Semi-supervised remote sensing image change detection method based on adaptive feature disturbance

Abstract

The invention relates to a semi-supervised remote sensing image change detection method based on adaptive feature disturbance, which comprises the following steps of: training a teacher model by adopting a weak enhancement training set, and obtaining a prediction deterministic value through maximum value normalization processing to relieve adverse effects caused by unreliability of a pseudo tag; then, a self-adaptive feature disturbance scheme is customized based on each prediction deterministic value, disturbance of sample features can be dynamically adjusted according to sample quality, and controllability is increased on the basis of random disturbance, so that confirmation deviation of a student model is remarkably reduced, and the performance of the student model for semi-supervised change detection of remote sensing image changes is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision and change detection, and in particular to a semi-supervised remote sensing image change detection method based on adaptive feature perturbation. Background Art

[0002] Remote sensing change detection (RSCD) is a task to identify changes in surface features by analyzing remote sensing images acquired at different times. This task is essentially a binary classification problem, and its purpose is to identify the target change areas of interest (such as buildings, waters, vegetation, and roads, etc.) in dual-phase image pairs taken at different times in the same area. RSCD has a wide range of applications in urban construction planning, forest environmental protection, rural land management, and natural disaster assessment, which makes RSCD an important research direction in the field of remote sensing. In recent years, change detection methods based on deep learning have made significant progress. However, such methods are highly dependent on a large amount of labeled training data, and their performance will significantly decrease when the labeled samples are insufficient. In addition, the data annotation process of change detection tasks is complex, usually requiring high-precision geometric image registration and pixel-level fine annotation, which is time-consuming and costly. At the same time, with the rapid development of earth observation technology, a large amount of unlabeled data has become easy to obtain, providing an opportunity for semi-supervised learning. Semi-supervised learning methods can effectively utilize these unlabeled samples, thereby significantly improving the performance of the model when the labeled data is limited, and are an effective way to solve the above problems.

[0003] Semi-supervised Change Detection (SSCD) aims to improve the change detection performance by using limited labeled training samples and a large number of unlabeled training samples for training at the same time. In this way, the model can not only learn the required semantic information, but also learn the semantic feature distribution more fully, thereby significantly improving the robustness of the model. The current mainstream methods include pseudo-label self-training methods, consistency regularization methods, and generative adversarial network methods. Among them, the pseudo-label self-training method first uses a prediction model to generate pseudo-labels for unlabeled samples, mixes these unlabeled samples and pseudo-labels with labeled training samples, and conducts supervised training together to provide expanded training data for the model; the consistency regularization method imposes different degrees of perturbation on the input data and uses the consistency of the model's output on these input data as a training constraint; the generative adversarial network-based method aims to use the generative model to model the distribution of data, thereby inferring the potential information of unlabeled data.

[0004] Although the above methods have achieved certain success, there are still some challenges: the method based on pseudo-label self-training has inevitable noise problems, which continue to accumulate during the training process and may seriously affect the performance of the model; in addition, the training process of the generative adversarial network is highly unstable, and the gradient vanishing problem often makes it difficult to continuously optimize the generator, making it difficult to achieve the ideal optimal training results. Summary of the invention

[0005] The problem solved by the present invention is how to effectively reduce the influence of noise in model prediction and prevent the problem of excessive confirmation bias caused by random disturbances, thereby improving the reliability and stability of model prediction.

