Small sample target detection method based on extended anchor point theory

Abstract

The application provides a small sample target detection method based on an extended anchor point theory, which comprises a pseudo-label initialization step, a pseudo-label refinement step and a target detection model fine-tuning step. Potential new class targets are mined from base class images, and the pseudo-labels of the mined new class targets are corrected. Based on the position relationship and semantic relationship between the candidate frames, the label noise transfer matrix related to the instance is decomposed into a plurality of label noise transfer matrices irrelevant to the instance, and the anchor point theory is used to estimate the label noise transfer matrix to obtain the real probability of the candidate frame belonging to a certain class. Compared with the standard pseudo-label generation method, the method effectively improves the quality of the pseudo-labels and greatly improves the detection accuracy of the target detection model on the small sample new class.

Description

Technical Field

[0001] This invention relates to Few-shot Object Detection (FSOD) technology, and particularly to a potential new type of object mining technology based on Extended Anchor Theory (EAT). Background Technology

[0002] With the development of deep learning technology, computer vision technology has been widely applied in people's daily lives. A large number of labeled images are used to train high-performance computer vision models. However, data is sometimes difficult to obtain. In the absence of sufficient data, it is difficult to train a high-performance computer vision model. To effectively train computer vision models under conditions of insufficient data, the few-shot problem has been proposed. Among these, the few-shot object detection task has received widespread attention. In this task, there are two parts of data: sufficient labeled data for the base class and limited labeled data for the new class, and the base class and the new class are disjoint.

[0003] Currently, existing few-shot object detection methods are mainly based on transfer learning, which can be roughly divided into two steps: First, train the object detection model using sufficient base class labeled data to obtain a base class object detection model; second, without updating the model parameters of the feature extraction backbone network in the object detection model, fine-tune the model using a small amount of new class labeled data to obtain a new class object detection model. Improved algorithms include extracting more accurate general prototypes from base class images to assist new class training during the fine-tuning stage; increasing inter-class distance to avoid inter-class confusion; and explicitly or implicitly constructing semantic relationships between base classes and new classes, using base classes to assist in new class classification. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a potential new class target mining method based on Extended Anchor Theory (EAT) in the Few-shot Object Detection (FSOD) task, which attempts to mine potential new classes in base class images and corrects the pseudo-labels of potential new classes.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is a small-sample target detection method based on extended anchor point theory, comprising the following steps:

[0006] 1. Pseudo-tag initialization steps:

[0007] The initial object detection model is used to predict new class candidate boxes in the input base class image to obtain the class confidence of the new class candidate box in each target category. When the class confidence of a new class candidate box is higher than the set threshold, the new class candidate box is retained and the target category corresponding to the highest class confidence is used as the initial pseudo-label of the new class candidate box.

[0008] 2. Steps to refine pseudo-tags:

[0009] 2-1 Set a first cross-union ratio (CUP) threshold and a second CUP threshold, and calculate the maximum CUP between the new class candidate boxes and each base class target; the second CUP threshold is greater than the first CUP threshold, and the first and second CUP thresholds are used to determine different types of candidate boxes;

[0010] When the maximum crossover ratio is greater than the second crossover ratio threshold, the pseudo-label of the new class candidate box is corrected to the base class label corresponding to the maximum crossover ratio, and then proceed to step 3;

[0011] If the maximum cross-union ratio is between the first cross-union ratio threshold and the second cross-union ratio threshold, then the new class of candidate boxes is regarded as the background candidate boxes, and step 2-2 is executed again;

[0012] If the maximum crossover ratio is less than the first crossover ratio threshold, then proceed to steps 2-3;

[0013] 2-2 Using anchor point theory and known anchor points, the label noise transition probability is estimated, where P(A|B) represents the probability of A when condition B is met, and ' is the pseudo-label. Then, using Bayes' theorem, the label y is obtained when the pseudo-label is of class i and the label is background y. b The probability P(y) b |i′) and the probability P(i|i′) that the label is of class i when the pseudo-label is of class i, and the probability P(y b |i′) is compared with the probability P(i|i′), when P(i|i′) <P(y b If |i′), then the pseudo-label of the new class candidate box will be corrected to the background label y. b Then proceed to step 3, when P(i|i′)≥P(y) b If |i′), then keep the pseudo-label of the new class candidate box as the i-th class and proceed to step 3;

[0014] 2-3 Find the neighboring candidate boxes of the new class candidate box and calculate the average class confidence of each target class of the neighboring candidate boxes; for the initial pseudo-label of the new class candidate box as class i, determine whether there are cases where the average class confidence of its neighboring candidate boxes in other new classes j is higher than the average class confidence of class i. If so, name the candidate box as an incorrect new class candidate box and execute step 2-4; otherwise, keep the pseudo-label of the new class candidate box as class i and proceed to step 3.

