Semi-automatic target detection pre-labeling method based on human-computer interaction
By introducing active learning and iterative optimization methods into semi-automatic object detection pre-labeling technology, the problem of large errors in complex scenarios is solved, the workload of manual marking is reduced, the labeling efficiency and accuracy are improved, and more efficient target detection automation is achieved .
Patent Information
- Application Number
- CN202411119084.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2024-11-15
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing semi-automatic target detection pre-labeling technology has a large error in complex scenarios, resulting in large workloads of manual labeling and affecting efficiency.
Using semi-automatic object detection pre-labeling method based on human-computer interaction, through data cleaning and preprocessing in the initialization stage, the most valuable samples are selected for manual annotation using active learning technology, and through model update and iterative optimization, combined with rigorous verification Rules and automatic verification methods gradually improve model performance and labeling accuracy.
The workload of manual labeling is reduced, data utilization efficiency and automation and accuracy of target detection are improved, and the expected labeling effect is achieved.
Smart Images

Figure CN118968030A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of human-computer interaction, and in particular to a semi-automatic target detection pre-labeling method based on human-computer interaction. Background Art
[0002] Human-Computer Interaction (HCI) is a technical field that studies the information exchange between people and computers. It refers to the process of information exchange between people and computers using a certain dialogue language and a certain interactive method to complete a certain task. Research in this field aims to enhance the interactive experience between users and systems, so that users can communicate with machines more conveniently and intuitively. In the field of human-computer interaction, target detection is an important task used to identify specific targets or objects in images or videos. However, manually annotating large amounts of image data is time-consuming and labor-intensive for humans. Therefore, semi-automatic target detection pre-annotation was introduced to reduce the manual burden. This method uses an automatic annotation algorithm to preliminarily annotate the data, which is then checked and corrected by humans, achieving a balance between annotation efficiency and accuracy.
[0003] For example, the patent document with application number 202110328124.2 discloses a semi-automatic labeling method and system based on human-computer interaction. The method of the invention first uses additional clicks as prior information to detect the target, thereby achieving the expected effect of semi-automatic labeling, and then reduces the false detection rate by penalizing the misdetected targets, thereby further improving the detection effect; and the semi-automatic labeling model of the invention not only greatly improves the accuracy of the original model, but also has a good migration effect on different data sets, and can get rid of the time-consuming and labor-intensive manual labeling disadvantages. It only needs to manually provide prior information about the location of a target, and the model can actively strengthen the detection near the current position; in addition, a correction part is added to correct the target with unsatisfactory prediction results to make it correct; at the same time, the original structure is improved, and a correction operation is added on the basis of interactive detection to achieve the correction purpose.
[0004] However, the existing semi-automatic annotation technology similar to the above application still has the following shortcomings:
[0005] At present, although semi-automatic target detection pre-labeling technology has been introduced, due to the limitations of the automatic labeling algorithm, errors are inevitably present in the labeling results. Therefore, most of the automatic labeling results need to be manually checked and corrected. Especially in complex scenarios, the automatic labeling algorithm is even more difficult to accurately identify the target, and the error of the labeling result is also greater. This leads to the fact that the workload of manual labeling in the existing technology is still large, and affects the labeling efficiency.
[0006] Therefore, there is an urgent need to improve this shortcoming. The present invention studies and improves the existing technology and its shortcomings, and provides a semi-automatic target detection pre-labeling method based on human-computer interaction. Summary of the invention
[0007] The purpose of the present invention is to provide a semi-automatic target detection pre-labeling method based on human-computer interaction to solve the problems raised in the above background technology.
