Active domain adaptive learning-based foggy day target semi-automatic labeling method

By aligning the normal weather and foggy day data sets and building a deep supervision network, selecting representative samples for annotation, the problem of low labeling efficiency and accuracy of the foggy day object detection data set is solved, and an efficient and low-cost annotation process is achieved.

CN118941892APending Publication Date: 2024-11-12CHINA FORESTRY STAR BEIJING TECH INFORMATION CO LTD
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410980899.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing foggy day object detection data set labeling methods have problems such as huge time and energy of manual labeling, poor inconsistency and accuracy, and inapplicable conventional methods, especially in the case of high image complexity in foggy days.

Method used

Using the method based on active domain adaptation learning, the foggy day simulation data set is generated by aligning the normal weather data set with the foggy day data set to be marked, and a deeply supervised target detection network is built, representative samples are selected for manual annotation, and fine-tuning is used. The network predicts the remaining samples and is calibrated.

Benefits of technology

It improves labeling efficiency and accuracy, reduces manual labeling costs, and reduces sample resampling of model prediction uncertainty, and is suitable for high-complex foggy images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941892A_ABST
    Figure CN118941892A_ABST
Patent Text Reader

Abstract

A foggy day target semi-automatic labeling method based on active domain adaptive learning belongs to the field of target detection, and comprises the following steps: constructing a foggy day simulation data set; constructing and training a deep supervised target detection network; obtaining a foggy day simulation data set feature set and a to-be-labeled foggy day data set feature set; obtaining the feature representation of the foggy day simulation data set and the feature representation of the to-be-labeled foggy day data set; selecting a sample set containing representative features in the foggy day data set to be labeled; manually annotating the selected representative sample set of the foggy day data set to be annotated, combining the manually annotated sample set with the foggy day simulation data set, and then finely adjusting the target detection network; and predicting a residual general sample of the foggy day data set to be labeled by using the fine-tuned target detection network, and manually analyzing and calibrating a prediction result of the foggy day data set to be labeled. The method has the advantages of being high in efficiency, low in cost, low in manual dependence, high in labeling consistency and accuracy and the like.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a semi-automatic labeling method for targets in foggy weather based on active domain adaptation learning. Background Art

[0002] The foggy weather target detection dataset is of great significance in the fields of computer vision, autonomous driving, intelligent transportation, and intelligent security systems. At present, the production of target detection datasets usually relies on manual annotation, which has many disadvantages. On the one hand, when the amount of data is large, the workload of the annotation personnel is huge and the annotation cycle is long; on the other hand, compared with the dataset of conventional weather images, the targets in foggy weather images are usually unclear and the edges are blurred, which makes it more difficult for the annotation personnel to identify and annotate the targets, and it is more likely to mislabel or miss the labels. Therefore, providing a semi-automatic annotation method for foggy weather target detection datasets has become a technical problem that needs to be solved urgently by those skilled in the art.

[0003] In the prior art, a Chinese patent with publication number CN114724015A discloses a target detection method based on active domain adaptive learning to reduce the need for labeling. This method combines the common characteristics of active learning and domain adaptive learning. First, a DAFaster RCNN target detection model with resnet50 as the backbone is constructed, and the network model is initialized using labeled source domain data and part of the unlabeled target domain data set; then the unlabeled target domain data set is divided into an unlabeled sample pool and a test set; according to the set active sampling strategy, samples are collected from the unlabeled sample pool, and after expert labeling, they are updated to the training set of the target detection model. Finally, the updated training set is used to iteratively fine-tune the target detection model, and the above target domain sample selection and model iteration steps are repeated until the target domain test set meets the preset indicators. This method adopts a sampling algorithm based on the minimum confidence strategy extension, resamples samples that cause high prediction uncertainty to the model, fails to deeply explore the characteristics of such samples, and needs to continuously mine samples through iteration.

[0004] The existing annotation methods for foggy object detection datasets have the following shortcomings:

[0005] (1) The edges of objects in foggy images are blurred. Manually labeling each image requires a huge amount of time and effort, which requires more manpower and time.

[0006] (2) In foggy conditions, different labelers may have different recognition of the same target, making it difficult to maintain consistency and accuracy in labeling.

