An active target detection method integrating classification and regression probability distribution
By improving the YOLOX object detection network structure and loss function, combined with active learning strategies of classification and regression probability distribution, the problem of regression uncertainty in the active object detection algorithm is solved, and the effect of reducing data annotation and improving model performance is achieved.
Patent Information
- Application Number
- CN202211553405.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-06
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2042-12-06
AI Technical Summary
The existing active object detection algorithm ignores the regression uncertainty of the target in the image in the object detection task, resulting in the inability to effectively select unlabeled samples with rich regression information, increasing the data annotation and cost.
Active object detection methods that integrate fusion classification and regression probability distribution are adopted. By improving the YOLOX target detection network structure and loss function, the probability distribution of the output feature map of the target detection network is estimated, the classification and regression uncertainty of the target in the image is considered, and the active learning sampling strategy is designed, and unlabeled samples with rich information in classification and regression are selected to reduce the amount of data labeling.
It effectively reduces the amount of data labeling required for the target detection task, saves labeling costs, and improves model performance.
Smart Images
Figure CN116051803B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of computer technology and proposes a target detection algorithm based on YOLOX, Gaussian YOLOX, and particularly a target detection method that estimates the probability distribution of Gaussian YOLOX output feature maps to reduce labeling requirements. Background Art
[0002] In the field of computer vision, object detection is the basis of many tasks, such as autonomous driving, pedestrian recognition, security systems, and face detection.
[0003] However, achieving high-performance object detection networks requires the support of thousands of labeled datasets. Unlike image classification tasks, object detection tasks require more complex labeling. Labeling thousands of object detection datasets is not only labor-intensive and time-consuming, but repetitive labeling inevitably leads to incorrect labeling. Therefore, in recent years, researchers have attempted to combine deep learning and active learning, a method called deep active learning.
[0004] Active learning aims to achieve high-performance networks using a small amount of richly labeled data. Active learning allows machine learning models to leverage their existing knowledge to proactively select rich, unlabeled data from the unlabeled dataset. This reduces the amount of labeling required while achieving high-performance networks.
[0005] Currently, research on active learning mainly focuses on image classification tasks, and there is relatively little research on target detection tasks. However, target detection tasks are the basis of many computer vision tasks, so research on target detection algorithms based on active learning is more meaningful.
[0006] Active learning aims to select information-rich samples from a large number of unlabeled samples for labeling to improve the detection performance of the network. Currently, the most commonly used technology for active learning is the sampling strategy based on uncertainty.
[0007] The most similar existing implementation to the present invention is an object detection method based on active adaptive learning to reduce annotation requirements. By combining active learning with domain adaptive learning, a new active domain adaptive learning strategy is proposed, which significantly reduces the amount of data annotation required for object detection tasks, saving annotation costs. A new sampling strategy is designed based on the global and local feature alignment process during domain adaptation, taking into account both image-level and object instance-level features.
[0008] Existing technologies only consider the classification uncertainty of objects in images. They do not consider the regression uncertainty of objects in images. Consequently, they can only select unlabeled samples with rich classification information for target detection tasks, while ignoring unlabeled samples with rich regression information. The present invention considers not only the classification uncertainty of objects in images but also the regression uncertainty of objects in images. This allows for the simultaneous selection of unlabeled samples with rich classification and regression information in images. Summary of the Invention
[0009] The present invention provides an active target detection method that integrates classification and regression probability distribution, which solves the problem that current active target detection algorithms can only select unlabeled samples with rich information based on the classification uncertainty of targets in images. Current active target detection algorithms ignore the regression uncertainty of targets in images and ignore many unlabeled samples with rich regression information.
[0010] The technical solution adopted is:
[0011] The present invention discloses an active target detection method that integrates classification and regression probability distribution, including: improvement of the YOLOX target detection network structure, rewriting of the YOLOX target detection network loss function, design of active learning sampling strategy and sampling method, expert labeling of selected unlabeled samples, retraining of the updated labeled data set to update the classification decision boundary of the detection network, and determining whether to perform iterative sampling in the next cycle based on whether the performance of the current optimal model on the test set meets the standard or whether the labeling budget is reached, and performing target detection using the optimal detection network obtained at last. The present invention proposes a new active target detection method by improving the network structure of the YOLOX target detection network and rewriting the loss function of the YOLOX target detection network. The method can estimate the probability distribution of the feature map output by the target detection network and simultaneously estimate the classification and regression uncertainty of the target in the image. The method can effectively select unlabeled samples with high classification and regression uncertainty (rich classification and regression information) in the image, greatly reducing the amount of data labeling required for the target detection task and saving labeling costs. The improvement of the YOLOX network structure and the rewriting of the loss function are conducive to improving the performance of the model.
