A weakly supervised learning-driven co-evolutionary method for large and small models and its terminal

Abstract

This invention provides a weakly supervised learning-driven method and terminal for co-evolution of large and small models. It introduces an existing large model to perform preliminary detection on the current task, obtaining initial detection results. The category of the initial detection results confirmed by the first assessment is obtained; the category is either high-quality or low-quality. High-quality results are those within the correct answer set, and low-quality results are those outside the correct answer set. It determines whether the category is high-quality; if so, a supervised model is constructed based on the high-quality result and replaces the first assessment for automatic detection of subsequent similar tasks; otherwise, proceed to step S4. A supervised model is then constructed based on the secondary annotation of the low-quality results and replaces the first assessment for automatic detection of subsequent similar tasks. This invention requires only two manual interventions throughout the entire process; subsequent task detection is fully automated, eliminating the need for further manual intervention, thus effectively reducing human involvement while ensuring detection quality.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a method and terminal for the co-evolution of large and small models driven by weakly supervised learning. Background Technology

[0002] Large models refer to neural network models in deep learning with a massive number of parameters and complex structures. These models typically contain billions to trillions of parameters, making them much larger than traditional small models (such as traditional neural networks). Due to their massive scale, large models possess more powerful representation and learning capabilities in many scenarios. However, in some niche domains, due to the high specificity of the detection objects, the difficulty in covering training materials, and the difficulty in accurately locating prompt words, their performance suffers. Fundamentally retraining or iteratively training large models would result in enormous resource overhead.

[0003] Small models are primarily based on supervised learning, using a small number of images to train on manually labeled categories of objects with lightweight training capabilities. In a specific domain, they can produce a high-quality model in a short period of time. However, small models have poor generalization performance, and their accuracy fluctuates greatly with changes in the target or environment, leading to frequent manual iterations.

[0004] Whether large or small, all models are essentially driven by high-quality data. This data must be manually labeled or annotated by human annotators for the model to learn and understand. However, because many tasks require accurate annotation, this process is very time-consuming, complex, and labor-intensive. From data collection and cleaning to actual annotation, a significant amount of human resources is required. Especially in tasks involving specific domains or domain knowledge, the complexity and specialization of annotation further increase the cost. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a weakly supervised learning-driven method and terminal for co-evolution of large and small models, so as to improve detection quality while reducing human intervention.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A weakly supervised learning-driven co-evolutionary method for large and small models includes the following steps:

[0008] S1. Introduce the existing large model to perform preliminary detection on the current task and obtain initial detection results;

[0009] S2. Obtain the category of the initial detection result confirmed by a single analysis, wherein the category is a high-quality result or a low-quality result, the high-quality result is the result in the set of correct answers, and the low-quality result is the result outside the set of correct answers;

[0010] S3. Determine whether the category is the high-quality result. If so, construct a supervised model based on the high-quality result and replace the first judgment to automatically detect subsequent tasks of the same type. Otherwise, proceed to step S4.

[0011] S4. Obtain and construct the supervised model based on the secondary annotation of the low-quality results, and replace the first judgment to automatically detect subsequent similar tasks.

[0012] To solve the above-mentioned technical problems, another technical solution adopted by the present invention is as follows:

[0013] A weakly supervised learning-driven co-evolutionary terminal for large and small models includes 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 steps in the weakly supervised learning-driven co-evolutionary method for large and small models as described above.

[0014] The beneficial effects of this invention are as follows: This invention provides a weakly supervised learning-driven method and terminal for the co-evolution of large and small models. It first performs preliminary detection of the task based on the existing large model, and then introduces a judgment to confirm the detection result. If the result is of high quality, a supervised model (i.e., a small model) is built based on the result, thereby replacing the first judgment for automatic detection of subsequent similar tasks. If the result is of low quality, a second annotation is introduced to initially modify the low-quality result into a high-quality result, thereby similarly building a supervised model to replace the first judgment for automatic detection of subsequent similar tasks. The whole process only requires two human interventions—the first judgment and the second annotation. Subsequent task detection is automatically achieved by the collaboration of large and small models, without the need for further human intervention, thereby effectively reducing human intervention while ensuring detection quality. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the main process of a weakly supervised learning-driven co-evolutionary method for large and small models according to an embodiment of the present invention.