[0008] To achieve the above object, the present invention provides the following technical solution: a semi-automatic target detection pre-labeling method based on human-computer interaction, comprising the following steps:
[0009] S1, initialization stage: prepare unlabeled image data to form a training data set, perform preliminary data cleaning and preprocessing, and then select or train an initial target detection model based on the training data set;
[0010] S2, the first round of prediction stage: the unlabeled data set is input into the initial model to predict the target detection. The model will output the prediction results of each image, including the location, size and category of the prediction box;
[0011] S3, sample selection stage: using active learning technology, according to the model's prediction results and uncertainty of the samples, the most valuable samples are selected for manual annotation, and the annotation content of the manual annotation includes but is not limited to the location (annotation box), size (bounding box size) and category (classification label) of the target;
[0012] S4, model update stage: add manually annotated samples to the training set, retrain or fine-tune the model, and gradually lower the uncertainty threshold or increase the representativeness threshold as the annotated data increases and the model performance improves, so as to select more valuable samples for annotation;
[0013] S5, loop iteration stage: repeat steps S2 to S4 to form a loop iteration process, and continue until the expected annotation effect and model performance are achieved;
[0014] S6. Data verification stage: Verify the automatically labeled and manually labeled data to ensure that the number of labeled boxes, the amount of labeled data, etc. meet the verification rules.
[0015] Furthermore, in step S1, the data cleaning specifically includes the following operations:
[0016] Data deduplication: compare the hash values or feature vectors of images, treat images with high similarity as duplicate images, delete them, and deduplicate them based on file names or image metadata (such as time, location, shooting device, etc.);
[0017] Data filtering: traverse the data set, use the is_image() function to determine whether each file is an image file, delete non-image files, and try to open each image file. Use the Image.open() and img.load() methods to detect whether the file is intact. If the file cannot be opened, it is considered a damaged image and deleted or moved to the damaged image directory. In addition, delete low-quality, non-standard, and unreasonable images, such as blurred, distorted, undersized, or oversized images;
[0018] Data restoration: For missing data, noise, and artifacts, image processing techniques including filtering, interpolation, and denoising are used to restore the data.
[0019] Furthermore, in step S1, data preprocessing specifically includes the following operations:
[0020] Format conversion: Convert image data into a format that is easier for deep learning models to process, such as converting images to RGB images or grayscale images;
[0021] Normalization: Scale the pixel values of an image to [0,1] or standardize them to a certain mean and standard deviation range;
[0022] Data augmentation: Increase the amount and diversity of data by rotating (rotating the image by a certain angle), translating (moving the image in the x or y direction on the image plane), scaling (changing the size of the image), and flipping (flipping the image horizontally or vertically).
[0023] Dataset segmentation: The dataset is divided into 80% training set, 10% validation set, and 10% test set according to the ratio.
[0024] Furthermore, in step S3, the sample selection strategy specifically includes:
[0025] Uncertainty sampling: Calculate the prediction probability or confidence of the model for each unlabeled sample, and select the sample with the lowest confidence for labeling; the uncertainty sampling is based on the model's prediction confidence for the sample, and selects the sample with the least model uncertainty for labeling. If the model's prediction confidence for a sample is very low, it means that the sample is very "difficult" for the model, so labeling such samples can provide more learning information for the model;
[0026] Query difficulty strategy: train several different models or use multiple versions of the same model as committee members, let each member predict the unlabeled samples, and select the samples with the largest difference in prediction results for labeling. By comparing the prediction results of different models, samples that are generally difficult for the models to handle can be selected, thereby improving the generalization ability of the model; the query difficulty strategy uses multiple models (or "committees") to predict the unlabeled samples, and selects the samples with the largest difference in predictions between committee members for labeling;
[0027] Representative sampling: Use clustering algorithms or feature selection methods to determine representative samples in unlabeled datasets. The specific operation is to first cluster the unlabeled samples, and then select a certain number of samples from each cluster for labeling, or select representative samples based on the feature distribution of the samples. The representative sampling strategy aims to select samples that can represent the overall distribution of the unlabeled dataset, which means that the selected samples should be able to cover various situations in the dataset, including different categories, different feature distributions, etc.
[0028] Hybrid strategy: In practical applications, uncertainty sampling, query difficulty strategy, and representative sampling can be combined to form a hybrid strategy; for example, uncertainty sampling can be used to select a part of samples for labeling, and then a representative sampling strategy can be used to supplement a part of samples that can cover more data distribution.