[0006] The present invention provides a semi-supervised remote sensing image change detection method based on adaptive feature perturbation, comprising: Step 1, collecting dual-phase image pairs and dividing them into labeled training sets and unlabeled training sets; Step 2, initialize the weights of the student model and the teacher model; Step 3: After weak enhancement processing on the labeled training set, input it into the student model and calculate the supervision loss; Step 4, weakly enhance the unlabeled training set to obtain a weakly enhanced training set, and strongly enhance the weakly enhanced training set to obtain a strongly enhanced training set; Step 5: Input the weakly enhanced training set into the teacher model to predict and obtain a confidence map, evaluate the confidence map to obtain a certainty value, and normalize all certainty values ​​to obtain a predicted certainty value; Step 6, customizing an adaptive feature perturbation scheme according to each prediction certainty value; Step 7: Input the strongly enhanced training set into the student model to obtain high-dimensional features, apply an adaptive feature perturbation scheme to the high-dimensional features, and calculate the unsupervised loss of the teacher model and the student model; Step 8: Use the exponential decay warm-up function to weight the unsupervised loss and add it to the supervised loss to get the overall training loss. Step 9: Use the stochastic gradient descent method to minimize the overall training loss to update the weights of the student model; update the weights of the teacher model based on the updated weights of the student model and the exponential moving average method.

[0007] Compared with the prior art, the present application has the following advantages: a weakly enhanced training set is used to train the teacher model, and a prediction certainty value is obtained through maximum normalization processing to alleviate the adverse effects of unreliable pseudo labels; then, an adaptive feature perturbation scheme is customized based on each prediction certainty value, which can dynamically adjust the perturbation of sample features according to the sample quality, and increase controllability on the basis of random perturbation, thereby significantly reducing the confirmation bias of the student model and improving the performance of the student model in semi-supervised change detection of remote sensing image changes.

[0008] In a possible implementation manner, the dual-phase image pair in step 1 includes two remote sensing images of the same area in different phases.

[0009] Compared with the existing technology, by making full use of the labeled data of the labeled training set, the value of the unlabeled training set is maximized, the dependence on manual labeling is significantly reduced, and the efficiency and practicality of semi-supervised learning are improved.

[0010] In a possible implementation, the calculation formula of the supervision loss in step 3 is: ; In the formula, represents the supervision loss, Represents the total number of labeled samples input into the student model; Indicates that the student model is For the dual-phase image pair The predicted change probability of pixels, Indicates For the dual-phase image pair The true label of each pixel; represents the length of the remote sensing image in the dual-temporal image pair, Represents the width of the remote sensing image in a dual-temporal image pair.

[0011] Compared with the existing technology, the method of calculating the supervision loss after the labeled training set is weakly enhanced can effectively utilize the annotation information of the labeled training set and enhance the student model's ability to learn the features of the labeled training set.

[0012] In a possible implementation, the weak enhancement processing performed on the unlabeled training set in step 4 includes random flipping, random scaling and random cropping; the strong enhancement processing performed on the weakly enhanced training set in step 4 includes random three of identity, contrast, automatic contrast, brightness, color, equalization, sharpness, tone separation, and inversion.

[0013] Compared with the existing technology, weak enhancement and strong enhancement are performed on the unlabeled training set to generate diversified training samples. The weak enhancement retains the core features and is used for the teacher model to generate reliable pseudo labels. The strong enhancement combines multiple image transformations to simulate complex scene changes, enhance the robustness of the student model to feature disturbances, and reduce the risk of overfitting.

[0014] In a possible implementation manner, the step 5 specifically includes: Step 501, dividing the weak enhancement training set and the strong enhancement training set into a plurality of weak enhancement sample sets and a plurality of strong enhancement sample sets in batches respectively; the corresponding samples in the weak enhancement sample sets and the strong enhancement sample sets of the same batch are derived from the same pair of dual-phase image pairs; Step 502, randomly select a weak enhancement sample set and input it into the teacher model to predict the change probability to obtain a confidence map; Step 503, evaluate the confidence map to obtain a certainty value, and the calculation formula is: ; In the formula, represents the certainty value, represents the confidence map, Represents the certainty value Taking the mean over all pixels, represents the predicted probability difference map, , Indicates absolute value operation; represents the weighted information entropy graph, , is the information entropy of the predicted probability, and Represent the information entropy of the unchanged class and the changed class respectively; ; In the formula, ; ; Indicates Predicted change probability for dual-phase image pairs; Step 504: Perform maximum normalization processing on each certainty value to obtain a predicted certainty value.

