[0002] With the advancement of technology, deep learning algorithms and models are gradually being applied to various industries. In the industrial sector, in order to better use deep learning methods to solve practical business problems, in addition to continuously optimizing deep learning models, the quality of datasets can also be improved by data cleaning and labeling for specific business purposes. The data labeling process often requires a lot of manpower and time, while scientific and efficient data labeling methods can greatly improve the efficiency of producing high-quality data and greatly reduce the cost of producing data.

[0003] Data annotation is usually done manually by professional data engineers or annotators. Data annotation is inefficient and time-consuming. Currently, there are also methods that use deep learning networks to process the raw data and output an uncalibrated annotation result, which is then manually calibrated by professional annotators. Even so, the manual calibration work by annotators is still time-consuming and labor-intensive, especially for the annotation of laser point cloud data in the field of autonomous driving. Due to the characteristics of laser point cloud data, a single frame of data often contains tens of thousands of point cloud data, which greatly reduces the efficiency of data annotation.

[0004] Furthermore, after data annotation is completed, traditional image processing or machine learning feature extraction methods are usually used to evaluate the annotation results output by the deep learning algorithm for automatic annotation. The accuracy of the evaluation results is low, which cannot provide accurate data support for data annotation and reduces the accuracy of data annotation.

[0005] Therefore, it is necessary to design a data annotation method that is efficient and accurate. Summary of the Invention

[0006] To address the problems of time-consuming, labor-intensive, inefficient, and low-accuracy data annotation methods, this invention designs a semi-supervised automatic data annotation method, computer equipment, and storage medium based on deep learning. This method uses a semi-supervised approach to automatically annotate the dataset, while employing a binary classification deep learning algorithm to evaluate and optimize the annotation results, which can greatly improve the accuracy of the annotation results output by the deep learning algorithm for automatic annotation.

[0007] The technical solution to achieve the purpose of the invention is as follows: A semi-supervised automatic data annotation method based on deep learning, comprising the following steps:

[0008] S1. Obtain the dataset to be labeled;

[0009] S2. Manually label a portion of the data in the dataset to be labeled, dividing the dataset into labeled and unlabeled datasets.

[0010] S3. Automatically label the unlabeled dataset based on the BEV deep learning algorithm to obtain labeled datasets and new unlabeled datasets;

[0011] S4. Determine whether the new unlabeled dataset is empty;

[0012] If the new unlabeled dataset is empty, then the automatic labeling of the dataset to be labeled is completed;

[0013] If the new unlabeled dataset is not empty, proceed to and execute S5;

[0014] S5. Based on a binary classification deep learning algorithm, evaluate new labeled datasets and new unlabeled datasets.

[0015] In one embodiment, step S3 above, which involves automatically labeling the unlabeled dataset based on the BEV deep learning algorithm to obtain a labeled dataset and a new unlabeled dataset, includes the following steps:

[0016] S31. Use the pre-trained BEV deep learning algorithm to perform transfer training on the labeled dataset to obtain the fine-tuned BEV deep learning algorithm.

[0017] S32. Automatically label the unlabeled dataset based on the fine-tuned BEV deep learning algorithm, and divide the unlabeled dataset into high-confidence dataset and low-confidence dataset based on the confidence threshold.

[0018] S33. Merge the labeled dataset with the high-confidence dataset to form a new labeled dataset, and use the low-confidence dataset as a new unlabeled dataset and proceed to step S4 to determine whether it is empty.

[0019] In another embodiment, step S3 above involves automatically labeling the unlabeled dataset based on the BEV deep learning algorithm to obtain a labeled dataset and a new unlabeled dataset, including the following steps:

[0020] S31. Use the pre-trained BEV deep learning algorithm to perform transfer training on the labeled dataset to obtain the fine-tuned BEV deep learning algorithm.

[0021] S32. Automatically label the unlabeled dataset based on the fine-tuned BEV deep learning algorithm, and divide the unlabeled dataset into high-confidence dataset and low-confidence dataset based on the confidence threshold.

[0022] S33. Merge the labeled dataset with the high-confidence dataset to form a new labeled dataset, and proceed to step S4 to determine whether the low-confidence dataset is empty;

[0023] If the low-confidence dataset is empty, then the automatic labeling of the dataset to be labeled is completed.

[0024] If the low-confidence dataset is not empty, proceed to and execute S34;

[0025] S34. Use the fine-tuned BEV deep learning algorithm to perform supervised training on the new labeled dataset to obtain a new fine-tuned BEV deep learning algorithm.