[0007] (3) The data distribution of foggy images is very different from that of normal weather images, and foggy images are highly complex, so conventional semi-automatic target detection and annotation methods are not applicable. Summary of the invention

[0008] In order to solve the problems existing in the prior art, the present invention provides a semi-automatic labeling method for foggy targets based on active domain adaptation learning.

[0009] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0010] The present invention provides a method for semi-automatic labeling of foggy targets based on active domain adaptation learning, which mainly includes the following steps:

[0011] Step 1: Align the categories of the normal weather dataset and the foggy dataset to be labeled; generate a foggy simulation dataset based on the atmospheric scattering model;

[0012] Step 2: Build and train a deep supervised object detection network;

[0013] Step 3: respectively obtain a feature set of the foggy weather simulation data set and a feature set of the foggy weather data set to be labeled; respectively obtain a feature representation of the foggy weather simulation data set and a feature representation of the foggy weather data set to be labeled; and select a sample set containing representative features in the foggy weather data set to be labeled;

[0014] Step 4: Manually label the representative sample set of the foggy data set to be labeled selected in step 3, merge the manually labeled sample set with the foggy simulation data set, and fine-tune the target detection network;

[0015] Step 5: Use the fine-tuned object detection network to predict the remaining general samples of the foggy dataset to be labeled, and manually analyze and calibrate the prediction results of the foggy dataset to be labeled.

[0016] The mathematical expression of the foggy day simulation data set is as follows:

[0017] I(x)=J(x)t(x)+A(1-t(x))

[0018] t(x)=e -βd(x)

[0019]

[0020] Where x represents the image sample in the dataset; J(x) represents the normal weather dataset; i(x) represents the foggy simulated dataset; A represents the global background brightness; t(x) represents the transmittance; β represents the scattering coefficient of the atmosphere; d(x) represents the distance from the object to the camera; ρ represents the Euclidean distance from the current pixel to the center pixel; w and h represent the width and height of the current image, respectively.

[0021] The target detection labels of the foggy weather simulation image samples in the foggy weather simulation data set are consistent with the target detection labels of the normal weather image samples in the normal weather data set.

[0022] Among them, the target detection part of the deep supervised target detection network includes: a non-fog feature extraction module, a regression module, a classification module and a post-processing module; the non-fog feature extraction module is used to extract the common features of the image, the regression module is used to locate the position of the target, the classification module is used to predict the category of the target, and the post-processing module is used to remove redundant prediction boxes.

[0023] Among them, a deep classification supervision module is added in the training stage of the deep supervised target detection network, and the deep classification supervision module includes: a gradient reversal layer and a fog classification module; the gradient reversal layer reverses the gradient of the fog classification module and propagates it to the feature extractor in the back propagation stage of network training, so as to supervise the feature extractor to learn the common features in the image except weather features; the fog classification module is used to predict whether the training image belongs to a normal weather data set or a foggy day simulation data set.

[0024] Among them, in the training stage of the deep supervised target detection network, a normal weather data set and a foggy day simulation data set are used for mixed training. Based on the target detection labels, the category labels of the normal weather image samples in the normal weather data set and the foggy day simulation image samples in the foggy day simulation data set are set to 0 and 1 respectively.

[0025] In step 3, the non-fog feature extraction module in the deep supervised target detection network is used to process all image samples in the foggy simulation dataset to obtain the feature set of the foggy simulation dataset. The non-fog feature extraction module in the deep supervised target detection network is used to process all image samples in the foggy dataset to be labeled, and the feature set of the foggy dataset to be labeled is obtained.

[0026]

[0027] Among them, F e represents the feature extractor, s represents the foggy simulated dataset, t represents the foggy dataset to be labeled, and N s Represents the number of samples in the foggy day simulation data set, N t Indicates the number of samples in the foggy dataset to be labeled, represents the i-th sample of the foggy day simulation data set, represents the jth sample of the foggy dataset to be labeled, represents the characteristics of the i-th sample of the foggy day simulation dataset, Represents the features of the jth sample in the foggy dataset to be labeled.

[0028] In step 3, the feature clustering center set of the foggy weather simulation data set is obtained based on the K-means feature clustering algorithm. And the feature cluster center set of the foggy dataset to be labeled C s , C t They respectively represent the number of cluster centers of the hyperparameters of the Kmeans feature clustering algorithm used to extract feature representations of the foggy simulation dataset and the foggy dataset to be labeled.