[0015] Compared with the existing technology, the certainty value is calculated through the confidence map, and the maximum value is normalized to obtain the predicted certainty value, which dynamically quantifies the credibility of the teacher model prediction, avoids the interference of low-quality samples, alleviates the negative impact of unreliable pseudo-labels, and standardizes the sample quality evaluation through maximum value normalization processing, providing a reliable basis for the subsequent customization of adaptive feature perturbation schemes.

[0016] In a possible implementation manner, the adaptive feature perturbation scheme customized in step 6 includes the perturbation quantity and perturbation intensity, and step 6 specifically includes: Step 601: As a scaling factor, the number of perturbations per predicted certainty value is calculated as: ; In the formula, Represents the prediction certainty value The number of disturbances, represents the total number of disturbances in the disturbance pool; Step 602, randomly select a perturbation number of perturbation methods from the perturbation pool, and calculate the perturbation intensity of each perturbation method, and the calculation formula is: ; In the formula, represents the disturbance intensity, For the Predefined parameters for each perturbation method.

[0017] Compared with the existing technology, the above technical solution is used to implement a differentiated strategy of low disturbance for high-quality samples and high disturbance for low-quality samples. This can not only avoid the blindness of random disturbance and optimize features in a targeted manner, but also reduce the confirmation bias of the student model and improve tolerance to noisy data.

[0018] In a possible implementation manner, the step 7 specifically includes: Step 701, inputting the strong enhancement sample set of the same batch as the weak enhancement sample set selected in step 5 into the student model for feature extraction and fusion to obtain high-dimensional features; Step 702, applying the adaptive feature perturbation scheme to the high-dimensional features, the expression is: ; In the formula, represents the high-dimensional features after the perturbation, Indicates that the student model is the first one from the current batch of strong enhancement samples For high-dimensional features extracted from unlabeled samples, Indicates the weakly enhanced sample set of the current batch A collection of adaptive feature perturbation schemes for unlabeled samples, Indicates the specific process of applying disturbance; Step 703: Use the student model decoder to decode the perturbed high-dimensional features. Decode and get the change probability, the calculation formula is: , Represents the probability of change of the student model on the strongly enhanced sample set; Step 704, calculate the unsupervised loss of the teacher model and the student model, and the calculation formula is: ; in, represents the unsupervised loss, is the total number of unlabeled samples in the current batch of weakly enhanced samples, represents the probability of change of the teacher model on the weakly enhanced sample set, Indicates that the The output change probability after the perturbation method is: represents the mean square error loss function.

[0019] Compared with the existing technology, an adaptive feature perturbation scheme is applied to the high-dimensional features extracted by the student model, and the unsupervised loss is calculated to enhance the sensitivity of the student model to feature changes. The reliability of pseudo-labels is further improved by predicting consistency constraints between the teacher model and the student model.

[0020] In a possible implementation manner, step 8 specifically includes: Step 801, using an exponential decay warm-up function to weight the unsupervised loss as: ; ; In the formula, represents the unsupervised loss The maximum achievable weight of Used to control the speed of the preheating process; Indicates the current iteration cycle; is the total step length of the preheating process, is the total number of iterations of the entire training process, and 0< <1; Step 802, calculate the overall training loss, the calculation formula is: ; In the formula, represents the overall training loss, Represents weight.

[0021] Compared with the existing technology, an exponential decay warm-up function is used to dynamically adjust the weight of the unsupervised loss. The student model focuses on supervised constraints in step 4, and introduces unsupervised constraints in step 7 and subsequent steps, which balances the optimization objectives of different training stages, avoids early unstable pseudo-labels interfering with the convergence of the student model, and improves training stability.

[0022] In a possible implementation, the calculation formula for updating the weight of the teacher model based on the updated student model weight and the exponential moving average method in step 9 is: ; In the formula, represents the initialization weight of the teacher model, represents the updated weight of the teacher model, represents the updated weight of the student model, is the momentum factor of the moving exponential average.

