A training data set annotation method and device for AI visual recognition

Abstract

The present invention relates to the technical field of image processing, and provides a training data set annotation method and device for AI visual recognition, the method comprising: obtaining and processing original image data, extracting feature mapping, inputting a target detection model to obtain preliminary annotation results and analyzing them, determining cluster centers and candidate target sets; classifying candidate target sets based on cluster centers, calculating the confidence of the classification annotation results, taking the annotation data with confidence higher than a threshold as high-confidence annotation data, and reviewing and checking the consistency of the data below the threshold to generate consistent annotation data; optimizing the target detection model in combination with high-confidence data, and re-annotating low-confidence data using the optimized model, and merging high-confidence and re-annotated data to generate a complete annotation training set to train the recognition model. Through the above scheme, the consistency and confidence of the data are ensured, and the problems of inconsistent annotation, missing or mislabeled annotation in complex scenes and multi-target recognition are effectively improved.

Description

Technical Field

[0001] The present application relates to the technical field of image processing, and in particular to a method and device for labeling training data sets for AI visual recognition. Background Art

[0002] With the rapid development of artificial intelligence technology, computer vision is increasingly being used in all walks of life. As an important branch of computer vision, visual recognition technology plays a key role in autonomous driving, face recognition, security monitoring, intelligent manufacturing and other fields. High-quality training data sets are the cornerstone for improving the performance of visual recognition algorithms, and their annotation quality directly affects the accuracy and stability of the model.

[0003] In related technical means, the labeler identifies the target with the naked eye and labels it. Although this method can achieve a certain degree of accuracy, it is extremely inefficient and prone to human errors when processing large amounts of image data. In addition, in recent years, automatic labeling technology based on deep learning has gradually emerged. By training detection models to perform preliminary labeling on images and combining manual review, the labeling efficiency and accuracy can be improved to a certain extent.

[0004] Regarding the above technical solutions, although the efficiency and accuracy of labeling can be effectively improved by combining automatic labeling with manual review, the existing methods still have problems such as inconsistent labeling, missed labels, and incorrect labeling when dealing with complex scenes and multi-target recognition. Summary of the invention

[0005] In order to improve the problems of inconsistent labeling, missed labels and mislabeling in existing methods when dealing with complex scenes and multi-target recognition, the present application provides a training data set labeling method and device for AI visual recognition, which optimizes the labeling process through preprocessing, feature extraction, cluster analysis and consistency check, improves the quality and confidence of the labeled data, and ensures the integrity and reliability of the training data set.

[0006] The present invention provides a training data set annotation method for AI visual recognition, comprising: obtaining original image data of a target and performing preprocessing to obtain preprocessed image data, performing feature extraction on the preprocessed image data to obtain a feature map; inputting the feature map and the preprocessed image data into a preset target detection model to obtain a preliminary annotation result, performing cluster analysis on the preliminary annotation result through a clustering algorithm to obtain a plurality of cluster centers and a corresponding candidate target set; classifying the candidate target set based on the cluster centers to obtain a classified annotation result, calculating the confidence of each annotation in the classified annotation result, screening the confidence according to a preset confidence threshold, taking the confidence above the threshold as high-confidence annotation data, and filtering the confidence below the threshold as low-confidence annotation data. The confidence value is used as low-confidence annotation data; the low-confidence annotation data is reviewed to obtain a reviewed annotation data set, the reviewed annotation data set is input into a preset consistency check model for consistency check, and the consistent annotation data is output, and the consistent annotation data and the high-confidence annotation data are fed back to the target detection model to obtain an optimized target detection model; the low-confidence annotation data is re-annotated using the optimized target detection model to obtain re-annotated data, the high-confidence annotation data is merged with the re-annotated data to generate a complete annotation training data set, and the complete annotation training data set is input into a preset artificial intelligence visual recognition model for model training to obtain a trained artificial intelligence visual recognition model.

[0007] As a preferred solution, the steps of acquiring the original image data of the target and preprocessing it to obtain the preprocessed image data, extracting features from the preprocessed image data, and obtaining feature mapping include: acquiring the original image data through an image acquisition device, and obtaining metadata related to the original image data, normalizing the original image data according to MinMax normalization to obtain normalized image data, applying a denoising filter to eliminate noise from the normalized image data to obtain denoised image data; performing color correction on the original image data through a color correction algorithm to obtain color corrected image data, and applying an edge enhancement algorithm to the color corrected image data. The data is edge enhanced to obtain enhanced image data, and the denoised image data and the enhanced image data are feature fused according to a feature fusion technology to obtain a feature set; the feature set is dimensionality reduced by using a principal component analysis algorithm to obtain combined feature data, and the combined feature data is segmented into a region of interest based on a preset standard to obtain a segmentation map and a corresponding region descriptor, and a convolutional neural network is used to perform feature extraction on the segmentation map and the region descriptor to obtain a refined feature map and key point data, and the metadata is associated with the refined feature map and the key point data by using a feature matching algorithm to generate a feature map for inputting a target detection model.

[0008] As a preferred scheme, the step of inputting the feature map and the preprocessed image data into a preset target detection model to obtain a preliminary labeling result, performing cluster analysis on the preliminary labeling result through a clustering algorithm, and obtaining multiple cluster centers and corresponding candidate target sets includes: inputting the feature map and the preprocessed image data into the target detection model to obtain a preliminary labeling result and the confidence of each label in the preliminary labeling result; using a Gaussian mixture model to estimate the position data and category probability of the preliminary labeling result to obtain the position data and category probability, applying a non-maximum suppression algorithm to filter the position data to obtain a filtered detection result, combining the filtered detection result with the confidence to obtain a detection set; grouping the detection set based on spatial proximity and category similarity through the DBSCAN algorithm to obtain several clusters, determining the cluster center of each cluster by calculating the centroid, and obtaining multiple cluster centers; applying the K-means algorithm to identify the candidate target set corresponding to each cluster center to obtain the candidate target set corresponding to each cluster center.