[0026] S35. Based on the new fine-tuned BEV deep learning algorithm, the low-confidence dataset is automatically labeled, and combined with the confidence threshold, the low-confidence dataset is divided into a new high-confidence dataset and a new low-confidence dataset.

[0027] S36. Merge the new labeled dataset with the new high-confidence dataset, and use the new low-confidence dataset as the new unlabeled dataset and proceed to step S4 to determine whether it is empty.

[0028] If the new unlabeled dataset is empty, then the automatic labeling of the dataset to be labeled is completed;

[0029] If the new unlabeled dataset is not empty, proceed to and execute S37;

[0030] S37. Repeat steps S34 to S36 until the new unlabeled dataset is not empty and the data volume does not change, then proceed to and execute S5.

[0031] Furthermore, the aforementioned dataset to be labeled includes a multimodal dataset formed by frame-aligned laser point clouds and multi-view images.

[0032] Furthermore, the labeled dataset in step S2 above consists of manually labeled and calibrated data.

[0033] Furthermore, in step S5 above, the evaluation of the new labeled dataset and the new unlabeled dataset based on the binary classification deep learning algorithm includes the following steps:

[0034] S51. Mark the data in the new labeled dataset as 0, and mark the data in the new unlabeled dataset as 1;

[0035] S52. Use a pre-trained binary classification deep learning algorithm to perform transfer training on a new labeled dataset and a new unlabeled dataset to obtain a fine-tuned binary classification deep learning algorithm.

[0036] S53. Input the new labeled dataset and the new unlabeled dataset into a fine-tuned binary classification deep learning algorithm, and output the binary classification confidence of each data in the new labeled dataset and the new unlabeled dataset to form a dataset with binary classification confidence.

[0037] S54. Sort the datasets with binary classification confidence scores in descending order according to their binary classification confidence scores, and extract the top n data points to form a new dataset.

[0038] S55. Calculate the overlap between the new dataset and the new unlabeled dataset;

[0039] S56. Compare the overlap ratio with the overlap threshold;

[0040] If the overlap is greater than the overlap threshold, it is determined that the new dataset is significantly different from the new unlabeled dataset. In this case, the data in the new unlabeled dataset is manually labeled and calibrated.

[0041] If the overlap is less than the overlap threshold, the data annotation results in the new labeled dataset are deemed unusable.

[0042] Compared with existing technologies, the beneficial effects of this invention are as follows: The semi-supervised automatic data annotation method based on deep learning designed in this invention is particularly suitable for multimodal datasets formed from frame-aligned laser point clouds and multi-view images. Its automatic annotation process, designed using the BEV deep learning algorithm and a transfer learning and semi-supervised strategy, solves the problem that the cost of automatic annotation for multimodal datasets is far higher than that for single-modal datasets when manual fusion annotation is required. Furthermore, this invention specifically designs a binary classification deep learning algorithm to evaluate the annotation results using the BEV deep learning algorithm, ensuring the quality of automatic annotation of the dataset by the BEV deep learning algorithm. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0044] Figure 1 The flowchart shows the semi-supervised automatic data annotation method based on deep learning of the present invention.

[0045] Figure 2 This is a block diagram illustrating the principle of the first fine-tuning of the pre-trained BEV deep learning algorithm in a specific implementation.

[0046] Figure 3 The flowchart shows the process of automatically labeling the unlabeled dataset D2 using a one-time fine-tuning BEV deep learning algorithm in a specific implementation.

[0047] Figure 4This is a block diagram illustrating the principle of a second fine-tuning of the BEV deep learning algorithm after a first fine-tuning in a specific implementation.

[0048] Figure 5 The flowchart shows the process of automatically labeling the low-confidence auto-labeling dataset D4 using the second-order fine-tuning BEV deep learning algorithm in a specific implementation.

[0049] Figure 6 This is a block diagram illustrating the principle of fine-tuning the pre-trained binary classification deep learning algorithm in a specific implementation.

[0050] Figure 7 This is a flowchart illustrating the process of using a fine-tuned binary classification deep learning algorithm to evaluate labeled and unlabeled datasets in a specific implementation. Detailed Implementation

[0051] The present invention will be further described below with reference to specific embodiments, and the advantages and features of the present invention will become clearer as a result. However, these embodiments are merely exemplary and do not constitute any limitation on the scope of the present invention. Those skilled in the art should understand that modifications or substitutions can be made to the details and form of the technical solutions of the present invention without departing from the spirit and scope of the present invention, but all such modifications and substitutions fall within the protection scope of the present invention.