[0023] Compared with the existing technology, the stochastic gradient descent method is used to update the weights of the student model, and the exponential moving average method is used to update the weights of the teacher model, inheriting the progressive optimization results of the student model, ensuring that the teacher model outputs more stable pseudo labels, forming a benign teacher-student interaction, and accelerating the convergence of the student model and the teacher model.

[0024] In a possible implementation manner, the dual-phase image pairs acquired in step 1 are further divided into a verification set and a test set; The step also includes step 10, verifying, testing and saving the student model after weight update based on the verification set and the test set; specifically including: Step 1001, after the training of the teacher model and the student model for the weakly enhanced sample set and the strongly enhanced sample set of the current batch is completed, another weakly enhanced sample set and the strongly enhanced sample set are randomly selected to train the teacher model and the student model through steps 5 to 9; Step 1002, using the average intersection-over-union ratio to evaluate the student model after each batch of training, and saving the weight of the student model with the highest average intersection-over-union ratio; Step 1003, use the test set to save the student model Tests are performed, the average intersection-over-union ratio is calculated, and the predicted change probability is visualized to obtain the change detection results of the dual-phase image pair.

[0025] Compared with the existing technology, the average intersection-over-union ratio is used as the evaluation indicator to directly reflect the change detection accuracy and ensure the practicality of the real scene. The visualization operation further verifies the robustness of the change detection method in complex scenes and provides an intuitive basis for actual deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the process framework of the present invention; Figure 2 This is a visual comparison chart of the 5% labeled LEVIR-CD dataset using different semi-supervised change detection methods in the embodiments of this application; Figure 3 This is a visual comparison chart of the 5% labeled WHU-CD dataset using different semi-supervised change detection methods in the embodiments of this application; Figure 4 In the embodiment of the present application, t-SNE technology is used to visualize the features on the data set when random feature perturbation and adaptive feature perturbation are used respectively. DETAILED DESCRIPTION

[0027] First, those skilled in the art should understand that these implementations are only used to explain the technical principles of the embodiments of the present application, and are not intended to limit the protection scope of the embodiments of the present application. Those skilled in the art can make adjustments to them as needed to adapt to specific application scenarios.

[0028] In the description of the embodiments of the present application, it should be noted that, unless otherwise clearly specified and limited, the terms "connected" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the embodiments of the present application can be understood according to specific circumstances.

[0029] In the embodiments of the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0030] The present application is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] See also Figure 1 As shown, the embodiment of the present application discloses a semi-supervised remote sensing image change detection method based on adaptive feature perturbation, comprising: Step 1: Collect dual-phase image pairs and divide them into labeled training sets , unlabeled training set , validation set and test set; the dual-phase image pair includes two remote sensing images of the same area at different phases, and the labeled training set , the unlabeled training set ,in, Indicates bi-temporal image pairs and their true labels, Indicates For unlabeled bi-temporal image pairs; represents the image of the front phase of the dual-phase image pair, represents the image of the posterior phase of the dual-phase image pair, Indicates a labeled training set The number of labeled samples, Represents the unlabeled training set The number of unlabeled samples, and ; In this specific embodiment, all the dual-phase image pairs are from 20 different locations in seven cities in Texas, USA, and were taken between 2002 and 2018. The time span of the dual-phase image pairs is 5 to 14 years. In this specific embodiment, the collected dual-phase images are cut into The labeled training set accounts for 3.5% (356 pairs), the unlabeled training set accounts for 66.5% (6764 pairs), the validation set accounts for 10% (1024 pairs), and the test set accounts for another 20% (2048 pairs).

[0032] Step 2: Build a student model and teacher model , and initialize the student model Weight and teacher model Weight In this specific embodiment, the student model and teacher model ResNet-50 is used as the backbone network, and the parameters of ResNet-50 pre-trained on ImageNet are used for initialization. and .