[0016] Figure 2 This is a flowchart illustrating a weakly supervised learning-driven co-evolutionary method for large and small models according to an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of the structure of a small-to-large co-evolutionary terminal driven by weakly supervised learning.

[0018] Label Explanation:

[0019] 1. A weakly supervised learning-driven co-evolutionary terminal for large and small models; 2. Memory; 3. Processor. Detailed Implementation

[0020] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0021] Please refer to Figure 1 and Figure 2 A weakly supervised learning-driven co-evolutionary method for large and small models includes the following steps:

[0022] S1. Introduce the existing large model to perform preliminary detection on the current task and obtain initial detection results;

[0023] S2. Obtain the category of the initial detection result confirmed by a single analysis, wherein the category is a high-quality result or a low-quality result, the high-quality result is the result in the set of correct answers, and the low-quality result is the result outside the set of correct answers;

[0024] S3. Determine whether the category is the high-quality result. If so, construct a supervised model based on the high-quality result and replace the first judgment to automatically detect subsequent tasks of the same type. Otherwise, proceed to step S4.

[0025] S4. Obtain and construct the supervised model based on the secondary annotation of the low-quality results, and replace the first judgment to automatically detect subsequent similar tasks.

[0026] As described above, the beneficial effects of this invention are as follows: by first performing preliminary detection of the task based on an existing large model, and then introducing a judgment to confirm the detection result, if the result is of high quality, a supervisory model (i.e., a small model) is built based on the result to replace the first judgment for automatic detection of subsequent similar tasks. If the result is of low quality, a second annotation is introduced to initially modify the low-quality result into a high-quality result, and a supervisory model is also built to replace the first judgment for automatic detection of subsequent similar tasks. The entire process only requires two human interventions—a first judgment and a second annotation. Subsequent task detection is automatically achieved by the collaboration of the large and small models without the need for further human intervention, thereby effectively reducing human intervention while ensuring detection quality.

[0027] Furthermore, the secondary annotation involves manually annotating the images of the low-quality results using Labelme or Labelimg annotation tools, obtaining a labeled dataset, and then using the labeled dataset together with the dataset obtained from the first assessment and the original dataset for training the supervised model.

[0028] As can be seen from the above description, manual annotation using annotation tools ensures annotation efficiency.

[0029] Furthermore, the construction of the supervised model in steps S3 and S4 specifically involves:

[0030] S41. Obtain the original dataset, the dataset of the first assessment, and / or the dataset of the second annotation, perform preprocessing such as image size adjustment, image normalization, and data augmentation, and divide the data into training set, validation set, and test set;

[0031] S42. Select different existing small models according to the task type, and select the stochastic gradient descent optimization algorithm and learning rate scheduler to control the update of small model parameters.

[0032] S43. Define a loss function according to the requirements of the task to measure the difference between the small model's prediction and the true label;

[0033] S44. Feed the data in the dataset into the small model for forward propagation, calculate the loss, then calculate the gradient through backpropagation, update the parameters of the small model using the selected optimization algorithm, and minimize the loss function;

[0034] S45. Repeat steps S43 and S44 multiple times, using different batches of training sets, until the preset number of iterations is reached or the training set batches are used up.

[0035] S46. At the end of each training cycle, the performance of the small model is evaluated using the validation set, the loss and metrics are monitored, and the learning rate and regularization parameters are adjusted based on the performance of the validation set to obtain the supervised model.

[0036] S47. After training is completed, the performance of the supervised model is evaluated using the test set.

[0037] As described above, before constructing a supervised model, the original dataset, the dataset of the first judgment and / or the second annotation are collected to ensure data quality and diversity. Then, according to the task, the corresponding initial small model is selected, and it is trained, validated and tested to obtain a supervised model that meets the current task, thereby achieving the accuracy of automatic detection of subsequent similar tasks.

[0038] Furthermore, step S3 also includes:

[0039] The high-quality results are fed back into the large model for iterative training.

[0040] Step S4 also includes:

[0041] The low-quality results obtained from the secondary annotation are taken as the high-quality results and fed back into the large model for iterative training.

