Object detection active learning method based on region-level information amount

The active learning method for object detection by region-level information division and dynamic weighting effectively reduces labeling costs and enhances performance by considering object relationships, improving scalability and addressing category imbalance.

CN119068261AActive Publication Date: 2024-12-03CHONGQING UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411233552.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-04
Publication Date
2024-12-03
Estimated Expiration
2044-09-04

AI Technical Summary

Technical Problem

Existing object detection methods based on deep neural networks require a large number of labeled samples for training, leading to high labeling costs, and fail to consider the connections between objects in an image.

Method used

An active learning method that divides images into regions, calculates regional information volume, and uses dynamic weights to correct the information amount, considering the relationship between objects, without adding additional model modules.

Benefits of technology

Reduces labeling costs by selecting the most informative samples for annotation, improving object detection performance and scalability, and addressing category imbalance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068261A_ABST
    Figure CN119068261A_ABST
Patent Text Reader

Abstract

The invention discloses an object detection active learning method based on region-level information amount. Comprising the steps of performing region division on image prediction according to a model; the three modules are used for calculating regional information amount; a region uncertainty calculation module; a region intra-class feature difference calculation module; an intra-region feature difference calculation module; and an image dynamic weight calculation module. According to the method, objects in an image are divided into different areas, the relation between the objects in the image is considered, and meanwhile, the method does not depend on a specific network structure or additional module construction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of object detection and active learning, and in particular to an object detection active learning method based on region-level information. Background Art

[0002] With the development of deep learning, object detection methods based on deep neural networks have also achieved good results. However, deep neural networks require a large number of labeled samples for training, which brings about the problem of labeling cost. Active learning is a technology used to solve the problem of labeling cost. By autonomously selecting samples by the model, active learning can greatly reduce the number of samples that need to be labeled. Therefore, it is necessary to apply active learning to the problem of object detection to reduce the problem of labeling cost in the problem of object detection.

[0003] For the combination of object detection and active learning, existing methods can be roughly divided into two categories. The first method regards the objects in the image as independent individuals, calculates the value of each object independently, and finally aggregates them into the image value. The other method regards the image as a whole and directly calculates the image value. However, both methods ignore the connection between objects in the image. Summary of the invention

[0004] 1. Technical issues to be resolved

[0005] In view of the shortcomings of the prior art, the present invention provides an active learning method for object detection based on region-level information volume, which divides objects in an image into different regions, calculates the region information volume for each region, and utilizes the relationship between objects in the image. It also has the characteristics of easy deployment and expansion.

[0006] (II) Technical solution

[0007] In order to solve the above problems, the present invention designs three modules to realize the final image value calculation, namely, the region division module, the region information calculation module and the dynamic weight module. The specific technical solutions adopted by the present invention are as follows:

[0008] An active learning method for object detection based on region-level information includes the following steps:

[0009] S0: Divide the given data set into two categories, namely Represents a labeled dataset and uses Represents an unlabeled dataset, where represents a sample with a label, Indicates the corresponding label, represents an unlabeled sample, and The number is much greater than the number of

[0010] use Represents an object detection model. For a given image A set of triplets will be predicted to represent the detected objects in, Indicates the category to which the object belongs. represents the classification score of the object, represents the bounding box coordinates of the object, where Is a four-dimensional vector that describes the object in the image Position in It can be any existing object detection model).