[0012] Compared to the closest existing technology, this invention not only considers the classification uncertainty of objects in an image, but also the regression uncertainty of objects in the image. This helps select unlabeled samples that contain rich classification and regression information, and can estimate the uncertainty of unlabeled images through a single forward inference. Existing technologies can only select unlabeled samples that contain rich classification information.
[0013] The present invention proposes an active target detection method based on YOLOX, which can simultaneously consider the classification and regression uncertainties of the target. The existing invention only considers the classification uncertainty of the target and ignores the unlabeled samples with rich regression information.
[0014] Its advantages are:
[0015] This paper addresses the inability of current active target detection algorithms to effectively estimate the probability distribution of the detection network output feature map. Based on YOLOX, this paper proposes a target detection method that reduces the amount of data annotation by estimating the probability distribution of the detection network output feature map. This method considers not only the classification uncertainty of the target in the image, but also the regression uncertainty of the target in the image, which helps to select unlabeled samples with rich classification and regression information. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] Figure 1 It is a schematic diagram of the overall structure of the present invention.
[0017] Figure 2 It is a schematic diagram of the Gaussian YOLOX structure in the present invention.
[0018] Figure 3 This is a schematic diagram of the active learning cycle sampling of the present invention. DETAILED DESCRIPTION
[0019] An active target detection method integrating classification and regression probability distribution includes the following steps:
[0020] Step 1: Build the active target detection network model architecture.
[0021] Step 2, initial data preparation: randomly label the initial data set in the unlabeled data, and then label the test set.
[0022] Determine the number of unlabeled images to be sampled per active learning iteration, and then determine the initial threshold for foreground objects. The remaining unlabeled data is used for active learning iteration sampling.
[0023] Step 3, update or initialize the model: use the initial data set to initialize the model or use the existing labeled data to update the classification decision boundary of the detection network.
[0024] Step 4: Model Evaluation: Determine whether to continue active learning based on whether the current optimal model's performance on the test set meets the target or whether the active learning annotation budget has been met. If so, proceed to Step 5. Otherwise, use the optimal model for the current cycle for testing.
[0025] Step 5, model forward reasoning: The remaining unlabeled data is forward inferred through the optimal detection network in step 4 to obtain the regression variance, classification, regression and confidence results of each feature point in the detection network output feature map.
[0026] Step 6, regression uncertainty: Compare the classification and confidence results in step 5 with the foreground target threshold to obtain the foreground target information, and use the regression variance of the foreground target as the regression uncertainty of the foreground target.
[0027] Step 7, classification uncertainty: The classification uncertainty is obtained by subtracting the second highest classification score from the highest classification score of the foreground target in step 5.
[0028] Step 8, image uncertainty: The uncertainty of each foreground target in the unlabeled image is the sum of the regression uncertainty of the foreground target in step 6 and the classification uncertainty of the foreground target in step 7, and the uncertainty of the unlabeled image is the sum of the uncertainty of each foreground target in the unlabeled image.
[0029] Step 9, active labeling: select the first k unlabeled images with the largest uncertainty in labeling step 8, where k is obtained by the number of unlabeled images sampled per cycle in step 2.
[0030] Step 10, data update: Update the dataset and foreground target threshold, remove the images annotated in step 9 from the unannotated set, add the images annotated in step 9 to the training set and return to step 3 to continue active learning.
[0031] In the present invention, step 1 comprises:
[0032] In step 1-1, for the image feature extraction part of YOLOX (surpassing the YOLO series in 2021, YOLO (You Only Look Once)), CSPDarknet-53 (a new backbone that improves the learning ability of convolutional neural networks) is selected as the feature extraction backbone network.
[0033] In step 1-2, for the feature fusion part, the output of the last three convolutional blocks of the YOLOX backbone network is extracted for feature fusion, and PAFPN (Path Aggregation Network) is selected as the Feature Pyramid Network (FPN) for detecting neck feature fusion.