[0042] As described above, high-quality results are fed back into the existing large model for iterative training, thereby optimizing the large model and enabling it to continuously evolve, ensuring that subsequent large models can obtain high-quality detection results in their initial detection. At the same time, low-quality results from secondary annotation can be equivalent to high-quality results, so these high-quality results can also be fed back into the existing large model for iterative training, thereby further optimizing the large model.

[0043] Furthermore, the iterative training specifically includes:

[0044] Step 1: Organize the original dataset, the dataset of the first analysis, and / or the dataset of the second annotation, and classify them according to specific tasks. Each category is further divided into training set, validation set, and test set.

[0045] Step 2: Load the parameters of the existing large model as initial weights, and select the stochastic gradient descent optimization algorithm and learning rate scheduler to control the update of the large model parameters;

[0046] Step 3: Define a loss function according to the requirements of the task to measure the difference between the prediction of the large model and the true label;

[0047] Step 4: Feed the data in the dataset into the large model for forward propagation, calculate the loss, then calculate the gradient through backpropagation, and update the parameters of the large model using the selected optimization algorithm to minimize the loss function;

[0048] Step 5: Iterate through Step 3 and Step 4 multiple times, using different batches of training sets, until the preset number of iterations is reached or the training set batches are used up.

[0049] Step 6: At the end of each training cycle, use the validation set to evaluate the performance of the large model, monitor the loss and metrics, and adjust the learning rate and regularization parameters based on the performance of the validation set.

[0050] Step 7: After training is complete, use the test set to evaluate the performance of the adjusted large model.

[0051] As described above, similar to building a supervised model, the large model constructs training, validation, and test sets by collecting the original dataset and the dataset obtained after each automatic detection task. The original large model is continuously iterated and trained, thereby continuously optimizing and improving the large model. This ensures that in the process of automatic detection of new tasks, human intervention is gradually reduced while the detection quality is improved.

[0052] Furthermore, the procedure prior to step S1 includes:

[0053] S0. Use a thesaurus to extract keywords for the current task and call the corresponding large model from the database.

[0054] As described above, the current task extracts keywords from the existing lexicon to trigger the corresponding large model for detection, ensuring the validity of the detection results from the introduced large model.

[0055] Further, step S0 specifically includes:

[0056] Keywords related to the task are predefined in the thesaurus, and the keywords are converted into word vectors. The corresponding large model is then called in the form of an interface deployed in the server database.

[0057] Furthermore, both step S3 and step S4 further include:

[0058] S34. Update the keyword extraction mechanism through the supervised model to optimize the lexicon.

[0059] As described above, each supervised model, i.e. the small model, is also used to update and optimize the lexicon, thereby ensuring the accuracy of detection triggered by the corresponding large model through keywords, and further improving the collaboration between the large and small models and the accuracy of task detection.

[0060] Furthermore, S34 specifically includes:

[0061] S341. Obtain annotation data containing keywords and corresponding tags, wherein the corresponding tags are tags indicating whether a keyword is present;

[0062] S342. Divide the labeled data into a training set, a validation set, and a test set;

[0063] S343. The supervised model is adopted, and the architecture of the supervised model is defined, including the vector sizes of the input layer, hidden layer and output layer;

[0064] S344. Convert the text data of the keywords into a format acceptable to the supervised model, the format including word segmentation and word vector representation;

[0065] S345. Train the supervised model using the training set, optimize the parameters of the supervised model and minimize the loss function, and evaluate the performance of the supervised model using the validation set;

[0066] S346. Adjust the hyperparameters of the supervised model, including the learning rate, batch size, and hidden layer dimension;

[0067] S347. Finally, the performance of the adjusted supervised model is evaluated using a test set.

[0068] S348. Analyze the performance of the supervised model on the validation set and test set, identify new keywords not included in the thesaurus, and include these keywords in the thesaurus;

[0069] S349. Based on the performance of the supervised model in practical applications and user feedback, continuously iterate and optimize, repeat the above steps periodically, update the labeled data, fine-tune the supervised model, and update the lexicon.

[0070] As described above, the keyword extraction mechanism is also optimized by iterative training to maintain its high efficiency.