[0009] As a preferred scheme, the step of classifying the candidate target set based on the cluster center to obtain a classification labeling result, calculating the confidence of each label in the classification labeling result, screening the confidence according to a preset confidence threshold, and taking the confidence above the threshold as high-confidence labeling data and the confidence below the threshold as low-confidence labeling data includes: assigning the candidate targets in the candidate target set to corresponding categories based on the cluster center to generate a classification labeling result, calculating the confidence of each label in the classification labeling result by Bayesian reasoning, and analyzing the feature diversity in each candidate target set by using information gain; adjusting the confidence by using the feature diversity to obtain an adjusted confidence, comparing the adjusted confidence with a preset confidence threshold, taking the confidence with a score higher than the confidence threshold as the initial high-confidence labeling data, and taking the confidence with a score lower than the confidence threshold as the initial low-confidence labeling data. confidence labeling data; identifying the overlapping area between the initial high-confidence labeling data and the initial low-confidence labeling data by the Dice coefficient, generating overlapping data, weighting the overlapping data by the entropy weight method, obtaining associated confidence, evaluating the associated confidence according to the F1 value, updating the initial high-confidence labeling data and the initial low-confidence labeling data based on the evaluation result, obtaining updated confidence labeling data and updated low-confidence labeling data; applying the Pearson correlation coefficient to evaluate the feature correlation between the updated confidence labeling data and the updated low-confidence labeling data, generating a correlation index, optimizing the adjusted confidence by the correlation index, obtaining an optimized confidence; comparing the optimized confidence with the confidence threshold, taking the confidence with a score higher than the confidence threshold as the high-confidence labeling data, and taking the confidence with a score lower than the confidence threshold as the low-confidence labeling data.

[0010] As a preferred scheme, the step of assigning the candidate targets in the candidate target set to corresponding categories based on the cluster center to generate classification labeling results includes: taking the cluster center as a reference point, using cosine similarity as a similarity measurement based on the reference point, calculating the similarity between the feature vector of each candidate target in the candidate target set and the feature vector of the cluster center to obtain a similarity measurement result, and according to the similarity measurement result, assigning each candidate target in the candidate target set to the category where the most similar cluster center is located to generate classification labeling results.

[0011] As a preferred scheme, the step of reviewing the low-confidence labeled data to obtain a reviewed labeled data set, inputting the reviewed labeled data set into a preset consistency check model for consistency check, outputting consistent labeled data, and feeding back the consistent labeled data and the high-confidence labeled data to the target detection model to obtain an optimized target detection model includes: manually reviewing the low-confidence labeled data by multiple experts to obtain a reviewed labeled data set, inputting the reviewed labeled data set into a preset consistency check model, parsing and comparing the reviewed labeled data set in the consistency check model using logistic regression analysis, and outputting consistent labeled data; merging the consistent labeled data with the high-confidence labeled data to form an enhanced labeled data set, feeding back the enhanced labeled data set to the target detection model, retraining the target detection model, and obtaining an optimized target detection model.

[0012] As a preferred scheme, the step of merging the high-confidence annotated data with the re-annotated data to generate a complete annotated training data set, inputting the complete annotated training data set into a preset artificial intelligence visual recognition model, performing model training, and obtaining a trained artificial intelligence visual recognition model includes: merging the re-annotated data with the high-confidence annotated data to generate a complete annotated training data set, using a preset generalization method to perform data enhancement on minority class samples in the complete annotated training data set to generate enhanced data, and integrating the enhanced data into the complete annotated training data set to obtain an enhanced training data set; wherein the generalization method includes any one of the following: rotation, flipping, scaling, cropping, and color transformation; inputting the enhanced training data set into a preset artificial intelligence visual recognition model for training to obtain an intermediate training model, verifying the intermediate training model using a preset verification data set to obtain performance indicators, and selecting the intermediate training model with the best performance from the performance indicators as the trained artificial intelligence visual recognition model.

[0013] The present application also provides a training data set annotation device for AI visual recognition, including: an acquisition module, used to acquire the original image data of the target and perform preprocessing to obtain preprocessed image data, perform feature extraction on the preprocessed image data, and obtain a feature map; an analysis module, used to input the feature map and the preprocessed image data into a preset target detection model to obtain a preliminary annotation result, perform cluster analysis on the preliminary annotation result through a clustering algorithm to obtain multiple cluster centers and corresponding candidate target sets; a classification module, used to classify the candidate target set based on the cluster centers to obtain a classified annotation result, calculate the confidence of each annotation in the classified annotation result, screen the confidence according to a preset confidence threshold, and use the confidence higher than the threshold as high-confidence annotation data , the confidence that is lower than the threshold is used as low-confidence annotation data; a review module is used to review the low-confidence annotation data to obtain a reviewed annotation data set, input the reviewed annotation data set into a preset consistency check model for consistency check, output consistency annotation data, and feed back the consistency annotation data and the high-confidence annotation data to the target detection model to obtain an optimized target detection model; a training module is used to re-annotate the low-confidence annotation data using the optimized target detection model to obtain re-annotated data, merge the high-confidence annotation data with the re-annotated data to generate a complete annotation training data set, input the complete annotation training data set into a preset artificial intelligence visual recognition model, perform model training, and obtain a trained artificial intelligence visual recognition model.

[0014] Compared with the prior art, the present application has the following beneficial effects: accurate labeling and high consistency. By acquiring and preprocessing the original image data, extracting the feature map, and inputting the target detection model to obtain the preliminary labeling results, the candidate target set is determined by combining cluster analysis. Based on confidence screening, the high-confidence labeled data is directly used for training, and the low-confidence data is optimized and fed back to the target detection model after review and consistency check to improve the model performance. The low-confidence data is re-labeled using the optimized model to generate a complete labeled training set for artificial intelligence visual recognition model training, thereby improving the quality and efficiency of labeling, ensuring the consistency and confidence of the data, and ensuring the confidence and consistency of the labeled data, thereby effectively improving the performance and stability of the visual recognition model, and improving the existing methods when dealing with complex scenes and multi-target recognition. There are still problems with inconsistent labeling, missing labels, and mislabeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0016] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0017] Figure 1 It is a flowchart of a method for labeling a training data set for AI visual recognition provided by an embodiment of the present invention;

[0018] Figure 2 It is a schematic block diagram of the structure of a training data set annotation device for AI visual recognition provided by an embodiment of the present invention.

[0019] Description of reference numerals:

[0020] 10. A training data set annotation device for AI visual recognition; 11. An acquisition module; 12. An analysis module; 13. A classification module; 14. A review module; 15. A training module. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0023] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0024] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0025] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0026] Embodiment 1:

[0027] like Figure 1 As shown, a training data set annotation method for AI visual recognition provided in an embodiment of the present application includes steps S100 to S500.

