A semantic segmentation method, device and storage medium for cross-domain remote sensing images

Abstract

The application provides a cross-domain remote sensing image semantic segmentation method and device and a storage medium, characterized in that remote sensing image data is input into a student network in a semantic segmentation model to output a semantic segmentation result; wherein the construction of the semantic segmentation model comprises: obtaining remote sensing image training data, wherein the training data comprises source domain data and target domain data; obtaining four loss functions based on the remote sensing image training data and a teacher-student network; performing back propagation derivation on the parameters of the student network based on the obtained four loss functions to continuously optimize the parameters of the student network; and continuously updating the parameters of the teacher network based on the parameters of the student network, repeatedly iterating until the training is completed, and obtaining the semantic segmentation model. The application learns a model from source domain data and target domain data, generalizes the model to a target domain by minimizing the difference between the feature distributions of the source domain and the target domain, and improves the segmentation accuracy of the target and the image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a semantic segmentation method, device, and storage medium for cross-domain remote sensing images. Background Technology

[0002] With the rapid development of computer science and technology, the technologies involved in remote sensing imagery are also constantly advancing. Traditional semantic segmentation methods for remote sensing imagery involve learning a model given training samples to obtain an objective function that minimizes the risk of test data, outputting the segmentation result of the remote sensing image. Supervised learning is the most common and successful method applicable to various tasks for training such a generalized model. The success of the generalized model is due to the large number of labeled training datasets and the assumption that the training and test datasets come from the same distribution and share similar joint probability distributions.

[0003] However, in real-world applications, test datasets may come from different feature spaces or distributions, leading to a failure to meet this requirement. More specifically, when the training dataset cannot accurately reflect the distribution of the test data—that is, when the source domain distribution differs from the target domain distribution—the trained model may experience a significant performance degradation when applied to the test data. Compared to natural imagery, remote sensing often suffers from various differences, such as variations in imaging sensors, image resolution, geographical location at the time of acquisition, and radiometric differences resulting from imaging time and angle. This often leads to poor segmentation accuracy when a trained model is transferred to other remote sensing images different from the training data, failing to achieve satisfactory results. This significantly hinders the widespread application of deep learning models in remote sensing. Summary of the Invention

[0004] This invention provides a semantic segmentation method, device, and storage medium for cross-domain remote sensing images, in order to solve the problems of low segmentation accuracy and poor applicability of existing semantic segmentation methods.

[0005] In a first aspect, the present invention provides a semantic segmentation method for cross-domain remote sensing images, which inputs remote sensing image data into a student network in a semantic segmentation model and outputs semantic segmentation results; The construction of this semantic segmentation model includes: Acquire remote sensing image training data, which includes source domain data and target domain data; the source domain data includes a one-to-one corresponding source domain image and source domain label; while the target domain data only includes target domain images and no labels. The source domain data is input into the student network for training to obtain a first loss function; The target domain data is input into the teacher network to generate pseudo-labels, and the student network is trained using the pseudo-labels to obtain the second loss function. The source domain data and the target domain data are mixed to obtain a mixed image and a corresponding mixed label. The mixed label is obtained by mixing the source domain label and the pseudo label. The mixed image is input into the student network for training, and a third loss function is obtained by using the mixed label and the mixed prediction result output by the student network. The source domain prediction result of the source domain data predicted by the student network and the target domain prediction result of the target domain data predicted by the student network are input into the discriminator network to obtain the fourth loss function. Based on the four loss functions obtained, the parameters of the student network are backpropagated and differentiated to continuously optimize the parameters of the student network; and the parameters of the teacher network are continuously updated based on the parameters of the student network. This process is repeated until the training ends, and the semantic segmentation model is obtained.

[0006] Furthermore, the source domain data includes a one-to-one correspondence of source domain images and source domain labels; the target domain data only includes target domain images.

[0007] Furthermore, The source domain image is input into the student network for training to obtain the source domain prediction result; a pixel-level cross-entropy function is obtained based on the source domain prediction result and the source domain label, and the pixel-level cross-entropy function is used as the first loss function. The target domain image is input into the student network and the teacher network respectively to obtain the target domain prediction result and the pseudo label corresponding to the target domain image; based on the target domain prediction result and the pseudo label, the second loss function is calculated. The mixed image is input into the student network to obtain a mixed prediction result; the source domain label and the pseudo label are mixed to obtain a mixed label; based on the mixed prediction result and the mixed label, a third loss function is calculated. Based on the source domain prediction results and the target domain prediction results, a fourth loss function is calculated.

