Remote sensing image annotation optimization method based on dynamic confidence coefficient and Gaussian clustering screening

Abstract

The invention particularly relates to a remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening, which comprises the following steps of: reading input image data, and respectively generating a strong enhancement image and a weak enhancement image from the image data; constructing a pre-trained classification network, inputting the strong enhancement image and the weak enhancement image into the pre-trained classification network for pre-training, preliminarily learning image information and constructing a dynamic confidence threshold; in the formal training stage, samples are divided into a clean set, a suspected set, a difficult set and a noise set according to a dynamic confidence threshold and Gaussian clustering screening; carrying out model training by adopting a differentiated regularization strategy aiming at different sample sets; updating a dynamic confidence coefficient threshold value and repeating iterative training; and a test stage: inputting the test set into the trained classification network and outputting a labeling result. According to the method, the problem that noise samples are difficult to extract is solved, the noise samples can be segmented more finely, misleading of the noise samples to the model is reduced, and therefore a better labeling result is achieved.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing image annotation, and in particular to a remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening. Background Art

[0002] With the widespread application of deep learning technologies in fields such as image recognition and natural language processing, the impact of data annotation quality on model performance has become increasingly prominent. However, in real-world scenarios, labeled data often contains noisy labels due to manual annotation errors, crowdsourced annotation noise, or limitations of automated annotation tools. Research has shown that deep neural networks, due to their powerful memory capacity, are prone to overfitting to noisy labels, resulting in a significant decrease in model generalization performance. Traditional supervised learning paradigms assume perfect label accuracy, making them difficult to cope with the noise interference that is prevalent in real-world scenarios.

[0003] To mitigate the negative impact of noisy labels on model training, researchers have proposed various noisy label learning methods. Noise-robust loss functions, such as symmetric loss and generalized cross-entropy loss, reduce the model's sensitivity to noisy labels by modifying the loss function. However, their performance relies heavily on assumptions about the noise type (e.g., symmetric noise or class-correlated noise), making them difficult to adapt to complex noisy scenarios. Noisy sample detection and purification methods, based on the assumption that clean samples have lower loss, dynamically identify potential noisy samples and exclude or downweight their training impact, such as low-loss sample screening and confidence threshold filtering. However, these methods may mistakenly delete difficult samples, exacerbating class imbalance. Noisy label correction techniques attempt to directly correct noisy labels, such as through label correction networks or meta-learning to dynamically adjust annotations. However, the correction process is susceptible to the uncertainty of the noise distribution and carries the risk of error accumulation. Although existing methods have made some progress, they still face key challenges. First, most methods require a pre-set noise distribution (e.g., uniform noise, flipped noise), making them difficult to adapt to the complex noise patterns found in real-world scenarios, such as mixed noise or sample-feature-dependent noise patterns. Second, the loss distributions of noisy samples overlap with those of truly difficult samples (such as boundary samples and long-tail category samples), and existing filtering strategies can lead to insufficient model learning for difficult samples. Furthermore, in online learning scenarios, noise can change over time or due to data distribution drift, making static noise processing strategies difficult to maintain and be effective.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0005] The present invention provides a remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening, which is used to solve the problem of difficulty in extracting noise samples during the annotation process. It can segment the noise samples more finely and reduce the misleading effect of noise samples on the model, thereby achieving better image annotation learning and prediction effects.

[0006] Other features and advantages of the present invention will become apparent from the following detailed description, or may be learned in part by practice of the present invention.

[0007] According to a first aspect of the present invention, a remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening is provided, the method comprising:

[0008] Read the input image data, and generate differentiated strongly enhanced images and weakly enhanced images through two data enhancement methods: strong enhancement and weak enhancement.

