AI model closed-loop iterative training method and system based on active learning training sample

Through active learning technology and closed-loop iterative training method, the training data set and model are automatically updated, which solves the expensive and time-consuming problems of obtaining high-quality labeled data, improves the training efficiency and performance of AI models, and is suitable for large-scale data sets.

CN118887488APending Publication Date: 2024-11-01DONGFENG MOTOR GRP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410857872.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-28
Publication Date
2024-11-01

AI Technical Summary

Technical Problem

Acquisition of high-quality labeled data is an expensive and time-consuming process, and it is difficult for the prior art to effectively select correctly distributed and effective data sets to improve the training efficiency and accuracy of AI models.

Method used

Using the closed-loop iterative training method of AI model based on active learning, select some pictures from the sample picture set for annotation, use the initial training model for semantic segmentation and entropy value calculation, and automatically update the training data set and model until the preset number or performance is reached. Require.

Benefits of technology

It improves the efficiency and performance of model training, reduces the annotation cost, is suitable for large-scale data sets, and improves the generalization ability and applicability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118887488A_ABST
    Figure CN118887488A_ABST
Patent Text Reader

Abstract

The invention discloses an AI model closed-loop iterative training method and system based on an active learning training sample. Selecting part of pictures from the sample picture set to mark to obtain a mark data set, and dividing the mark data set into a training data set and a test data set; training the deep learning model by using the labeled data set to obtain an initial training model; selecting part of picture labels from the sample picture set by using the initial training model and combining the part of picture labels into a training data set; updating the initial training model by using the updated training data set and the test data set; and repeatedly executing the operations of updating the training data set and updating the initial training model until the scale of the training data set reaches a preset number or the model performance meets the requirement. According to the method, part of samples are selected for labeling through active learning, time and resource waste caused by labeling of the whole data set is avoided, the method is suitable for an AI model needing training of a large-scale data set, mass data can be processed, and the model performance can be rapidly improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence model training technology, specifically relating to a closed-loop iterative training method and system for AI models based on active learning training samples. Background Technology

[0002] With the development and widespread application of artificial intelligence technology, more and more AI models require training with large-scale training data. However, obtaining high-quality labeled data is an expensive and time-consuming process. Furthermore, the distribution and size of the training data are closely related to model performance. Therefore, selecting a correctly distributed and effective dataset is crucial for improving the efficiency and accuracy of model training. Summary of the Invention

[0003] To address the problems described in the background section, this invention proposes a closed-loop iterative training method and system for AI models based on active learning training samples.

[0004] One of the objectives of this invention is a closed-loop iterative training method for an AI model based on actively learned training samples, comprising:

[0005] A portion of images from the sample image set are selected and labeled to obtain a labeled dataset. The labeled dataset is then divided into a training dataset and a test dataset. The deep learning model is trained using the labeled dataset to obtain an initial training model.

[0006] Update the training dataset: Select a portion of images from the sample image set using the initial training model, label them, and merge them into the training dataset to obtain the updated training dataset; Update the initial training model: Update the initial training model using the updated training dataset and the test dataset;

[0007] Repeat the above operations of updating the training dataset and updating the initial training model until the amount of data in the training dataset reaches the preset amount or the model performance meets the requirements.

[0008] Furthermore, methods for selecting a subset of data from the sample dataset include:

[0009] Select a subset of images from the sample image set as prediction images and input them into the initial training model;

[0010] The semantic segmentation model in the initial training model is used to perform semantic segmentation on each predicted image; the predicted images are sorted according to the semantic segmentation results of each predicted image, and a portion of the predicted images are selected from the sorted images and added to the training dataset.

[0011] Furthermore, the semantic segmentation method includes:

[0012] The predicted image is input into the initial training model, and the model inference outputs multiple matrices; the size of each matrix is ​​the same as the pixel size of the input predicted image;

[0013] Each matrix is ​​transformed into a binary image to obtain multiple binary images; the pixel size of each binary image is the same as the pixel size of the input prediction image.