[0034] In steps 1-3, for the classification and regression prediction part, the mixture density network (MDN) is selected as the detection head of YOLOX, and the Gaussian distribution function and the binary cross entropy function are selected as the regression and classification loss functions respectively. The above improved YOLOX is renamed as Gaussian YOLOX.
[0035] The regression loss function of Gaussian YOLOX is:
[0036]
[0037] Where N represents SimOTA (Optimal Transmission Allocation for Object Detection), which is the number of positive and negative sample prediction boxes that match the feature points corresponding to the foreground target center in the output feature map. represents the Gaussian distribution function, k represents the number of regression mixture density networks, π reg Represents the weight of the Gaussian mixture model fitted by the regression mixture density network, μ iou Represents the mean of the intersection over union (IOU) between the predicted box and the real box, ∑ reg Represents the variance of the IOU between the predicted box and the true box, ε represents a small positive number to ensure the stability of the logarithmic value, and log is the natural logarithm.
[0038] The classification loss function of Gaussian YOLOX is:
[0039]
[0040] Where N represents SimOTA (optimal transmission allocation for object detection), which is the number of classified positive and negative samples matched with the feature points corresponding to the foreground target center in the output feature map, k represents the number of classification mixed density networks, and π cls Represents the weight of the Gaussian mixture model fitted by the classification mixture density network, y cls Represents the classification prediction target.
[0041] represents the Gaussian distribution function, μ cls represents the mean of the classification Gaussian mixture model, ∑ cls Represents the variance of the classification Gaussian mixture model, and log is the natural logarithm.
[0042] The confidence loss function of Gaussian YOLOX is:
[0043]
[0044] Where N represents SimOTA (optimal transmission allocation for object detection), which is the number of confidence positive and negative samples of feature point matching corresponding to the foreground target center in the output feature map, k represents the number of confidence mixed density networks, and π obj Represents the weight of the Gaussian mixture model fitted by the confidence mixture density network, y obj Represents the confidence prediction target. represents the Gaussian distribution function, μ objrepresents the mean of the confidence Gaussian mixture model, ∑ obj Represents the variance of the confidence Gaussian mixture model, and log is the natural logarithm.
[0045] The total loss function of Gaussian YOLOX is:
[0046] L=L reg +L cls +L obj
[0047] In the present invention, step 2 comprises:
[0048] In step 2-1, for the existing data set, according to the characteristics of active learning, 6% of the unlabeled data is randomly labeled for the first cycle of active learning to initialize the network knowledge.
[0049] In step 2-2, 30% of the remaining unlabeled data in step 2-1 are randomly labeled as a test set, and the remaining unlabeled data set is used for cyclic sampling in the active learning cycle.
[0050] Step 2-3, set the initial threshold of the foreground target to: Thresh = 0.1
[0051] In the present invention, step 3 comprises:
[0052] Step 3-1: Initialize the parameters of the backbone network of the network model constructed in Step 1 using the model trained on the COCO dataset (a general-purpose dataset) using YOLOX. Use random initialization for the neck and head detection parts.
[0053] Step 3-2: Use the existing labeled data to update the classification decision boundary of the detection network based on step 3-1.
[0054] In the present invention, step 4 comprises:
[0055] Step 4-1: For the optimal network model initialized or updated in step 3, test Gaussian YOLOX with the test set labeled in step 2. If the test performance meets the detection requirements, use this model for target detection. Otherwise, proceed to step 5 for the next stage of active learning.
[0056] Updating or initializing the network model can be expressed as:
[0057] f θ (x)←f θ (D X_L (x,y))
[0058] where f θ (x) represents Gaussian YOLOX in the present invention, DX_L (x,y) represents the labeled dataset.
[0059] In the present invention, step 5 comprises:
[0060] Step 5-1: For the model with the best performance tested in step 4, input all the remaining unlabeled data into the model for forward inference, and retain the regression variance, classification, regression and confidence results corresponding to the foreground target feature points in each unlabeled data output feature map.
[0061] Forward reasoning can be expressed as:
[0062] ∑ reg ,cls,reg,obj←f θ (D X_U (x))
[0063] Among them D X_U (x) indicates unlabeled data.
[0064] The regression variance can be expressed as:
[0065]
[0066] Classification prediction can be expressed as:
[0067]
[0068] Regression prediction can be expressed as:
[0069]
[0070] where μ reg Represents the mean of the prediction box center and width and height adjustment parameters.