[0009] Build and initialize a pre-trained classification network, where the classification network is a ResNet model;

[0010] Input the strongly enhanced image and the weakly enhanced image into the pre-trained classification network for pre-training, preliminarily learn the image information and construct a dynamic confidence threshold;

[0011] In the formal training phase, the samples are divided into clean set C, suspected set S, difficult set H and noise set N according to the dynamic confidence threshold;

[0012] Use differentiated regularization strategies for model training for different sample sets;

[0013] Update the dynamic confidence threshold and repeat the iterative training;

[0014] In the testing phase, the test set is input into the trained classification network to output the label prediction results.

[0015] In some exemplary embodiments, the specific steps of initializing the pre-trained classification network include:

[0016] Set the classification category, set the optimizer to SGD optimizer, and initialize the number of pre-training rounds and training rounds;

[0017] Initialize cross entropy loss, adaptive weighted loss function, dynamic penalty generalized cross entropy loss and binary cross loss.

[0018] In some exemplary embodiments, the specific steps of preliminarily learning image information and constructing a dynamic confidence threshold include:

[0019] Preliminary learning is performed on strongly enhanced images and weakly enhanced images in batches, and the initial cross entropy loss is used for loss correction;

[0020] The threshold is dynamically updated according to the prediction confidence of the sample. The formula is:

[0021] τ(t)=λτ(t-1)+(1-λ)P(t),τ(0)=0

[0022] Among them, τ(t) is the current round threshold, P(t) is the confidence of the predicted category given by the current round model, and λ is a hyperparameter.

[0023] In some exemplary embodiments, the specific step of dividing the samples into a clean set C, a suspected set S, a difficult set H, and a noise set N according to the dynamic confidence threshold includes:

[0024] The samples are divided into the initial screening clean set C' and the difficult set H according to the dynamic confidence threshold;

[0025] The initial clean set C' is further divided into a clean set C and a suspected set S through the mixed Gaussian clustering redistribution module;

[0026] The remaining samples are pseudo-labeled to generate a noise set N.

[0027] In some exemplary embodiments, the specific steps of the mixture Gaussian clustering redistribution module include:

[0028] Gaussian distribution modeling is performed on the samples in the initial screening clean set by category, and the mean and variance of the sample distance from the cluster center are calculated;

[0029] A threshold is set to filter out samples that deviate from the cluster center and divide them into the suspected set S, and the rest are divided into the clean set C.

[0030] In some exemplary embodiments, the specific steps of using differentiated regularization strategies for different sample sets to perform model training include:

[0031] The clean set C is trained using standard cross entropy loss;

[0032] Adaptive weighted loss function is used to train the suspected set S;

[0033] The difficult set H is trained using dynamic penalty generalized cross entropy loss;

[0034] The Mixup strategy is used to generate mixed samples for the noise set N, and the training is performed using binary cross loss.

[0035] In some exemplary embodiments, the specific steps of the Mixup strategy include:

[0036] Mix the noise set N with other sample sets (C, S, H) to generate mixed samples;

[0037] Train on mixed samples, calculate binary cross loss and update model parameters.

[0038] According to a second aspect of the present invention, a storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening described in the first aspect is implemented.

[0039] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored. When the computer program is executed by a processor, the remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening described in the first aspect is implemented.

[0040] According to a fourth aspect of the present invention, there is provided an electronic device, comprising:

[0041] processor; and

[0042] a memory for storing executable instructions of the processor;

[0043] Wherein, the processor is configured to implement the remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening described in the first aspect by executing the executable instructions.

[0044] The remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening provided by the embodiments of the present invention utilizes a Gaussian clustering redistribution module. This initial clean set, initially screened using threshold partitioning, is further divided into a clean set and a suspected set based on the intrinsic properties of the samples. This reduces the potential for misleading the model caused by noise samples in the clean set due to incorrect partitioning. Furthermore, a new dynamic loss is set for the suspected set, and a more adaptive dynamic robust loss is adopted for the difficult set. The resulting effect directly impacts metrics such as the accuracy of clean labels and final precision.