[0011] S1: let a=1;

[0012] S2: On training Count the number of objects in each category in the labeled data set, align and normalize to get the category distribution of the labeled data set

[0013] S3: Used on each image Perform inference to obtain the object detection results of the corresponding image Represents the detection result of the i-th object in the image;

[0014] S4: According to the obtained right Perform regional division on the objects in the corresponding image and obtain the regional division result represents the jth region after region division;

[0015] Calculate the regional information of each region after the region is divided

[0016] The maximum value of the regional information in all regions is selected as the information amount representing the image

[0017] The information of all regions in the comprehensive image is obtained The amount of information corresponding to the image

[0018] right Each image in the image is calculated to have its corresponding information amount

[0019] S5: Calculation The dynamic weight ω of each image in the image is used to correct the amount of information corresponding to each image to

[0020] S6: Arrange the corrected information amount corresponding to each image in descending order, take the first K images corresponding to the corrected information amount for annotation, remove the annotated images from the unlabeled dataset, and add them to the labeled dataset to update the labeled dataset;

[0021] S7: If a>T, execute the next step, otherwise return to S2;

[0022] S8: Supervised training using current labeled datasets Finally, the optimal detection model is obtained;

[0023] S9: For a new image, use the optimal detection model to predict the new image and output the object detection result of the new image.

[0024] Specifically, the steps of obtaining the object detection result in S3 are as follows:

[0025] For a given image Feed it into the object detection model , the detection result of the model is obtained; for the detection result of the model, the category probability value of the object in the detection result is used to represent the confidence of the object, and the confidence is lower than the threshold θ 1 The objects are filtered out to obtain the final object detection result.

[0026] Specifically, the region division process of S4 is as follows:

[0027] S41: For a given test result Create a new bounding box set that has been selected Empty, create a new final area division result set is empty;

[0028] S42: let t=1;

[0029] S43: Create a new candidate region set If it is empty, take the tth result in the test result. If the bounding box is middle, Then go to S411, otherwise go to S44;

[0030] S44: Add to the set of selected bounding boxes and candidate region set middle;

[0031] S45: set q=1;

[0032] S46: Get the qth result in the detection result If the bounding box is middle, Then go to S49, otherwise go to S47;

[0033] S47: Calculate the closeness of the two bounding boxes according to the following formula Indicates finding the area of ​​the region, and min() indicates finding the minimum value;

[0034]

[0035] S48: If the tightness Greater than the threshold θ 2 ,Will Add to and Temporarily save subscript t and subscript q, set t = q to jump to S45, and execute recursively; when the recursion returns, restore subscript t and subscript q;

[0036] S49: Let q=q+1. If q is less than or equal to the number of test results, go to S46. Otherwise, if it is in recursion, return to S48. If it is not in recursion, go to S410.

[0037] S410: Add to middle;

[0038] S411: Let t = t + 1. If t is less than or equal to the number of test results, go to S43. Otherwise, output Specifically, in S4, the regional information amount is calculated for each region. The process is as follows:

[0039] S41: Calculate the regional uncertainty according to the following formula

[0040]

[0041] in represents the classification score of the kth object in the jth region, represents the kth object in the jth region;

[0042] S42: Get image feature map It is predicted by the backbone network, for the feature map area Perform average pooling to obtain the object feature vector f of each object in the image j,k ; Among them, the coordinate (x 1 ,y 1 ),(x 2 ,y 2 ) The object coordinates obtained by prediction The coordinates are changed by the following formula: is the original image size, is the feature map size, and (x 1 ,y 1 ),(x 2 ,y 2 ) are substituted into the formula for x and y to calculate the transformed coordinates;

[0043]

[0044] S43: Calculate the feature difference δ within the region class according to the following formula j,k ; c represents the object category, f j,k , represents the feature vector of the kth object in the jth region, represents the category of the kth object in the jth region, V represents the set that satisfies the corresponding conditions, represents the mean value of the eigenvector;

[0045]

[0046] S44: Calculate the feature difference δ within the region class according to the following formula j :

[0047]

[0048] S44: Calculate the regional information amount according to the following formula

[0049]

[0050] Among them, λ represents the coefficient that adjusts the proportion of uncertainty and characteristic variability.

[0051] Specifically, in S5, the process of calculating the dynamic weight ω of the sample is as follows, where Represents an indicator function, which takes 1 if the expression in the brackets is true, and 0 otherwise:

[0052]

[0053] in, Represents the category distribution of the labeled dataset , the probability of category c.