[0028] Step S100: Acquire original image data of the target and perform preprocessing to obtain preprocessed image data, perform feature extraction on the preprocessed image data to obtain feature mapping.

[0029] In this step, the original image data is first processed using image enhancement technology to improve the contrast and clarity of the image; specifically, the noise and low-contrast areas in the image are processed through histogram equalization and denoising algorithms to obtain preprocessed image data. Subsequently, the convolutional neural network (CNN) is used to extract features from the preprocessed image data to obtain the feature map of the image.

[0030] For example, by using the ResNet-50 network architecture, high-level features of the image can be extracted and input as feature maps to the next step of processing.

[0031] Step S200: input the feature map and the preprocessed image data into a preset target detection model to obtain a preliminary labeling result, and perform cluster analysis on the preliminary labeling result through a clustering algorithm to obtain multiple cluster centers and corresponding candidate target sets.

[0032] In this step, the preset target detection model can use the Faster R-CNN model for preliminary target detection; specifically, the feature map is input into Faster R-CNN to obtain preliminary annotation results, including the target's bounding box and category information. Subsequently, the preliminary annotation results are clustered using the K-means clustering algorithm to determine multiple cluster centers and generate a corresponding set of candidate targets.

[0033] For example, by using K-means clustering, similar targets can be clustered together to obtain multiple cluster centers for subsequent classification and labeling.

[0034] Step S300: classify the candidate target set based on the cluster center to obtain the classification labeling results, calculate the confidence of each label in the classification labeling results, filter the confidence according to the preset confidence threshold, and use the confidence higher than the threshold as high-confidence labeling data, and the confidence lower than the threshold as low-confidence labeling data.

[0035] In this step, the cluster centers are first classified using a support vector machine (SVM) to obtain the classification and annotation results; specifically, the candidate target set is classified through the SVM model, the category label of each target is output, and the confidence of the classification result is calculated. Then, the confidence in the classification and annotation results is screened according to the preset confidence threshold, and the confidence above the threshold is used as high-confidence annotation data, while the confidence below the threshold is used as low-confidence annotation data.

[0036] For example, if the preset confidence threshold is 0.8, the labeled data with a confidence higher than 0.8 is considered as high-confidence labeled data, and the labeled data with a confidence lower than 0.8 is considered as low-confidence labeled data.

[0037] Step S400: review the low-confidence labeled data to obtain a reviewed labeled data set, input the reviewed labeled data set into a preset consistency check model for consistency check, output consistent labeled data, feed the consistent labeled data and high-confidence labeled data back to the target detection model to obtain an optimized target detection model.

[0038] In this step, low-confidence labeled data needs to be manually reviewed to ensure the accuracy of the labeling; specifically, the manual labeler reviews the low-confidence labeled data one by one, confirms and modifies the incorrect labels. The reviewed labeled data set is checked by the consistency check model to ensure the consistency of all labeled data and output the consistent labeled data. Finally, the consistent labeled data and high-confidence labeled data are fed back to the target detection model for model optimization.

[0039] For example, by introducing a manual review step, the accuracy of low-confidence annotated data can be greatly improved, thereby optimizing the object detection model.

[0040] Step S500: Use the optimized target detection model to re-label the low-confidence labeled data to obtain re-labeled data, merge the high-confidence labeled data with the re-labeled data to generate a complete labeled training data set, input the complete labeled training data set into the preset artificial intelligence visual recognition model, perform model training, and obtain a trained artificial intelligence visual recognition model.

[0041] In this step, the previously screened low-confidence annotated data is first re-annotated using the optimized target detection model to improve the accuracy of the annotation; specifically, the optimized target detection model can more accurately identify and annotate low-confidence data to generate re-annotated data. The high-confidence annotated data is merged with the re-annotated data to generate a complete annotated training data set. Finally, the complete annotated training data set is input into the preset artificial intelligence visual recognition model for model training to obtain a trained artificial intelligence visual recognition model.

[0042] For example, by merging high-confidence annotated data with re-annotated data, the completeness and accuracy of the training dataset is ensured, improving the performance of the visual recognition model.

[0043] In this embodiment, the original image data of the target is obtained and preprocessed to obtain the preprocessed image data. Then, the preprocessed image data is subjected to feature extraction to obtain a feature map. The feature map and the preprocessed image data are input into a preset target detection model to obtain a preliminary annotation result. The preliminary annotation result is clustered by a clustering algorithm to obtain a plurality of cluster centers and a corresponding candidate target set. The candidate target set is classified based on the cluster center to obtain a classification annotation result, and the confidence of each annotation in the classification annotation result is calculated. The confidence is screened according to a preset confidence threshold, and the confidence above the threshold is used as high-confidence annotation data, and the confidence below the threshold is used as low-confidence annotation data. The low-confidence annotation data is reviewed to obtain a reviewed annotation data set, and the reviewed annotation data set is input into a preset consistency check model for consistency check, and the consistent annotation data is output. Finally, the consistent annotation data and the high-confidence annotation data are fed back to the target detection model to obtain an optimized target detection model. The optimized target detection model is used to re-label the low-confidence labeled data to obtain the re-labeled data, and the high-confidence labeled data is merged with the re-labeled data to generate a complete labeled training data set. Finally, the complete labeled training data set is input into the preset artificial intelligence visual recognition model for model training to obtain a trained artificial intelligence visual recognition model.

[0044] Through preprocessing, feature extraction, cluster analysis and consistency check, an efficient and high-precision training data set annotation method is provided. While improving the annotation efficiency, it also significantly improves the quality and confidence of the annotated data, solves the problems of inconsistent annotation, missing annotations and mislabeling in the existing technology, ensures the integrity and reliability of the training data set, and thus provides a solid data foundation for improving the performance of artificial intelligence visual recognition models.

[0045] Embodiment 2:

[0046] In step S100, original image data is acquired through an image acquisition device, and metadata related to the original image data is acquired, the original image data is normalized according to MinMax normalization to obtain normalized image data, and a denoising filter is applied to the normalized image data to eliminate noise and obtain denoised image data.