[0033] Step 3: For the labeled training set After weak enhancement, input into the student model , and calculate the supervision loss ; There will be a labeled training set Weak enhancement processing is performed, including random flipping, random scaling, and random cropping; the labeled weak enhancement training set obtained by processing is expressed as ; In this specific embodiment, each round randomly extracts a sample from the labeled weakly enhanced sample data set. For labeled samples, the student model Training is performed; in this specific embodiment ; The calculation formula of supervision loss is: ; In the formula, Indicates the input student model in each round The total number of labeled samples in ; Represents the student model In the For the dual-phase image pair The predicted change probability of pixels, Indicates For the dual-phase image pair The true label of each pixel; represents the length of the remote sensing image in the dual-temporal image pair, Represents the width of the remote sensing image in a dual-temporal image pair.

[0034] Step 4: Unlabeled training set Weak enhancement processing is performed to obtain a weak enhancement training set, and strong enhancement processing is performed on the weak enhancement training set to obtain a strong enhancement training set; among them, for the unlabeled training set The weak enhancement processing includes random flipping, random scaling and random cropping; the weak enhancement training set obtained by the processing is expressed as ; The strong enhancement processing of the weak enhancement training set in step 4 includes three random ones of identity, contrast, auto contrast, brightness, color, equalize, sharpness, posterize, and solarize. The strong enhancement training set obtained by the processing is represented as .

[0035] Step 5: Input the weakly enhanced training set into the teacher model Predict the probability of each pixel in the dual-phase image pair belonging to the change category or the unchanged category, and obtain the confidence map , and evaluate the confidence map Get the certainty value , the maximum value normalizes all deterministic values Get the prediction certainty value ; Specifically include: Step 501, dividing the weak enhancement training set and the strong enhancement training set into batches respectively to obtain a plurality of weak enhancement sample sets and a plurality of strong enhancement sample sets; the corresponding samples in the weak enhancement sample sets and the strong enhancement sample sets of the same batch are derived from the same pair of dual-phase image pairs; Step 502: randomly select a batch of weakly enhanced sample sets and input them into the teacher model Predict the probability of change and get the confidence map ; Step 503: Confidence map Evaluate to get certainty value , the calculation formula is: ; In the formula, Represents the certainty value Taking the mean over all pixels, represents the predicted probability difference map, , Indicates absolute value operation; represents the weighted information entropy graph, , The information entropy representing the prediction probability, and Represent the information entropy of the unchanged class and the changed class respectively; ; In the formula, ; ; Step 504, for each certainty value Perform maximum value normalization to obtain the predicted certainty value , the calculation formula is: ; In the formula, Represents all the certainty values ​​corresponding to the weakly enhanced sample set of the current batch The maximum value of Represents all the certainty values ​​corresponding to the weakly enhanced sample set of the current batch The minimum value of .

[0036] Step 6: Based on each prediction certainty value Customize an adaptive feature perturbation scheme; the adaptive feature perturbation scheme includes the perturbation quantity and perturbation intensity, specifically including: Step 601: As a scaling factor, each prediction certainty value is calculated The number of disturbances is calculated as: ; In the formula, Represents the prediction certainty value The number of disturbances, represents the total number of disturbances in the disturbance pool; the parameters of the disturbance pool are shown in Table 1; Table 1 Disturbance pool parameters ; Step 602: Randomly select the number of disturbances from the disturbance pool perturbation methods, calculate the perturbation intensity of each perturbation method , the calculation formula is: ; For the Predefined parameters for each perturbation method.