[0014] The entropy value of each pixel in the predicted image and the entropy value of each pixel in each binary image corresponding to the predicted image are calculated using the semantic segmentation model in the initial training model.

[0015] Iterate through each binary image corresponding to the predicted image, calculate the difference between the entropy value of each pixel in the predicted image and the entropy value of the corresponding pixel in the binary image, and sum the differences to obtain the sum value corresponding to each binary image.

[0016] The summation of the values ​​corresponding to each binary image is used to obtain the semantic segmentation result of the predicted image.

[0017] Furthermore, methods for performing binary transformations on each matrix to obtain multiple binary images include:

[0018] The element values ​​of each matrix are compared with a set threshold for that matrix. Based on the comparison results, each matrix is ​​converted into a binary matrix whose element values ​​contain only two values. The binary graph corresponding to each matrix is ​​obtained from the converted binary matrix.

[0019] Furthermore, when the predicted image is a correctly labeled image, a different threshold is set for each matrix corresponding to the predicted image; otherwise, the same threshold is set for each matrix corresponding to the predicted image.

[0020] A closed-loop iterative training system for AI models based on active learning training samples, which achieves the second objective of this invention, includes:

[0021] The annotation module is used to select a portion of sample images from the collected sample image set for annotation.

[0022] Model building module: used to select a portion of images from the sample image set for annotation to obtain an annotated dataset, divide the annotated dataset into a training dataset and a test dataset; and use the annotated dataset to train the deep learning model to obtain the initial training model;

[0023] Training dataset update module: Used to select a portion of images from the sample image set using the initial training model, label them, and merge them into the training dataset to obtain the updated training dataset;

[0024] Model update module: used to update the initial training model using the updated training dataset and the test dataset;

[0025] Iterative training module: This module repeatedly executes the operations described above in the update training dataset module and update initial training model module until the amount of data in the training dataset reaches the preset amount or the model performance meets the requirements.

[0026] In the above system, the training dataset update module includes:

[0027] Predicted image selection module: Used to select a subset of images from the sample image set as predicted images to be input into the initial training model;

[0028] Semantic segmentation module: Used to perform semantic segmentation on each predicted image using the semantic segmentation model in the initial training model;

[0029] Data filtering module: used to sort the predicted images according to the semantic segmentation results of each predicted image; and select a portion of the predicted images from the sorted images and merge them into the training dataset.

[0030] In the above system, the semantic segmentation module includes:

[0031] The binary image generation module is used to input the predicted image into the initial training model, and the model inference outputs multiple matrices; the size of each matrix is ​​the same as the pixel size of the input predicted image; binary transformation is performed on each matrix to obtain multiple binary images; the pixel size of each binary image is the same as the pixel size of the input predicted image.

[0032] Entropy calculation module: used to calculate the entropy value of each pixel in the predicted image and the entropy value of each pixel in each binary image corresponding to the predicted image using the semantic segmentation model in the initial training model;

[0033] Difference calculation module: used to traverse each binary image corresponding to the predicted image, calculate the difference between the entropy value of each pixel in the predicted image and the entropy value of the corresponding pixel in the binary image; and sum the differences to obtain the sum value corresponding to each binary image;

[0034] Semantic segmentation result calculation module: sums the sums of the values ​​corresponding to each binary image to obtain the semantic segmentation result of the predicted image.

[0035] In the above system, the binary graph generation module compares the element values ​​of each matrix with a set threshold for that matrix. Based on the comparison result, each matrix is ​​converted into a binary matrix whose element values ​​contain only two values. The binary graph corresponding to each matrix is ​​obtained based on the converted binary matrix.

[0036] In the above system, when the predicted image is a correctly labeled image, a different threshold is set for each matrix corresponding to the predicted image; otherwise, the same threshold is set for each matrix corresponding to the predicted image.

[0037] The beneficial effects of this invention include:

[0038] 1. Efficient model training: By using active learning techniques to select the most informative samples for labeling, the time and resources wasted on labeling the entire dataset are avoided, thus improving the efficiency of model training.