[0052] This specific implementation discloses a semi-supervised data automatic annotation method based on deep learning, see [link to relevant documentation]. Figure 1 As shown, the semi-supervised data automatic annotation method based on deep learning includes the following steps:

[0053] S1. Obtain the dataset to be labeled.

[0054] In this step, the labeled dataset is preferably a multimodal dataset, such as a multimodal dataset formed by frame-aligned laser point clouds and multi-view images.

[0055] S2. Manually label a portion of the data in the dataset to be labeled, dividing the dataset into labeled and unlabeled datasets.

[0056] In this step, the data in the labeled dataset is manually labeled and calibrated.

[0057] S3. Automatically label the unlabeled dataset based on the BEV deep learning algorithm to obtain labeled datasets and new unlabeled datasets.

[0058] In one embodiment of this step, the unlabeled dataset is automatically labeled based on the BEV deep learning algorithm (i.e., Bird Eye View) to obtain labeled datasets and new unlabeled datasets, including the following steps:

[0059] S31. Use the pre-trained BEV deep learning algorithm to perform transfer training on the labeled dataset to obtain the fine-tuned BEV deep learning algorithm.

[0060] At this point, the weights of the pre-trained BEV deep learning algorithm are W0, and the weights of the fine-tuned BEV deep learning algorithm are W1.

[0061] S32. Automatically label the unlabeled dataset based on the fine-tuned BEV deep learning algorithm, and divide the unlabeled dataset into high-confidence dataset and low-confidence dataset based on the confidence threshold.

[0062] S33. Merge the labeled dataset with the high-confidence dataset to form a new labeled dataset, and use the low-confidence dataset as a new unlabeled dataset and proceed to step S4 to determine whether it is empty.

[0063] If the new unlabeled dataset is empty, then the automatic labeling of the dataset to be labeled is completed.

[0064] If the new unlabeled dataset is not empty, proceed to and execute S5.

[0065] In another embodiment of this step, step S3 above, which involves automatically labeling the unlabeled dataset based on the BEV deep learning algorithm to obtain a labeled dataset and a new unlabeled dataset, includes the following steps:

[0066] S31. Use the pre-trained BEV deep learning algorithm to perform transfer training on the labeled dataset to obtain the fine-tuned BEV deep learning algorithm. At this time, the weights of the pre-trained BEV deep learning algorithm are W0, and the weights of the fine-tuned BEV deep learning algorithm are W1.

[0067] S32. Automatically label the unlabeled dataset based on the fine-tuned BEV deep learning algorithm, and divide the unlabeled dataset into high-confidence dataset and low-confidence dataset based on the confidence threshold.

[0068] S33. Merge the labeled dataset with the high-confidence dataset to form a new labeled dataset, and proceed to step S4 to determine whether the low-confidence dataset is empty.

[0069] If the low-confidence dataset is empty, then the automatic labeling of the dataset to be labeled is completed.

[0070] If the low-confidence dataset is not empty, proceed to and execute S34.

[0071] S34. Use the fine-tuned BEV deep learning algorithm to perform supervised training on the new labeled dataset to obtain a new fine-tuned BEV deep learning algorithm. At this time, the weight of the new fine-tuned BEV deep learning algorithm is W2.

[0072] S35. Based on the new fine-tuned BEV deep learning algorithm, the low-confidence dataset is automatically labeled, and combined with the confidence threshold, the low-confidence dataset is divided into a new high-confidence dataset and a new low-confidence dataset.

[0073] S36. Merge the new labeled dataset with the new high-confidence dataset, and use the new low-confidence dataset as the new unlabeled dataset and proceed to step S4 to determine whether it is empty.

[0074] If the new unlabeled dataset is empty, then the automatic labeling of the dataset to be labeled is completed.

[0075] If the new unlabeled dataset is not empty, proceed to and execute S37.

[0076] S37. Repeat steps S34 to S36 until the new unlabeled dataset is not empty and the data volume does not change, then proceed to and execute S5.

[0077] S4. Determine whether the new unlabeled dataset is empty;

[0078] If the new unlabeled dataset is empty, then the automatic labeling of the dataset to be labeled is completed;

[0079] If the new unlabeled dataset is not empty, proceed to and execute S5;

[0080] S5. Based on a binary classification deep learning algorithm, evaluate new labeled datasets and new unlabeled datasets.