[0054] Compared with the prior art, the present invention has at least the following advantages:

[0055] 1. The present invention uses a simple and effective region division algorithm, which divides the results of object detection into different regions. In this way, the present invention utilizes the relationship between objects in the image. Then, for each region, the present invention uses regional uncertainty and regional feature differences to measure the value of the region, taking into account the uncertainty, similarity and diversity of objects in the region, and better represents the amount of information in the region. Finally, the present invention uses dynamic weights to correct the amount of image information obtained, alleviating the problem of category imbalance. The three modules jointly improve the final performance.

[0056] 2. The object detection active learning method based on region-level information invented in this paper proposes a new method that considers the connection between objects in the image by dividing the objects into regions. In this way, the method does not need to build additional model modules or rely on a specific network structure, which makes our method have good scalability and portability. The present invention has conducted a large number of experiments on different data sets and different object detection models. In all experimental environments, the present invention has obvious performance improvements compared with traditional object detection active learning algorithms. Experiments show that this method has certain advantages and good scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to more clearly illustrate the embodiments of the present invention and the technical solutions in the prior art, the drawings required for use in the embodiments 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 work.

[0058] Figure 1 It is a schematic diagram of the method of the present invention.

[0059] Figure 2 It is a flow chart of the method of the present invention.

[0060] Figure 3 It is an instance after object area division.

[0061] Figure 4 It is a visual display of the difference in features within a class.

[0062] Figure 5 It is the method dynamic weight module display.

[0063] Figure 6 This is the result of the method on the VOC 07 dataset.

[0064] Figure 7 This is the result of the method on the VOC 07+12 dataset.

[0065] Figure 8The experimental results of the present invention under the Faster RCNN network framework are shown.

[0066] Fig. 9 The experimental results of the present invention under the Retina network framework are shown. DETAILED DESCRIPTION

[0067] The present invention is described in further detail below.

[0068] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in the field should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0069] This paper designs an active learning method for object detection based on region-level information, referred to as RL-I. This method can consider the connection between objects in the image to a certain extent without adding additional network modules and without relying on a specific network structure. The designed method contains three main modules, namely region division, region information calculation, and dynamic weight module.

[0070] After dividing the image into different regions, the information content of the region is calculated from three aspects: regional uncertainty, regional intra-class feature difference, and regional feature difference. Among them, regional uncertainty reflects the overall uncertainty of all objects in the region. Regional feature difference reflects the diversity of features in the region. Regional intra-class feature difference can ignore the redundant features in the region and capture the misclassification in the region while selecting as many diverse features as possible.

[0071] The present invention will now be further described in conjunction with the accompanying drawings and specific implementation methods. Figure 2 As shown in FIG. 1 , the active learning method for object detection based on region-level information includes the following steps:

[0072] Step S1: Define the object detection problem and active learning problem as follows:

[0073] The given data set is divided into two categories, namely Represents a labeled dataset and uses Represents an unlabeled dataset, where represents a sample with a label, Indicates the corresponding label, represents an unlabeled sample, and The number is much greater than The number of

[0074] For a given image Model A set of triplets will be predicted to represent the detected objects in, Indicates the category to which the object belongs. represents the classification score of the object, Represents the bounding box coordinates of the object, which is a four-dimensional vector that describes the location of the object in the image.

[0075] For a given image According to the model Predicted object detection results The active learning method calculates the information content of the image, which expresses the learning value of the image. The higher the information content of the image, the more valuable it is to learn. Finally, the images are sorted according to their information content, and the most valuable samples are selected for annotation to realize the active learning process.

[0076] Step S2: On a small number of labeled datasets Training model For different detection models, the training process is different and can be completed according to the requirements of different detection models. At the same time, at the beginning of each round of active learning, the category distribution of the current labeled data set is calculated It is obtained by normalizing the number of objects covered by each category in the current labeled data set. If the probability distribution of a category is high, it means that there are more objects of this category in the data set.

