Cross-scale building extraction method based on pseudo-label refinement and related device

Abstract

The application discloses a cross-scale building extraction method based on pseudo-label refining and a related device, and relates to the technical field of machine learning. The method comprises the following steps: training a segmentation network based on labeled source domain data; generating initial pseudo-labels by using unlabeled target domain images based on the trained segmentation network; adaptively adjusting the weights of uncertain regions of the initial pseudo-labels based on a double-prediction branch; geometrically correcting the initial pseudo-labels based on an overlap rate judgment, generating refined pseudo-labels based on the geometric correction result and the adaptive weight adjustment result; iteratively training the trained segmentation network based on the refined pseudo-labels, extracting cross-scale buildings from target remote sensing images based on the segmentation network after the iterative training, and obtaining a cross-scale building extraction result. The application can effectively improve the building extraction precision and cross-domain adaptability of the model in cross-platform and cross-resolution remote sensing images under the condition of no target domain labeling.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and apparatus for cross-scale building extraction based on pseudo-label refinement. Background Technology

[0002] With the rapid development of high-resolution remote sensing image acquisition technology, building footprint data has become a core foundational data for applications such as smart city construction, disaster emergency assessment, real estate management, and energy planning. Deep learning-based semantic segmentation methods have made significant progress in building extraction; however, these methods typically rely on large-scale, homogeneous, and uniformly labeled data for supervised training, leading to a significant decrease in generalization ability across platforms, sensors, or scales. In practical applications, differences in resolution, spectral response, scene style, and building morphology are common among different data sources. For example, when migrating from 0.3m resolution aerial imagery to 0.03m UAV imagery, building texture details are significantly enhanced, and the target scale distribution changes, resulting in severe performance degradation of the model in the target domain.

[0003] To address these issues, unsupervised domain adaptation techniques have been extensively studied. Existing methods typically employ pseudo-label self-training mechanisms or adversarial learning mechanisms for inter-domain feature alignment. However, in cross-scale building extraction tasks, pseudo-label noise, boundary blurring, and error accumulation are particularly prominent. Especially when the target domain is unlabeled, the quality of the initial pseudo-labels directly affects subsequent training results. Some studies have proposed uncertainty weighting based on prediction inconsistency, reducing training weights for high-uncertainty regions by varying the degree of difference between different prediction branches. However, these methods still only suppress noise at the semantic probability level and do not correct building boundaries at the geometric structure level. Meanwhile, visual baseline models possess excellent geometric segmentation capabilities but lack semantic recognition capabilities, making them difficult to directly apply to building footprint extraction tasks. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art. This invention provides a cross-scale building extraction method and related apparatus based on pseudo-label refinement, which can effectively improve the building extraction accuracy and cross-domain adaptability of the model in cross-platform and cross-resolution remote sensing images under the condition of no target domain annotation.

[0005] To address the aforementioned technical problems, this invention provides a cross-scale building extraction method based on pseudo-label refinement, the method comprising: Acquire labeled source domain data and train the segmentation network based on the source domain data to obtain a trained segmentation network; Initial pseudo-labels are generated from unlabeled target domain images based on a trained segmentation network. Based on the main prediction branch and the auxiliary prediction branch, the initial pseudo-label is adaptively weighted in the uncertain region to obtain the adaptive weighting result. The initial pseudo-labels are geometrically corrected based on the overlap rate determination to obtain the geometric correction result, and refined pseudo-labels are generated based on the geometric correction result and the adaptive weight adjustment result. The trained segmentation network is iteratively self-trained based on the refined pseudo-labels to obtain the segmentation network after iterative self-training. The target remote sensing image is then acquired, and cross-scale building extraction is performed on the target remote sensing image based on the segmentation network after iterative self-training to obtain the cross-scale building extraction result.