[0081] Specifically, this step evaluates the new labeled dataset and the new unlabeled dataset based on a binary classification deep learning algorithm, including the following steps:

[0082] S51. Mark the data in the new labeled dataset as 0, and mark the data in the new unlabeled dataset as 1.

[0083] S52. Use a pre-trained binary classification deep learning algorithm to perform transfer training on a new labeled dataset and a new unlabeled dataset to obtain a fine-tuned binary classification deep learning algorithm.

[0084] At this point, the pre-trained binary classification deep learning algorithm has weights of B0, and the first fine-tuned binary classification deep learning algorithm has weights of B1.

[0085] S53. Input the new labeled dataset and the new unlabeled dataset into a fine-tuned binary classification deep learning algorithm, and output the binary classification confidence of each data in the new labeled dataset and the new unlabeled dataset to form a dataset with binary classification confidence.

[0086] S54. Sort the datasets with binary classification confidence scores in descending order according to their binary classification confidence scores, and extract the top n data points to form a new dataset.

[0087] S55. Calculate the overlap between the new dataset and the new unlabeled dataset.

[0088] S56. Compare the overlap degree with the overlap degree threshold.

[0089] If the overlap is greater than the overlap threshold, it is determined that the new dataset is significantly different from the new unlabeled dataset. In this case, the data in the new unlabeled dataset is manually labeled and calibrated.

[0090] If the overlap is less than the overlap threshold, the data annotation results in the new labeled dataset are deemed unusable.

[0091] The following specific examples illustrate the semi-supervised data automatic annotation method based on deep learning described above in this invention:

[0092] Step 1: See Figure 2 As shown, the prepared dataset D0 (a dataset of frame-aligned multi-view images and point clouds) is manually labeled and refined. The dataset D0 is then divided into labeled dataset D1 and unlabeled dataset D2. The pre-trained BEV (Bird EyeView) deep learning algorithm weights W0 are then transferred to the labeled dataset D1 for training, resulting in a finely tuned BEV deep learning algorithm weight W1.

[0093] Step 2: See Figure 3 As shown, the unlabeled dataset D2 is automatically labeled using the BEV deep learning algorithm weights W1 with one fine-tuning. Based on the automatic labeling results and a manually set confidence threshold, the unlabeled dataset D2 is divided into a high-confidence automatically labeled dataset D3 and a low-confidence automatically labeled dataset D4. The high-confidence dataset D3 means that the automatic labeling results of the BEV deep learning algorithm weights W1 with one fine-tuning are closer to the results of manual labeling and fine-tuning.

[0094] At this point, we can check whether the low-confidence automatically labeled dataset D4 is empty. If it is empty, it means that the data has been labeled. If it is not empty, it means that the data labeling is not complete. At this point, we can proceed to the third step to label again, or we can directly proceed to the ... step to evaluate the labeling results using a binary classification deep learning algorithm.

[0095] Step 3: See Figure 4As shown, the weights W1 of the BEV deep learning algorithm are trained in a supervised manner using the high-confidence automatically labeled dataset D3 and the labeled dataset D1, and the weights W2 of the BEV deep learning algorithm are generated in a second-order fine-tuning. The purpose of this is to improve the performance of the second-order fine-tuned weights W2 in the automatic labeling on the previously unlabeled dataset, so as to have better performance in the next step on the low-confidence automatically labeled dataset D4.

[0096] Step 4: See Figure 5 As shown, the low-confidence automatically labeled dataset D4 is automatically labeled using the BEV deep learning algorithm weights W2 with secondary fine-tuning. After labeling, a new high-confidence automatically labeled dataset D is generated based on the confidence threshold manually set in the second step. ＇ 3 and the new low-confidence automatically labeled dataset D ＇ 4.

[0097] The new low-confidence automatically labeled dataset D ＇ 4. A new low-confidence auto-labeling dataset D4 is used in subsequent processes, replacing the previous low-confidence dataset. Simultaneously, high-confidence auto-labeling datasets D3 and D4 are also included. ＇ 3. The datasets are merged into a high-confidence automatically labeled dataset D3, which is then used in subsequent processes.