[0029] In step 3, the difference set between the feature cluster center of the foggy weather dataset to be labeled and the feature cluster center of the foggy weather simulation dataset is calculated.

[0030]

[0031] in, represents the i-th feature cluster center of the foggy day simulation data set; Represents the jth feature cluster center of the foggy data set to be labeled; Indicates the difference between the jth feature cluster center of the foggy data set to be labeled and the feature cluster center of the simulated foggy data set;

[0032] Calculate the number of samples N to be selected for the foggy weather dataset to be labeled:

[0033]

[0034] in, Indicates the number of samples in the category corresponding to the feature cluster center with the largest difference between the foggy weather dataset to be labeled and the foggy weather simulation dataset;

[0035] According to the difference value set, calculate the number of samples to be selected from the sample set corresponding to the cluster center of each foggy dataset to be labeled:

[0036]

[0037] in, Indicates the number of samples in the category corresponding to the j-th cluster center of the foggy data set to be labeled;

[0038] Finally, in the sample set corresponding to the feature cluster center of each foggy data set to be labeled, a corresponding number of samples are selected from large to small according to the difference between the sample and the cluster center of the current set.

[0039] Among them, in step 4, the deep classification supervision module is not used in the process of fine-tuning the target detection network.

[0040] The beneficial effects of the present invention are:

[0041] The present invention provides a semi-automatic labeling method for foggy weather targets based on active domain adaptation learning, which realizes the labeling of unlabeled foggy weather data with minimal manual cost based on the labeled normal weather data set. The present invention first aligns the categories of the normal weather data set and the foggy weather data set to be labeled; generates a foggy weather simulation data set based on the atmospheric scattering model; constructs and trains a deep supervised target detection network; obtains a feature set of the foggy weather simulation data set and a feature set of the foggy weather data set to be labeled respectively; obtains a feature representation of the foggy weather simulation data set and a feature representation of the foggy weather data set to be labeled respectively; selects a sample set containing representative features in the foggy weather data set to be labeled; manually labels the selected representative sample set of the foggy weather data set to be labeled, and fine-tunes the deep supervised target detection network after merging the manually labeled sample set with the foggy weather simulation data set; uses the fine-tuned deep supervised target detection network to predict the remaining general samples of the foggy weather data set to be labeled, and manually analyzes and calibrates the prediction results of the foggy weather data set to be labeled to obtain the final foggy weather data set.

[0042] Compared with the prior art, the semi-automatic labeling method for foggy targets based on active domain adaptation learning provided by the present invention has the following advantages:

[0043] 1. High efficiency: In addition to a small amount of labeling work, labelers only need to correct a small number of samples that are incorrectly predicted by the model, which saves a lot of time compared to labeling all samples.

[0044] 2. Low cost: It reduces the human resource investment required for pure manual labeling, thereby reducing the labeling cost.

[0045] 3. Low manual dependence, improved consistency and accuracy of annotation: According to the characteristics of the existing normal weather data set and the foggy day data set to be annotated, samples that need to be manually annotated are automatically selected without relying on expert experience, which improves the consistency and accuracy of annotation.

[0046] 4. The present invention is applicable to highly complex foggy images. It does not require iterative sample mining, thus avoiding sample resampling that causes high prediction uncertainty in the model and improving the accuracy of sample annotation. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow chart of a method for semi-automatic labeling of targets in foggy weather based on active domain adaptation learning provided by the present invention.

[0048] Figure 2 A block diagram of the structure of a deeply supervised object detection network. DETAILED DESCRIPTION

[0049] The present invention provides a method for semi-automatic labeling of foggy targets based on active domain adaptation learning, which mainly includes the following steps:

[0050] 1) Construct a foggy day simulation data set;

[0051] 2) Build and train a deep supervised object detection network;

[0052] 3) Select representative samples from the foggy dataset to be labeled;

[0053] 4) Dataset aggregation and fine-tuning of deep supervised object detection networks;

[0054] 5) Manual verification.

[0055] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] like Figure 1 As shown, the present invention provides a semi-automatic labeling method for foggy targets based on active domain adaptation learning, which specifically includes the following steps:

[0057] Step 1: Construct a foggy day simulation dataset;

[0058] The specific operation process is as follows:

[0059] S1.1: Align the categories of the normal weather dataset and the foggy weather dataset to be labeled.