[0008] Furthermore, the first loss function The calculation methods include: ; in, Represents the source domain image; Indicates the source domain tag; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; A one-hot vector representing the source domain label; This represents the softmax probability of the prediction result in the source domain.

[0009] Furthermore, the second loss function is calculated using pixels whose confidence level in the target domain prediction results is higher than a certain threshold, and the second loss function... The calculation methods include: ; in, Represents the target domain image; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; A one-hot vector representing pixels whose confidence level in the target domain prediction result is higher than a certain threshold; This represents the softmax probability of the prediction result in the target domain; Indicates the threshold; This represents the one-hot vector of the pseudo-label.

[0010] Furthermore, the process of obtaining the mixed image and the corresponding mixed label includes: A mask is generated for half of the source domain labels corresponding to the source domain image, and pixels belonging to that half of the category are marked as 1, while pixels of other categories are marked as 0. The pixel corresponding to the source domain image marked as 1 is pasted onto the target domain image to obtain the hybrid image; the pixel corresponding to the source domain label marked as 1 is pasted onto the pseudo label to obtain the hybrid label.

[0011] Furthermore, a weight mask is generated using the proportion of pixels whose confidence level in the target domain prediction results is higher than a certain threshold, and the third loss function is calculated based on the weight mask. The third loss function The calculation methods include: ; in, Represents a mixed image; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; Represents a weighted mask image. ; Indicates a mask. Represents a one-hot vector. Indicates pixel ratio; Indicates mixed tags a one-hot vector; This represents the softmax probability of the mixed domain prediction result.

[0012] Furthermore, the fourth loss function The calculation methods include: ; in, Represents the source domain image; Represents the target domain image; These represent the parameters of the discriminator network; Indicates the length of the image; Indicates the width of the image; Represents the first loss function; This represents the entropy graph of the source domain prediction results. Input to the discriminator network middle; This represents the entropy graph of the prediction results for the target domain. Input to the discriminator network middle.

[0013] Secondly, the present invention provides a semantic segmentation device for cross-domain remote sensing images, the device comprising at least one processing unit and at least one storage unit; wherein, the storage unit stores a computer program, which, when executed by the processing unit, causes the processing unit to perform the steps of any of the above methods.

[0014] Thirdly, the present invention provides a storage medium storing a computer program executable by an access authentication device, which, when run on the access authentication device, causes the access authentication device to perform the steps of any of the above methods.

[0015] In general, the technical solution conceived in this invention can achieve the following beneficial effects compared with the prior art: (1) This invention provides a semantic segmentation method, device and storage medium for cross-domain remote sensing images. The model is learned from source domain data and target domain data. The model is generalized to the target domain by minimizing the difference between the feature distributions of the source domain and the target domain, so as to improve the segmentation accuracy of the target and the image.

[0016] (2) This invention provides a semantic segmentation method, device and storage medium for cross-domain remote sensing images. By using a supervision strategy based on pseudo-labels with self-training consistency and consistency regularization of mixed samples, it can not only mine the potential data distribution of the target domain, but also constrain the network to extract domain-invariant features.

[0017] (3) This invention provides a semantic segmentation method, device and storage medium for cross-domain remote sensing images. By combining self-training and adversarial training, the transferability of the network model is further improved, which can better address the domain offset problem caused by differences in imaging position and band combination in remote sensing image data. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the steps of a semantic segmentation method, device, and storage medium for cross-domain remote sensing images provided by the present invention. Figure 2 This is a schematic diagram illustrating the construction principle of a semantic segmentation model for a cross-domain remote sensing image semantic segmentation method, device, and storage medium provided by the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0021] It should be noted that, in the description of the embodiments of the present invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element.

[0022] Firstly, this invention proposes a semantic segmentation method for cross-domain remote sensing images. By directly inputting the remote sensing image data into the student network of the semantic segmentation model, the semantic segmentation results can be output.