[0029] Furthermore, in step S5, during the iteration process, the validation set is regularly used to evaluate the performance of the model to adjust the labeling strategy and model parameters, which specifically includes the following process steps:
[0030] Set the iteration cycle: determine the frequency of iterations, for example, perform a validation after each round of iterative model training, or perform a validation every few rounds of iterations;
[0031] Model training: Use the training set to train the model, select the loss function and optimization algorithm according to the task, and optimize the model parameters through multiple iterations;
[0032] Verify the model: Use the validation set to evaluate the model and calculate the performance indicators of the model on the validation set, such as accuracy, recall, precision, F1 value, etc.
[0033] Analyze the evaluation results: Analyze the evaluation results of the validation set to find out where the model performs poorly;
[0034] Adjust the labeling strategy: adjust the labeling strategy based on the evaluation results of the validation set;
[0035] Adjust model parameters: According to the evaluation results of the validation set, adjust the model's hyperparameters, for example, adjust the learning rate, batch size, number of iterations and other parameters to optimize the model's performance, or try different model structures and algorithms to improve the model's performance;
[0036] Repeated iteration: Reuse the adjusted labeled data for model training, repeat the above steps, and continuously iterate and optimize the model until satisfactory performance is achieved.
[0037] Furthermore, in step S5, during the iteration process, the training time, evaluation results, and parameter adjustment information of each iteration are recorded, and a monitoring tool or a log system is used to track the performance changes and parameter adjustments of the model.
[0038] Furthermore, in step S6, the verification adopts an automated verification method, specifically: using programming languages and data processing tools including Python and Pandas, a script is written to automatically check whether various indicators of the data meet the verification rules.
[0039] Furthermore, in step S6, the verification rules specifically include verification of six aspects: number of annotation boxes, amount of annotation data, data format and consistency, logical relationship, boundary value and abnormal value, and data distribution, as follows:
[0040] Verification rule 1: Verification of the number of annotation boxes
[0041] Image dataset: Ensure that the number of annotation boxes for each label exceeds the specified number, for example, at least 10 annotation boxes. This rule is used to ensure that the model can fully learn the characteristics of the target object;
[0042] Verification rule 2: Annotation data volume verification
[0043] Text Dataset:
[0044] The amount of labeled data in the dataset should exceed the set threshold, such as 600, to ensure that the model has enough labeled data to train;
[0045] The amount of data for each annotated label should reach a set number, such as more than 50, which helps the model better learn the characteristics of each category;
[0046] The amount of unlabeled data must meet the set requirements, such as more than 600, and these data can be used for subsequent model training and evaluation;
[0047] Verification rule three: data format and consistency verification
[0048] Verify that the data follows a unified format and annotation standards, such as the coordinate format of the annotation box, the naming rules of the labels, etc.
[0049] Ensure data consistency, for example, check whether the same object is labeled consistently in different images;
[0050] Verification rule 4: logical relationship verification
[0051] Verify the logical relationships in the annotated data, such as the subject-verb-object relationship in a sentence, the association between words, etc.
[0052] Verification rule 5: Boundary value and outlier verification
[0053] Check whether the boundary values in the labeled data are reasonable, such as the size range of numbers, the legitimacy of time, etc.;
[0054] Detect and process outliers, such as abnormal text content, abnormal images, etc.
[0055] Verification rule six: data distribution verification
[0056] For numerical data, check whether its distribution is reasonable to avoid the problem of imbalance in labeled data.
[0057] Furthermore, the method further comprises the steps of:
[0058] S7, labeling data output stage: Save the final labeling data that has passed the verification into a standard file format, such as VOC, COCO, etc., or use JSON, XML and other data structures according to specific needs, and output the file to the specified location.
[0059] Furthermore, in step S7, the content of the output file includes: the path of the image file, the image size, the target category ID, and the bounding box coordinates.