[0060] S1.2: Generate a foggy day simulation data set based on the atmospheric scattering model. The specific implementation process is as follows:

[0061] I(x)=J(x)t(x)+A(1-t(x))

[0062] t(x)=e -βd(x)

[0063]

[0064] Where x represents the image sample in the dataset; J(x) represents the normal weather dataset; I(x) represents the foggy simulated dataset; A represents the global background brightness; t(x) represents the transmittance; β represents the scattering coefficient of the atmosphere; d(x) represents the distance from the object to the camera; ρ represents the Euclidean distance from the current pixel to the center pixel; w and h represent the width and height of the current image, respectively.

[0065] In the process of generating the foggy weather simulation data set, the global background light brightness A is preferably taken as 0.5, the atmospheric scattering coefficient β is preferably taken as a random number between 0.07 and 0.12, and the target detection labels of the foggy weather simulation image samples in the foggy weather simulation data set are consistent with the target detection labels of the normal weather image samples in the normal weather data set.

[0066] Step 2: Build and train a deep supervised object detection network;

[0067] Build a deep supervised object detection network. Figure 2 As shown, the target detection part of the deep supervised target detection network constructed by the present invention mainly includes four parts: non-fog feature extraction module, regression module, classification module, and post-processing module; wherein, the non-fog feature extraction module is mainly used to extract the general features of the image, the regression module is mainly used to locate the position of the target, the classification module is mainly used to predict the category of the target, and the post-processing module is mainly used to remove the redundant prediction boxes of the prediction. A deep classification supervision module is added in the training stage, and the deep classification supervision module is mainly used to supervise the non-fog feature extraction module to have the ability to learn the non-fog features in the image for target detection. The deep classification supervision module mainly includes: a gradient reversal layer (GRL) and a foggy classification module; wherein, the gradient reversal layer reverses the gradient of the foggy classification module in the back propagation stage of the network training and propagates it to the feature extractor to supervise the feature extractor to learn the general features in the image except the weather features; the foggy classification module is used to identify whether the image belongs to a normal weather data set or a foggy simulation data set.

[0068] The object detection network constructed by the present invention has a deep supervision, and the specific implementation process of the object detection part is as follows:

[0069] Firstly, the common features of foggy simulated image samples in the foggy simulated data set and the common features of normal weather image samples in the normal weather data set are extracted using the non-fog feature extraction module. Then, the position of the target is located using the regression module, and the category of the target is predicted using the classification module. Finally, the redundant prediction boxes are removed through the post-processing module to obtain the target detection results.

[0070] The specific implementation process of the deep supervised target detection network constructed by the present invention in the training phase is as follows:

[0071] Based on the target detection labels, the category labels of the normal weather image samples in the normal weather dataset and the foggy weather simulation image samples in the foggy weather simulation dataset are set to 0 and 1 respectively.

[0072] In the training phase, the target detection network is trained based on the mixed target detection labels of the normal weather dataset and the foggy weather simulation dataset. On this basis, a deep supervised classification module is introduced to supervise the feature extractor of the target detection network to learn the general features in the image except for weather features based on the category labels of the normal weather dataset and the foggy weather simulation dataset. The foggy weather classification module predicts whether the image belongs to the normal weather dataset or the foggy weather simulation dataset based on the features obtained by the feature extraction module, and calculates the gradient based on the loss of the prediction result and the label; the gradient reversal layer reverses the gradient of the foggy weather classification module and propagates it to the feature extraction module during the back propagation process of model optimization.

[0073] Step 3: Select representative samples from the foggy data set to be labeled;

[0074] S3.1: Obtain a feature set of a foggy weather simulation data set and a feature set of a foggy weather data set to be labeled respectively.