[0023] This invention is based on a self-training cross-domain remote sensing image semantic segmentation method. The entire framework uses a teacher-student network, where the network structures of the student network and the teacher network are the same.

[0024] In the training phase, to obtain more robust domain-invariant features, the method of this invention mainly designs the entire semantic segmentation model from four aspects: source domain supervised training, pseudo-label supervision strategy and hybrid supervision strategy based on self-training consistency, and adversarial training based on uncertainty. Therefore, the semantic segmentation model can be divided into four parts: source domain image supervision branch, target domain image pseudo-label supervision branch, cross-domain image hybrid branch, and adversarial training branch.

[0025] During the testing phase, the semantic segmentation results can be obtained simply by inputting the test data into the student network. It should be noted that during the testing phase, neither the teacher network nor the discriminator network participates in the reasoning process.

[0026] like Figure 1 As shown, the construction of the semantic segmentation model includes: Step 101: Obtain remote sensing image training data, which includes source domain data and target domain data.

[0027] It should be noted that the source domain data includes a one-to-one correspondence of the source domain image and source domain label; while the target domain data only includes the target domain image and does not include a label.

[0028] Step 102: Input the source domain data into the student network for training to obtain the first loss function.

[0029] In the source domain image supervised branch, the source domain image is input into the student network. Similar to other semantic segmentation methods, the student network is trained in a supervised learning manner, using the pixel-level cross-entropy function as the first loss function to obtain the first loss function.

[0030] More specifically, such as Figure 2 As shown, the source domain image is input into the student network for training to obtain the source domain prediction result; based on the source domain prediction result and the source domain label, the pixel-level cross-entropy function is obtained, and the pixel-level cross-entropy function is used as the first loss function.

[0031] Wherein, the first loss function The calculation methods include: ; in, Represents the source domain image; Indicates the source domain tag; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; A one-hot vector representing the source domain label; This represents the softmax (normalization exponent) probability of the prediction result in the source domain.

[0032] It should be noted that, , Represents the space of real numbers. Indicates the image dimension. Indicates the length of the image. This indicates the width of the image, and 3 represents the three color channels of RGB. , Indicates the image dimension. Indicates the length of the image. Indicates the width of the image. Indicates the number of categories.

[0033] Step 103: Input the target domain data into the teacher network to generate pseudo-labels, train the student network using the pseudo-labels, and obtain the second loss function.

[0034] In the target domain image pseudo-label supervision branch, since the target domain data has no label, the target domain image is input into the teacher network to obtain pseudo-labels corresponding to the target domain image, and the student network is supervised and trained by the pseudo-labels generated by the teacher network.

[0035] More specifically, such as Figure 2 As shown, the target domain image is input into the student network and the teacher network respectively, and the target domain prediction result and the pseudo label corresponding to the target domain image are obtained. Based on the target domain prediction result and the pseudo label, the second loss function is calculated.

[0036] In order to extract knowledge from the target domain, the target domain image is used during training. Input to teacher network Obtaining pseudo-labels Specifically ;in, Indicates the category of the source domain tag; Represents the target domain image; Indicates the length of the image; Indicates the width of the image.

[0037] Since the target domain data lacks labels, the student network is trained using pseudo-labels generated by the teacher network. Therefore, as an embodiment of the present invention, the second loss function... The calculation method is as follows: ; However, due to the differences in data distribution between the source and target domains, noise is inevitably introduced into the pseudo-labels. Therefore, this invention employs a self-training-based pseudo-label supervision strategy, using pixels whose confidence level in the target domain prediction results is higher than a certain threshold to calculate the pseudo-label supervision loss, thereby obtaining a second loss function.

[0038] Therefore, in a preferred embodiment of the present invention, a second loss function is calculated using pixels whose confidence level in the target domain prediction result is higher than a certain threshold. The calculation methods include: ; in, Represents the target domain image; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; A one-hot vector representing pixels whose confidence level in the target domain prediction result is higher than a certain threshold; This represents the softmax probability of the prediction result in the target domain; Indicates the threshold; This represents the one-hot vector of the pseudo-label.

[0039] It should be noted that, , Represents the space of real numbers. Indicates the image dimension. Indicates the length of the image. 3 represents the width of the image, and 3 represents the three color channels of RGB.