[0006] Optionally, training the segmentation network based on the source domain data to obtain a trained segmentation network includes: The source domain data is subjected to noise reduction processing to obtain the noise-reduced source domain data; Image enhancement processing is performed on the preprocessed source domain data to obtain image-enhanced source domain data. The segmentation network is then trained based on the image-enhanced source domain data to obtain a trained segmentation network.

[0007] Optionally, the adaptive weight adjustment of the initial pseudo-label for uncertain regions based on the main prediction branch and the auxiliary prediction branch to obtain the adaptive weight adjustment result includes: Based on the trained segmentation network, feature extraction is performed on unlabeled target domain images to obtain deep semantic features and shallow semantic features; The main prediction branch predicts uncertain regions based on the deep semantic features to obtain a first prediction result; The auxiliary prediction branch predicts uncertain regions based on the shallow semantic features to obtain a second prediction result; Based on the first and second prediction results, adaptive weight adjustment is performed on the uncertain region to obtain the adaptive weight adjustment result.

[0008] Optionally, the adaptive weight adjustment of the uncertain region based on the first prediction result and the second prediction result to obtain the adaptive weight adjustment result includes: The average distribution is determined based on the first and second prediction results, and the expression for the average distribution is: , Where M is the average distribution, This is the first prediction result. This is the second prediction result; The JS divergence is determined based on the average distribution, and the expression for the JS divergence is: , in, Let JS divergence be the metric. This is the first prediction result. This is the second prediction result, where M is the average distribution. Let KL divergence be a metric. The weighted pseudo-label loss is determined based on the JS divergence, and the adaptive weight adjustment is performed on the uncertain region based on the weighted pseudo-label loss to obtain the adaptive weight adjustment result.

[0009] Optionally, the expression for the weighted pseudo-label loss is: , in, For weighted pseudo-label loss, Let JS divergence be the metric. Let exp be the standard classification loss, and exp be the exponential function.

[0010] Optionally, the geometric correction of the initial pseudo-label based on the overlap rate determination to obtain the geometric correction result includes: Several fine boundary masks are generated based on the trained segmentation network, and the intersection-union ratio of each fine boundary mask with the initial pseudo-label is calculated. The intersection-union ratio of each fine boundary mask with the initial pseudo label is used as the overlap rate. The overlap rate is compared with a preset threshold to obtain the comparison result. Based on the comparison results, the initial pseudo-labels are geometrically corrected to obtain the geometrically corrected results.

[0011] Optionally, the step of acquiring the target remote sensing image and extracting buildings across scales from the target remote sensing image based on an iteratively self-trained segmentation network to obtain the cross-scale building extraction results includes: Acquire remote sensing images of targets from different source domains; Preprocessing is performed on target remote sensing images from different source domains to obtain preprocessed target remote sensing images; The preprocessed remote sensing effects of targets from different source domains are input into the iteratively self-trained segmentation network to extract buildings across scales, thus obtaining the results of the building extraction.

[0012] In addition, the present invention also provides a cross-scale building extraction device based on pseudo-label refinement, the device comprising: Network training module: used to acquire labeled source domain data and train the segmentation network based on the source domain data to obtain a trained segmentation network; Initial pseudo-label module: used to generate initial pseudo-labels based on the trained segmentation network using unlabeled target domain images; Weight adjustment module: used to adaptively adjust the weights of the initial pseudo-labels in uncertain regions based on the main prediction branch and the auxiliary prediction branch, and obtain the adaptive weight adjustment result; The refined pseudo-label module is used to perform geometric correction on the initial pseudo-labels based on the overlap rate determination, obtain the geometric correction result, and generate refined pseudo-labels based on the geometric correction result and the adaptive weight adjustment result. Building extraction module: used to iteratively self-train the trained segmentation network based on the refined pseudo-labels to obtain the segmentation network after iterative self-training, acquire the target remote sensing image, and extract buildings across scales from the target remote sensing image based on the segmentation network after iterative self-training to obtain the cross-scale building extraction results.