[0075] The specific operation process is as follows:

[0076] The non-fog feature extraction module in the deep supervised target detection network is used to process all image samples in the foggy simulation dataset to obtain the feature set of the foggy simulation dataset. The non-fog feature extraction module in the deep supervised target detection network is used to process all image samples in the foggy dataset to be labeled, and the feature set of the foggy dataset to be labeled is obtained. The specific implementation process is as follows:

[0077]

[0078] Among them, F e represents the feature extractor, s represents the foggy simulated dataset, t represents the foggy dataset to be labeled, and N s Represents the number of samples in the foggy day simulation data set, N t Indicates the number of samples in the foggy dataset to be labeled, represents the i-th sample of the foggy day simulation data set, represents the jth sample of the foggy dataset to be labeled, represents the characteristics of the i-th sample of the foggy day simulation dataset, Represents the features of the jth sample in the foggy dataset to be labeled.

[0079] S3.2: Obtain feature representation of the foggy weather simulation data set and feature representation of the foggy weather data set to be labeled respectively.

[0080] The specific operation process is as follows:

[0081] Based on the K-means feature clustering algorithm, the feature clustering center set of the foggy weather simulation data set is obtained. And the feature cluster center set of the foggy data set to be labeled Among them C s , C t They represent the number of cluster centers of the hyperparameters of the Kmeans feature clustering algorithm used to extract the feature representation of the foggy simulation dataset and the foggy dataset to be labeled, and their values ​​are C s =10, C t =20.

[0082] S3.3: Select a set of samples containing representative features in the foggy weather dataset to be labeled.

[0083] The specific operation process is as follows:

[0084] First, calculate the difference set between the feature clustering center of the foggy weather dataset to be labeled and the feature clustering center of the foggy weather simulation dataset The specific implementation process is as follows:

[0085]

[0086] in, represents the i-th feature cluster center of the foggy day simulation data set; Represents the jth feature cluster center of the foggy data set to be labeled; It represents the difference between the jth feature cluster center of the foggy dataset to be labeled and the feature cluster center of the simulated foggy dataset.

[0087] Then, the number of samples N to be selected for the foggy weather dataset to be labeled is calculated. The specific calculation formula is as follows:

[0088]

[0089] in, Indicates the number of samples in the category corresponding to the feature cluster center with the largest difference between the foggy weather dataset to be labeled and the foggy weather simulation dataset.

[0090] Then, the number of samples to be selected from the sample set corresponding to the cluster center of each foggy dataset to be labeled is calculated based on the difference value set. The specific implementation process is as follows:

[0091]

[0092] in, Indicates the number of samples in the category corresponding to the j-th cluster center of the foggy data set to be labeled.

[0093] Finally, in the sample set corresponding to the feature cluster center of each foggy data set to be labeled, a corresponding number of samples are selected from large to small according to the difference between the sample and the cluster center of the current set.

[0094] Step 4: Dataset aggregation and fine-tuning of target detection network;

[0095] The representative sample set of the foggy weather data set to be labeled selected in step 3 is manually labeled, and the manually labeled sample set is merged with the foggy weather simulation data set to fine-tune the target detection network. In the process of fine-tuning the target detection network, the deep classification supervision module is not used.

[0096] Step 5: Manual verification;

[0097] The fine-tuned object detection network is used to predict the general samples of the remaining foggy dataset to be labeled, and the prediction results of the foggy dataset to be labeled are manually analyzed and calibrated.

[0098] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A semi-automatic labeling method for foggy targets based on active domain adaptation learning, characterized in that: The following steps are involved: Step 1: Align the categories of the normal weather dataset and the foggy dataset to be labeled; generate a foggy simulation dataset based on the atmospheric scattering model; Step 2: Build and train a deep supervised object detection network; Step 3: respectively obtain a feature set of the foggy weather simulation data set and a feature set of the foggy weather data set to be labeled; respectively obtain a feature representation of the foggy weather simulation data set and a feature representation of the foggy weather data set to be labeled; Select a sample set containing representative features in the foggy data set to be labeled; Step 4: Manually label the representative sample set of the foggy data set to be labeled selected in step 3, merge the manually labeled sample set with the foggy simulation data set, and fine-tune the target detection network; Step 5: Use the fine-tuned object detection network to predict the remaining general samples of the foggy dataset to be labeled, and manually analyze and calibrate the prediction results of the foggy dataset to be labeled.