[0077] Step S3: On the unlabeled dataset Using the model Perform inference to obtain the object detection results of the image Represents the detection result of the i-th object in the image.

[0078] S3-1: For a given image Feed it into the object detection model In the process, the detection results of the model are obtained. For different detection networks, the process is consistent. After the image is preprocessed, feature extracted, classified and regressed, the prediction results of the objects that may exist in the image are obtained.

[0079] S3-2: For the detection results of the model, there are interference items with low confidence. For example, objects with a classification score of 0.01 often have no reference value. These objects may be interference in the image. Retaining these objects will increase the amount of calculation and interfere with the results of the method. For this reason, a threshold θ is used 1Filter out objects whose classification scores are lower than the threshold to obtain the final object detection results The best threshold θ for different networks and datasets 1 The value of θ is different, but it does not significantly affect the model performance. 1 The value of can be simply set to 0.3.

[0080] Step S4: Based on the obtained bounding box coordinates Divide the objects in the image into regions and obtain the region division results j represents the jth region after region division.

[0081] S4-1: For a given test result Create a new bounding box set that has been selected Empty, create a new final area division result set is empty;

[0082] S4-2: let t=1;

[0083] S4-3: Create a new candidate region set If it is empty, take the tth result in the test result. If the bounding box is middle, Then go to S4-11, otherwise go to S4-4;

[0084] S4-4: Add to the set of selected bounding boxes and candidate region set middle;

[0085] S4-5: Let q=1;

[0086] S4-6: Take the qth result in the detection result If the bounding box is middle, Then go to S49, otherwise go to S4-7;

[0087] S4-7: Calculate the closeness of the two bounding boxes according to the following formula Indicates finding the area of ​​the region, and min() indicates finding the minimum value;

[0088]

[0089] S4-8: If the tightness Greater than the threshold θ 2 ,Will Add to and Temporarily save subscript t and subscript q, set t = q to jump to S4-5, and execute recursively; when the recursion returns, restore subscript t and subscript q;

[0090] S4-9: Let q=q+1. If q is less than or equal to the number of test results, go to S4-6. Otherwise, if in recursion, return to S48. If not in recursion, go to S4-10.

[0091] S4-10: Add to middle;

[0092] S4-11: Let t = t + 1. If t is less than or equal to the number of test results, go to S4-3. Otherwise, output Specifically, Figure 3 Different results are obtained after the region segmentation module in the display method segments the objects in the image.

[0093] Step S5: Calculate the regional information of each region after the region division The present invention calculates the information content of the region from three aspects, namely, regional uncertainty, regional intra-class feature difference, and regional feature difference. Among them, regional uncertainty reflects the overall uncertainty of all objects in the region. Regional feature difference reflects the diversity of features in the region. Regional intra-class feature difference can ignore the redundant features existing in the region and capture the misclassification existing in the region while selecting as many diverse features as possible.

[0094] S5-1: Calculate the regional uncertainty according to the following formula j represents the jth region, and k represents the kth object in the jth region.

[0095]

[0096] in represents the classification score of the kth object in the jth region, represents the kth object in the jth region; Indicates the uncertainty of an object. The larger its value, the more uncertain the model's prediction of the object is. The above formula calculates the mean uncertainty of each object in the region, which reflects the overall uncertainty of the objects in the region and reflects the value of the region from the perspective of uncertainty.

[0097] S5-2: Get image feature map It is predicted by the backbone network, for the feature map area Perform average pooling to obtain the object feature vector f of each object in the image j,k . Among them, the coordinate (x 1 ,y 1),(x 2 ,y 2 ) The object coordinates obtained by prediction It is obtained by the following coordinate transformation. is the original image size, is the feature map size, and (x 1 ,y 1 ),(x 2 ,y 2 ) are substituted into the formula for x and y to calculate the transformed coordinates. Since the feature map scale is generally different from the original image scale, the original image coordinates are converted to feature map coordinates through transformation, thereby extracting the features of the object in the feature map.