[0015] 2-4 Use the anchor point theory and known anchor points to estimate the label noise transfer probability, and then obtain the probability P(j|i′) that the label is the j-th class when the pseudo-label is the i-th class through Bayes' formula. Compare the probability P(j|i′) with the probability P(i|i′) that the label is the i-th class when the pseudo-label is the i-th class. When P(i|i′) < P(j|i′), revise the pseudo-label of the new class candidate box to the j-th class and then enter step 3. When P(i|i′) ≥ P(j|i′), keep the pseudo-label of the new class candidate box as the i-th class and then enter step 3;

[0016] 3. Steps for fine-tuning the object detection model:

[0017] Input the new class candidate box and the corresponding pseudo-label into the object detection model for model fine-tuning, so as to obtain the fine-tuned object detection model.

[0018] The present invention mines potential new class objects from the base class images and revises the pseudo-labels of the mined new class objects. Group the new class candidate boxes extracted by the model from the base class images, convert the label noise related to the instance in the pseudo-label into label noise unrelated to the instance in each group of candidate boxes, and adopt the anchor point theory to improve the accuracy of the estimated pseudo-label of the candidate box on each group of candidate boxes to facilitate the revision of the pseudo-label.

[0019] The beneficial effect of the present invention is that based on the positional relationship and semantic relationship between candidate boxes, the label noise transfer matrix related to the instance is decomposed into several label noise transfer matrices unrelated to the instance, and the anchor point theory is used to estimate the label noise transfer matrix to obtain the true probability that the candidate box belongs to a certain class. Compared with the standard pseudo-label generation method, this method effectively improves the quality of the pseudo-label and greatly improves the detection accuracy of the object detection model on small-sample new classes; further, an adaptive threshold strategy is provided, which can obtain sufficient pseudo-labels on difficult classes while avoiding the background being labeled as the foreground. The method for mining potential new class objects based on the extended anchor point theory proposed by the present invention has the characteristics of high accuracy and wide algorithm applicability. Brief Description of the Drawings

[0020] Figure 1 : Computational example of candidate box division and pseudo-label revision;

[0021] Figure 2 : Flowchart of the embodiment. Detailed Embodiment

[0022] Because the number of labeled images in few-sample object detection tasks is extremely small, the model accuracy on the current rare class is improved by utilizing a large amount of existing labeled data. The classes in the existing labeled data are called base classes, and the current rare class is called the new class. Simultaneously, the base class images contain a large number of potential new class objects. Potential new class candidate boxes and their pseudo-labels can be automatically mined from the base class images to expand the number of labels for new class objects. When generating pseudo-labels for new class candidate boxes, a class confidence threshold needs to be set. When the class confidence of a new class candidate box for a certain class is greater than the threshold, the class with the highest class confidence is used as the pseudo-label for that candidate box.

[0023] However, the applicant found that in the existing methods, the category confidence threshold used is usually fixed. If the difficulty of predicting a certain type of target is high and the category confidence is low, it is impossible to obtain the pseudo-label of the candidate box for that type. If the difficulty of predicting a certain type of target is low and the category confidence is high, it is easy to set the pseudo-label of the background as a new class.

[0024] Preferably, in this embodiment, the sum of the mean and variance of each confidence level is used as the confidence threshold, i.e.:

[0025] τ=μ+s,

[0026] Where μ represents the mean class confidence score of the candidate box in a certain class, s represents the standard deviation of the class confidence score of the candidate box in a certain class, and τ represents the class confidence threshold. Using this strategy, sufficient pseudo-labels can be obtained on hard classes while avoiding the background being labeled as foreground.

[0027] In the process of discovering new class candidate boxes and their corresponding pseudo-labels, inaccurate pseudo-labels cannot be obtained, leading to a decrease in model accuracy. We correct the pseudo-labels of the new class candidate boxes.