[0039] 2. Improve model performance and generalization ability: Through a closed-loop feedback mechanism, the training dataset is automatically updated, enabling the model to have better performance and generalization ability when faced with new data.

[0040] 3. Reduce labeling costs: Compared with traditional methods, only the most informative samples need to be labeled, which can reduce labeling costs while achieving training results comparable to labeling the entire dataset.

[0041] 4. Suitable for large-scale datasets: This system is suitable for AI models that require large-scale datasets for training, and can handle massive amounts of data and quickly improve model performance. Attached Figure Description

[0042] Figure 1 This is a flowchart illustrating an embodiment of the method described in this invention;

[0043] Figure 2 This is a block diagram of the system described in this invention. Detailed Implementation

[0044] The following detailed embodiments are provided to explain the technical solutions of the claims of this invention, so that those skilled in the art can understand the claims. The scope of protection of this invention is not limited to the following specific embodiments. Any modifications made by those skilled in the art that incorporate the technical solutions of the claims but differ from the following detailed embodiments are also within the scope of protection of this invention.

[0045] This invention provides a closed-loop iterative training method for AI models based on actively learned training samples, such as... Figure 1 As shown, it includes the following steps:

[0046] 1. Basic Sample Collection

[0047] The lane line samples in the RGB image primarily come from dashcam footage, open datasets, internet data, or data collected by professional road data collection vehicles. To ensure the generalization and applicability of the AI ​​model, the sources of the basic sample data should be diverse and complex. The obtained basic data undergoes preprocessing, including data cleaning, noise reduction, deduplication, and format conversion, to ensure data quality and consistency, thus creating the basic data sample pool. If mobile devices continuously collect data, the basic data sample pool is constantly expanding.

[0048] 2. Initial model training:

[0049] First, select some diverse datasets from the basic data sample pool, use a portion of them as the initial training dataset, and use the other portion as the test dataset.

[0050] Next, appropriate annotation techniques are used to annotate all data in the selected training and test datasets to obtain the labeled dataset. Annotation methods can include manual annotation, crowdsourced annotation, semi-automatic annotation, etc. It is necessary to ensure the accuracy and consistency of all annotation results.

[0051] Then, using the labeled dataset, a suitable deep learning model (such as the semantic segmentation algorithm mask2former) is selected for training to obtain the initial training model;

[0052] Finally, the obtained initial training model is used to evaluate the test dataset, and the evaluation results, iteration rounds, and number of targets are saved for subsequent comparative testing.

[0053] 3. Active learning:

[0054] Based on the evaluation results of the initial training model on unlabeled or incorrectly labeled data, multiple samples with the most information content from the unlabeled data are selected for annotation (manual annotation, crowdsourced annotation, semi-automatic annotation, etc.), thereby increasing the diversity and coverage of the training dataset. The most information-rich samples are those that are most helpful in improving model performance.

[0055] In some embodiments, the active learning strategy employs uncertainty sampling: this strategy selects samples based on the model's uncertainty, typically showing greater interest in samples with higher uncertainty because these samples may have a greater impact on the model's predictions. For example, model entropy can be used to measure sample uncertainty, and samples with the highest uncertainty are selected for labeling, using the same labeling methods and specifications as the base sample labeling methods.

[0056] In some embodiments, entropy can be used to measure the uncertainty of the output of the initially trained model in a semantic segmentation task. For each pixel, the semantic segmentation model in the initial training model typically outputs a probability distribution p1 to p2. n , representing the probability that the pixel belongs to each of the n categories. Entropy is a commonly used concept in information theory; it measures the uncertainty of a random variable. For each pixel, the probability distribution is p1~p... n The entropy H(p) corresponding to pixel p can be calculated using the following formula:

[0057]

[0058] In semantic segmentation, the probability distribution of each pixel can be calculated, and its entropy can be calculated using the formula mentioned above. A higher entropy value indicates less certainty in the semantic segmentation model's prediction for that pixel, meaning the model cannot definitively determine which category the pixel belongs to. Conversely, a lower entropy value indicates more certainty in the model's prediction for that pixel, meaning the model more clearly identifies the pixel as belonging to a particular category.