[0013] In addition, the present invention also provides an electronic device, which includes a processor and a memory. The memory is used to store instructions, and the processor is used to call the instructions in the memory to cause the electronic device to execute the above-described cross-scale building extraction method based on pseudo-label refinement.

[0014] In addition, the present invention provides a computer-readable storage medium that stores computer instructions that, when executed on an electronic device, cause the electronic device to perform the above-described cross-scale building extraction method based on pseudo-label refinement.

[0015] In this embodiment of the invention, a segmentation network is trained based on labeled source domain data; initial pseudo-labels are generated using unlabeled target domain images based on the trained segmentation network; adaptive weight adjustment of uncertain regions is performed on the initial pseudo-labels based on dual prediction branches; geometric correction is performed on the initial pseudo-labels based on overlap rate determination; refined pseudo-labels are generated based on the geometric correction results and adaptive weight adjustment results; the trained segmentation network is iteratively self-trained based on the refined pseudo-labels; and cross-scale building extraction is performed on the target remote sensing images based on the iteratively self-trained segmentation network. This achieves synergistic optimization of semantic probability information and geometric structure information, effectively improving the model's building extraction accuracy and cross-domain adaptability in cross-platform and cross-resolution remote sensing images under unlabeled target domain conditions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1This is a flowchart illustrating the cross-scale building extraction method based on pseudo-label refinement in an embodiment of the present invention. Figure 2 This is a flowchart illustrating a cross-scale building extraction method based on pseudo-label refinement in another embodiment of the present invention. Figure 3 This is a schematic diagram of the structural composition of the cross-scale building extraction device based on pseudo-label refinement in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structural composition of the electronic device in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the building extraction effects of various networks in different scenarios in embodiments of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating the cross-scale building extraction method based on pseudo-label refinement in an embodiment of the present invention. The method includes: S11: Obtain labeled source domain data, and train the segmentation network based on the source domain data to obtain a trained segmentation network; In the specific implementation of this invention, labeled source domain data is acquired, and the source domain data is denoised to obtain denoised source domain data. Image enhancement processing is then performed on the preprocessed source domain data to obtain image-enhanced source domain data. The segmentation network is then trained based on the image-enhanced source domain data to obtain a trained segmentation network, providing a good starting point for the model's parameters. The model learns general building features, avoiding the blindness of training from scratch in the target domain.

[0020] S12: Generate initial pseudo-labels based on the trained segmentation network using unlabeled target domain images; In the specific implementation of this invention, the initial pseudo-labels are generated based on the trained segmentation network using unlabeled target domain images. In the absence of manual annotation, preliminary supervision signals are created for the unlabeled data, making subsequent self-training possible.

[0021] S13: Based on the main prediction branch and the auxiliary prediction branch, perform adaptive weight adjustment on the uncertain region of the initial pseudo-label to obtain the adaptive weight adjustment result; In the specific implementation of this invention, feature extraction is performed on unlabeled target domain images based on the trained segmentation network to obtain deep semantic features and shallow semantic features; the main prediction branch predicts uncertain regions based on the deep semantic features to obtain a first prediction result; the auxiliary prediction branch predicts uncertain regions based on the shallow semantic features to obtain a second prediction result; adaptive weight adjustment of uncertain regions is performed based on the first and second prediction results to obtain an adaptive weight adjustment result. In the refining stage, a probability weighting mechanism based on the inconsistency of the two prediction branches is introduced to adaptively adjust the weights of the predicted uncertain regions, which can suppress noise propagation.

[0022] S14: Based on the overlap rate determination, the initial pseudo-labels are geometrically corrected to obtain the geometric correction result, and a refined pseudo-label is generated based on the geometric correction result and the adaptive weight adjustment result; In the specific implementation of this invention, several fine boundary masks are generated based on the trained segmentation network, and the intersection-union ratio (IU) of each fine boundary mask with the initial pseudo-label is calculated. The IU of each fine boundary mask with the initial pseudo-label is used as the overlap rate, which is compared with a preset threshold to obtain a comparison result. Based on the comparison result, the initial pseudo-label is geometrically corrected to obtain a geometrically corrected result. Refined pseudo-labels are generated based on the geometrically corrected result and the adaptive weight adjustment result. By introducing geometric priors, the shortcomings of pure pixel-level classification in spatial consistency are compensated, resulting in higher-quality refined pseudo-labels and providing more reliable supervision signals for subsequent iterative training.