[0037] Step 7: Input the strong enhancement training set into the student model Extract and fuse high-dimensional features, apply an adaptive feature perturbation scheme to the high-dimensional features, use the student model decoder to decode the perturbed high-dimensional features, and obtain the change probability; calculate the teacher model With student models The unsupervised loss ; Specifically include: Step 701: Input the strong enhancement sample set from the same batch as the weak enhancement sample set selected in step 5 into the student model Perform feature extraction and fusion to obtain high-dimensional features; Step 702, applying the adaptive feature perturbation scheme to the high-dimensional features, the expression is: ; In the formula, Represents the student model From the current batch of strong enhancement samples For high-dimensional features extracted from unlabeled samples, Indicates the number of weakly enhanced samples in the current batch A collection of methods for adaptive feature perturbation schemes for unlabeled samples, Indicates the specific process of applying disturbance; Step 703: Use the student model decoder to decode the perturbed high-dimensional features. Decode and get the change probability, the calculation formula is: ; Step 704, calculate the teacher model With student models The unsupervised loss , the calculation formula is: ; ; in is the total number of unlabeled samples in the current batch of weakly enhanced samples, Model for teachers The probability of change on the weakly enhanced sample set, Indicates that the The output change probability after the perturbation method.

[0038] Step 8: Use exponential decay warm-up function to adjust the unsupervised loss Weighted and compared with the supervision loss Add together to get the overall training loss; specifically: Step 801: Use exponential decay warm-up function to adjust the unsupervised loss The weighted expression is: ; ; In the formula, represents the maximum achievable weight of the unsupervised loss, Used to control the speed of the preheating process; Indicates the current iteration cycle; is the total step length of the preheating process, is the total number of iterations of the entire training process, and 0< <1; Step 802, calculate the overall training loss , the calculation formula is: ; In the formula, Represents weight.

[0039] Step 9: Use stochastic gradient descent to minimize the overall training loss Update Student Model Weight ; Based on the updated student model Weight Update the teacher model using the exponential moving average method Weight ; ; In the formula, Represents the learning rate, which is used to control the update step size. represents the gradient calculated with respect to the overall training loss.

[0040] Based on the updated student model Weight Update the teacher model using the exponential moving average method Weight The calculation formula is: ; In the formula, is the momentum factor of the moving exponential average.

[0041] Step 10: Update the weights of the student model based on the validation set and test set Verify, test and save, including: Step 1001: train the teacher model with the weakly enhanced sample set and the strongly enhanced sample set of the current batch and student models After the end, randomly select another batch of weakly enhanced sample sets and strongly enhanced sample sets to train the teacher model through steps 5 to 9 and student models ; In this specific embodiment, verification is performed every 5 epochs. When it is the current best, update the saved best parameters; Step 1002, using average intersection-over-union ratio Evaluate the student model after each batch of training , save the average intersection-union ratio Top student model The weight of The calculation formula is: ; in , and They represent the number of changed pixels correctly identified, the number of unchanged pixels incorrectly classified as changed pixels, and the number of changed pixels incorrectly identified as unchanged pixels; Step 1003, use the test set to save the student model Test and calculate the average intersection-union ratio The predicted change probability is visualized to obtain the change detection results of the dual-phase image pair.

[0042] Data Experiment A dataset consisting of randomly collected dual-phase image pairs is input into the semi-supervised remote sensing image change detection method based on adaptive feature perturbation of the present application. Table 2 is the average quantitative comparison of different datasets in different semi-supervised change detection methods. The blue highlighted part in Table 2 represents the optimal result, and the underlined part represents the suboptimal result. In summary, the semi-supervised remote sensing image change detection method based on adaptive feature perturbation of the present application achieves the best change detection effect on all datasets consisting of dual-phase image pairs, and all detection results are better than the RCR change detection effect. This shows that the adaptive feature perturbation consistency of the present application is more suitable for semi-supervised change detection than the random feature perturbation consistency used in RCR.

[0043] Table 2 Average quantitative comparison of different datasets in different semi-supervised change detection methods ; In addition, since the resolution of the GZ-CD dataset is very low and the manual annotations are relatively rough, it poses a huge challenge to FPA, a method based on pseudo-label self-training. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation in this application does not rely on binary hard pseudo-labels, but uses the perturbation consistency in a controllable search space as a training constraint, overcoming the difficulty of inaccurate pseudo-labels and achieving better experimental results.