[0060] The present invention provides a semi-automatic target detection pre-marking method based on human-computer interaction, which has the following beneficial effects:
[0061] By introducing active learning technology, the present invention can select the most valuable samples, that is, the samples that most need manual inspection, for manual labeling according to the model's prediction results and uncertainties for the samples, thereby reducing the workload of manual labeling, which is beneficial to reducing labeling costs and improving data utilization efficiency. The model is iterated cyclically based on the sample selection results until the expected labeling effect and model performance are achieved. In combination with strict verification rules and automatic verification methods, it helps to further improve the degree of automation and accuracy of target detection pre-labeling. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 A schematic diagram of the operation flow of a semi-automatic target detection pre-labeling method based on human-computer interaction according to the present invention;
[0063] Figure 2A schematic diagram of verification rules of a semi-automatic target detection pre-labeling method based on human-computer interaction of the present invention. DETAILED DESCRIPTION
[0064] The following embodiments of the present invention are described in further detail in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.
[0065] like Figure 1-Figure 2 As shown, a semi-automatic target detection pre-labeling method based on human-computer interaction includes the following steps:
[0066] S1, initialization stage: prepare unlabeled image data to form a training data set, perform preliminary data cleaning and preprocessing, and then select or train an initial target detection model based on the training data set;
[0067] Data cleaning specifically includes the following operations:
[0068] Data deduplication: compare the hash values or feature vectors of images, treat images with high similarity as duplicate images, delete them, and deduplicate them based on file names or image metadata (such as time, location, shooting device, etc.);
[0069] Data filtering: traverse the data set, use the is_image() function to determine whether each file is an image file, delete non-image files, and try to open each image file. Use the Image.open() and img.load() methods to detect whether the file is intact. If the file cannot be opened, it is considered a damaged image and deleted or moved to the damaged image directory. In addition, delete low-quality, non-standard, and unreasonable images, such as blurred, distorted, undersized, or oversized images;
[0070] Data restoration: Use image processing techniques including filtering, interpolation, and denoising to restore missing, noisy, or artifact data.
[0071] Data preprocessing specifically includes the following operations:
[0072] Format conversion: Convert image data into a format that is easier for deep learning models to process, such as converting images to RGB images or grayscale images (depending on the needs of the model);
[0073] Normalization: Scale the pixel values of the image to [0, 1] or standardize them to a certain mean and standard deviation range so that the model can extract features and learn more easily.
[0074] Data augmentation: Increase the amount and diversity of data by rotating (rotating the image by a certain angle), translating (moving the image in the x or y direction on the image plane), scaling (changing the size of the image), and flipping (flipping the image horizontally or vertically) to improve the generalization ability of the model;
[0075] Dataset segmentation: The dataset is divided into 80% training set, 10% validation set, and 10% test set in proportion to facilitate subsequent model training, parameter adjustment, and evaluation;
[0076] S2, the first round of prediction stage: the unlabeled data set is input into the initial model to predict the target detection. The model will output the prediction results of each image, including the location, size and category of the prediction box;
[0077] S3, sample selection stage: using active learning technology, according to the model's prediction results and uncertainty of the samples, the most valuable samples are selected for manual annotation, and the annotation content of the manual annotation includes but is not limited to the location (annotation box), size (bounding box size) and category (classification label) of the target;
[0078] In step S3, the sample selection strategy specifically includes:
[0079] Uncertainty sampling: Calculate the prediction probability or confidence of the model for each unlabeled sample, and select the sample with the lowest confidence for labeling, so as to efficiently select the sample that is most helpful to improve the model performance; Uncertainty sampling is based on the model's prediction confidence of the sample, and selects the sample with the least model uncertainty for labeling. If the model's prediction confidence for a sample is very low, it means that the sample is very "difficult" for the model, so labeling such samples can provide more learning information for the model;
[0080] Query difficulty strategy: Train several different models or use multiple versions of the same model as committee members, let each member predict the unlabeled samples, and select the samples with the largest difference in prediction results for labeling. By comparing the prediction results of different models, you can select those samples that are generally difficult for the model to handle, thereby improving the generalization ability of the model; The query difficulty strategy uses multiple models (or "committees") to predict unlabeled samples, and selects those samples with the largest difference in predictions between committee members for labeling;
[0081] Representative sampling: Use clustering algorithms or feature selection methods to determine representative samples in unlabeled data sets. The specific operation is to first cluster the unlabeled samples, and then select a certain number of samples from each cluster for labeling, or select representative samples based on the feature distribution of the samples. By selecting representative samples, it can be ensured that the labeled data set can fully reflect the actual situation of the unlabeled data set, thereby improving the adaptability of the model to the entire data set; the representative sampling strategy aims to select samples that can represent the overall distribution of the unlabeled data set, which means that the selected samples should be able to cover various situations in the data set, including different categories, different feature distributions, etc.