[0040] Step 104: Mix the source domain data and the target domain data to obtain a mixed image and a corresponding mixed label. Input the mixed image into the student network for training, and use the mixed label and the mixed prediction result output by the student network to obtain the third loss function.

[0041] Because there is a domain offset between the source and target domains, implicitly learning from these two types of data alone may not be optimal.

[0042] Therefore, in the cross-domain image mixing branch, this invention employs the ClassMix algorithm to introduce new data through cross-domain mixing. Specifically, the source domain image and the target domain image are mixed to obtain a mixed image and a corresponding mixed label. Then, the mixed image is input into the student network, and finally, the cross-entropy loss is calculated on the network output using the corresponding mixed label.

[0043] As an embodiment of the present invention, the mixed image is input into the student network to obtain the mixed prediction result; the source domain label and the pseudo label are mixed to obtain the mixed label; and a third loss function is calculated based on the mixed prediction result and the mixed label.

[0044] The process of obtaining the hybrid image and the corresponding hybrid label includes: generating a mask for half of the categories of the source domain label corresponding to the source domain image, and marking pixels belonging to the half category as 1 and pixels of other categories as 0; pasting the pixels of the source domain image corresponding to the pixels marked as 1 onto the target domain image to obtain the hybrid image; and pasting the pixels of the source domain label corresponding to the pixels marked as 1 onto the pseudo label to obtain the hybrid label.

[0045] The ClassMix algorithm selects half of the predicted classes from a given image to generate a mask, and then pastes the corresponding pixels onto a second image to obtain highly perturbed samples. As a specific embodiment of this invention, the ClassMix algorithm is used to mix the source domain image and the target domain image to obtain a mixed image and corresponding mixed labels.

[0046] Specifically, in the cross-domain image fusion branch, a mask is generated from half of the categories of the source domain image labels. , Pixels belonging to this half of the category are labeled as 1, and pixels from other categories are labeled as 0, that is: .

[0047] Then the mask The corresponding source domain image with a pixel of 1 Paste the pixels into the target domain image The above yields a mixed image. , mask The corresponding source domain label with a pixel value of 1 The pixels are pasted into the pseudo-labels of the target domain predicted by the teacher network. Above, we obtain mixed tags. That is, to obtain a one-to-one corresponding mixed image. and mixed tags for: ; .

[0048] After obtaining the blended image and blended labels, the blended image is input into the student network, and the cross-entropy loss, which is also the third loss function, is calculated on the output of the student network using the corresponding blended labels. .

[0049] Similar to the target domain pseudo-label supervision branch, the cross-domain image fusion branch introduces a self-training-based pseudo-label supervision strategy.

[0050] Therefore, in a preferred embodiment of the present invention, a weight mask map is generated using the proportion of pixels with a confidence level higher than a certain threshold in the target domain prediction results, and the third loss function is calculated based on the weight mask map; the third loss function The calculation methods include: ; in, Represents a mixed image; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; Represents a weighted mask image. ; Indicates a mask. Represents a one-hot vector. Indicates pixel ratio; Indicates mixed tags a one-hot vector; This represents the softmax probability of the mixed domain prediction result.

[0051] It should be noted that the pixel ratio The calculation method is as follows: ; in, Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; A one-hot vector representing pixels whose confidence level in the target domain prediction result is higher than a certain threshold; This represents the softmax probability of the prediction result in the target domain; This represents the threshold.

[0052] Step 105: Input the source domain prediction results of the source domain data predicted by the student network and the target domain prediction results of the target domain data predicted by the student network into the discriminator network to obtain the fourth loss function.

[0053] In the adversarial training branch, a fully convolutional discriminator network is added. The entropy maps of the source domain image predicted by the student network and the target domain image predicted by the student network are input into the discriminator network. The adversarial training loss is calculated by the discriminator network and used as the fourth loss function, thereby optimizing the student network.

[0054] During training, the student network takes either a source domain image or a target domain image as input, while the fully convolutional discriminator network extracts feature maps from the student network and attempts to predict whether the input feature map belongs to the source domain or the target domain. The student network tries to deceive the discriminator network so that the features from the two domains have similar distributions.