[0047] The original image data of the target is acquired through an image acquisition device, such as a digital camera, a video camera or a professional industrial camera, and metadata related to the original image data, such as shooting time, shooting location and equipment parameters, etc. are simultaneously acquired. Specifically, the image pixel values ​​are normalized to the range of [0, 1] by applying the MinMax normalization algorithm to the original image data to reduce the impact of illumination changes on image quality, thereby obtaining normalized image data. Then, by applying a denoising filter, such as a Gaussian filter or a median filter, the noise in the normalized image data is processed and eliminated to obtain denoised image data.

[0048] For example, by using a digital camera to shoot a scene, high-resolution original image data is obtained, and metadata such as the shooting time and location are recorded; the pixel value is adjusted to between 0 and 1 through MinMax normalization, and then the noise in the image is eliminated through a Gaussian filter.

[0049] The original image data is color corrected by a color correction algorithm to obtain color corrected image data, the edge enhancement algorithm is applied to the color corrected image data to obtain enhanced image data, and the denoised image data and the enhanced image data are feature fused according to the feature fusion technology to obtain a feature set.

[0050] The original image data is color corrected by color correction algorithms, such as white balance adjustment and gamma correction, to correct color distortion and color cast, and obtain color corrected image data. Specifically, white balance adjustment can balance the color temperature of the image, making the image present a more natural color; gamma correction can adjust the brightness and contrast of the image. Then, edge enhancement algorithms, such as Laplacian filter or Sobel filter, are applied to enhance the edge features in the color corrected image data to obtain enhanced image data. According to the feature fusion technology, the denoised image data and the enhanced image data are feature fused, and the advantages of both are combined to obtain a feature set containing rich feature information.

[0051] For example, the yellowish or bluish tones in the image can be corrected through white balance adjustment, and the edge details can be highlighted through the Sobel filter. Finally, the features of the denoised image and the enhanced image can be fused to extract richer feature information.

[0052] The principal component analysis algorithm is used to reduce the dimension of the feature set to obtain combined feature data. The combined feature data is segmented into regions of interest based on preset standards to obtain segmentation maps and corresponding region descriptors. Convolutional neural networks are used to extract features from the segmentation maps and region descriptors to obtain refined feature maps and key point data. The feature matching algorithm is used to associate metadata with the refined feature maps and key point data to generate a feature map for input into the target detection model.

[0053] The principal component analysis (PCA) algorithm is used to reduce the dimension and redundancy of the feature set, retain the main feature information, and obtain the combined feature data. Specifically, the covariance matrix of the feature set is calculated, and the eigenvalue decomposition is performed, and the main eigenvectors are selected as the new basis to map the high-dimensional feature data to the low-dimensional space. Then, the combined feature data is segmented into regions of interest (ROI) based on preset standards (such as color, shape or texture features) to obtain segmentation maps and corresponding region descriptors. Convolutional neural networks (CNNs) are used to extract features from segmentation maps and region descriptors to obtain more refined feature maps and key point data. Finally, the feature matching algorithm is used to associate metadata with the refined feature map and key point data to generate feature maps for input into the target detection model.

[0054] For example, PCA is used to reduce the dimensionality of high-dimensional feature data containing color and texture information, retain the main components, and reduce computational complexity. Region segmentation is then performed using preset color standards, and ResNet is used to extract refined features and key point data, which are then associated with metadata.

[0055] In step S200, the feature map and the preprocessed image data are input into the target detection model to obtain a preliminary labeling result and the confidence of each label in the preliminary labeling result.

[0056] By inputting the feature map and preprocessed image data into a preset target detection model, such as FasterR-CNN, YOLO or SSD model, target detection processing is performed to obtain preliminary annotation results. Specifically, the target detection model generates preliminary annotation results, including the bounding box position, category label and confidence score of each target, by performing feature analysis on the input feature map.

[0057] For example, by inputting feature maps and preprocessed image data into the YOLO model, preliminary annotation results including target bounding boxes, category labels, and confidence scores are quickly generated.

[0058] The Gaussian mixture model is used to estimate the position data and category probability of the preliminary annotation results to obtain the position data and category probability. The non-maximum suppression algorithm is applied to filter the position data to obtain the filtered detection results. The filtered detection results are combined with the confidence level to obtain the detection set.

[0059] The Gaussian mixture model (GMM) is used to estimate the location data and category probability of the preliminary annotation results to improve the accuracy of target detection. Specifically, the spatial distribution and category distribution of the target are modeled by GMM to estimate the precise location data and category probability of each target. Then, the non-maximum suppression (NMS) algorithm is applied to filter the location data to eliminate redundant target detection boxes and obtain the filtered detection results. The filtered detection results are combined with the confidence score to generate a detection set.

[0060] For example, GMM is used to estimate the probability of the target position and category of the preliminary annotation results, NMS is applied to remove overlapping detection boxes, and the results with high confidence are retained to form a detection set.

[0061] The detection set is grouped based on spatial proximity and category similarity through the DBSCAN algorithm to obtain several clusters. The cluster center of each cluster is determined by calculating the centroid to obtain multiple cluster centers.

[0062] The DBSCAN (density-based clustering) algorithm is used to group the detection set based on the spatial proximity and category similarity of the targets in the detection set to obtain several clusters. Specifically, the DBSCAN algorithm aggregates targets that are adjacent to each other and have similar categories into a cluster by setting the neighborhood radius and the minimum number of points. Then, the centroid position of each cluster is calculated to determine the cluster center of each cluster, and multiple cluster centers are obtained.

[0063] For example, DBSCAN is used to identify objects that are close in space and belong to similar categories, and they are divided into a cluster. The cluster center is determined by calculating the centroid of each cluster.

[0064] The K-means algorithm is applied to identify the candidate target set corresponding to each cluster center, and the candidate target set corresponding to each cluster center is obtained.

[0065] By applying the K-means clustering algorithm, the targets in the detection set are further clustered to identify the candidate target set corresponding to each cluster center. Specifically, the target points in the detection set are used as input, the K-means algorithm is iterated, the distance between the target point and the cluster center is calculated, the target point is assigned to the nearest cluster center, and the candidate target set corresponding to each cluster center is obtained.

[0066] For example, the target points in the detection set are assigned to the nearest cluster center through the K-means algorithm to form a candidate target set for each cluster center.

[0067] In step S300, the candidate targets in the candidate target set are assigned to corresponding categories based on the cluster center, and a classification labeling result is generated. The confidence of each label in the classification labeling result is calculated by Bayesian reasoning, and the feature diversity in each candidate target set is analyzed using information gain.