[0028] First, assume the label of the new class candidate box x is y, and the pseudo-label is y′. The noise transition probability P(y′=j|y=i,x) for the new class candidate box x refers to the probability that the candidate box x is labeled i and the pseudo-label y′ is j, where j and i are two different semantic categories. The following relationship exists:

[0029]

[0030] P(y′=j|x) is the noisy posterior probability of candidate box x (i.e., the class confidence of the candidate box), P(y=k|x) is the probability that the actual label of candidate box x is k, and C is the total number of labels. Assuming that the new class candidate box x in the above equation is an anchor sample and its label is i, then we know that P(y′=j|x)=P(y′=j|y=i,x).

[0031] According to the above formula, P(y′|i,x) can be estimated using anchor point theory. Anchor points refer to samples with known labels. In practice, it is impossible to obtain the label noise transition probability on every candidate box. Therefore, existing methods average the noisy posterior probability of all anchor point samples to obtain P(y′=j|y=i) as a model of label noise.

[0032] Considering that label noise in pseudo-labels is instance-dependent, directly using the instance-independent label noise transition matrix P(y′=j|y=i) for modeling would result in significant label noise errors in pseudo-labels. In this invention, candidate boxes are divided into different groups based on their positional and semantic relationships, such as... Figure 1 As shown, the purpose is to make the label noise transition probability in each group approximately independent of the instance, thereby improving the estimation accuracy of the label noise transition matrix. This invention calls it the extended anchor theory. Its key is to transform the instance-related transition probability estimation problem into an instance-independent transition probability estimation problem through appropriate sample grouping, thereby reducing the transition probability estimation error. Figure 1 The base class tags include: Person, Cat, Chair; the new class tags include: Horse, Cow, Sofa.

[0033] We use Intersection over Union (IoU) to represent the positional relationship between candidate boxes and class confidence to represent the semantic relationship between them. Since the base class target is known, the IoU between the new class candidate box and the base class target can be obtained. A first IoU threshold γ is set. Candidate boxes with IoU values ​​higher than the first threshold have a high probability of being pseudo-labeled and contain noise, which is the focus of this invention. Candidate boxes with IoU values ​​lower than the first threshold are likely to be new class candidate boxes. In this embodiment, γ is set to 0.1.

[0034] Furthermore, a second Cross-Union Ratio (CUI) threshold is set; in this embodiment, the second CUI threshold is set to 0.5. We found that if the CUI is greater than 0.5, according to the candidate box label assignment criterion proposed by Faster R-CNN, these new class candidate boxes can be considered to belong to the base class, and they are named base class candidate boxes (Base Proposals). Figure 1 In the example, when u = 0.71, the pseudo-label "sofa" of the initial classification of the new class candidate box is corrected to "chair"; when u = 0.92, the pseudo-label "horse" of the initial classification of the new class candidate box is corrected to "person".

[0035] When the applicant finds that the intersection-union ratio (IU) of the candidate bounding box and the base class target is in the range [0.1, 0.5], the probability P(y′=i|y=i,x) that the pseudo-label of candidate bounding box x belongs to class i and the label belongs to class i is less than 0.5, and the probability P(y′=i|y=y) that the pseudo-label of candidate bounding box x belongs to class i and the label belongs to the background is less than 0.5. b If x) > 0.5, then y b This represents the background label. That is, when the intersection-union ratio (IU) of a candidate box with its base class is in the range [0.1, 0.5], the candidate box x is highly likely to be background. Therefore, candidate boxes with IU of [0.1, 0.5] are named Possible Background Proposals. The labels of these candidate boxes are highly likely to be background, but still have a certain probability of being foreground targets, thus requiring further estimation and correction using extended anchor point theory.

[0036] Assuming that the intersection-union ratio (IU) of a candidate box x with its neighboring candidate boxes is less than the first IU threshold of 0.1, we further observe the neighboring candidate boxes. If the confidence level of a neighboring candidate box belonging to class j is high, but the candidate box is labeled as class i, then the probability of the candidate box being mislabeled is relatively high. When the IU between a candidate box and the base class target is in the range [0.0, 0.1), the candidate box has a certain probability of being mislabeled as another new class. This invention requires correction in this case. The new class label of a candidate box can only be determined when the IU between a candidate box and the base class target is less than 0.1 and its label type predicted by the neighboring candidate boxes is the same new class. When a candidate box is inconsistent with the prediction of its neighboring candidate boxes, we name the candidate box an "Incorrect Novel Proposal" and then proceed with subsequent correction steps.