[0071] The confidence prediction can be expressed as:
[0072]
[0073] Step 5-2: For the foreground target determination in step 5-1, the classification, confidence score and foreground target threshold after non-maximum suppression in the inference result of step 5-1 are compared to obtain the foreground target.
[0074] Non-Maximum Suppression (NMS) can be expressed as:
[0075] keep←NMS(∑ reg ,cls,reg,obj)
[0076] Where keep represents the target information retained after non-maximum suppression.
[0077] The classification, confidence and foreground target threshold comparison after non-maximum suppression can be expressed as:
[0078]
[0079] Where pos represents the feature point in the Gaussian YOLOX output feature map that meets the foreground target threshold condition.
[0080] In the present invention, step 6 includes:
[0081] For the regression uncertainty of the foreground target, the regression variance of the foreground target is used as the regression uncertainty of the target. The regression uncertainty of the foreground target can be expressed as:
[0082] U reg =∑ reg [pos]
[0083] In the present invention, step 7 includes:
[0084] Step 7-1: For the classification prediction results in step 5, first map the classification prediction results to the probability space, which can be expressed as:
[0085]
[0086] Where C represents the number of target categories.
[0087] The present invention uses the VOC dataset (Pascal Visual Object Class Challenge), which has 20 categories.
[0088] Map the classification scores to probability space and sort them from largest to smallest:
[0089] cls←(ProbSpace(cls[Pos]))·sort()
[0090] The classification uncertainty can be expressed as:
[0091] U cls ←1-(cls[0]-cls[1])
[0092] In the present invention, step 8 includes:
[0093] Regarding the uncertainty of the image, the uncertainty of the image is the sum of the regression uncertainty of each foreground target in step 6 and the classification uncertainty of each foreground target in step 7, which can be expressed as:
[0094] U img ←U reg +U cls
[0095] In the present invention, step 9 includes:
[0096] Step 9-1: Sort the uncertainty of each unlabeled image obtained in step 8 from large to small and retain the top k images with the largest uncertainty, which can be expressed as:
[0097] x top1 ,...,x topk ←sort(U img )
[0098] Step 9-2: Perform expert annotation on the k images obtained in step 9-1, which can be expressed as:
[0099] (x i ,y i ); i∈(1,…,k)←Oracle x top1 ,…,x topk
[0100] Oracle represents expert annotation.
[0101] Step 10 in the present invention comprises:
[0102] Step 10-1: Add the labeled dataset obtained in step 9 to the training set, which can be expressed as:
[0103] D X_L (x,y)←D X_L (x,y)∪(x i ,y i )
[0104] Step 10-2: Remove the labeled dataset in step 9 from the unlabeled dataset, which can be expressed as:
[0105] D X_U (x)←D X_U (x)\(x i )
[0106] Step 10-3: Update the foreground target threshold, which can be expressed as:
[0107] Thresh←Thresh+0.05
[0108] Step 10-4: Use the existing labeled dataset to return to step 3 to update the classification decision boundary of the network. The dataset used in the present invention is the VOC dataset (Pascal Visual Object Class Challenge).
[0109] The experimental environment of the present invention is: system: win10; processor: i7-10700F 2.9GHz; memory: 32G; GPU: RTX3060 (12G); deep learning framework: Pytorch1.9, cuda11.1; software: PyCharm.
[0110] The experimental parameters in the present invention are set as follows: the number of image sampling per cycle in the present invention is 6%, that is, the number of image iteration sampling in the active learning cycle follows: 6%, 12%, 18%, 24%, 30%, 36%, 42%, 48%, 54%, 60%...., and the number of mixed density networks of classification, regression, and confidence branches in the present invention is k=4 (deep active target detection based on probability modeling).
[0111] The training parameters in the present invention are set as follows: epoch (number of training iterations): 50, batch_size (batch data size): 8, lr (learning rate) = 0.001, the learning rate reduction strategy follows YOLOX, min_lr_ratio (minimum learning rate ratio) = 0.05, that is, as the number of training iterations increases, the learning rate will eventually drop to 0.00005.