[0068] Based on the cluster centers, the candidate targets in the candidate target set are assigned to the corresponding categories to generate classification labeling results. Specifically, based on the feature vector of each cluster center, the confidence of the labeling in each classification labeling result is calculated by the Bayesian reasoning method. Then, the feature diversity in the candidate target set is analyzed using the information gain algorithm to evaluate the feature differences between different candidate targets.

[0069] For example, the Bayesian reasoning method is used to calculate the confidence of each classification labeling result, and then the information gain is used to analyze the feature diversity of different candidate targets and evaluate the feature differences.

[0070] The confidence is adjusted using feature diversity to obtain the adjusted confidence, and the adjusted confidence is compared with the preset confidence threshold. The confidence with a score higher than the confidence threshold is used as the initial high-confidence annotation data, and the confidence with a score lower than the confidence threshold is used as the initial low-confidence annotation data.

[0071] By utilizing feature diversity, the confidence of each annotation is adjusted to obtain an adjusted confidence. Specifically, by analyzing the feature diversity in the candidate target set, the feature gain is calculated, and the annotation confidence is adjusted according to the feature gain. Then, the adjusted confidence is compared with the preset confidence threshold, and the annotation data with a confidence score higher than the threshold is used as the initial high-confidence annotation data, the annotation data with a confidence score lower than the threshold is used as the initial low-confidence annotation data, and the annotation data with a confidence score lower than the threshold is used as the initial low-confidence annotation data.

[0072] For example, by analyzing feature diversity, the confidence is adjusted according to the feature gain, and the confidence annotation data above the preset threshold (such as 0.8) is used as the initial high-confidence data.

[0073] The Dice coefficient is used to identify the overlapping area between the initial high-confidence labeled data and the initial low-confidence labeled data, and overlapping data is generated. The overlapping data is weighted using the entropy weight method to obtain the associated confidence. The associated confidence is evaluated according to the F1 value. Based on the evaluation results, the initial high-confidence labeled data and the initial low-confidence labeled data are updated to obtain updated confidence labeled data and updated low-confidence labeled data.

[0074] The Dice coefficient is used to identify the overlapping area between the initial high-confidence annotated data and the initial low-confidence annotated data, and the overlapping data is generated. Specifically, the Dice coefficient is used to measure the similarity between two data sets, and the overlapping part of the data is determined by calculating the overlapping area. The overlapping data is weighted using the entropy weight method to obtain the associated confidence. Then, the associated confidence is evaluated according to the F1 value, and the recall rate and precision rate are combined to determine the final confidence score. Based on the evaluation results, the initial high-confidence annotated data and the initial low-confidence annotated data are updated to obtain the updated confidence annotated data and the updated low-confidence annotated data.

[0075] For example, the Dice coefficient is used to measure the overlap of high-confidence and low-confidence data, the overlapping data is weighted by the entropy weight method, and the F1 value is used to evaluate the results.

[0076] The Pearson correlation coefficient is applied to evaluate the feature correlation between the updated confidence labeled data and the updated low confidence labeled data, and a correlation index is generated. The adjusted confidence is optimized using the correlation index to obtain the optimized confidence.

[0077] By applying the Pearson correlation coefficient, the feature correlation between the updated confidence annotated data and the updated low confidence annotated data is evaluated to generate a correlation index. Specifically, the Pearson correlation coefficient is used to measure the degree of linear correlation between two variables. By calculating the correlation coefficient, the correlation index between the features is obtained. Using this correlation index, the adjusted confidence is further optimized to obtain the optimized confidence.

[0078] For example, the linear correlation between two data sets is analyzed through the Pearson correlation coefficient, a correlation index is generated, and the confidence level is optimized.

[0079] The optimized confidence score is compared with the confidence threshold, and the confidence score with a score higher than the confidence threshold is taken as high-confidence annotation data, and the confidence score with a score lower than the confidence threshold is taken as low-confidence annotation data.

[0080] By comparing the optimized confidence with the preset confidence threshold, the confidence annotation data with a score higher than the confidence threshold is regarded as high-confidence annotation data, and the confidence annotation data with a score lower than the confidence threshold is regarded as low-confidence annotation data. Specifically, by comparing with the confidence threshold, the high-confidence and low-confidence annotation data are screened out, and further processed according to the classification results.

[0081] For example, the optimized confidence level is compared with a preset threshold (such as 0.85) to filter out high-confidence labeled data and low-confidence labeled data.

[0082] Among them, the step of assigning candidate targets in the candidate target set to corresponding categories based on the cluster center and generating classification labeling results includes: taking the cluster center as a reference point, using cosine similarity as a similarity measurement based on the reference point, calculating the similarity between the feature vector of each candidate target in the candidate target set and the feature vector of the cluster center, and obtaining a similarity measurement result.

[0083] Specifically, it is achieved through the following steps:

[0084] Set cluster centers as reference points: Set the feature vector of each cluster center as the reference point of the corresponding category. The selection of reference points is based on the cluster centers that have been determined by the K-means algorithm in the early stage. These center points represent the typical characteristics of each category.

[0085] Compare the feature vector of each target in the candidate target set with the feature vectors of all cluster centers and calculate the cosine similarity between them. The cosine similarity formula is:

[0086]

[0087] in, and Represents the feature vector of the candidate target and the feature vector of the cluster center, dot product Compute the inner product of the eigenvectors, and are the modulus lengths of the eigenvectors.

[0088] By calculating the cosine similarity between the candidate target and each cluster center, the similarity measurement result corresponding to each candidate target is obtained. The similarity measurement result indicates the similarity between the candidate target and different cluster centers.

[0089] According to the calculated similarity measurement results, each candidate target in the candidate target set is assigned to the category where the cluster center with the highest similarity is located. In this way, it can ensure that the candidate target is assigned to the most matching category and improve the accuracy of classification labeling.

[0090] For example, by calculating the cosine similarity between the feature vector of a candidate target and the feature vectors of three cluster centers, assuming that its similarities with cluster centers A, B, and C are 0.85, 0.75, and 0.90 respectively, the candidate target will be assigned to the category where cluster center C is located because its similarity is the highest.

[0091] According to the similarity measurement results, each candidate target in the candidate target set is assigned to the category where the most similar cluster center is located, and the classification labeling results are generated.