[0098]

[0099] S5-3: Calculate the intra-region feature difference δ according to the following formula j,k . c represents the object category, f j,k , represents the feature vector of the kth object in the jth region, represents the category of the kth object in the jth region, V represents the set that satisfies the corresponding conditions, represents the mean of the eigenvectors.

[0100]

[0101] The above formula calculates the feature difference of each category in the region through cosine similarity and converts the value to [0,1]. By calculating the feature difference of each category in the region, the redundant features in the region are ignored, making the algorithm more inclined to select regions with rich features in the category. Objects in these regions may be misclassified or have strong differences and are more valuable to learn. Specifically, Figure 4 It shows the ability of regional intra-class feature differences to ignore redundant features and capture misclassifications. For features in the same category in the same region, if there are redundant features, the object features are highly similar and the intra-class feature differences are low. On the contrary, if there are misclassified objects in the same category, the feature differences are large and the intra-class feature differences are high. S5-4: Calculate the regional intra-class feature difference δ according to the following formula j .

[0102]

[0103] The above formula calculates the feature differences between objects in the region through cosine similarity and converts the values ​​to [0,1]. By calculating the feature differences between objects in the region, the algorithm is more inclined to select regions with rich features, which considers the value of the region from the perspective of the overall feature diversity in the region.

[0104] S5-4: Calculate the regional information according to the following formula

[0105]

[0106] The final regional information is obtained by adding regional uncertainty, regional intra-class feature differences, and regional feature differences. λ represents the coefficient for adjusting the proportion of uncertainty and feature differences. When λ is 5, the best result will be obtained.

[0107] Step S6: Comprehensively calculate the value of all regions in the image to obtain the information content of the image

[0108] The maximum value of all the information in the region is used to represent the information content of the image. The calculation is as follows:

[0109]

[0110] Step S7: Calculate the dynamic weight ω of the sample, and the information content of the corrected image is

[0111] The object detection dataset contains multiple categories. Due to the large amount of data, each image contains different object categories. Therefore, there is a serious problem of imbalanced category distribution in the object detection dataset. Some object categories have a large number of training samples, while other object categories only have a small number of training samples. Therefore, the present invention designs a dynamic weight module, such as Figure 5 As shown, the sample weights are dynamically adjusted according to the category distribution of the current labeled data set, so that the weights of images containing rare categories are increased, while the weights of images containing common categories are reduced. Specifically, ω can be calculated by the following formula. Represents an indicator function, which takes the value 1 if the expression in the brackets is true, otherwise it takes the value 0.

[0112]

[0113] in, Represents the category distribution of the labeled dataset , the probability of category c.

[0114] Step S8: Arrange the corrected information volume corresponding to each image in descending order, take the first K images corresponding to the corrected information volume for annotation, remove the annotated images from the unlabeled dataset, and add them to the labeled dataset to update the labeled dataset.

[0115] S8-1: Sort the data in the unlabeled dataset according to the amount of image information, and select the K images with the largest amount of image information. Depending on different problems and different datasets, the value of K can be given arbitrarily. It is worth noting that in large-scale datasets, the value of K should not be too small, which will bring efficiency problems. In the PASCAL VOC 2007 dataset, we set K to 400. In the PASCAL VOC 07+12 dataset, we set K to 1000.

[0116] S8-2: The selected K images are handed over to humans for labeling, and the objects and object categories in the images are labeled. The labeling is done manually. It is worth noting that when testing the algorithm, in order to avoid the cost of manual labeling, a part of the labeled data set can be manually divided as an unlabeled data set to test the algorithm.

[0117] S8-3: Delete the K labeled images from the unlabeled dataset and add them to the labeled dataset.

[0118] Step S9: Repeat the process of step S2 to step S8 T times. In this way, the selection process is carried out for a total of T rounds and T*K samples are labeled.