[0059] Multiple prediction images are selected from the basic data sample pool. These prediction images may be labeled or unlabeled.

[0060] (1) When the predicted image has ground truth, it means that the predicted image has been correctly labeled and can be considered real and reliable:

[0061] After the predicted image is input into the initial training model, the initial training model inferences and outputs 11 matrices for each image. The size of each matrix is ​​the same as the pixel size of the input predicted image (e.g., if the pixel size of the input predicted image is 300*500, then the matrix size is also 300 rows*500 columns).

[0062] For the 11 matrices, different thresholds are set. In this embodiment, the thresholds are (0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.99).

[0063] The matrix elements are transformed according to their respective thresholds, and the transformation rules are as follows:

[0064] When the element value is greater than the set threshold corresponding to the matrix, the element value is updated to the first value; otherwise, it is updated to the second value. In this embodiment, the first value is 0 and the second value is 255. This invention does not limit this.

[0065] Based on the above rules for transforming matrix elements, each element value of each matrix is ​​transformed to obtain 11 binary matrices (the element values ​​of the matrix only include the first and second values).

[0066] Each binary matrix yields a corresponding binary image, with the pixel size of each binary image being the same as the matrix size.

[0067] Calculate the entropy value of each pixel in each binary image separately. The corresponding image in the input prediction image The difference in entropy values ​​of pixels is calculated, and the difference in entropy values ​​of all pixels in each binary image is summed to obtain 11 sums. The sum of these 11 sums is then used as the uncertainty output for the predicted image.

[0068] The uncertainty output values ​​of the predicted images are sorted from largest to smallest, and the top N predicted images are selected as samples for the active learning algorithm in a new round of training. These samples are then incorporated into the training dataset to improve the model's performance.

[0069] (2) When the predicted image has no ground truth, it means that the predicted image has not been correctly labeled or has not been labeled at all:

[0070] For each predicted image, the initial training model also infers and outputs 11 matrices, each of which is the same size as the pixel size of the input predicted image;

[0071] Set the threshold for all matrices to the same value; in this embodiment, it is set to 0.5.

[0072] Each element in each matrix is ​​transformed according to the threshold, and the transformation rules are as follows:

[0073] When the element value is greater than the set threshold, the element value is updated to the first value; otherwise, it is updated to the second value. In this embodiment, the first value is 0 and the second value is 255. This invention does not limit this.

[0074] Based on the above rules for transforming matrix elements, each element value of each matrix is ​​transformed to obtain 11 binary matrices (the element values ​​of the matrix only include the first and second values).

[0075] Each binary matrix yields a corresponding binary image, with the pixel size of each binary image being the same as the matrix size.

[0076] Calculate the entropy value of each pixel in each binary image separately. The corresponding image in the input prediction image The difference in entropy values ​​of pixels is calculated, and the difference in entropy values ​​of all pixels in each binary image is summed to obtain 11 sums. The sum of these 11 sums is then used as the uncertainty output for the predicted image.

[0077] 4. Model Update:

[0078] The samples selected using active learning in the previous step are added to the existing labeled dataset. The optimal initial training model from the previous iteration is loaded as a pre-trained model and retrained to improve the model's performance. After the model training is complete, the aforementioned test dataset is always used for evaluation. The evaluation results, iteration number, and number of targets are saved for subsequent comparative testing.

[0079] 5. Repeat steps 3 and 4:

[0080] The process of sample selection, labeling, and model updating is repeated until a preset number of labeled samples or performance requirements are met, resulting in an AI model based on the active learning training samples. Based on the performance evaluation results and areas for improvement, the data collection and active learning strategies are automatically adjusted, the training dataset is updated, and the AI ​​model is retrained.