[0092] By taking the cluster center as the reference point and using cosine similarity as the similarity metric based on the reference point, the feature vector of each candidate target in the candidate target set is similar to the feature vector of the cluster center to obtain the similarity metric result. Specifically, cosine similarity is used to measure the angle between two vectors to determine their similarity. According to the similarity metric result, each candidate target in the candidate target set is assigned to the category where the most similar cluster center is located, and the classification labeling result is generated.

[0093] In step S400, multiple experts manually review the low-confidence annotation data to obtain a reviewed annotation data set, which is then input into a preset consistency check model. In the consistency check model, logistic regression analysis is used to parse and compare the reviewed annotation data set to output consistent annotation data.

[0094] The low-confidence annotation data is manually reviewed by multiple experts to ensure the accuracy and consistency of the annotation data. Specifically, multiple experts independently review the low-confidence annotation data and summarize the review results to obtain the reviewed annotation data set. Then, the reviewed annotation data set is input into the preset consistency check model for consistency check. In the consistency check model, logistic regression analysis is used to parse and compare the reviewed annotation data set, and the consistent annotation data is output.

[0095] For example, low-confidence data can be independently reviewed by multiple professional annotators, and a logistic regression model can be applied to check data consistency and output consistent annotated data.

[0096] The consistent labeled data and the high-confidence labeled data are merged to form an enhanced labeled data set, and the enhanced labeled data set is fed back to the target detection model, and the target detection model is retrained to obtain an optimized target detection model.

[0097] By merging the consistent annotated data with the high-confidence annotated data, an enhanced annotated data set is formed. Specifically, the two types of annotated data are combined to supplement and improve the diversity and completeness of the data set. Then, the enhanced annotated data set is fed back to the target detection model, the target detection model is retrained, and the performance of the model is further optimized to obtain an optimized target detection model.

[0098] For example, the consistently labeled data can be merged with the high-confidence data to form a more comprehensive labeled dataset, and the object detection model can be retrained to optimize performance.

[0099] In step S500, the re-annotated data is merged with the high-confidence annotated data to generate a complete annotated training data set, and a preset generalization method is used to perform data enhancement on the minority class samples in the complete annotated training data set to generate enhanced data, and the enhanced data is integrated into the complete annotated training data set to obtain an enhanced training data set; wherein the generalization method includes any one of the following: rotation, flipping, scaling, cropping, and color transformation.

[0100] By merging the re-labeled data with the high-confidence labeled data, a complete labeled training data set is generated. Specifically, the complete labeled training data set includes all high-confidence labeled data and the verified low-confidence labeled data. Then, the minority class samples in the complete labeled training data set are enhanced using a preset generalization method to increase the diversity and representativeness of the data and generate enhanced data. The enhanced data is integrated into the complete labeled training data set to obtain an enhanced training data set. Among them, the generalization method can include a variety of technical means such as rotation, flipping, scaling, cropping and color transformation.

[0101] For example, by rotating, flipping, and color-changing minority class samples, more varied sample data can be generated to enhance data diversity.

[0102] The enhanced training data set is input into the preset artificial intelligence visual recognition model for training to obtain an intermediate training model. The intermediate training model is verified using the preset verification data set to obtain performance indicators. The intermediate training model with the best performance is selected from the performance indicators as the trained artificial intelligence visual recognition model.

[0103] The enhanced training data set is input into the preset artificial intelligence visual recognition model for training to improve the recognition ability and generalization ability of the model. Specifically, the enhanced training data set is input into the model, and multiple rounds of training are performed to obtain an intermediate training model. Then, the intermediate training model is verified using a preset verification data set, the performance of the model is evaluated, and various performance indicators are obtained. According to the performance indicators, the best performing model is selected from the intermediate training models as the final trained artificial intelligence visual recognition model.

[0104] For example, by using a validation dataset to validate the model, you can ensure that the final model performs well in real-world environments.

[0105] In this embodiment, the original image data is normalized by using MinMax normalization, and the noise is eliminated by using a noise reduction filter to obtain noise-reduced image data. Then, the image data is color-corrected by using a color correction algorithm, and the edge enhancement algorithm is used to enhance the edge. The noise-reduced image data and the enhanced image data are feature-fused by using a feature fusion technique to obtain a feature set. Then, the feature set is dimensionally reduced by using a principal component analysis algorithm, the region of interest is segmented based on a preset standard, a convolutional neural network is used to extract features, and metadata is associated with a refined feature map and key point data to generate a feature map for input into a target detection model.

[0106] In the target detection process, the feature map and preprocessed image data are input into the target detection model, and the Gaussian mixture model and non-maximum suppression algorithm are used to estimate and filter the location data and category probability of the preliminary annotation results. The detection set is grouped by the DBSCAN algorithm, and the candidate target set is identified by the K-means algorithm. Candidate targets are assigned based on cluster centers, and classification annotation results are generated. The confidence and feature diversity of the classification annotation results are calculated by Bayesian reasoning and information gain. The Dice coefficient and entropy weight method are used to identify overlapping areas, evaluate and update the confidence annotation data. The updated confidence data is evaluated by the Pearson correlation coefficient to obtain the optimized confidence data, and the confidence data is screened.

[0107] During the review and optimization process, multiple experts manually review the low-confidence annotated data, input the reviewed annotated data set into the consistency check model, and output the consistent annotated data. The consistent annotated data is merged with the high-confidence annotated data to form an enhanced annotated data set, and the target detection model is retrained. The re-annotated data is merged with the high-confidence annotated data to generate a complete annotated training data set, and the data is enhanced through the generalization method. Finally, the enhanced training data set is input into the artificial intelligence visual recognition model for training and verification, and an artificial intelligence visual recognition model with excellent performance is obtained. Through the above method, the annotation quality and efficiency of the training data set can be significantly improved, the confidence and consistency of the annotated data can be ensured, and the performance and stability of the visual recognition model can be effectively improved.

[0108] Embodiment 3:

[0109] like Figure 2 As shown, the present application also provides a training data set annotation device 10 for AI visual recognition, including an acquisition module 11, an analysis module 12, a classification module 13, a review module 14 and a training module 15.

[0110] The acquisition module 11 is mainly used to acquire the original image data of the target and perform preprocessing to obtain the preprocessed image data, and perform feature extraction on the preprocessed image data to obtain a feature map.