[0119] First, retrain the model on the labeled dataset to obtain the model under the current labeled dataset At the same time, the category distribution of the labeled data set is calculated Use the classification results predicted by the current model to calculate the amount of information for each image According to the uncertainty of each image, select the most uncertain K images for annotation and update the data set; repeat the above process until the number of iterations T is reached, at which time T*K samples are annotated. For different models and different data sets, the choice of T is also arbitrary. T*K represents the final number of annotated samples, which represents the cost of annotation. Appropriately increasing the K value and reducing the T value can speed up the model training process and reduce the time cost of model training.

[0120] Step S10: Use the current labeled dataset to train in a supervised manner Finally, the optimal detection model is obtained. The current labeled data set includes the original labeled images and the newly added T*K labeled images.

[0121] Step S11: For a new image, use the optimal detection model to predict the new image and output the object detection result of the new image.

[0122] experiment:

[0123] The present invention uses the PASCAL VOC dataset as experimental data. The PASCAL VOC dataset contains 20 object categories. We use VOC 2007trainval and VOC 07+12trainval as training sets and test on VOC 2007. For fair comparison, we use the existing network framework to train the model. We compared a series of active learning methods. Including random selection, entropy-based selection, core set-based selection, and probability model-based selection. Among them, random selection randomly selects samples for labeling. Entropy-based selection selects samples with the largest entropy for labeling based on the predicted category probability vector of the object. Core set-based selection selects those samples that are farthest from the labeled dataset. The probability model-based method models the classification uncertainty and positioning uncertainty of the object, and then aggregates them for selection. Figure 6 The results of our method on the VOC 2007 dataset are shown. Figure 7 The results of the present invention on the VOC 07+12 dataset are shown. The present invention has achieved good results on different datasets.

[0124] pass Figure 6 It can be seen that the method of the present invention shows a more obvious performance improvement compared with the existing traditional active learning method under the VOC 2007 dataset, which verifies the effectiveness of our method.

[0125] pass Figure 7 It can be seen that the present invention shows a more obvious performance improvement compared with the existing traditional active learning method under the more complex VOC 07+12 dataset, which further verifies the effectiveness of our method.

[0126] In order to further verify the effectiveness of the method, the present invention is compared with other existing methods under two other different network frameworks, including a method based on positioning uncertainty and positioning closeness, a sub-supervised sample mining method, and a variational adversarial active learning method. Figure 8 The experimental results of the present invention under the Faster RCNN network framework are shown. Fig. 9 The experimental results of the present invention under the Retina network framework are shown. Under different detection models, the present invention has achieved good results.

[0127] pass Figure 8 It can be seen that the present invention has achieved a certain degree of leadership compared with the existing active learning methods under the Faster RCNN network framework, which verifies the effectiveness and scalability of our method.

[0128] pass Fig. 9It can be seen that the present invention has also achieved a certain degree of leadership compared with the existing active learning methods under the Retina network framework, which further verifies the effectiveness and scalability of our method.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. An active learning method for object detection based on region-level information, characterized in that: The steps include: S0: Divide the given data set into two categories, using Represents a labeled dataset and uses Represents an unlabeled dataset, where represents a sample with a label, Indicates the corresponding label, represents an unlabeled sample, and The number is much greater than the number of use Represents an object detection model. For a given image A set of triplets will be predicted to represent the detected objects in, Indicates the category to which the object belongs. represents the classification score of the object, represents the bounding box coordinates of the object, where Is a four-dimensional vector that describes the object in the image Position in S1: let a=1; S2: On training Count the number of objects in each category in the labeled data set, align and normalize to get the category distribution of the labeled data set S3: Used on each image Perform inference to obtain the object detection results of the corresponding image Represents the detection result of the i-th object in the image; S4: According to the obtained right Perform regional division on the objects in the corresponding image and obtain the regional division result represents the jth region after region division; Calculate the regional information of each region after the region is divided The maximum value of the regional information in all regions is selected as the information amount representing the image The information of all regions in the comprehensive image is obtained The amount of information corresponding to the image right Each image in the image is calculated to have the corresponding information S5: Calculation The dynamic weight ω of each image in the image is used to correct the amount of information corresponding to each image to S6: Arrange the corrected information amount corresponding to each image in descending order, take the first K images corresponding to the corrected information amount for annotation, remove the annotated images from the unlabeled dataset, and add them to the labeled dataset to update the labeled dataset; S7: If a>T, execute the next step, otherwise return to S2; S8: Supervised training using current labeled datasets Finally, the optimal detection model is obtained; S9: For a new image, use the optimal detection model to predict the new image and output the object detection result of the new image.