[0037] Abbreviate \(P(y' = j|y = i, x)\) as \(P(j'|i, x)\), where \(P(A|B)\) represents the probability of \(A\) when the condition \(B\) holds, and the prime symbol \(' \) is for pseudo-label marking; abbreviate \(P(y' = j|y = i)\) as \(P(j'|i)\), and abbreviate \(P(y' = i|y = j)\) as \(P(i'|j)\). \(P(j|i')\) represents the probability that the label is \(j\) when the pseudo-label of the candidate box is \(i\), and \(P(j'|i)\) represents the probability that the pseudo-label is \(j\) when the label of the candidate box is \(i\). Abbreviate the probability \(P(y' = j)\) that the pseudo-label of the candidate box is of the \(j\)-th class as \(P(j')\), and abbreviate the probability \(P(y = j)\) that the label of the candidate box is of the \(j\)-th class as \(P(j)\). After dividing different types of candidate boxes, since the label noise transition probability \(P(j'|i, x)\) of candidate boxes of the same type is similar, it is assumed that the transition probability is independent of the candidate box \(x\), that is, for any candidate box of the same type, the probability that the label is \(i\) and the pseudo-label is \(j\) is the same, \(P(j'|i, x)\approx P(j'|i)\), and \(P(j'|i)\) can be estimated by the anchor theory. Then, based on Bayes' formula, we have:

[0038]

[0039] Among them, \(P(j)\) is approximately \(P(j')\), and the value of \(P(j')\) can be obtained through estimation. Therefore, \(P(j|i')\) can be obtained, that is, the probability that the label of the candidate box is of the \(j\)-th class when the pseudo-label is of the \(i\)-th class. When \(P(i|i') < P(j|i')\), it means that the probability that the label of the candidate box is of the \(j\)-th class is higher than the probability that the label is of the \(i\)-th class. Therefore, the pseudo-label is corrected from \(i\) to \(j\).

[0040] 1. Pseudo-label initialization step:

[0041] Use the initial object detection model to predict the new-class candidate boxes in the input base-class images, and obtain the class confidence of the new-class candidate boxes for each object class; preferably, calculate the mean and variance of the class confidence of the new-class candidate boxes for each object class respectively, and use the sum of the mean and variance as the adaptive threshold of the class confidence; when the class confidence of the new-class candidate box is higher than the adaptive threshold, then retain the new-class candidate box, and use the object class corresponding to the highest class confidence as the pseudo-label of the new-class candidate box;

[0042] 2. Pseudo-label refinement step:

[0043] 2-1 Set the first intersection over union (IoU) threshold and the second IoU threshold, and calculate the maximum IoU \(u\) between the new-class candidate box and each base-class object; the second IoU threshold is greater than the first IoU threshold, and the first and second IoU thresholds are used to determine different types of candidate boxes;

[0044] When the maximum intersection over union (IoU) is greater than the second IoU threshold, correct the pseudo-label of the new class candidate box to the base class label corresponding to the maximum IoU, and proceed to step 3;

[0045] When the maximum IoU is between the first IoU threshold and the second IoU threshold, consider the new class candidate box as a background candidate box, and then execute step 2-2;

[0046] When the maximum IoU is less than the first IoU threshold, execute step 2-3;

[0047] 2-2 Use the anchor theory and known anchors to estimate the label noise transition probabilities P(i′|y b ) and P(i′|i), where P(A|B) represents the probability of A when condition B holds, ′ is the pseudo-label mark, and then obtain the probability P(y b |i′) that the label is the background y when the pseudo-label is the i-th class and the probability P(i|i′) that the label is the i-th class when the pseudo-label is the i-th class through Bayes' formula. Compare the probabilities P(y b |i′) and P(i|i′). When P(i|i′) < P(y b |i′), correct the pseudo-label of the new class candidate box to the background label y b and then proceed to step 3. When P(i|i′) ≥ P(y b |i′), keep the pseudo-label of the new class candidate box as the i-th class and then proceed to step 3; b |i′), keep the pseudo-label of the new class candidate box as the i-th class and then proceed to step 3;