[0045] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The accompanying drawings are incorporated into and constitute a part of this specification, illustrate embodiments consistent with the present invention, and together with the description, serve to explain the principles of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and it is clear that those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0047] Figure 1 This is a flow chart of the remote sensing image annotation optimization method based on dynamic confidence and mixed Gaussian;

[0048] Figure 2 This is a block diagram of a remote sensing image annotation optimization method based on dynamic confidence and mixed Gaussian.

[0049] Figure 3 This is a schematic diagram of the Gaussian clustering redistribution module structure;

[0050] Figure 4 Line graph showing the changes in the number of symmetric noise labels in each subset of the FGSC-23 dataset: (a) 20% symmetric noise labels; (b) 40% symmetric noise labels;

[0051] Figure 5 Line graph showing the changes in the number of asymmetric noise labels in each subset of the FGSC-23 dataset: (a) 20% asymmetric noise labels; (b) 40% asymmetric noise labels;

[0052] Figure 6 Schematic diagrams schematically illustrating the composite transmit and receive patterns of antennas according to exemplary embodiments of the present invention: (a) 20% symmetrical noise signature; (b) 40% symmetrical noise signature; (c) 20% asymmetrical noise signature; (b) 40% asymmetrical noise signature. DETAILED DESCRIPTION

[0053] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be embodied in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0054] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the blocks shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0055] In noisy label learning tasks, traditional robust loss functions have limitations when dealing with noisy samples. They struggle to effectively distinguish clean from noisy samples, resulting in low label prediction accuracy. Relying solely on a unified loss strategy to adjust the model's learning across all samples can lead to model misfitting of noisy samples, impacting overall training performance and lacking a sophisticated sample selection mechanism. The DISC method partitions the dataset based on confidence scores. However, in the early stages of training, when model performance is not yet stable, directly selecting clean sets based on confidence scores can mistakenly select noisy samples that are easy to fit but mislabeled. This selection approach reduces the purity of the clean set, which is crucial for model generalization. Therefore, this method leaves significant room for improvement. Furthermore, while the DISC method introduces a robust loss in the difficult set to mitigate misleading effects of noisy labels on the model, it still suffers from insufficient information utilization. The difficult set may contain samples with some noisy labels but still possess some learning value. Existing robust loss strategies fail to fully exploit the potential information in these samples, resulting in the model failing to effectively utilize this data for optimization. This paper aims to address the shortcomings of noisy label learning tasks by proposing a more sophisticated method for noise sample screening and processing, thereby improving label prediction for remote sensing images. By accurately screening noise samples, optimizing the clean set screening strategy, and improving the learning mechanism for the difficult set, the robustness of the model is improved, the negative impact of noisy data on the training process is reduced, and the model's generalization ability and classification performance are enhanced.

[0056] The present invention proposes a remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening to alleviate the impact of noisy labels on model training. Specifically, in each training round, the model generates a dynamic confidence threshold based on the historical confidence distribution of the sample, and then predicts and classifies the annotation situation according to the confidence threshold. The conventional approach is to predict the entire data set as a clean set and an error set. However, there may still be some ambiguous samples in the data set that do not belong to the above two categories. Therefore, the method of the present invention combines the Gaussian distribution redistribution strategy to predict the data into four categories of samples: clean set, suspected set, difficult set, and error set. Among them, the suspected set is the part of the data set with simple features but incorrect annotations, which is easily misfitted by the model; the difficult set is the part of the data set that is difficult to judge correct or incorrect, and contains data with correct annotations and noise labels. Then, according to the model's prediction of the sample category, the method designs different optimization strategies. For clean set samples, standard supervised learning is directly adopted. For suspected set samples, an adaptive weighted loss function is used to reduce the interference of potential noise labels on the model. For difficult set samples, the negative impact of noise labels on training is alleviated by dynamically penalizing generalized cross entropy loss. For error set samples, the Mixup strategy is adopted to effectively utilize their image information and mix them with other samples for training, thereby reducing the negative impact of noise labels on training. At the same time, the potential feature information in the error samples is fully explored to further improve the robustness and generalization ability of the model.