2. The active learning method for object detection based on region-level information as claimed in claim 1, characterized in that: The steps of obtaining the object detection result in S3 are as follows: For a given image Feed it into the object detection model The detection result of the model is obtained; for the detection result of the model, the category probability value of the object in the detection result is used to represent the confidence of the object, and the objects with confidence lower than the threshold θ1 are filtered out to obtain the final object detection result.

3. The active learning method for object detection based on region-level information as claimed in claim 2, characterized in that: The process of region division performed by S4 is as follows: S41: For a given test result Create a new bounding box set that has been selected Empty, create a new final area division result set is empty; S42: let t=1; S43: Create a new candidate region set If it is empty, take the tth result in the test result. If the bounding box is middle, Then go to S411, otherwise go to S44; S44: Add to the set of selected bounding boxes and candidate region set middle; S45: set q=1; S46: Get the qth result in the detection result If the bounding box is middle, Then go to S49, otherwise go to S47; S47: Calculate the closeness of the two bounding boxes according to the following formula Indicates finding the area of ​​the region, and min() indicates finding the minimum value; S48: If the tightness is greater than the threshold θ2, Add to and Temporarily save subscript t and subscript q, set t = q to jump to S45, and execute recursively; when the recursion returns, restore subscript t and subscript q; S49: Let q=q+1. If q is less than or equal to the number of test results, go to S46. Otherwise, if it is in recursion, return to S48. If it is not in recursion, go to S410. S410: Add to middle; S411: Let t = t + 1. If t is less than or equal to the number of test results, go to S43. Otherwise, output 4. The active learning method for object detection based on region-level information as claimed in claim 3, characterized in that: In S4, the regional information amount is calculated for each region. The process is as follows: S41: Calculate the regional uncertainty according to the following formula in represents the classification score of the kth object in the jth region, represents the kth object in the jth region; S42: Get image feature map It is predicted by the backbone network, for the feature map area Perform average pooling to obtain the object feature vector f of each object in the image j,k ; Among them, the coordinates (x1, y1), (x2, y2) are the predicted object coordinates The coordinates are changed by the following formula: is the original image size, is the feature map size, and (x1, y1) and (x2, y2) are substituted into the formula for x and y to calculate the transformed coordinates; S43: Calculate the feature difference δ within the region class according to the following formula j,k ; c represents the object category, f j,k , represents the feature vector of the kth object in the jth region, represents the category of the kth object in the jth region, V represents the set that satisfies the corresponding conditions, represents the mean value of the eigenvector; S44: Calculate the feature difference δ within the region class according to the following formula j : S44: Calculate the regional information amount according to the following formula Among them, λ represents the coefficient that adjusts the proportion of uncertainty and characteristic variability.

5. The active learning method for object detection based on region-level information as claimed in claim 4, characterized in that: In S5, the process of calculating the dynamic weight ω of the sample is as follows, where Represents the indicator function: in, Represents the category distribution of the labeled dataset , the probability of category c.

Citation Information

Patent Citations

  • Remote sensing target detection active sampling method guided by category semantic information

    CN115861738A

  • Object detection active learning method based on classification consistency before and after object regression

    CN117372822A

  • Localization method and system based on deep learning

    WO2020173036A1