Claims
1. An active target detection method that integrates classification and regression probability distribution, characterized by The steps include: Step 1: Build an active target detection network model architecture; Step 2, initial data preparation: select and label the initial data set from the unlabeled data, and then label the test set; Determine the number of unlabeled images to be sampled in each active learning iteration, and then determine the initial threshold for foreground objects; the remaining unlabeled data is used for active learning iteration sampling; Step 3, update or initialize the model: use the initial data set to initialize the model or use the existing labeled data to update the classification decision boundary of the detection network; Step 4: Model evaluation: Determine whether to continue active learning based on whether the performance of the existing optimal model on the test set meets the standard or whether the active learning annotation amount meets the annotation budget. If so, proceed to step 5; otherwise, use the optimal model of the current cycle for testing. Step 5, model forward reasoning: The remaining unlabeled data is forward-inferred through the optimal detection network in step 4 to obtain the regression variance, classification, regression, and confidence results of each feature point in the detection network output feature map; Step 6, regression uncertainty: Compare the classification and confidence results in step 5 with the foreground target threshold to obtain the foreground target information, and use the regression variance of the foreground target as the regression uncertainty of the foreground target; Step 7, classification uncertainty: the classification uncertainty is obtained by subtracting the second highest classification score from the highest classification score of the foreground object in step 5; Step 8, image uncertainty: the uncertainty of each foreground object in the unlabeled image is calculated by adding the regression uncertainty of the foreground object in step 6 and the classification uncertainty of the foreground object in step 7. The uncertainty of the unlabeled image is calculated by adding the uncertainty of each foreground object in the unlabeled image. Step 9, active labeling: select the first k unlabeled images with the largest uncertainty in labeling step 8, where k is obtained by the number of unlabeled images sampled per cycle in step 2; Step 10, data update: Update the dataset and foreground target threshold, remove the images annotated in step 9 from the unannotated set, add the images annotated in step 9 to the training set and return to step 3 to continue active learning; Step 1-1: For the image feature extraction part, select CSPDarknet-53 as the feature extraction backbone network; In step 1-2, for the feature fusion part, the output of the last three convolutional blocks of the YOLOX backbone network is extracted for feature fusion, and PAFPN is selected as the FPN for detecting neck feature fusion; In steps 1-3, for the classification and regression prediction part, the mixture density network (MDN) is selected as the YOLOX detection head, and the Gaussian distribution function and the binary cross entropy function are selected as the regression and classification loss functions respectively. The improved YOLOX is renamed as Gaussian YOLOX. The Gaussian YOLOX regression loss function is: Where N represents the number of positive and negative sample prediction boxes that SimOTA matches with the feature points corresponding to the foreground target center in the output feature map. represents the Gaussian distribution function, k represents the number of regression mixture density networks, π reg Represents the weight of the Gaussian mixture model fitted by the regression mixture density network, μ iou Represents the mean IOU of the predicted box and the real box, ∑ reg Represents the variance of the intersection of the predicted box and the true box, ε represents a small positive number to ensure the stability of the logarithmic value, and log is the natural logarithm; The Gaussian YOLOX classification loss function is: Where N represents the number of positive and negative samples matched by SimOTA for the feature points corresponding to the foreground target center in the output feature map; k represents the number of classification mixture density networks, π cls Represents the weight of the Gaussian mixture model fitted by the classification mixture density network, y cls Represents the classification prediction target, represents the Gaussian distribution function, μ cls represents the mean of the classification Gaussian mixture model, ∑ cls Represents the variance of the classification Gaussian mixture model, log is the natural logarithm; The confidence loss function is: Where N represents the number of confidence positive and negative samples matched by SimOTA for the feature point corresponding to the foreground target center in the output feature map, k represents the number of confidence mixed density networks, and π obj Represents the weight of the Gaussian mixture model fitted by the confidence mixture density network, y obj represents the confidence prediction target, μ obj represents the mean of the confidence Gaussian mixture model, ∑ obj Represents the variance of the confidence Gaussian mixture model, log is the natural logarithm; The total loss function is: L=L reg +L cls +L obj 。 2. The active target detection method according to claim 1, characterized in that The method comprises the following steps: Step 2 comprises: Step 2-1: For the existing dataset, based on the characteristics of active learning, 6% of the unlabeled data is randomly labeled to initialize the network knowledge in the first cycle of active learning; Step 2-2: Randomly label 30% of the remaining unlabeled data in step 2-1 as a test set, and the remaining unlabeled data set is used for cyclic sampling in the active learning cycle; In step 2-3, the initial threshold of the foreground object is set to Thresh = 0.
1.