[0057] The remote sensing image annotation optimization framework based on dynamic confidence and Gaussian clustering screening proposed in this invention is shown in the figure Figure 2As shown. Specifically, for each batch of input samples, differentiated sample views are first generated through two data enhancement methods, strong enhancement and weak enhancement, and the enhanced samples are input into the model for learning and prediction to obtain corresponding confidence information. These confidence information are then sent to the dynamic threshold generation and sample screening module for processing. The dynamic threshold generation module will be dynamically updated throughout the training process, and the dynamic confidence threshold of each sample will be comprehensively generated based on the confidence distribution of the sample in the current training round and the historical training round. The sample screening module takes effect after pre-training, and the current batch of samples is screened and divided according to the dynamic threshold of the sample generated after the end of the previous round. It will be divided into three subsets: the initial screening clean set (C'), the difficult set (H), and the error set (N). However, due to the limited predictive ability of the model in the initial learning stage, the samples in the "initial screening clean set" may still contain simple samples with some noise labels, and are therefore not completely pure. To this end, the present invention further introduces a mixed Gaussian clustering redistribution module to refine the initial screening clean set. By performing Gaussian clustering modeling on the sample feature vector, screening is performed by threshold setting. This module can filter out samples that deviate greatly from the distribution of cluster centers and divide these samples into suspected sets (S). This part includes simple samples with noise labels and difficult samples with clean labels. The remaining samples are considered to be clean sets (C). For the four types of samples after division, this method designs differentiated robust optimization strategies to achieve targeted processing. For the final clean set (C), the standard supervised learning loss function is directly used for optimization to make full use of its high-quality label information; for the suspected set (S), an improved robust loss function is designed to reduce the interference of noise label samples on the training process; for the difficult set (H), by dynamically adjusting its loss weight, the negative impact of high-loss samples on the model optimization process is effectively reduced; and for the error set (N), a mixup strategy is introduced to enhance its sample information to further improve the robustness and generalization ability of the model.

[0058] like Figure 3 As shown in the figure, the present invention designs a label screening submodule of mixed Gaussian clustering to further process the initial clean set. Specifically, based on the mixed Gaussian clustering model, the samples in the clean set are modeled by category, their distribution in the feature space is estimated, and the mean u and variance of each sample from the cluster center in the Gaussian distribution are calculated. And set a threshold T as For samples outside the threshold It is considered as a doubtful sample S, that is, a simple sample that may be a noise label; and the sample distributed within the threshold It is used as the final clean set C. This process further improves the purity of the clean set and reduces the interference of noise labels on the model.

[0059] refer to Figure 1 As shown, this example embodiment provides a remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening, which can specifically include the following steps:

[0060] Step 1: Obtain image dataset and initialize classification network;

[0061] Step 101: The input image data is read and subjected to two data augmentation methods: strong augmentation and weak augmentation, generating differentiated sample views img_s and img_w. Weak augmentation employs two classic data augmentation methods: cropping and horizontal flipping. Strong augmentation utilizes the AugMix algorithm, which combines multiple image augmentation operations to form an enhancement chain. Three of these operations are randomly selected from the following: automatic contrast adjustment, histogram equalization, tone separation, rotation, daylight exposure adjustment, X-axis shearing, Y-axis shearing, X-axis translation, and Y-axis translation. These operations are randomly combined to form an enhancement chain, and one of the generated enhancement chains is randomly selected and applied to the image.

[0062] Step 102: The two differentiated views are normalized and converted to the torch.tensor format. The image tensor data and label information are packaged and put into the dataloader. The training data generates the train_dataloader and the test data generates the test_dataloader for use in training and testing.

[0063] Step 103: Initialize the network structure, use the resnet network model, set the classification category c, set the optimizer to SGD optimizer, and initialize the number of pre-training rounds E w and the number of training rounds E.