[0082] Hybrid strategy: In practical applications, uncertainty sampling, query difficulty strategy, and representative sampling can be combined to form a hybrid strategy. For example, uncertainty sampling can be used to select a part of samples for labeling, and then representative sampling strategy can be used to supplement a part of samples that can cover more data distribution.
[0083] S4, model update stage: add manually annotated samples to the training set, retrain or fine-tune the model, and gradually lower the uncertainty threshold or increase the representativeness threshold as the annotated data increases and the model performance improves, so as to select more valuable samples for annotation;
[0084] S5, loop iteration phase: Repeat steps S2 to S4 to form a loop iteration process, and continue until the expected labeling effect and model performance are achieved; during the iteration process, the validation set is regularly used to evaluate the performance of the model to adjust the labeling strategy and model parameters. The specific process steps include the following:
[0085] Set the iteration cycle: determine the frequency of iterations, for example, perform a validation after each round of iterative model training, or perform a validation every few rounds of iterations;
[0086] Model training: In the data cleaning and preprocessing stage, the data set has been divided into training set, validation set and test set. The model is trained using the training set, and the loss function and optimization algorithm are selected according to the task. The model parameters are optimized through multiple iterations.
[0087] Verify the model: Use the validation set to evaluate the model and calculate the performance indicators of the model on the validation set, such as accuracy, recall, precision, F1 value, etc.
[0088] Analyze the evaluation results: Analyze the evaluation results of the validation set to find out where the model performs poorly, such as which categories of samples have a high prediction error rate or which features are not well utilized;
[0089] Adjust the labeling strategy: Adjust the labeling strategy based on the evaluation results of the validation set. For example, if you find that the model does not accurately predict certain categories of samples, increase the labeling of these categories of samples; or if you find that certain features are not well utilized, improve the labeling method to better extract these features.
[0090] Adjust model parameters: According to the evaluation results of the validation set, adjust the model's hyperparameters, for example, adjust the learning rate, batch size, number of iterations and other parameters to optimize the model's performance, or try different model structures and algorithms to improve the model's performance;
[0091] Repeated iteration: Use the adjusted labeled data again for model training, repeat the above steps, and continuously iterate and optimize the model until satisfactory performance is achieved;
[0092] During the iteration process, record the training time, evaluation results, and parameter adjustment information of each iteration, and use monitoring tools or log systems to track the performance changes and parameter adjustments of the model;
[0093] S6, data verification stage: Verify the automatically annotated and manually annotated data to ensure that the number of annotated boxes, the amount of annotated data, etc. meet the verification rules; the verification rules specifically include verification of the number of annotated boxes, the amount of annotated data, data format and consistency, logical relationship, boundary values and abnormal values, and data distribution in six aspects, as follows:
[0094] Verification rule 1: Verification of the number of annotation boxes
[0095] Image dataset: Ensure that the number of annotation boxes for each label exceeds the specified number, for example, at least 10 annotation boxes. This rule is used to ensure that the model can fully learn the characteristics of the target object;
[0096] Verification rule 2: Annotation data volume verification
[0097] Text Dataset:
[0098] The amount of labeled data in the dataset should exceed the set threshold, such as 600, to ensure that the model has enough labeled data to train;
[0099] The amount of data for each annotated label should reach a set number, such as more than 50, which helps the model better learn the characteristics of each category;
[0100] The amount of unlabeled data must meet the set requirements, such as more than 600, and these data can be used for subsequent model training and evaluation;
[0101] Verification rule three: data format and consistency verification
[0102] Verify that the data follows a unified format and annotation standards, such as the coordinate format of the annotation box, the naming rules of the labels, etc.