[0023] S15: Based on the refined pseudo-labels, the trained segmentation network is iteratively self-trained to obtain the segmentation network after iterative self-training, the target remote sensing image is acquired, and the target remote sensing image is extracted across scales based on the segmentation network after iterative self-training to obtain the cross-scale building extraction results.

[0024] In the specific implementation of this invention, the trained segmentation network is iteratively self-trained based on the refined pseudo-labels to obtain the segmentation network after iterative self-training, and target remote sensing images from different source domains are acquired; the target remote sensing images from different source domains are pre-processed to obtain pre-processed target remote sensing images; the pre-processed target remote sensing images from different source domains are input into the segmentation network after iterative self-training to extract cross-scale buildings, and cross-scale building extraction results are obtained, realizing the synergistic optimization of semantic probability information and geometric structure information, thereby significantly improving the accuracy and robustness of cross-scale building footprint extraction under the condition of no target domain labeling.

[0025] In this embodiment of the invention, a segmentation network is trained based on labeled source domain data; initial pseudo-labels are generated using unlabeled target domain images based on the trained segmentation network; adaptive weight adjustment of uncertain regions is performed on the initial pseudo-labels based on dual prediction branches; geometric correction is performed on the initial pseudo-labels based on overlap rate determination; refined pseudo-labels are generated based on the geometric correction results and adaptive weight adjustment results; the trained segmentation network is iteratively self-trained based on the refined pseudo-labels; and cross-scale building extraction is performed on the target remote sensing images based on the iteratively self-trained segmentation network. This achieves synergistic optimization of semantic probability information and geometric structure information, effectively improving the model's building extraction accuracy and cross-domain adaptability in cross-platform and cross-resolution remote sensing images under unlabeled target domain conditions.

[0026] Example 2 Please see Figure 2 , Figure 2 This is a flowchart illustrating a cross-scale building extraction method based on pseudo-label refinement according to another embodiment of the present invention, the method comprising: S201: Obtain labeled source domain data, and train the segmentation network based on the source domain data to obtain a trained segmentation network; In a specific implementation of the present invention, the step of training the segmentation network based on the source domain data to obtain a trained segmentation network includes: performing noise reduction processing on the source domain data to obtain noise-reduced source domain data; performing image enhancement processing on the preprocessed source domain data to obtain image-enhanced source domain data; and training the segmentation network based on the image-enhanced source domain data to obtain a trained segmentation network.

[0027] Specifically, labeled source domain data is acquired by collecting one or more remote sensing image datasets with precise pixel-level labels as the source domain. These datasets typically contain diverse building styles, but their distribution may differ from the target domain of the final application. The source domain data is then denoised to obtain denoised source domain data. The denoising process can employ bilateral filtering, which considers both spatial proximity and pixel value similarity during filtering, thus better preserving edge details while reducing noise.

[0028] Image enhancement processing is performed on the preprocessed source domain data to obtain enhanced source domain data. Image enhancement processing may include geometric transformation and color-photometric transformation. Geometric transformation includes random rotation and random scaling, while color-photometric transformation includes adjustments to brightness, contrast, and saturation. The segmentation network is then trained based on the enhanced source domain data to obtain a trained segmentation network. The segmentation network can be any segmentation model. In the source domain supervision stage, the basic segmentation network is trained using labeled source domain data, enabling the model to learn the initial semantic representation ability of buildings.