[0064] Step 104: Initialize the cross entropy loss to L c , the formula is

[0065]

[0066] Among them, x i represents the i-th image sample, y i represents the true label of the i-th sample, N represents the total number of samples, and p s (y i ,x i ) and p w (y i ,x i ) represent the confidence of the model for strong and weak enhancement samples respectively.

[0067] Step 105: Initialize the adaptive weighted loss function to L s , the formula is

[0068]

[0069] Among them, q is 0.7, N represents the total number of samples, c represents the number of categories, and p s (c j ,x i ) and p w (c j ,x i ) represent the model for x i The predicted probability of the sample for the jth class under strong and weak enhancement conditions.

[0070] Step 106: Initialize the dynamic penalty generalized cross entropy loss L H , the formula is

[0071]

[0072] Among them, q is set to q = 0.6 + 0.2 * p, where p is the p of the current predicted value of each sample w (y i ; x) or p s (y i ; x) After softmax normalization, it is exponentially amplified Where c is the number of categories, p(y i ; x) represents the current sample x in y i The confidence level of the prediction on .

[0073] Step 107: Initialize Binary Cross Loss (BCE) L BCE , the formula is

[0074]

[0075] in, is the image generated by mixup, is the tag generated by mixup, The model is The confidence of the prediction for class c on the image.

[0076] Step 2: Pre-training stage, preliminary learning of image information and construction of thresholds.

[0077] Step 201: Using the resnet network initialized in step 103, in this embodiment, the resnet50 network is used to perform preliminary learning on the strongly enhanced images and weakly enhanced images of all samples in batches, using the cross entropy loss L initialized in step 104 c Make loss corrections.

[0078] Step 202: Process the prediction confidence given by each sample model and perform threshold iteration according to the formula τ(t)=λτ(t-1)+(1-λ)P(t),τ(0)=0, where τ(t) is the current round threshold, P(t) is the confidence of the prediction category given by the current round model, P(t)=max(P(c;X)), c=0,1,...,k-1, which means taking the maximum value P of the prediction probability of sample X in each category c, and λ is a hyperparameter. According to this formula, the thresholds of strong enhancement and weak enhancement samples are updated as τ and τ, respectively. s (t) and τ w (t).

[0079] Step 203: Repeat steps 201 and 202 for pre-training learning, with an iteration round of E w .

[0080] Step 3: In the formal training phase, all samples are divided into a clean set C, a suspected set S, a difficult set H, and a noise set P.

[0081] Step 301: Using the resnet50 network initialized in step 103, perform dual-view learning on the strongly enhanced images and weakly enhanced images of all samples in batches, and obtain the model prediction confidence p of the two perspectives respectively. s (y i ,x i ) and p w (y i ,x i ).

[0082] Step 302: Use the confidence threshold τ at the end of the previous iteration s (t) and τ w (t) Divide the samples. When the confidence of the labels of the two groups of images with strong enhancement and weak enhancement in the model prediction is greater than the dynamic threshold, the samples are divided into the initial screening clean set C'. The process can be expressed as:

[0083] C`={x i ,y i |p w (y i ;x i )>τ w (t)}∩{x i ,y i |p s (y i ;x i )>τ s (t)}

[0084] Step 303: When the confidence of at least one of the two groups of images is lower than the dynamic threshold, the sample is divided into the difficult set. The process can be expressed as:

[0085] H={x i ,y i |p w (y i ;x i )>τ w (t)}∪{x i ,y i |p s (y i ;x i )>τ s (t)}-C`

[0086] Step 304: For the samples selected from the initial screening clean set C', send them to the mixed Gaussian redistribution module, model the samples in the clean set by category, estimate their distribution in the feature space, and calculate the mean u and variance of each sample from the cluster center in the Gaussian distribution. And set a threshold T as For samples outside the threshold It is considered as a doubtful sample S, that is, a simple sample that may be a noise label; and the sample distributed within the threshold It is taken as the final clean set C.