[0055] As an embodiment of the present invention, the fourth loss function The calculation methods include: ; in, Represents the source domain image; Represents the target domain image; These represent the parameters of the discriminator network; Indicates the length of the image; Indicates the width of the image; Represents the first loss function; This represents the entropy graph of the source domain prediction results. Input to the discriminator network middle; This represents the entropy graph of the prediction results for the target domain. Input to the discriminator network middle.

[0056] It should be noted that, firstly, Shannon entropy is used as a measure of the uncertainty of network predictions to calculate the entropy graph. For example, for a given image, its entropy graph is calculated and normalized to the range [0,1]; specifically, the entropy graph... The calculation method is as follows: ; in, Indicates the length of the input image. Indicates the width of the input image. Indicates the number of categories; This represents the softmax probability of the prediction result for the student network or the teacher network.

[0057] Next, the entropy map is input into a fully convolutional discriminator network. In this context, its parameters are expressed as The discriminator network consists of five convolutional layers with a kernel size of 4 and a stride of 2, with channel numbers of 64, 128, 256, 512, and 1 respectively. Except for the last layer, each convolutional layer is followed by a Leaky ReLU activation function. The discriminator network predicts whether the input belongs to the source or target domain; that is, a label of 1 indicates the input belongs to the source domain, and a label of 0 indicates the input belongs to the target domain. This allows for the acquisition of a fourth loss function.

[0058] It should be noted that the students' network adversarial training loss for: .

[0059] Step 106: Based on the four loss functions obtained, backpropagate and differentiate the parameters of the student network to continuously optimize the parameters of the student network; and continuously update the parameters of the teacher network based on the parameters of the student network, iterating repeatedly until the training ends, and obtain the semantic segmentation model.

[0060] In a specific embodiment of the present invention, the loss values ​​in steps 102-105 are summed to obtain the total loss; the derivative of the total loss with respect to the student network parameters is obtained by backpropagation using the gradient descent algorithm; the parameters of the student network are updated using the obtained derivative; and the parameters of the teacher network are then updated using the parameters of the student network through an exponential moving average. This process is repeated iteratively until the training is completed, resulting in a model for predicting remote sensing image segmentation results.

[0061] Finally, during the testing phase, a window was set up to slide across the image, and the image blocks of each window were input into the student network to predict the prediction results for each window, ultimately obtaining the segmentation results of the remote sensing image.

[0062] Secondly, the present invention provides a semantic segmentation device for cross-domain remote sensing images, the device comprising at least one processing unit and at least one storage unit; wherein, the storage unit stores a computer program, which, when executed by the processing unit, causes the processing unit to perform the steps of any of the above methods.

[0063] Thirdly, the present invention provides a storage medium storing a computer program executable by an access authentication device, which, when run on the access authentication device, causes the access authentication device to perform the steps of any of the above methods.

[0064] The device or storage medium has the same technical features as the method described above, and will not be repeated here.

[0065] In summary, this invention proposes a pseudo-label supervision strategy based on self-training consistency, which calculates the loss of high-quality pseudo-labels generated from target domain images to supervise network training. Furthermore, this invention proposes a consistency regularization method based on mixed samples, implicitly learning from both source and target domain data simultaneously through ClassMix to obtain highly perturbed mixed samples. In addition, it proposes entropy-based adversarial training, which positions the model's decision boundary in a low-density region, reducing the entropy of predictions in both the source and target domains, and further extracting domain-invariant features.

[0066] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0067] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0068] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some service interfaces; indirect couplings or communication connections between devices or units may be electrical or other forms.

[0069] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0070] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0071] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0072] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0073] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0074] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A semantic segmentation method for cross-domain remote sensing images, characterized in that, Remote sensing image data is input into the student network of the semantic segmentation model, and semantic segmentation results are output. The construction of this semantic segmentation model includes: Acquire remote sensing image training data, which includes source domain data and target domain data; the source domain data includes a one-to-one corresponding source domain image and source domain label; while the target domain data only includes target domain images and no labels. The source domain data is input into the student network for training to obtain a first loss function; The target domain data is input into the teacher network to generate pseudo-labels, and the student network is trained using the pseudo-labels to obtain the second loss function. The source domain data and the target domain data are mixed to obtain a mixed image; the source domain labels and the pseudo labels are mixed to obtain a mixed label; the mixed image is input into the student network for training, and a third loss function is obtained by using the mixed label and the mixed prediction result output by the student network. The source domain prediction result of the source domain data predicted by the student network and the target domain prediction result of the target domain data predicted by the student network are input into the discriminator network to obtain the fourth loss function. Based on the four loss functions obtained, the parameters of the student network are backpropagated and differentiated to continuously optimize the parameters of the student network; and the parameters of the teacher network are continuously updated based on the parameters of the student network. This process is repeated until the training ends, and the semantic segmentation model is obtained.