[0071] Please refer to Figure 3 A weakly supervised learning-driven co-evolutionary terminal for large and small models includes a storage memory, a processor, and a computer program stored in the storage memory and executable on the processor. When the processor executes the computer program, it implements the steps in the weakly supervised learning-driven co-evolutionary method for large and small models as described above.

[0072] As can be seen from the above description, the beneficial effects of the present invention are as follows: Based on the same technical concept, and in conjunction with the aforementioned weakly supervised learning-driven co-evolution method for large and small models, a weakly supervised learning-driven co-evolution terminal for large and small models is provided. This involves first performing preliminary task detection based on the existing large model, then introducing a judgment to confirm the detection result. If the result is high-quality, a supervisory model (i.e., a small model) is constructed based on the result, thereby replacing the first judgment for automatic detection of subsequent tasks of the same type. If the result is low-quality, a second annotation is introduced to initially modify the low-quality result into a high-quality result, thus similarly constructing a supervisory model to replace the first judgment for automatic detection of subsequent tasks of the same type. The entire process only requires two human interventions—a first judgment and a second annotation. Subsequent task detection is automatically achieved through the collaboration of large and small models, eliminating the need for further human intervention, thereby effectively reducing human involvement while ensuring detection quality.

[0073] This invention provides a weakly supervised learning-driven method and terminal for co-evolution of large and small models, which can be applied to task detection in various scenarios. The following is a description with reference to specific embodiments.

[0074] Please refer to Figure 1 Embodiment 1 of the present invention is as follows:

[0075] A weakly supervised learning-driven co-evolutionary method for large and small models, such as Figure 1 As shown, the steps include:

[0076] S1. Introduce the existing large model to perform preliminary detection on the current task and obtain the initial detection results.

[0077] S2. Obtain the category of the initial test results confirmed by the first assessment. The categories include high-quality results and low-quality results.

[0078] High-quality results are those within the set of correct answers, while low-quality results are those outside this set. For example, in tasks involving the detection of people and vehicles, a high-quality result is one where the large model accurately identifies vehicles as vehicles and people as people. Conversely, a low-quality result is one where the large model identifies trees as people and trash cans as vehicles, indicating a clear misidentification.

[0079] S3. Determine whether the category is a high-quality result. If so, build a supervised model based on the high-quality result and replace the first judgment to automatically detect subsequent similar tasks. Otherwise, proceed to step S4.

[0080] In this embodiment, the main steps in building a supervised model are to convert the detection results of the large model into a label file for model training, thus obtaining a labeled dataset, which is then added to an existing dataset for model training.

[0081] S4. Obtain and construct the supervised model based on the secondary annotation of the low-quality results, and replace the first judgment to automatically detect subsequent similar tasks.

[0082] In this embodiment, secondary annotation involves manually annotating low-quality images using Labelme or Labelimg annotation tools, obtaining an annotated dataset, and then using the annotated dataset, along with the dataset obtained from the first evaluation and the original dataset, to supervise the training of the model.

[0083] In this embodiment, preliminary task detection is first performed based on an existing large model. Then, an initial assessment is introduced to confirm the detection results. If the results are of high quality, a supervisory model (i.e., a small model) is built based on the results to replace the initial assessment for automatic detection of subsequent similar tasks. If the results are of low quality, a secondary annotation is introduced to initially modify the low-quality results into high-quality results. Similarly, a supervisory model is built to replace the initial assessment for automatic detection of subsequent similar tasks. The entire process only requires two human interventions—an initial assessment and a secondary annotation. Subsequent task detection is automatically achieved by the collaboration of the large and small models without the need for further human intervention. This effectively reduces human involvement while ensuring detection quality. In particular, the use of annotation tools for manual annotation effectively ensures annotation efficiency.

[0084] Please refer to Figure 2 Embodiment two of the present invention is as follows:

[0085] A weakly supervised learning-driven co-evolutionary method for large and small models, based on the above embodiment one, in this embodiment, as follows: Figure 2 As shown, the construction of the supervised model in steps S3 and S4 is specifically as follows:

[0086] S41. Data collection, annotation and preprocessing: Obtain the original dataset, the dataset from the first analysis and / or the dataset from the second annotation, perform preprocessing such as image resizing, image normalization and data augmentation, and divide the data into training sets, validation sets and test sets for specific tasks to ensure data quality and diversity.