2. According to claim 1, a method for semi-automatic labeling of foggy targets based on active domain adaptation learning is characterized in that: The mathematical expression of the foggy day simulation data set is as follows: I(x)=J(x)t(x)+A(1-t(x)) t(x)=e -βd(x) Where x represents the image sample in the dataset; J(x) represents the normal weather dataset; I(x) represents the foggy simulated dataset; A represents the global background brightness; t(x) represents the transmittance; β represents the scattering coefficient of the atmosphere; d(x) represents the distance from the object to the camera; ρ represents the Euclidean distance from the current pixel to the center pixel; w and h represent the width and height of the current image, respectively.

3. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 1 is characterized in that: The target detection labels of the foggy weather simulation image samples in the foggy weather simulation data set are consistent with the target detection labels of the normal weather image samples in the normal weather data set.

4. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 1 is characterized in that: The target detection part of the deep supervised target detection network includes: a non-fog feature extraction module, a regression module, a classification module and a post-processing module; the non-fog feature extraction module is used to extract common features of the image, the regression module is used to locate the position of the target, the classification module is used to predict the category of the target, and the post-processing module is used to remove redundant prediction boxes.

5. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 4 is characterized in that: A deep classification supervision module is added during the training phase of the deep supervised target detection network, and the deep classification supervision module includes: a gradient reversal layer and a fog classification module; during the back propagation phase of network training, the gradient reversal layer reverses the gradient of the fog classification module and propagates it to the feature extractor to supervise the feature extractor to learn common features in the image except weather features; the fog classification module is used to predict whether the training image belongs to a normal weather data set or a foggy day simulation data set.

6. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 5 is characterized in that: During the training phase of the deeply supervised target detection network, a normal weather dataset and a foggy day simulation dataset are used for mixed training. Based on the target detection labels, the category labels of the normal weather image samples in the normal weather dataset and the foggy day simulation image samples in the foggy day simulation dataset are set to 0 and 1, respectively.

7. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 1 is characterized in that: In step 3, the non-fog feature extraction module in the deep supervised target detection network is used to process all image samples in the foggy simulation dataset to obtain the feature set of the foggy simulation dataset. The non-fog feature extraction module in the deep supervised target detection network is used to process all image samples in the foggy dataset to be labeled, and the feature set of the foggy dataset to be labeled is obtained. Among them, F e represents the feature extractor, s represents the foggy simulated dataset, t represents the foggy dataset to be labeled, and N s Represents the number of samples in the foggy day simulation data set, N t Indicates the number of samples in the foggy dataset to be labeled, represents the i-th sample of the foggy day simulation data set, represents the jth sample of the foggy dataset to be labeled, represents the characteristics of the i-th sample of the foggy day simulation dataset, Represents the features of the jth sample in the foggy dataset to be labeled.

8. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 1, characterized in that: In step 3, the feature clustering center set of the foggy weather simulation data set is obtained based on the K-means feature clustering algorithm. And the feature cluster center set of the foggy dataset to be labeled C s , C t They respectively represent the number of cluster centers of the hyperparameters of the Kmeans feature clustering algorithm used to extract feature representations of the foggy simulation dataset and the foggy dataset to be labeled.

9. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 1, characterized in that: In step 3, the difference set between the feature cluster center of the foggy weather dataset to be labeled and the feature cluster center of the foggy weather simulation dataset is calculated. in, represents the i-th feature cluster center of the foggy day simulation data set; Represents the jth feature cluster center of the foggy data set to be labeled; Indicates the difference between the feature cluster center of the foggy weather dataset to be labeled and the feature cluster center of the simulated foggy weather dataset; Calculate the number of samples N to be selected for the foggy weather dataset to be labeled: in, Indicates the number of samples in the category corresponding to the feature cluster center with the largest difference between the foggy weather dataset to be labeled and the foggy weather simulation dataset; According to the difference value set, calculate the number of samples to be selected from the sample set corresponding to the cluster center of each foggy dataset to be labeled: in, Indicates the number of samples in the category corresponding to the j-th cluster center of the foggy data set to be labeled; Finally, in the sample set corresponding to the feature cluster center of each foggy data set to be labeled, a corresponding number of samples are selected from large to small according to the difference between the sample and the cluster center of the current set.

10. The method for semi-automatic labeling of foggy targets based on active domain adaptation learning according to claim 1, characterized in that: In step 4, the deep classification supervision module is not used in the process of fine-tuning the object detection network.

Citation Information

Cited By

  • Semi-automatic labeling method suitable for single object

    CN121259825A