[0111] The analysis module 12 is mainly used to input the feature map and preprocessed image data into a preset target detection model to obtain a preliminary labeling result, and perform cluster analysis on the preliminary labeling result through a clustering algorithm to obtain multiple cluster centers and corresponding candidate target sets.

[0112] The classification module 13 is mainly used to classify the candidate target set based on the cluster center, obtain the classification labeling results, calculate the confidence of each label in the classification labeling results, filter the confidence according to the preset confidence threshold, and use the confidence above the threshold as high-confidence labeling data, and the confidence below the threshold as low-confidence labeling data.

[0113] The review module 14 is mainly used to review the low-confidence labeled data to obtain a reviewed labeled data set, input the reviewed labeled data set into a preset consistency check model for consistency check, output consistent labeled data, and feed back the consistent labeled data and high-confidence labeled data to the target detection model to obtain an optimized target detection model.

[0114] The training module 15 is mainly used to use the optimized target detection model to re-label the low-confidence labeled data to obtain the re-labeled data, merge the high-confidence labeled data with the re-labeled data to generate a complete labeled training data set, input the complete labeled training data set into the preset artificial intelligence visual recognition model, perform model training, and obtain a trained artificial intelligence visual recognition model.

[0115] In this embodiment, the original image data of the target is obtained by the acquisition module 11, and preprocessed, including normalization, noise elimination, color correction and edge enhancement, so as to obtain high-quality preprocessed image data. These image data are extracted to obtain detailed feature maps, providing rich information for subsequent analysis.

[0116] The analysis module 12 receives feature mapping and preprocessed image data, generates preliminary annotation results using a preset target detection model, and performs cluster analysis on the preliminary annotation results through a clustering algorithm to obtain multiple cluster centers and corresponding candidate target sets, thereby ensuring accurate positioning and classification of targets in the image.

[0117] The classification module 13 classifies the candidate target set in detail based on the cluster center, generates classification annotation results, and calculates the confidence of each annotation. Through the preset confidence threshold screening, the confidence annotation data above the threshold is used as high-confidence annotation data, and the annotation data below the threshold is used as low-confidence annotation data to ensure data quality.

[0118] The review module 14 manually reviews the low-confidence labeled data, generates a reviewed labeled data set, and checks it through a consistency check model to output consistent labeled data. These consistent labeled data and high-confidence labeled data are fed back to the target detection model to optimize model performance and improve detection accuracy.

[0119] The training module 15 uses the optimized target detection model to re-label the low-confidence labeled data and generate accurate re-labeled data. The high-confidence labeled data is merged with the re-labeled data to generate a complete labeled training data set. The data set is further input into the preset artificial intelligence visual recognition model for comprehensive model training, and finally a trained artificial intelligence visual recognition model is obtained. The entire process is modularly designed, which not only improves the efficiency and accuracy of data labeling, but also optimizes the detection performance and stability of the model, ensuring the successful implementation of the visual recognition task.

[0120] It should be noted that technicians in the relevant technical field can clearly understand that for the convenience and conciseness of description, the specific working process of the above-described device and each module can refer to the corresponding process in the aforementioned embodiment of the training data set annotation method for AI visual recognition, and will not be repeated here.

[0121] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training data set annotation method for AI visual recognition, characterized in that: include: Acquire original image data of the target and perform preprocessing to obtain preprocessed image data, and perform feature extraction on the preprocessed image data to obtain a feature map; Inputting the feature map and the preprocessed image data into a preset target detection model to obtain a preliminary annotation result, and performing cluster analysis on the preliminary annotation result by a clustering algorithm to obtain a plurality of cluster centers and corresponding candidate target sets; Taking the cluster center as a reference point, using cosine similarity as a similarity measurement based on the reference point, calculating the similarity between the feature vector of each candidate target in the candidate target set and the feature vector of the cluster center to obtain a similarity measurement result, assigning each candidate target in the candidate target set to the category where the most similar cluster center is located according to the similarity measurement result, generating a classification labeling result, calculating the confidence of each label in the classification labeling result by Bayesian reasoning, and analyzing the feature diversity in each candidate target set by using information gain; adjusting the confidence by using the feature diversity to obtain an adjusted confidence, comparing the adjusted confidence with a preset confidence threshold, taking the confidence with a score higher than the confidence threshold as initial high-confidence labeling data, and taking the confidence with a score lower than the confidence threshold as initial low-confidence labeling data; Identify the overlapping area between the initial high-confidence labeled data and the initial low-confidence labeled data by using the Dice coefficient, generate overlapping data, perform weighted processing on the overlapping data by using the entropy weight method to obtain associated confidence, evaluate the associated confidence according to the F1 value, and update the initial high-confidence labeled data and the initial low-confidence labeled data based on the evaluation result to obtain updated high-confidence labeled data and updated low-confidence labeled data; Applying the Pearson correlation coefficient to evaluate the feature correlation between the updated high-confidence annotated data and the updated low-confidence annotated data, generating a correlation index, optimizing the adjusted confidence using the correlation index to obtain an optimized confidence; comparing the optimized confidence with the confidence threshold, taking the confidence with a score higher than the confidence threshold as the high-confidence annotated data, and taking the confidence with a score lower than the confidence threshold as the low-confidence annotated data; The low-confidence labeled data is reviewed to obtain a reviewed labeled data set, the reviewed labeled data set is input into a preset consistency check model for consistency check, and consistent labeled data is output, and the consistent labeled data and the high-confidence labeled data are fed back to the target detection model to obtain an optimized target detection model; The low-confidence labeled data is re-labeled using the optimized target detection model to obtain re-labeled data, the high-confidence labeled data is merged with the re-labeled data to generate a complete labeled training data set, and the complete labeled training data set is input into a preset artificial intelligence visual recognition model to perform model training to obtain a trained artificial intelligence visual recognition model.