[0087] S42. Model Selection and Optimization: Select different existing small models based on the task type. For example, select the YOLOv5 model for object detection and the U-Net model for image segmentation. Then, select the stochastic gradient descent optimization algorithm and the learning rate scheduler to control the update of the small model parameters.

[0088] S43. Define the loss function: Define the loss function according to the requirements of the task to measure the difference between the prediction of the small model and the true label.

[0089] S44, Backpropagation and Parameter Update: Feed the data from the dataset into the small model for forward propagation, calculate the loss, then calculate the gradient through backpropagation, update the parameters of the small model using the selected optimization algorithm, and minimize the loss function.

[0090] S45. Iterative training: Repeat steps S43 and S44 multiple times, using different batches of training sets, until the preset number of iterations is reached or the training set batches are used up.

[0091] S46. Validation and Hyperparameter Tuning: At the end of each training cycle, the performance of the small model is evaluated using the validation set. The loss and other metrics are monitored to avoid overfitting. Based on the performance of the validation set, parameters such as the learning rate and regularization are adjusted to further improve the performance of the small model, resulting in a supervised model.

[0092] S47. Testing and Evaluation: After training is completed, the performance of the supervised model is evaluated using a test set, which helps to estimate the performance of the supervised model in real-world scenarios.

[0093] In this embodiment, before constructing the supervised model, the original dataset, the dataset of the first judgment and / or the second annotation are collected to ensure data quality and diversity. Then, according to the task, the corresponding initial small model is selected, and it is trained, validated and tested to obtain a supervised model that meets the current task, thereby achieving the accuracy of automatic detection of subsequent similar tasks.

[0094] Step S3 also includes:

[0095] High-quality results are fed back into a large model for iterative training.

[0096] Step S4 also includes:

[0097] The low-quality results that undergo secondary annotation are treated as high-quality results and fed back into the large model for iterative training.

[0098] High-quality results are fed back into the existing large model for iterative training, thereby optimizing the large model and ensuring its continuous evolution. This ensures that subsequent large models can obtain high-quality detection results in their initial detection. At the same time, low-quality results from secondary annotation can be considered equivalent to high-quality results, so these high-quality results can also be fed back into the existing large model for iterative training, thereby further optimizing the large model.

[0099] In this embodiment, iterative training in a large model refers to further optimizing the performance of the large model or adapting it to new data or tasks by retraining it based on the already trained large model. The specific steps of iterative training are as follows:

[0100] Step 1: Prepare data: Organize the original dataset, the dataset from the first analysis, and / or the dataset from the second annotation, and classify them according to specific tasks. Each category is further divided into training set, validation set, and test set to ensure data quality and diversity.

[0101] Step 2, Loading the Model and Optimization Selection: Load the parameters of the existing large model as initial weights, and select the stochastic gradient descent optimization algorithm and learning rate scheduler to control the updating of the large model parameters.

[0102] Step 3: Define the loss function: Define the loss function according to the requirements of the task to measure the difference between the prediction of the large model and the true label.

[0103] Step 4, Backpropagation and Parameter Update: Feed the data from the dataset into the large model for forward propagation, calculate the loss, then calculate the gradient through backpropagation, and update the parameters of the large model using the selected optimization algorithm to minimize the loss function.

[0104] Step 5, iterative training: Repeat steps 3 and 4 multiple times, using different batches of training sets, until the preset number of iterations is reached or the training set batches are used up.

[0105] Step 6, Validation and Evaluation and Hyperparameter Tuning: At the end of each training cycle, use the validation set to evaluate the performance of the large model, monitor the loss and other metrics to avoid overfitting, and adjust parameters such as learning rate and regularization based on the performance of the validation set to further improve the performance of the large model.

[0106] Step 7: After training is complete, use the test set to evaluate the performance of the adjusted large model. This helps to estimate the performance of the large model in real-world scenarios.