[0103] Ensure data consistency, for example, check whether the same object is labeled consistently in different images;
[0104] Verification rule 4: logical relationship verification
[0105] Verify the logical relationships in the annotated data, such as the subject-verb-object relationship in a sentence, the association between words, etc.
[0106] Verification rule 5: Boundary value and outlier verification
[0107] Check whether the boundary values in the labeled data are reasonable, such as the size range of numbers, the legitimacy of time, etc.;
[0108] Detect and process outliers, such as abnormal text content, abnormal images, etc.
[0109] Verification rule six: data distribution verification
[0110] For numerical data, check whether its distribution is reasonable to avoid the problem of imbalanced labeled data;
[0111] S7, labeling data output stage: save the final labeling data that has passed the verification into a standard file format, such as VOC, COCO, etc., or use JSON, XML and other data structures according to specific needs, and output the file to the specified location; the content of the output file includes: image file path, image size, target category ID, bounding box coordinates;
[0112] Taking the VOC format as an example, each target includes one or more <object>Tags, each of which contains a category name (such as <name> person< / name> ) and bounding box coordinates (such as <bndbox> <xmin> 10< / xmin> <ymin> 20< / ymin> <xmax> 50< / xmax> <ymax> 60< / yma x>< / ymax> < / bndbox> ).
[0113] The embodiments of the present invention are given for the purpose of illustration and description, and are not intended to be exhaustive or to limit the invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.< / object>
Claims
1. A semi-automatic target detection pre-labeling method based on human-computer interaction, characterized in that: The following steps are involved: S1, initialization stage: prepare unlabeled image data to form a training data set, perform preliminary data cleaning and preprocessing, and then select or train an initial target detection model based on the training data set; S2, first round prediction stage: input the unlabeled data set into the initial model to predict the target detection, and the model will output the prediction results for each image; S3, sample selection stage: using active learning technology, according to the model's prediction results and uncertainty of the samples, the most valuable samples are selected for manual labeling; S4, model update stage: add manually annotated samples to the training set, retrain or fine-tune the model, and gradually lower the uncertainty threshold or increase the representativeness threshold as the annotated data increases and the model performance improves, so as to select more valuable samples for annotation; S5, loop iteration stage: repeat steps S2 to S4 to form a loop iteration process, and continue until the expected annotation effect and model performance are achieved; S6. Data verification stage: Verify the automatically labeled and manually labeled data to ensure compliance with the verification rules.
2. The semi-automatic target detection pre-marking method based on human-computer interaction according to claim 1 is characterized in that: In step S1, data cleaning specifically includes the following operations: Data deduplication: compare the hash values or feature vectors of images, treat images with high similarity as duplicate images, delete them, and deduplicate them based on file names or image metadata; Data filtering: traverse the data set, use the is_image() function to determine whether each file is an image file, delete non-image files, and try to open each image file. Use the Image.open() and img.load() methods to detect whether the file is intact. If the file cannot be opened, it is considered a damaged image and deleted or moved to the damaged image directory. In addition, delete low-quality, non-standard, and unreasonable images; Data restoration: For missing data, noise, and artifacts, image processing techniques including filtering, interpolation, and denoising are used to restore the data.
3. The semi-automatic target detection pre-marking method based on human-computer interaction according to claim 1 is characterized in that: In step S1, data preprocessing specifically includes the following operations: Format conversion: Convert image data into a format that is easier for deep learning models to process; Normalization: Scale the pixel values of an image to [0,1] or standardize them to a certain mean and standard deviation range; Data augmentation: Increase the amount and diversity of data by rotating, translating, scaling, and flipping; Dataset segmentation: The dataset is divided into 80% training set, 10% validation set, and 10% test set according to the ratio.
4. The semi-automatic target detection pre-marking method based on human-computer interaction according to claim 1 is characterized in that: In step S3, the sample selection strategy specifically includes: Uncertainty sampling: Calculate the model’s predicted probability or confidence for each unlabeled sample, and select the sample with the lowest confidence for labeling; Query difficulty strategy: Train several different models or use multiple versions of the same model as committee members, let each member make predictions on unlabeled samples, and select the samples with the largest difference in prediction results for labeling; Representative sampling: Use clustering algorithms or feature selection methods to determine representative samples in unlabeled data sets. The specific operation is to first cluster the unlabeled samples, and then select a certain number of samples from each cluster for labeling, or select representative samples based on the feature distribution of the samples; Hybrid strategy: In practical applications, uncertainty sampling, query difficulty strategy, and representative sampling can be combined to form a hybrid strategy.