[0044] Finally, this experiment also visualizes the 5% labeled LEVIR-CD dataset and 5% labeled WHU-CD dataset using different semi-supervised change detection methods in Table 2, as shown in Figure 2 and Figure 3 As shown, the t-SNE technology is used to reduce the dimension of the features on the data set when using random feature perturbation and adaptive feature perturbation respectively and project them to two-dimensional coordinates for visualization, as shown in the figure. According to the comparison results, compared with the random perturbation used in RCR, the semi-supervised method using adaptive feature perturbation has a more obvious decision boundary in the extracted feature space, and the features of the changing foreground and background categories are more compact within the class and more dispersed between classes, which intuitively verifies the superiority of the adaptive perturbation method.

[0045] In the description of the embodiments of the present application, it should be noted that in the description of the present application, terms such as "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the drawings. This is only for the convenience of description, and does not indicate or imply that the device or component must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it cannot be understood as a limitation on the present application.

[0046] In the description of the present application, the description with reference to the terms "one embodiment", "some embodiments", "in the present embodiment", "specific example", or "some examples" etc. means that the specific features, mechanisms, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, mechanisms, materials or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0047] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A semi-supervised remote sensing image change detection method based on adaptive feature perturbation, characterized in that: include: Step 1, collecting dual-phase image pairs and dividing them into labeled training sets and unlabeled training sets; Step 2, initialize the weights of the student model and the teacher model; Step 3: After weak enhancement processing on the labeled training set, input it into the student model and calculate the supervision loss; Step 4, weakly enhance the unlabeled training set to obtain a weakly enhanced training set, and strongly enhance the weakly enhanced training set to obtain a strongly enhanced training set; Step 5: Input the weakly enhanced training set into the teacher model to obtain a confidence map, evaluate the confidence map to obtain a certainty value, and normalize all certainty values ​​to obtain a predicted certainty value; Step 6, matching an adaptive feature perturbation scheme according to each predicted certainty value; Step 7, input the strongly enhanced training set into the student model to obtain high-dimensional features, and apply an adaptive feature perturbation scheme to the high-dimensional features; Calculate the unsupervised loss of the teacher model and the student model; Step 8: Use the exponential decay warm-up function to weight the unsupervised loss and add it to the supervised loss to get the overall training loss. Step 9: Use the stochastic gradient descent method to minimize the overall training loss to update the weights of the student model; update the weights of the teacher model based on the updated weights of the student model and the exponential moving average method.

2. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 1 is characterized in that: The dual-phase image pair in step 1 includes two remote sensing images of the same area in different phases.

3. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 2 is characterized in that: The calculation formula of the supervision loss in step 3 is: ； In the formula, represents the supervision loss, Represents the total number of labeled samples input into the student model; Indicates that the student model is For the dual-phase image pair The predicted change probability of pixels, Indicates For the dual-phase image pair The true label of each pixel; represents the length of the remote sensing image in the dual-temporal image pair, Represents the width of the remote sensing image in the dual-temporal image pair.

4. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 2 is characterized in that: The weak enhancement processing performed on the unlabeled training set in step 4 includes random flipping, random scaling and random cropping; the strong enhancement processing performed on the weakly enhanced training set in step 4 includes three random ones among identity, contrast, automatic contrast, brightness, color, equalization, sharpness, tone separation and inversion.

5. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 2, characterized in that: The step 5 specifically includes: Step 501, dividing the weak enhancement training set and the strong enhancement training set into a plurality of weak enhancement sample sets and a plurality of strong enhancement sample sets in batches respectively; the corresponding samples in the weak enhancement sample sets and the strong enhancement sample sets of the same batch are derived from the same pair of dual-phase image pairs; Step 502, randomly select a weak enhancement sample set and input it into the teacher model to predict the change probability to obtain a confidence map; Step 503: Evaluate the confidence map to obtain a certainty value , the calculation formula is: ； In the formula, represents the certainty value, represents the confidence map, Represents the certainty value Taking the mean over all pixels, represents the predicted probability difference map, , Indicates absolute value operation; represents the weighted information entropy graph, , is the information entropy of the predicted probability, and Represent the information entropy of the unchanged class and the changed class respectively; ; In the formula, ; ; Indicates Predicted change probability for dual-phase image pairs; Step 504: Perform maximum normalization processing on each certainty value to obtain a predicted certainty value.

6. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 5, characterized in that: The adaptive characteristic perturbation scheme calculated in step 6 includes the perturbation quantity and perturbation intensity. Step 6 specifically includes: Step 601: As a scaling factor, the number of perturbations per predicted certainty value is calculated as: ； In the formula, Represents the prediction certainty value The number of disturbances, represents the total number of disturbances in the disturbance pool; Step 602, randomly select a perturbation number of perturbation methods from the perturbation pool, and calculate the perturbation intensity of each perturbation method, and the calculation formula is: ； In the formula, represents the disturbance intensity, For the Predefined parameters for each perturbation method.

7. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 6, characterized in that: The step 7 specifically includes: Step 701, inputting the strongly enhanced sample set of the same batch as the weakly enhanced sample set in step 5 into the student model for feature extraction and fusion to obtain high-dimensional features; Step 702, applying the adaptive feature perturbation scheme to the high-dimensional features, the expression is: ； In the formula, represents the high-dimensional features after the perturbation, Indicates that the student model is the first one from the current batch of strong enhancement samples For high-dimensional features extracted from unlabeled samples, Indicates the weakly enhanced sample set in the current batch A collection of adaptive feature perturbation schemes for unlabeled samples, Indicates the specific process of applying disturbance; Step 703: Use the student model decoder to decode the perturbed high-dimensional features. Decode and get the change probability, the calculation formula is: , Represents the probability of change of the student model on the strongly enhanced sample set; Step 704, calculate the unsupervised loss of the teacher model and the student model, and the calculation formula is: ； in, represents the unsupervised loss, is the total number of unlabeled samples in the current batch of weakly enhanced samples, represents the probability of change of the teacher model on the weakly enhanced sample set, Indicates that the The output change probability after the perturbation method is: represents the mean square error loss function.

8. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 7, characterized in that: The step 8 specifically includes: Step 801, using an exponential decay warm-up function to weight the unsupervised loss as: ； ； In the formula, represents the unsupervised loss The maximum achievable weight of Used to control the speed of the preheating process; Indicates the current iteration cycle; is the total step length of the preheating process, is the total number of iterations of the entire training process, and 0 < < 1; Step 802, calculate the overall training loss, the calculation formula is: ； In the formula, represents the overall training loss, Represents weight.

9. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 8, characterized in that: The calculation formula for updating the weight of the teacher model based on the updated student model weight and the exponential moving average method in step 9 is: ； In the formula, represents the initial weight of the teacher model, represents the updated weight of the teacher model, represents the updated weight of the student model, is the momentum factor of the moving exponential average.

10. The semi-supervised remote sensing image change detection method based on adaptive feature perturbation according to claim 2, characterized in that: The dual-phase image pairs collected in step 1 are further divided into a verification set and a test set; The step also includes step 10, verifying, testing and saving the student model after weight update based on the verification set and the test set; specifically including: Step 1001, after the training of the teacher model and the student model for the weakly enhanced sample set and the strongly enhanced sample set of the current batch is completed, another weakly enhanced sample set and the strongly enhanced sample set are randomly selected to train the teacher model and the student model through steps 5 to 9; Step 1002, using the average intersection-over-union ratio to evaluate the student model after each batch of training, and saving the weight of the student model with the highest average intersection-over-union ratio; Step 1003, using the test set to test the saved student model, calculating the average intersection-over-union ratio and visualizing the predicted change probability to obtain the change detection result of the dual-phase image pair.

Images