[0107] Similar to building a supervised model, the large model constructs training, validation, and test sets by collecting the original dataset and the dataset obtained after each automatic detection task. The original large model is continuously iterated and trained to continuously optimize and improve the large model, ensuring that the detection quality is improved while gradually reducing human intervention in the automatic detection process of subsequent new tasks.

[0108] Please refer to Figure 2 Embodiment 3 of the present invention is as follows:

[0109] A weakly supervised learning-driven co-evolutionary method for large and small models, based on the above embodiment one or two, in this embodiment, as follows: Figure 2 As shown, the procedure before step S1 also includes:

[0110] S0. Extract keywords for the current task using a thesaurus, and retrieve the corresponding large model from the database, specifically:

[0111] Keywords related to the task are predefined in the thesaurus. For example, for the tasks of human and vehicle detection, the keywords are: human, fishing, swimming, vehicle, truck, fog cannon truck, sedan, etc. Then, the keywords are converted into word vectors and the corresponding large model deployed in the server database is called in the form of an interface.

[0112] In this embodiment, the task extracts keywords from the existing lexicon to trigger the corresponding large model for detection, ensuring the validity of the detection results from the introduced large model.

[0113] Both steps S3 and S4 include:

[0114] S34. Update the keyword extraction mechanism through a supervised model to optimize the thesaurus.

[0115] Each time a supervised model is built, the smaller model is also used to update and optimize the lexicon, thereby ensuring the accuracy of detection triggered by keywords in the corresponding larger model, and further improving the collaboration between the larger and smaller models and the accuracy of task detection.

[0116] In this embodiment, step S34 specifically includes:

[0117] S341. Collect labeled data: Obtain labeled data containing keywords and corresponding labels. The corresponding labels are labels indicating whether the keywords are keywords. The labeled data includes manually labeled datasets and data generated by some existing labeling tools.

[0118] S342. Prepare data: Divide the labeled data into training set, validation set and test set to ensure that the distribution of the dataset and the application scenario are consistent with the actual application.

[0119] S343. Define the model architecture: Use a supervised model and define the architecture of the supervised model, including the vector sizes of the input layer, hidden layer, and output layer.

[0120] S344. Data Preprocessing: Transform the text data of keywords into a format acceptable to the supervised model. The format includes word segmentation and word vector representations. A pre-trained word vector model can be used, or word vectors can be learned along with the model's heart rate.

[0121] S345. Model Training: Train the supervised model using the training set, optimize the parameters of the supervised model and minimize the loss function (such as cross-entropy), and use the validation set to evaluate the performance of the supervised model to avoid overfitting.

[0122] S346. Validation, evaluation, and hyperparameter optimization: Adjust the hyperparameters of the supervised model, including learning rate, batch size, and hidden layer dimension, to obtain optimal performance.

[0123] S347. Testing and Evaluation: Finally, the performance of the adjusted supervised model is evaluated using a test set.

[0124] S348. Lexicon Update: During the training process, the supervised model gradually learns keyword features and context. By analyzing the performance of the supervised model on the validation and test sets, new keywords not included in the lexicon can be identified and added to the lexicon to better capture new keywords.

[0125] S349. Continuous Iteration: Based on the performance of the supervised model in practical applications and user feedback, continuous iterative optimization can be carried out. The above steps can be repeated periodically to update the labeled data, fine-tune the supervised model, and update the lexicon, etc., in order to maintain the high efficiency of the keyword extraction mechanism.

[0126] Please refer to Figure 3 Embodiment 3 of the present invention is as follows:

[0127] A weakly supervised learning-driven co-evolutionary terminal 1 for large and small models includes a memory 2, a processor 3, and a computer program stored on the memory 2 and executable on the processor 3. When the processor 3 executes the computer program, it implements the steps of a weakly supervised learning-driven co-evolutionary method for large and small models in any of the embodiments 1 to 3 described above.

[0128] In summary, the present invention provides a weakly supervised learning-driven method and terminal for co-evolution of large and small models. This method first performs preliminary task detection based on an existing large model, then introduces a judgment to confirm the detection results. If the results are high-quality, a supervised model (i.e., a small model) is built based on the results to replace the initial judgment for automatic detection of subsequent similar tasks. If the results are low-quality, a secondary annotation is introduced to initially modify the low-quality results into high-quality results, thus again building a supervised model to replace the initial judgment for automatic detection of subsequent similar tasks. The entire process requires only two human interventions—an initial judgment and a secondary annotation. Subsequent task detection is automatically achieved through the collaboration of large and small models, eliminating the need for further human intervention. This effectively reduces human involvement while ensuring detection quality.