3. The active target detection method according to claim 2, characterized in that The method comprises the following steps: Step 3 comprises: Step 3-1: For the network model constructed in step 1, initialize the parameters of the backbone network part. Use the model trained by YOLOX on the COCO dataset to initialize the parameters; use random initialization for the neck and head detection parts; Step 3-2: Use the existing labeled data to update the network's classification decision boundary based on step 3-1.
4. The active target detection method according to claim 3, characterized in that The method comprises the following steps: Step 4 comprises: Step 4-1: For the optimal network model initialized or updated in step 3, test Gaussian YOLOX with the test set marked in step 2. If the test performance meets the detection requirements, use this model for target detection. Otherwise, proceed to step 5 for the next stage of active learning. Updating or initializing the network model is expressed as: f θ (x)←f θ (D X_L (x,y)) where f θ (x) represents Gaussian YOLOX in the present invention, D X_L (x,y) represents the labeled dataset.
5. The active target detection method integrating classification and regression probability distribution according to claim 4 is characterized in that The method comprises the following steps: Step 5 comprises: Step 5-1: For the model with the best performance in step 4, all remaining unlabeled data are input into the model for forward reasoning, and the regression variance, classification, regression and confidence results corresponding to the foreground target feature points in each unlabeled data output feature map are retained; expressed as: ∑ reg ,cls,reg,obj←f θ (D X_U (x)) Among them D X_U (x) indicates unlabeled data; The regression variance is expressed as: The classification prediction is expressed as: The regression prediction is expressed as: where μ reg Represents the mean of the prediction box center and width and height adjustment parameters; The confidence prediction is expressed as: Step 5-2: For the foreground target determination in step 5-1, the classification, confidence score and foreground target threshold of the inference result of step 5-1 after non-maximum suppression are compared to obtain the foreground target; Non-maximum suppression NMS is expressed as: keep←NMS(∑ reg ,cls,reg,obj) Among them, keep represents the target information retained after non-maximum suppression; The classification and confidence after non-maximum suppression are compared with the foreground target threshold as follows: Where pos represents the feature point in the Gaussian YOLOX output feature map that meets the foreground target threshold condition.
6. The active target detection method integrating classification and regression probability distribution according to claim 5 is characterized in that The method comprises the following steps: Step 6 comprises: For the regression uncertainty of the foreground target, the regression variance of the foreground target is used as the regression uncertainty of the target; the regression uncertainty of the foreground target is expressed as: U reg =∑ reg [pos]。 7. The active target detection method integrating classification and regression probability distribution according to claim 6 is characterized in that The steps include: Step 7 includes: Step 7-1: for the classification prediction result in step 5, first map the classification prediction result to the probability space, which can be expressed as: Where C represents the number of target categories; Step 7-2: After mapping the classification scores to the probability space, subtract the second highest classification score from the highest classification score to obtain the classification uncertainty; Map the classification scores to probability space and sort them from largest to smallest: cls←(Pr obSpace(cls[pos])).sort() The classification uncertainty is expressed as: U cls ←1-(cls[0]-cls[1])。 8. The active target detection method integrating classification and regression probability distribution according to claim 7 is characterized in that The method comprises the following steps: Step 8 comprises: For the uncertainty of the image, the uncertainty of the image is expressed as the sum of the regression uncertainty of each foreground target in step 6 and the classification uncertainty of each foreground target in step 7: IN img ←U reg +U cls 。 9. The active target detection method integrating classification and regression probability distribution according to claim 8 is characterized in that The method comprises the following steps: Step 9 comprises: Step 9-1: Sort the uncertainty of each unlabeled image obtained in step 8 from large to small and retain the top k images with the largest uncertainty, expressed as: x top1 ,...,x topk ←sort(U img ) Step 9-2: Perform expert annotation on the k images obtained in step 9-1, expressed as: (x i ,y i );i∈(1,…,k)←Oracle x top1 ,…,x topk Among them, Oracle represents expert annotation; Step 10 includes: Step 10-1: Add the labeled dataset obtained in step 9 to the training set, expressed as: D X_L (x,y)←D X_L (x,y)∪(x i ,and i ) Step 10-2: Remove the labeled dataset in step 9 from the unlabeled dataset, expressed as: D X_U (x)←D X_U (x)\(x i ) Step 10-3: Update the foreground target threshold, expressed as: Thresh←Thresh+0.05 Step 10-4: Use the existing labeled dataset to return to step 3 to update the network's classification decision boundary.