[0087] Step 305: For the remaining noise samples, perform pseudo-labeling strategy. Set the threshold of pseudo-labeling label to τ'(t) = min(τ ws (t)+σ, 0.99), where σ is a positive offset value that sets the upper limit to 0.99 to limit the τ'(t) value. ws (t) represents τ w (t) and τ s The average value of (t).

[0088] Step 306: Calculate the fusion confidence of the noise sample, and perform confidence fusion on the confidence of the strongly enhanced sample and the weakly enhanced sample according to the following formula.

[0089]

[0090] Where γ is the fusion coefficient, in this embodiment of the present invention, γ=0.5, Y refers to the total category, p s (c;xi) and p w (c; xi) refer to the prediction confidence of the model on the cth category under the strong enhancement and weak enhancement perspectives, respectively.

[0091] Step 307: Filter the noise set to N and assign pseudo labels. The formula is That is, samples outside the initial screening clean set and difficult set are removed, and the maximum value of the confidence score greater than the threshold is set as the pseudo label.

[0092] Step 4: For different sets, adopt different regularization strategies to make the model more inclined to fit clean samples and reduce the misleading of the model by noise samples.

[0093] Step 401: For the clean set C, the cross entropy loss defined in step 104 is used to calculate the loss as L s .

[0094] Step 402: For the suspected set S, the adaptive weighted loss function defined in step 105 is used to calculate the loss as L s .

[0095] Step 403: For the difficult set H, the dynamic penalty generalized cross entropy loss function defined in step 106 is used to calculate the loss as L H .

[0096] Step 404: For the noise set N, perform a Mixup operation on it and all other images and then learn. Perform Mixup on the mixed set M consisting of C, S, H, and P: M = {C∪S∪H∪P}. Then, the Mixup of images and labels in M can be expressed as λ~Beta(α,α), where x m(i) is the sample selected from the M set, y m(i) represents the label of the sample selected in the M set, and Represent the samples and labels after Mixup respectively.

[0097] Step 405: After training the mixed-up image, obtain the model for the mixed sample The confidence level of , the strong enhancement and weak enhancement samples are corrected respectively using the binary cross loss (BCE) in step 107 to obtain and The final Mixup loss is where N m is the total number of Mixup samples.

[0098] Step 406: Add the losses by weight to obtain the total loss of the model, which is L = λ c L c +λ s L s +λ H L H +L M , where λ c is 1.0, λ s is 0.5, λH is 0.8.

[0099] Step 407: Back-propagate the loss and update the model parameters.

[0100] Step 5: Update the screening threshold and repeat the iterative training rounds.

[0101] Step 501: Using the threshold momentum update method of step 202, the prediction confidence of strong and weak enhancements in the current round is updated according to the formula. The updated threshold will be used for the next training.

[0102] Step 502: Repeat training steps 3 to 5, re-dividing the clean set, suspected set, difficult set, and noise set in each round, iteratively updating the filtering threshold, and performing E rounds of training to obtain the final model.

[0103] Step 6: During the testing phase, obtain the annotation results of the model on the test set.

[0104] The present invention conducted multiple experiments on two data sets to verify the superiority of the method of the present invention.

[0105] The experimental results on the FGSC-23 dataset are as follows Figure 3-4 As shown, it can be seen that the subset data in the two noise label ratio datasets have similar changing trends during the training process.

[0106] The following conclusions can be drawn from the experimental results:

[0107] 1. Under different noise label conditions, the number of error sets identified by the dynamic threshold screening method is generally consistent with the set noise label ratio. For example, under 20% and 40% symmetric noise label conditions, the error set ratio is close to 20% and 40%, respectively, indicating that the proposed method has high accuracy in noise label detection.

[0108] 2. Reasonable subset division ratio: The dynamic threshold screening method can reasonably divide samples into clean, suspected, and difficult sets based on the difficulty of the data. The results show that the clean set accounts for the majority of the samples, while the proportions of suspected and difficult sets gradually converge as training progresses. This trend is logical because as the model is gradually fitted, its classification ability gradually improves, and the proportion of difficult and noisy samples gradually decreases.