[0081] This invention utilizes an active learning strategy, combined with data preprocessing and annotation techniques, to collect high-quality foundational samples from various data sources, providing a reliable data foundation for training AI models. By actively selecting information-rich samples for annotation, annotation costs and time are reduced, while ensuring the diversity and coverage of training data. This method is applicable to various AI application scenarios requiring large-scale datasets for training.

[0082] This invention also provides a closed-loop iterative training method for AI models based on active learning training samples, comprising:

[0083] A labeled dataset is obtained by selecting a portion of images from the sample image set and labeling them. This labeled dataset is then divided into a training dataset and a test dataset. The deep learning model is trained using the labeled dataset to obtain an initial training model. In this embodiment, the sample image set comes from preprocessed data collected by dashcams, open datasets, internet data, or professional road data collection vehicles. To ensure the generalization and applicability of the AI ​​model, the data sources of the sample image set should be diverse and complex. The preprocessing includes data cleaning, noise reduction, deduplication, and format conversion to ensure data quality and consistency. If data is continuously collected by mobile devices, the sample image set will continuously increase.

[0084] Update the training dataset: Select a portion of images from the sample image set using the initial training model, label them, and merge them into the training dataset to obtain the updated training dataset; Update the initial training model: Update the initial training model using the updated training dataset and the test dataset;

[0085] Repeat the above operations of updating the training dataset and updating the initial training model until the amount of data in the training dataset reaches the preset amount or the model performance meets the requirements.

[0086] In some embodiments, the method for selecting a subset of data from a sample dataset includes:

[0087] Select a subset of images from the sample image set as prediction images and input them into the initial training model;

[0088] The semantic segmentation model in the initial training model is used to perform semantic segmentation on each predicted image; the predicted images are sorted according to the semantic segmentation results of each predicted image, and a portion of the predicted images are selected from the sorted images and added to the training dataset.

[0089] In some embodiments, the semantic segmentation method includes:

[0090] The predicted image is input into the initial training model, and the model inference outputs multiple matrices; the size of each matrix is ​​the same as the pixel size of the input predicted image; the generation process of the matrix is ​​existing technology and will not be described in detail here.

[0091] Each matrix is ​​transformed into a binary image to obtain multiple binary images; the pixel size of each binary image is the same as the pixel size of the input prediction image.

[0092] The entropy value of each pixel in the predicted image and the entropy value of each pixel in each binary image corresponding to the predicted image are calculated using the semantic segmentation model in the initial training model.

[0093] Iterate through each binary image corresponding to the predicted image, calculate the difference between the entropy value of each pixel in the predicted image and the entropy value of the corresponding pixel in the binary image, and sum the differences to obtain the sum value corresponding to each binary image.

[0094] The summation of the values ​​corresponding to each binary image is used to obtain the semantic segmentation result of the predicted image.

[0095] In some embodiments, the method of performing a binary transformation on each matrix to obtain multiple binary images includes:

[0096] The element values ​​of each matrix are compared with a set threshold for that matrix. Based on the comparison results, each matrix is ​​converted into a binary matrix whose element values ​​contain only two values. Based on the binary matrix, a binary graph corresponding to each matrix is ​​obtained.

[0097] In some embodiments, when the predicted image is a correctly labeled image, a different threshold is set for each matrix corresponding to the predicted image; otherwise, the same threshold is set for each matrix corresponding to the predicted image.

[0098] This invention also includes an AI model closed-loop iterative training system based on active learning training samples, such as... Figure 2 As shown, it includes:

[0099] The annotation module is used to select a portion of sample images from the collected sample image set for annotation.

[0100] Model building module: used to select a portion of images from the sample image set for annotation to obtain an annotated dataset, divide the annotated dataset into a training dataset and a test dataset; and use the annotated dataset to train the deep learning model to obtain the initial training model;

[0101] Training dataset update module: Used to select a portion of images from the sample image set using the initial training model, label them, and merge them into the training dataset to obtain the updated training dataset;

[0102] Model update module: used to update the initial training model using the updated training dataset and the test dataset;

[0103] Iterative training module: This module repeatedly executes the operations described above in the update training dataset module and update initial training model module until the amount of data in the training dataset reaches the preset amount or the model performance meets the requirements.