[0129] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A weakly supervised learning driven size model co-evolution method, characterized in that, The method comprises the steps of: S1, introducing an existing large model to preliminarily detect a current task to obtain an initial detection result; S2, obtaining a category of the initial detection result confirmed by a first research and judgment, the category being a high-quality result or a low-quality result, the high-quality result being a result in a correct answer set, and the low-quality result being a result outside the correct answer set; S3, judging whether the category is the high-quality result, if yes, constructing a supervised model according to the high-quality result, and replacing the first research and judgment to automatically detect a subsequent same-type task, otherwise, entering step S4; S4, obtaining and constructing the supervised model according to a second labeling of the low-quality result, and replacing the first research and judgment to automatically detect a subsequent same-type task, the same-type task including a target detection task or an image segmentation task; The second labeling is that a data set labeled by manual labeling of an image of the low-quality result by using a Labelme or Labelimg labeling tool is used together with a data set obtained by the first research and judgment and an original data set for training of the supervised model.

2. The weakly supervised learning driven size model co-evolution method of claim 1, wherein, The constructing of the supervised model in steps S3 and S4 is specifically: S41, obtaining the original data set, the data set of the first research and judgment and / or the data set of the second labeling, performing preprocessing of image size adjustment, image normalization and data enhancement, and dividing data into a training set, a verification set and a test set; S42, selecting different existing small models according to a task type, and selecting a stochastic gradient descent optimization algorithm and a learning rate scheduler to control updating of parameters of the small model; S43, defining a loss function according to a requirement of the task to measure a gap between a prediction of the small model and a real label; S44, feeding data in the data set into the small model to perform forward propagation, calculating a loss, then calculating a gradient by back propagation, and updating parameters of the small model by using the selected optimization algorithm, and minimizing the loss function; S45, performing steps S43 and S44 for multiple iterations, using different batches of the training set, until a preset iteration number is reached or the training set batch is used up; S46, at the end of each training period, evaluating performance of the small model by using the verification set, monitoring a loss and a metric index, and adjusting a learning rate and a regularization parameter according to a performance of the verification set to obtain the supervised model; S47, after training is completed, evaluating performance of the supervised model by using the test set.

3. The weakly supervised learning driven size model co-evolution method of claim 1, wherein, The step S3 further comprises: feeding back the high-quality result to the large model to perform iterative training of the large model; The step S4 further comprises: feeding back the low-quality result subjected to the second labeling as the high-quality result to the large model to perform iterative training of the large model.

4. The weakly supervised learning driven size model co-evolution method of claim 3, wherein, The iterative training is specifically: Step 1, sorting the original data set, the data set of the first research and judgment and / or the data set of the second labeling, and classifying according to a specific task, and dividing each category into a training set, a verification set and a test set; Step2, load the parameters of the existing large model as the initial weights, and select the optimization algorithm of stochastic gradient descent and the learning rate scheduler to control the update of the parameters of the large model; Step3, define the loss function according to the requirements of the task, measure the gap between the prediction of the large model and the real label; Step4, feed the data in the data set into the large model for forward propagation, calculate the loss, then calculate the gradient by back propagation, and update the parameters of the large model using the selected optimization algorithm to minimize the loss function; Step5, execute steps Step3 and Step4 multiple times using different batches of training sets until the preset number of iterations is reached or the training set batch is used up; Step6, at the end of each training cycle, evaluate the performance of the large model using the validation set, monitor the loss and metric indicators, and adjust the learning rate and regularization parameters according to the performance of the validation set; Step7, after training is completed, evaluate the performance of the adjusted large model using the test set.

5. A weakly supervised learning driven size model co-evolution terminal, characterized in that, A computer program product comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, the processor implementing the steps of the weakly supervised learning driven large model co-evolution method according to any one of claims 1 to 4 when executing the computer program.

Images