2. The training data set annotation method for AI visual recognition according to claim 1, characterized in that: The steps of acquiring the original image data of the target and preprocessing it to obtain the preprocessed image data, and extracting features from the preprocessed image data to obtain feature mapping include: Acquire original image data through an image acquisition device, and acquire metadata related to the original image data, perform normalization processing on the original image data according to MinMax normalization to obtain normalized image data, and apply a noise reduction filter to eliminate noise from the normalized image data to obtain noise-reduced image data; Performing color correction processing on the original image data by using a color correction algorithm to obtain color corrected image data, performing edge enhancement on the color corrected image data by using an edge enhancement algorithm to obtain enhanced image data, and performing feature fusion on the denoised image data and the enhanced image data according to a feature fusion technology to obtain a feature set; The principal component analysis algorithm is used to perform dimensionality reduction processing on the feature set to obtain combined feature data, the combined feature data is segmented into regions of interest based on preset standards to obtain segmentation maps and corresponding region descriptors, a convolutional neural network is used to perform feature extraction on the segmentation maps and the region descriptors to obtain refined feature maps and key point data, the metadata is associated with the refined feature maps and the key point data using a feature matching algorithm, and a feature map is generated for input into a target detection model.

3. The training data set annotation method for AI visual recognition according to claim 1, characterized in that: The step of inputting the feature map and the preprocessed image data into a preset target detection model to obtain a preliminary labeling result, and performing cluster analysis on the preliminary labeling result by a clustering algorithm to obtain a plurality of cluster centers and corresponding candidate target sets includes: Inputting the feature map and the preprocessed image data into a target detection model to obtain a preliminary labeling result and a confidence level of each label in the preliminary labeling result; Using a Gaussian mixture model to estimate the position data and category probability of the preliminary labeling results, obtaining the position data and category probability, applying a non-maximum suppression algorithm to filter the position data to obtain a filtered detection result, and combining the filtered detection result with the confidence of each label in the preliminary labeling results to obtain a detection set; The detection set is grouped based on spatial proximity and category similarity by using a DBSCAN algorithm to obtain a plurality of clusters, and a cluster center of each cluster is determined by calculating a centroid to obtain a plurality of cluster centers; The K-means algorithm is applied to identify the candidate target set corresponding to each cluster center, and the candidate target set corresponding to each cluster center is obtained.

4. The method for labeling training data sets for AI visual recognition according to claim 1, characterized in that: The step of reviewing the low-confidence labeled data to obtain a reviewed labeled data set, inputting the reviewed labeled data set into a preset consistency check model for consistency check, outputting consistent labeled data, and feeding back the consistent labeled data and the high-confidence labeled data to the target detection model to obtain an optimized target detection model includes: Manually reviewing the low-confidence annotated data by multiple experts to obtain an annotated data set after review, inputting the annotated data set after review into a preset consistency check model, and using logistic regression analysis in the consistency check model to parse and compare the annotated data set after review, and output consistent annotated data; The consistent labeled data and the high-confidence labeled data are merged to form an enhanced labeled data set, the enhanced labeled data set is fed back to the target detection model, the target detection model is retrained, and an optimized target detection model is obtained.

5. The method for labeling training data sets for AI visual recognition according to claim 1, characterized in that: The step of merging the high-confidence annotated data with the re-annotated data to generate a complete annotated training data set, inputting the complete annotated training data set into a preset artificial intelligence visual recognition model, performing model training, and obtaining a trained artificial intelligence visual recognition model includes: The re-annotated data is merged with the high-confidence annotated data to generate a complete annotated training data set, and a preset generalization method is used to perform data enhancement on minority class samples in the complete annotated training data set to generate enhanced data, and the enhanced data is integrated into the complete annotated training data set to obtain an enhanced training data set; wherein the generalization method includes any one of the following: rotation, flipping, scaling, cropping, and color transformation; The enhanced training data set is input into a preset artificial intelligence visual recognition model for training to obtain an intermediate training model, the intermediate training model is verified using a preset verification data set to obtain performance indicators, and the intermediate training model with the best performance is selected from the performance indicators as the trained artificial intelligence visual recognition model.

6. A training data set annotation device for AI visual recognition, characterized in that: include: An acquisition module is used to acquire the original image data of the target and perform preprocessing to obtain preprocessed image data, and perform feature extraction on the preprocessed image data to obtain a feature map; An analysis module, used for inputting the feature map and the preprocessed image data into a preset target detection model to obtain a preliminary annotation result, and performing cluster analysis on the preliminary annotation result by a clustering algorithm to obtain a plurality of cluster centers and corresponding candidate target sets; A classification module, used to take the cluster center as a reference point, use cosine similarity as a similarity measurement based on the reference point, calculate the similarity between the feature vector of each candidate target in the candidate target set and the feature vector of the cluster center, obtain a similarity measurement result, assign each candidate target in the candidate target set to the category where the most similar cluster center is located according to the similarity measurement result, generate a classification labeling result, calculate the confidence of each label in the classification labeling result by Bayesian reasoning, and use information gain to analyze the feature diversity in each candidate target set; use the feature diversity to adjust the confidence to obtain an adjusted confidence, compare the adjusted confidence with a preset confidence threshold, use the confidence with a score higher than the confidence threshold as initial high-confidence labeling data, and use the confidence with a score lower than the confidence threshold as initial low-confidence labeling data; Identify the overlapping area between the initial high-confidence labeled data and the initial low-confidence labeled data by using the Dice coefficient, generate overlapping data, perform weighted processing on the overlapping data by using the entropy weight method to obtain associated confidence, evaluate the associated confidence according to the F1 value, and update the initial high-confidence labeled data and the initial low-confidence labeled data based on the evaluation result to obtain updated high-confidence labeled data and updated low-confidence labeled data; Applying the Pearson correlation coefficient to evaluate the feature correlation between the updated high-confidence annotated data and the updated low-confidence annotated data, generating a correlation index, optimizing the adjusted confidence using the correlation index to obtain an optimized confidence; comparing the optimized confidence with the confidence threshold, taking the confidence with a score higher than the confidence threshold as the high-confidence annotated data, and taking the confidence with a score lower than the confidence threshold as the low-confidence annotated data; A review module is used to review the low-confidence labeled data to obtain a reviewed labeled data set, input the reviewed labeled data set into a preset consistency check model for consistency check, output consistent labeled data, and feed the consistent labeled data and the high-confidence labeled data back to the target detection model to obtain an optimized target detection model; A training module is used to use the optimized target detection model to re-label the low-confidence labeled data to obtain re-labeled data, merge the high-confidence labeled data with the re-labeled data to generate a complete labeled training data set, input the complete labeled training data set into a preset artificial intelligence visual recognition model, perform model training, and obtain a trained artificial intelligence visual recognition model.

Images