[0098] Through the four steps described above, we will eventually obtain a high-confidence automatically labeled dataset D3 and a low-confidence automatically labeled dataset D4. At this point, we need to determine if the low-confidence automatically labeled dataset D4 is empty. If it is empty, then all unlabeled datasets have completed the automatic labeling process and become the high-confidence automatically labeled dataset D3. Finally, we obtain the manually labeled and refined dataset D1 and the high-confidence automatically labeled dataset D3, and we do not need to perform the subsequent binary classification evaluation process. Otherwise, we jump to step three to further fine-tune the BEV algorithm weights, and then use the fine-tuned BEV algorithm weights and the low-confidence automatically labeled dataset to execute step four and continue the evaluation.

[0099] If, during the iterations of steps three and four, a low-confidence automatically labeled dataset D4 is found to be non-empty and its data size does not decrease after multiple iterations, then the low-confidence automatically labeled dataset D4 will be used as the new unlabeled dataset D. ＇ 2. Simultaneously, the labeled dataset D1 and the high-confidence automatically labeled dataset D3 are merged into a new labeled dataset D. ＇ 1. Proceed to the next step of binary classification.

[0100] Step 5: See Figure 6As shown, all data in the labeled dataset D1 are assigned the label 'labeled' with a value of 0, and all data in the unlabeled dataset D2 are assigned the label 'unlabeled' with a value of 1. The pre-trained binary classification deep learning algorithm weights B0 are then transferred to both the labeled dataset D1 and the unlabeled dataset D2 to generate a fine-tuned binary classification deep learning algorithm weight B1. The purpose of this fine-tuning is to ensure that the fine-tuned binary classification deep learning algorithm weights B1 can effectively distinguish between the labeled and unlabeled datasets.

[0101] Step 6: See Figure 7 As shown, the labeled dataset D1 and the unlabeled dataset D2 are input together into the fine-tuned binary classification deep learning algorithm weights B1. For each input data point, a binary classification confidence score is generated. The closer the confidence score is to 1, the more likely the fine-tuned binary classification deep learning algorithm weights B1 consider it to belong to the unlabeled data, and vice versa. The final generated dataset D1 contains binary classification confidence scores. ＇ 0. Sort all data in descending order based on their binary classification confidence scores to generate dataset D. ＇＇ 0.

[0102] Choose D ＇＇ 0 Middle D ＇ Given two data points (i.e., the first n data points), we can say that this D ＇ Two data points are the ones in the unlabeled dataset D2 that differ most from the labeled dataset D1. ＇ Two data points, then use the selected D ＇ Two datasets were used to create a new unlabeled dataset D, which was output by the previous semi-supervised automatic annotation process. ＇ 2. Calculate and generate the overlap results between the two.

[0103] By performing steps five and six above, we will eventually obtain D, which has the largest difference from the labeled dataset. ＇ Two unlabeled datasets and a new unlabeled dataset D output from the previous semi-supervised automatic labeling process. ＇ The overlap between 2 is calculated and compared with the manually selected overlap threshold.

[0104] If the overlap is higher than a manually selected overlap threshold, it indicates that a new unlabeled dataset D has been generated from the semi-supervised automatic annotation process. ＇ If the dataset is indeed significantly different from the labeled dataset, the weights of the BEV deep learning algorithm trained through transfer learning will not be able to automatically label this new unlabeled dataset. Manual labeling and refining of these data are required. After manual refining and labeling of these data, the original partially labeled dataset becomes the fully labeled dataset.

[0105] If the overlap is lower than a manually selected overlap threshold, it means that the high-confidence automatically labeled dataset results after the weight transfer training of the BEV deep learning algorithm are unusable. It is necessary to readjust the transfer training strategy and re-execute the entire semi-supervised automatic labeling process and binary classification evaluation process.

[0106] This invention presents a semi-supervised automatic data annotation method based on deep learning, particularly suitable for multimodal datasets formed from frame-aligned laser point clouds and multi-view images. Utilizing the BEV deep learning algorithm, and employing transfer learning and semi-supervised strategies, the automatic annotation process addresses the issue that the cost of automatic annotation for multimodal datasets is significantly higher than that for single-modal datasets when manual fusion annotation is required. Furthermore, this invention specifically designs a binary classification deep learning algorithm to evaluate the annotation results using the BEV deep learning algorithm, ensuring the quality of automatic annotation of the dataset by the BEV deep learning algorithm.

[0107] This specific embodiment also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described semi-supervised data automatic annotation method based on deep learning.