[0048] 2-3 Find the neighborhood candidate boxes of the new class candidate box and calculate the mean class confidence on each target class of the neighborhood candidate boxes; for the initial pseudo-label of the new class candidate box being the i-th class, determine whether there is a situation where the mean class confidence of its neighborhood candidate boxes on other new class j is higher than the mean class confidence on the i-th class. If so, name the candidate box as an incorrect new class candidate box and execute step 2-4. Otherwise, keep the pseudo-label of the new class candidate box as the i-th class and then proceed to step 3;

[0049] 2-4 Use the anchor theory and known anchors to estimate the label noise transition probability, and then obtain the probability P(j|i′) that the label is the j-th class when the pseudo-label is the i-th class through Bayes' formula. Compare the probability P(j|i′) with the probability P(i|i′) that the label is the i-th class when the pseudo-label is the i-th class. When P(i|i′) < P(j|i′), correct the pseudo-label of the new class candidate box to the j-th class and then proceed to step 3. When P(i|i′) ≥ P(j|i′), keep the pseudo-label of the new class candidate box as the i-th class and then proceed to step 3;

[0050] 3. Fine-tuning step of the object detection model: Input the new class candidate box and the corresponding pseudo-label into the object detection model for model fine-tuning.

[0051] The examples were conducted on the TITAN experimental platform, such as... Figure 2 The process mainly includes: using an initial object detection model to generate initial pseudo-labels, then performing pseudo-label refinement, and finally using the refined labels to complete fine-tune training through the object detection model.

[0052] Step 1:

[0053] Step 1: Obtain the initial object detection model (Initial detector) using the existing base image set and the DeFRCN algorithm. Proceed to Step 2.

[0054] Step Two:

[0055] Step 2: Use the initial detector model to predict novel class proposals in the base class image and obtain the confidence score of class i for the novel class proposals. Proceed to Step 3;

[0056] Step 3: Calculate the mean and variance of the class confidence scores for the candidate boxes across different classes. Use the sum of the mean and variance as the adaptive threshold for the class confidence scores. Proceed to Step 4;

[0057] Step 4: If the class confidence of a candidate box in a new class is higher than the adaptive threshold, retain the candidate box and use the class with the highest class confidence as the initial pseudo-label for that candidate box. Proceed to Step 5;

[0058] Step 5: Calculate the maximum intersection-union ratio (U) between the new class candidate box and a certain base class target: If it is greater than the second intersection-union threshold of 0.5, refine the pseudo-label of the new class candidate box to the label of the base class, and then proceed to step 9; If the maximum intersection-union ratio U is between the first and second intersection-union thresholds, that is, greater than or equal to the threshold of 0.1 and less than or equal to 0.5, then name the candidate box as the background candidate box, and then proceed to step 6; If the maximum intersection-union ratio U between the new class candidate box and the base class annotation box is less than the first intersection-union threshold of 0.1, then proceed to step 7.

[0059] Step 6: Estimate the label noise transition probability P(i′|j) using the anchor theory and known anchors, and obtain P(j|i′) through Bayes' formula. If, for a candidate box, the probability that the pseudo-label is class i when the label is class i is less than the probability that the pseudo-label is class i when the label is class j, i.e., P(i|i′) < P(j|i′), then refine the pseudo-label of the candidate box to class j. Here, j in this step is y b , if P(i|i′) ≥ P(j|i′), then refine the pseudo-label of the candidate box to class i; after completing the refinement of the pseudo-label, execute Step 9;

[0060] Step 7: Find the neighborhood candidate boxes of the candidate box and calculate the mean class confidence on different classes of the neighborhood candidate boxes. If, for a candidate box, the following conditions are met: the initial pseudo-label of the candidate box is class i; the mean class confidence of its neighborhood candidate boxes on a certain other new class j is higher than the mean class confidence on class i, then name the candidate box as an incorrect new class candidate box. Execute Step 8; if the conditions are not met, then refine the pseudo-label of the candidate box to class i;

[0061] Step 8: Estimate the label noise transition probability P(i′|j) using the anchor theory and known anchors, and obtain P(j|i′) through Bayes' formula. If, for a candidate box, P(i|i′) < P(j|i′), then refine the pseudo-label of the candidate box to class j; otherwise, keep the initial pseudo-label class and refine the pseudo-label of the candidate box to class i, and execute Step 9.