[0109] 3. Dynamic Changes During Training: In the early stages of training (after pre-training), the number of clean set samples is at its lowest, while the number of suspected and difficult sets is relatively high. This is because the model's ability to discriminate data is weak in the early stages, and some clean samples may be mistakenly classified as suspected or difficult sets. However, as training progresses, the number of clean set samples gradually increases, indicating that the model's ability to discriminate data is gradually improving, which verifies the effectiveness of the Gaussian clustering redistribution module. At the same time, the number of error sets remains stable, verifying the robustness of the dynamic threshold screening method.

[0110] The present invention conducted comparative experiments with the original method DISC on the FGSC-23 dataset under different ratios and types of noise labels. The final prediction accuracy on the test set is as follows:

[0111] Table 1 Comparative experiments on the FGSC-23 dataset

[0112]

[0113] Experiments compared the classification accuracy of the DISC method and the proposed method under different noise conditions. The results showed that under 20% symmetrical noise, the DISC method achieved an accuracy of 80.91%, while the proposed method achieved an accuracy of 82.88%, an improvement of 1.97%. Under 40% symmetrical noise, the DISC method achieved an accuracy of 68.78%, while the proposed method achieved an accuracy of 70.90%, an improvement of 2.12%.

[0114] Under 20% asymmetric noise conditions, the DISC method achieved an accuracy of 79.54%, while the proposed method achieved 81.27%, an improvement of 1.73%. Under the more challenging 40% asymmetric noise conditions, the DISC method achieved an accuracy of 58.43%, while the proposed method achieved 60.07%, an improvement of 1.64%.

[0115] The proposed method outperforms the DISC method under different noise types and noise ratios, and is particularly effective in high-noise environments (40% noise), where it can still effectively improve classification accuracy. This demonstrates that the proposed method can better adapt to noisy data and improve the robustness and generalization ability of the model.

[0116] At the same time, the present invention compares the prediction accuracy of the clean set on the cifar-10 dataset to prove the effectiveness of the Gaussian clustering redistribution module we proposed. Under the same optimizer, learning rate, and learning round settings, we compare the model prediction accuracy of the clean set during the learning process, average the results of every 20 rounds, and draw a line graph as shown in the figure. Figure 5 shown.

[0117] Experimental results show that the proposed method significantly outperforms the original DISC method in the accuracy of clean samples under various noise settings. In particular, under high-noise conditions, such as a 40% asymmetric noise scenario, the proposed method achieves an average accuracy 0.3% higher than the original method. This improvement has a positive effect on better fitting the model to clean samples in the early stages of training and reducing the interference of noisy samples, thereby improving the model's robustness and generalization performance.

[0118] Table 2 Comparative experiments on the cifar-10 dataset

[0119]

[0120]

[0121] This experiment compares the performance of our method with the DISC method on the CIFAR-10 dataset, focusing on the accuracy of clean set samples under different noise label conditions. Four noise conditions were set: 20% symmetric noise, 40% symmetric noise, 20% asymmetric noise, and 40% asymmetric noise. The experimental results show that our method outperforms the DISC method under all noise conditions. Specifically, under 20% symmetric noise, the accuracy of our method is 96.41%, a 0.13% improvement over the DISC method's 96.28%. Under 40% symmetric noise, the accuracy of our method is 95.67%, a 0.22% improvement over the DISC method's 95.45%. Under 20% asymmetric noise, the accuracy of our method is 95.78%, a 0.27% improvement over the DISC method's 95.51%. Under 40% asymmetric noise, the accuracy of our method is 94.52%, a 0.22% improvement over the DISC method's 94.30%. The performance improvement of our method is particularly significant under high-noise conditions (e.g., 40% asymmetric noise), demonstrating its enhanced robustness when dealing with noisy labels. This improvement has a positive effect on better fitting clean samples and reducing the interference of noisy samples in the early stages of model training, thereby improving the model's generalization ability and stability.