5. The semi-automatic target detection pre-marking method based on human-computer interaction according to claim 1 is characterized in that: In step S5, during the iteration process, the validation set is regularly used to evaluate the performance of the model to adjust the labeling strategy and model parameters, which specifically includes the following process steps: Set the iteration cycle: determine the frequency of iteration; Model training: Use the training set to train the model, select the loss function and optimization algorithm according to the task, and optimize the model parameters through multiple iterations; Verify the model: Use the validation set to evaluate the model and calculate the performance indicators of the model on the validation set; Analyze the evaluation results: Analyze the evaluation results of the validation set to find out where the model performs poorly; Adjust the labeling strategy: adjust the labeling strategy based on the evaluation results of the validation set; Adjust model parameters: According to the evaluation results of the validation set, adjust the model's hyperparameters, or try different model structures and algorithms to improve the model's performance; Repeated iteration: Reuse the adjusted labeled data for model training, repeat the above steps, and continuously iterate and optimize the model until satisfactory performance is achieved.
6. The semi-automatic target detection pre-marking method based on human-computer interaction according to claim 5, characterized in that: In step S5, during the iteration process, the training time, evaluation results, and parameter adjustment information of each iteration are recorded, and a monitoring tool or a log system is used to track the performance changes and parameter adjustments of the model.
7. The semi-automatic target detection pre-marking method based on human-computer interaction according to claim 1 is characterized in that: In step S6, the verification adopts an automated verification method, specifically: using programming languages and data processing tools including Python and Pandas, a script is written to automatically check whether various indicators of the data meet the verification rules.
8. The method for semi-automatic target detection and pre-marking based on human-computer interaction according to claim 7, characterized in that: In step S6, the verification rules specifically include verification of the number of annotation boxes, the amount of annotation data, data format and consistency, logical relationship, boundary value and abnormal value, and data distribution, as follows: Verification rule 1: Verification of the number of annotation boxes Image dataset: Ensure that the number of annotation boxes for each label exceeds the specified number. This rule is used to ensure that the model can fully learn the characteristics of the target object; Verification rule 2: Annotation data volume verification Text dataset: The amount of labeled data in the dataset should exceed the set threshold; The amount of data for each labeled label should reach the set amount; the amount of unlabeled data must meet the set requirements; Verification rule three: data format and consistency verification Verify that data follows a unified format and annotation standard; ensure data consistency; Verification rule 4: logical relationship verification Verify the logical relationship in the labeled data; Verification rule 5: Boundary value and outlier verification Check whether the boundary values in the labeled data are reasonable; Detect and handle outliers; Verification rule six: data distribution verification For numerical data, check whether its distribution is reasonable.
9. The semi-automatic target detection pre-marking method based on human-computer interaction according to claim 1, characterized in that: Also includes the steps: S7, annotation data output stage: save the final annotation data that has passed the verification into a standard file format, or use a specified data structure according to specific needs, and output the file to a specified location.
10. The method for semi-automatic target detection and pre-marking based on human-computer interaction according to claim 9, characterized in that: In step S7, the content of the output file includes: the path of the image file, the image size, the target category ID, and the bounding box coordinates.
Citation Information
Patent Citations
Semi-automatic labeling method and system based on human-computer interaction
CN112966777A
Target detection method for reducing labeling demand based on active domain adaptive learning
CN114724015A
Object detection active learning method based on classification consistency before and after object regression
CN117372822A
Pipeline material picture data online cleaning and management system and method
CN118132775A
Cited By
Sample processing method and processing system
CN120071053A
Metalearning anomaly detection method and system under data full life cycle
CN120579100A
Large model labeling method based on data labeling rule
CN120596535A