[0104] In some embodiments, the training dataset update module includes:

[0105] Predicted image selection module: Used to select a subset of images from the sample image set as predicted images to be input into the initial training model;

[0106] Semantic segmentation module: Used to perform semantic segmentation on each predicted image using the semantic segmentation model in the initial training model;

[0107] Data filtering module: used to sort the predicted images according to the semantic segmentation results of each predicted image; and select a portion of the predicted images from the sorted images and merge them into the training dataset.

[0108] In some embodiments, the semantic segmentation module includes:

[0109] The binary image generation module is used to input the predicted image into the initial training model, and the model inference outputs multiple matrices; the size of each matrix is ​​the same as the pixel size of the input predicted image; binary transformation is performed on each matrix to obtain multiple binary images; the pixel size of each binary image is the same as the pixel size of the input predicted image.

[0110] Entropy calculation module: used to calculate the entropy value of each pixel in the predicted image and the entropy value of each pixel in each binary image corresponding to the predicted image using the semantic segmentation model in the initial training model;

[0111] Difference calculation module: used to traverse each binary image corresponding to the predicted image, calculate the difference between the entropy value of each pixel in the predicted image and the entropy value of the corresponding pixel in the binary image; and sum the differences to obtain the sum value corresponding to each binary image;

[0112] Semantic segmentation result calculation module: sums the sums of the values ​​corresponding to each binary image to obtain the semantic segmentation result of the predicted image.

[0113] In some embodiments, the binary graph generation module compares the element values ​​of each matrix with a set threshold for that matrix, converts each matrix into a binary matrix whose element values ​​contain only two values, and obtains the binary graph corresponding to each matrix based on the converted matrix.

[0114] In some embodiments, when the predicted image is a correctly labeled image, a different threshold is set for each matrix corresponding to the predicted image; otherwise, the same threshold is set for each matrix corresponding to the predicted image.

[0115] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0116] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A closed-loop iterative training method for AI models based on active learning training samples, It is characterized in that include: Select some pictures from the sample picture set for annotation to obtain an annotation data set, and divide the annotation data set into a training data set and a test data set; Use the labeled data set to train the deep learning model to obtain the initial training model; Update the training data set: Use the initial training model to select some images from the sample image set for annotation and merge them into the training data set to obtain an updated training data set; Updating the initial training model: updating the initial training model using the updated training data set and the test data set; The above operations of updating the training data set and updating the initial training model are repeatedly performed until the amount of data in the training data set reaches a preset amount or the model performance meets the requirements.

2. The closed-loop iterative training method for an AI model based on active learning training samples as claimed in claim 1, It is characterized in that Methods for selecting some data from the sample dataset include: Select a portion of images from the sample image set as prediction images and input them into the initial training model; The semantic segmentation model in the initial training model is used to perform semantic segmentation on each predicted image; the predicted images are sorted according to the semantic segmentation results of each predicted image, and some images are selected from the predicted images according to the sorting results and incorporated into the training data set.

3. The closed-loop iterative training method for an AI model based on active learning training samples as claimed in claim 2, It is characterized in that The semantic segmentation method comprises: Input the predicted image into the initial training model, and the model inference output obtains multiple matrices; the size of each matrix is ​​the same as the pixel size of the input predicted image; Perform binary conversion on each matrix to obtain multiple binary images; the pixel size of each binary image is the same as the pixel size of the input prediction image; The semantic segmentation model in the initial training model is used to calculate the entropy value of each pixel of the predicted image and the entropy value of each pixel of each binary image corresponding to the predicted image; Traversing each binary image corresponding to the predicted image, calculating the difference between the entropy value of each pixel point of the predicted image and the entropy value of the pixel point at the corresponding position on the binary image one by one; and summing the differences to obtain the sum value corresponding to each binary image; The sum values ​​corresponding to each of the two-value images are summed to obtain the semantic segmentation result of the predicted image.