[0062] Step 3:

[0063] Step 9: Use the new class candidate boxes and their corresponding refined pseudo labels to fine-tune the model, and execute Step 10; the pseudo-labels that are new classes and base classes both belong to the foreground. In the embodiment, for the case where the pseudo-label is the background, it is divided according to the source into those from the background, base classes, and new classes. The image part belonging to the background corrected from the new class uses the symmetric loss to constrain the object detection model, and the image part belonging to the background corrected from the base class and the image part determined to be the background itself use the cross-entropy loss to constrain the model; the image part belonging to the foreground also uses the cross-entropy loss to constrain the model;

[0064] Step 10: Loop back to Step 2 to process the next image batch.

[0065] The innovation of this invention lies in step two. In step 9, those skilled in the art can use the differentiated loss commonly used in small sample object detection to fine-tune the object detection model with new samples (the image parts of the candidate boxes) and pseudo-labels as needed.

Claims

1. A small-sample target detection method based on extended anchor point theory, characterized in that, Includes the following steps:

1. Pseudo-tag initialization steps: The initial object detection model is used to predict new class candidate boxes in the input base class image to obtain the class confidence of the new class candidate box in each target category. When the class confidence of a new class candidate box is higher than the set threshold, the new class candidate box is retained and the target category corresponding to the highest class confidence is used as the initial pseudo-label of the new class candidate box.

2. Steps to refine pseudo-tags: 2-1 Set a first cross-union ratio (CUP) threshold and a second CUP threshold, and calculate the maximum CUP between the new class candidate boxes and each base class target; the second CUP threshold is greater than the first CUP threshold, and the first and second CUP thresholds are used to determine different types of candidate boxes; When the maximum crossover ratio is greater than the second crossover ratio threshold, the pseudo-label of the new class candidate box is corrected to the base class label corresponding to the maximum crossover ratio, and then proceed to step 3; If the maximum cross-union ratio is between the first cross-union ratio threshold and the second cross-union ratio threshold, then the new class of candidate boxes is regarded as the background candidate boxes, and step 2-2 is executed again; If the maximum crossover ratio is less than the first crossover ratio threshold, then proceed to steps 2-3; 2-2 Estimating the tag noise transition probability P(i) using anchor point theory and known anchor points ′ |y b ) and P(i ′ |i), where P(A|B) represents the probability of A given condition B. ′ We mark the pseudo-labels, and then use Bayes' theorem to determine if the pseudo-label is of class i and the label is background y. b The probability P(y) b |i′) and the probability P(i|i′) that the label is of class i when the pseudo-label is of class i, and the probability P(y b |i′) is compared with the probability P(i|i′), when P(i|i′) <P(y b If |i′), then the pseudo-label of the new class candidate box will be corrected to the background label y. b Then proceed to step 3, when P(i|i′)≥P(y) b If |i′), then keep the pseudo-label of the new class candidate box as the i-th class and proceed to step 3; 2-3 Find the neighboring candidate boxes of the new class candidate box and calculate the average class confidence of each target class of the neighboring candidate boxes; for the initial pseudo-label of the new class candidate box as class i, determine whether there are cases where the average class confidence of its neighboring candidate boxes in other new classes j is higher than the average class confidence of class i. If so, name the candidate box as an incorrect new class candidate box and execute step 2-4; otherwise, keep the pseudo-label of the new class candidate box as class i and proceed to step 3. 2-4 Use the anchor theory and known anchors to estimate the label noise transition probabilities P(i ′ |j) and P(i ′ |i), and then use Bayes' formula to obtain the probability P(j|i′) that the label is the j-th class when the pseudo-label is the i-th class, and the probability P(i|i′) that the label is the i-th class when the pseudo-label is the i-th class. Compare the probability P(j|i′) with the probability P(i|i′). When P(i|i′) < P(j|i′), revise the pseudo-label of the new class candidate box to the j-th class and then enter step 3. When P(i|i′) ≥ P(j|i′), keep the pseudo-label of the new class candidate box as the i-th class and then enter step 3; 3. Fine-tuning steps for the object detection model: The new class candidate boxes and their corresponding pseudo-labels are input into the object detection model for fine-tuning, thus obtaining the fine-tuned object detection model.

2. The method as described in claim 1, characterized in that, In the pseudo-label initialization step, the threshold is set to an adaptive threshold; the adaptive threshold is the sum of the mean and variance of the class confidence scores of the new class candidate boxes on each target class.

3. The method as described in claim 1, characterized in that, The extended anchor theory is used to obtain the label noise transition probability of different groups of samples, and the pseudo-labels are corrected based on the probability.

Images