2. The semantic segmentation method for cross-domain remote sensing images as described in claim 1, characterized in that, The source domain data includes a one-to-one correspondence between the source domain image and the source domain label; the target domain data includes only the target domain image.

3. The semantic segmentation method for cross-domain remote sensing images as described in claim 2, characterized in that, The source domain image is input into the student network for training to obtain the source domain prediction result; a pixel-level cross-entropy function is obtained based on the source domain prediction result and the source domain label, and the pixel-level cross-entropy function is used as the first loss function. The target domain image is input into the student network and the teacher network respectively to obtain the target domain prediction result and the pseudo label corresponding to the target domain image; based on the target domain prediction result and the pseudo label, the second loss function is calculated. The mixed image is input into the student network to obtain the mixed prediction result; the source domain label and the pseudo label are mixed to obtain the mixed label; Based on the hybrid prediction results and the hybrid labels, a third loss function is calculated. Based on the source domain prediction results and the target domain prediction results, a fourth loss function is calculated.

4. The semantic segmentation method for cross-domain remote sensing images as described in claim 2, characterized in that, The first loss function The calculation methods include: ； in, Represents the source domain image; Indicates the source domain tag; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; A one-hot vector representing the source domain label; This represents the softmax probability of the prediction result in the source domain.

5. The semantic segmentation method for cross-domain remote sensing images as described in claim 2, characterized in that, The second loss function is calculated using pixels whose confidence level in the target domain prediction results is higher than a certain threshold, and the second loss function... The calculation methods include: ； in, Represents the target domain image; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; A one-hot vector representing pixels whose confidence level in the target domain prediction result is higher than a certain threshold; This represents the softmax probability of the prediction result in the target domain; Indicates the threshold; This represents the one-hot vector of the pseudo-label.

6. The semantic segmentation method for cross-domain remote sensing images as described in claim 2, characterized in that, The process of obtaining the hybrid image and the corresponding hybrid label includes: A mask is generated for half of the source domain labels corresponding to the source domain image, and pixels belonging to that half of the category are marked as 1, while pixels of other categories are marked as 0. The pixel corresponding to the source domain image marked as 1 is pasted onto the target domain image to obtain the hybrid image; the pixel corresponding to the source domain label marked as 1 is pasted onto the pseudo label to obtain the hybrid label.

7. The semantic segmentation method for cross-domain remote sensing images as described in claim 6, characterized in that, A weight mask is generated using the proportion of pixels whose confidence level in the target domain prediction results is higher than a certain threshold, and the third loss function is calculated based on the weight mask. The third loss function The calculation methods include: ； in, Represents a mixed image; Indicates a student network; Indicates the length of the image; Indicates the width of the image; Indicates the category of the source domain tag; Represents a weighted mask image. ; Indicates a mask. Represents a one-hot vector. Indicates pixel ratio; Indicates mixed tags a one-hot vector; This represents the softmax probability of the mixed domain prediction result.

8. The semantic segmentation method for cross-domain remote sensing images as described in claim 2, characterized in that, The fourth loss function The calculation methods include: ； in, Represents the source domain image; Represents the target domain image; These represent the parameters of the discriminator network; Indicates the length of the image; Indicates the width of the image; Represents the first loss function; This represents the entropy graph of the source domain prediction results. Input to the discriminator network middle; This represents the entropy graph of the prediction results for the target domain. Input to the discriminator network middle.

9. A semantic segmentation device for cross-domain remote sensing images, characterized in that, The device includes at least one processing unit and at least one storage unit; wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of the method according to any one of claims 1 to 8.

10. A storage medium, characterized in that, It stores a computer program executable by an access authentication device, which, when run on the access authentication device, causes the access authentication device to perform the steps of the method according to any one of claims 1 to 8.

Images