4. The closed-loop iterative training method for an AI model based on active learning training samples as claimed in claim 3, It is characterized in that Methods for performing binary value conversion on each matrix to obtain multiple binary value graphs include: The element value of each matrix is ​​compared with the set threshold value of the matrix, and each matrix is ​​converted into a dual-value matrix whose element value only contains two values ​​according to the comparison result, and the dual-value map corresponding to each matrix is ​​obtained according to the dual-value matrix.

5. The closed-loop iterative training method for an AI model based on active learning training samples as claimed in claim 3, It is characterized in that When the predicted picture is a correctly labeled picture, a different threshold is set for each matrix corresponding to the predicted picture; otherwise, the same threshold is set for each matrix corresponding to the predicted picture.

6. A closed-loop iterative training system for AI models based on active learning training samples, It is characterized in that include: Annotation module: used to select some sample images from the collected sample image set for annotation; Model building module: used to select some pictures from the sample picture set for annotation to obtain an annotated data set, and divide the annotated data set into a training data set and a test data set; Use the labeled data set to train the deep learning model to obtain the initial training model; Training data set update module: used to select some pictures from the sample picture set by using the initial training model for annotation and merge them into the training data set to obtain an updated training data set; Model update module: used to train the initial training model using the updated training data set, and update the parameters of the trained initial training model using the test data set; Iterative training module: used to repeatedly perform the operations of updating the training dataset module and updating the initial training model module, until the amount of data in the training dataset reaches a preset amount or the model performance meets the requirements.

7. The AI ​​model closed-loop iterative training system based on active learning training samples as claimed in claim 6, It is characterized in that The training data set updating module comprises: Prediction picture selection module: used to select a part of pictures from the sample picture set as prediction pictures and input them into the initial training model; Semantic segmentation module: used to perform semantic segmentation on each predicted image using the semantic segmentation model in the initial training model; Data screening module: used to sort the predicted images according to the semantic segmentation results of each predicted image; select some images from the predicted images according to the sorting results and merge them into the training data set.

8. The AI ​​model closed-loop iterative training system based on active learning training samples as claimed in claim 7, It is characterized in that The semantic segmentation module includes: A binary image generation module is used to input the prediction image into the initial training model, and the model inference output obtains multiple matrices; the size of each matrix is ​​the same as the pixel size of the input prediction image; each matrix is ​​converted into binary values ​​to obtain multiple binary images; the pixel size of each binary image is the same as the pixel size of the input prediction image; Entropy calculation module: used to calculate the entropy value of each pixel of the predicted image and the entropy value of each pixel of each binary image corresponding to the predicted image using the semantic segmentation model in the initial training model; The difference calculation module is used to traverse each binary image corresponding to the predicted image, calculate the difference between the entropy value of each pixel point of the predicted image and the entropy value of the pixel point at the corresponding position on the binary image one by one, and sum the differences to obtain the sum value corresponding to each binary image; Semantic segmentation result calculation module: sums the sum values ​​corresponding to each of the two-value images to obtain the semantic segmentation result of the predicted image.

9. The AI ​​model closed-loop iterative training system based on active learning training samples as claimed in claim 7, It is characterized in that In the binary graph generation module, the element value of each matrix is ​​compared with the set threshold value of the matrix, and each matrix is ​​converted into a binary matrix whose element values ​​only contain two values ​​according to the comparison result, and the binary graph corresponding to each matrix is ​​obtained according to the converted matrix.

10. The AI ​​model closed-loop iterative training system based on active learning training samples as claimed in claim 9, It is characterized in that When the predicted image is a correctly labeled image, a different threshold is set for each matrix corresponding to the predicted image; otherwise, the same threshold is set for each matrix corresponding to the predicted image.

Citation Information

Patent Citations

  • Active learning method for data-driven modeling

    CN110070131A

  • Earthquake first arrival training set generation method and device

    CN115481710A