[0122] Furthermore, the above-described figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention and are not intended to be limiting. It is readily understood that the processes illustrated in the above-described figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0123] Other embodiments of the present invention will readily occur to those skilled in the art after considering the specification and practicing the invention herein. This application is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the invention being indicated by the claims.

[0124] It should be understood that the present invention is not limited to the exact construction described above and shown in the drawings and that various modifications and variations can be made without departing from the scope thereof, which is limited only by the appended claims.

Claims

1. A remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening, characterized in that: The method comprises: Read the input image data, and generate differentiated strongly enhanced images and weakly enhanced images through two data enhancement methods: strong enhancement and weak enhancement. Build and initialize a pre-trained classification network, where the classification network is a ResNet model; Input the strongly enhanced image and the weakly enhanced image into the pre-trained classification network for pre-training, preliminarily learn the image information and construct a dynamic confidence threshold; In the formal training phase, the samples are divided into clean set C, suspected set S, difficult set H and noise set N according to the dynamic confidence threshold and Gaussian clustering screening; Use differentiated regularization strategies for model training for different sample sets; Update the dynamic confidence threshold and repeat the iterative training; In the testing phase, the test set is input into the trained classification network to output the label prediction results.

2. The method according to claim 1, characterized in that The specific steps of initializing the pre-trained classification network include: Set the classification category, set the optimizer to SGD optimizer, and initialize the number of pre-training rounds and training rounds; Initialize cross entropy loss, adaptive weighted loss function, dynamic penalty generalized cross entropy loss and binary cross loss.

3. The method according to claim 2, characterized in that The specific steps of preliminarily learning image information and constructing a dynamic confidence threshold include: Preliminary learning is performed on strongly enhanced images and weakly enhanced images in batches, and the initial cross entropy loss is used for loss correction; The threshold is dynamically updated according to the prediction confidence of the sample. The formula is: τ(t)=λτ(t-1)+(1-λ)P(t),τ(0)=0 Among them, τ(t) is the current round threshold, P(t) is the confidence of the predicted category given by the current round model, and λ is a hyperparameter.

4. The method according to claim 3, characterized in that The specific steps of dividing the samples into a clean set C, a suspected set S, a difficult set H, and a noise set N according to the dynamic confidence threshold and Gaussian clustering screening include: The samples are divided into the initial screening clean set C' and the difficult set H according to the dynamic confidence threshold; The initial clean set C' is further divided into a clean set C and a suspected set S through the mixed Gaussian clustering redistribution module; The remaining samples are pseudo-labeled to generate a noise set N.

5. The method according to claim 4, characterized in that The specific steps of the mixed Gaussian clustering redistribution module include: Gaussian distribution modeling is performed on the samples in the initial screening clean set by category, and the mean and variance of the sample distance from the cluster center are calculated; A threshold is set to filter out samples that deviate from the cluster center and divide them into the suspected set S, and the rest are divided into the clean set C.

6. The method according to claim 5, characterized in that The specific steps of using differentiated regularization strategies for model training for different sample sets include: The clean set C is trained using standard cross entropy loss; Adaptive weighted loss function is used to train the suspected set S; The difficult set H is trained using dynamic penalty generalized cross entropy loss; The Mixup strategy is used to generate mixed samples for the noise set N, and the training is performed using binary cross loss.

7. The method according to claim 6, characterized in that The specific steps of the Mixup strategy include: Mix the noise set N with other sample sets (C, S, H) to generate mixed samples; Train on mixed samples, calculate binary cross loss and update model parameters.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening according to any one of claims 1 to 7 is implemented.

9. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening according to any one of claims 1 to 7 is implemented.

10. An electronic device, characterized in that: include: processor; as well as a memory for storing executable instructions of the processor; The processor is configured to execute the remote sensing image annotation optimization method based on dynamic confidence and Gaussian clustering screening according to any one of claims 1 to 7 by executing the executable instructions.

Images