[0108] This specific embodiment also discloses a computer-readable storage medium storing the above-described semi-supervised data automatic annotation method based on deep learning, which has the right to execute the method.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0110] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A semi-supervised automatic data annotation method based on deep learning, characterized in that, Includes the following steps: S1. Obtain the dataset to be labeled, wherein the dataset to be labeled includes a multimodal dataset formed by frame-aligned laser point clouds and multi-view images; S2. Manually label a portion of the data in the dataset to be labeled, and divide the dataset into labeled dataset and unlabeled dataset; S3. Automatically label the unlabeled dataset based on the BEV deep learning algorithm to obtain a labeled dataset and a new unlabeled dataset, including: S31. Using a pre-trained BEV deep learning algorithm to perform transfer training on the labeled dataset to obtain a fine-tuned BEV deep learning algorithm; S32. Automatically label the unlabeled dataset according to the fine-tuned BEV deep learning algorithm, and combine it with a confidence threshold to divide the unlabeled dataset into a high-confidence dataset and a low-confidence dataset; S33. Merge the labeled dataset with the high-confidence dataset to form a new labeled dataset, and use the low-confidence dataset as a new unlabeled dataset and proceed to step S4 to determine whether it is empty; S4. Determine whether the new unlabeled dataset is empty; If the new unlabeled dataset is empty, then the automatic labeling of the dataset to be labeled is completed; If the new unlabeled dataset is not empty, proceed to and execute S5; S5. Based on a binary classification deep learning algorithm, evaluate the new labeled dataset and the new unlabeled dataset, including: S51. Mark the data in the new labeled dataset as 0 and the data in the new unlabeled dataset as 1; S52. Use a pre-trained binary classification deep learning algorithm to perform transfer training on the new labeled dataset and the new unlabeled dataset to obtain a fine-tuned binary classification deep learning algorithm. S53. Input the new labeled dataset and the new unlabeled dataset into the first fine-tuned binary classification deep learning algorithm, and output the binary classification confidence of each data in the new labeled dataset and the new unlabeled dataset to form a dataset with binary classification confidence. S54. Sort the dataset with binary classification confidence values ​​in descending order, and extract the top n data points to form a new dataset. S55. Calculate the overlap between the new dataset and the new unlabeled dataset; S56. Compare the overlap ratio with the overlap threshold; If the overlap is greater than the overlap threshold, it is determined that the difference between the new dataset and the new unlabeled dataset is small, and then the data in the new unlabeled dataset are manually labeled and calibrated. If the overlap is less than the overlap threshold, the data annotation results in the new labeled dataset are deemed unusable.

2. The semi-supervised automatic data annotation method according to claim 1, characterized in that: In step S3, the unlabeled dataset is automatically labeled using the BEV deep learning algorithm to obtain labeled datasets and new unlabeled datasets, including the following steps: S31. Use the pre-trained BEV deep learning algorithm to perform transfer training on the labeled dataset to obtain the fine-tuned BEV deep learning algorithm. S32. The unlabeled dataset is automatically labeled according to the fine-tuned BEV deep learning algorithm, and the unlabeled dataset is divided into high-confidence dataset and low-confidence dataset based on the confidence threshold. S33. Merge the labeled dataset with the high-confidence dataset to form a new labeled dataset, and proceed to step S4 to determine whether the low-confidence dataset is empty; If the low-confidence dataset is empty, then the automatic labeling of the dataset to be labeled is completed; If the low-confidence dataset is not empty, proceed to and execute S34; S34. The new labeled dataset is trained in a supervised manner using the fine-tuned BEV deep learning algorithm to obtain a new fine-tuned BEV deep learning algorithm. S35. The low-confidence dataset is automatically labeled according to the new fine-tuned BEV deep learning algorithm, and the low-confidence dataset is divided into a new high-confidence dataset and a new low-confidence dataset based on the confidence threshold. S36. Merge the new labeled dataset with the new high-confidence dataset, and use the new low-confidence dataset as a new unlabeled dataset and proceed to step S4 to determine whether it is empty; If the new unlabeled dataset is empty, then the automatic labeling of the dataset to be labeled is completed; If the new unlabeled dataset is not empty, proceed to and execute S37; S37. Repeat steps S34 to S36 until the new unlabeled dataset is not empty and the data volume does not change, then proceed to and execute S5.

3. The semi-supervised automatic data annotation method according to claim 1 or 2, characterized in that: The labeled dataset in step S2 consists of manually labeled and calibrated data.

4. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the semi-supervised data automatic annotation method based on deep learning as described in any one of claims 1 to 3.

5. A computer-readable storage medium, characterized in that: A computer-readable storage medium stores a semi-supervised data automatic annotation method based on deep learning, which performs any one of claims 1 to 3.

Images