[0029] S202: Generate initial pseudo-labels based on unlabeled target domain images using a trained segmentation network; In the specific implementation of this invention, initial pseudo-labels are generated based on the trained segmentation network using unlabeled target domain images. In the target domain stage, the trained model generates initial pseudo-labels and probability distribution maps for the unlabeled target domain images. The trained network is then forward-propagated on the unlabeled target domain images to obtain a probability map of each pixel belonging to either a building or the background. Label generation: Through the argmax operation (taking the category with the highest probability) or by setting a threshold (e.g., probability > 0.5), the probability map is converted into hard category labels, i.e., initial pseudo-labels. This creates preliminary supervision signals for the unlabeled data without manual annotation, making subsequent self-training possible.

[0030] S203: Based on the trained segmentation network, feature extraction is performed on unlabeled target domain images to obtain deep semantic features and shallow semantic features; In the specific implementation of this invention, a pre-trained segmentation network is used as a feature extractor to perform forward propagation on the target domain image, and feature maps are extracted at different stages of the network. Shallow semantic features are usually extracted from the shallow layers of the network, such as feature maps after the first or second downsampling block. These features have high resolution and contain rich spatial detail information, such as edges, textures, corners, and colors. However, their semantic level is low, and their ability to discriminate the overall category of objects is weak. Deep semantic features are extracted from the deep layers of the network, such as feature maps after the bottleneck layer or the last downsampling block. These features have low resolution but contain rich semantic information. After multiple convolutions and downsampling, neurons have a large receptive field, enabling them to capture the overall structure and contextual relationships of buildings, and have a strong ability to discriminate the category of objects. However, their spatial detail is severely lost, and edge localization may be inaccurate.

[0031] S204: The main prediction branch predicts uncertain regions based on the deep semantic features and obtains the first prediction result; In the specific implementation of this invention, the cross-scale geometrically refined building footprint extraction framework (CSG-Net) is adopted. This framework aims to improve the domain adaptability of buildings in remote sensing images under cross-platform and cross-resolution conditions by constructing a probabilistically weighted, geometrically corrected, concatenated pseudo-label refinement mechanism, effectively solving the model performance degradation problem caused by scale and imaging differences. The framework includes two core stages: source domain supervised training and target domain pseudo-label refinement. The main prediction branch predicts uncertain regions based on the deep semantic features, obtaining a first prediction result. A prediction head is designed and connected to the deep feature output of the network. This branch typically needs to upsample the feature map to the size of the input image, and then output a probability map of each pixel belonging to a building or background through a Softmax layer, obtaining the predicted uncertain regions. The main branch receives the deep semantic features output by the encoder, providing stronger semantic discrimination capabilities.

[0032] S205: The auxiliary prediction branch predicts the uncertain region based on the shallow semantic features to obtain a second prediction result; In the specific implementation of this invention, the auxiliary prediction branch predicts uncertain regions based on the shallow semantic features, obtains a second prediction result, and designs another independent prediction head, which is connected to the shallow feature output of the network. Since the shallow feature resolution is already high, usually only a few convolutional layers are needed to adjust the number of channels, and then a Softmax layer is needed to output a probability map, obtaining the uncertain region prediction corresponding to the auxiliary prediction branch. The auxiliary branch receives relatively shallow features and focuses on the expression of local structural information.

[0033] S206: Based on the first prediction result and the second prediction result, perform adaptive weight adjustment for the uncertain region to obtain the adaptive weight adjustment result; In a specific implementation of this invention, the adaptive weight adjustment of the uncertain region based on the first and second prediction results to obtain the adaptive weight adjustment result includes: determining the average distribution based on the first and second prediction results, wherein the expression for the average distribution is: , Where M is the average distribution, This is the first prediction result. This is the second prediction result; The JS divergence is determined based on the average distribution, and the expression for the JS divergence is: , in, Let JS divergence be the metric. This is the first prediction result. This is the second prediction result, where M is the average distribution. Let KL divergence be a metric. The weighted pseudo-label loss is determined based on the JS divergence, and the adaptive weight adjustment is performed on the uncertain region based on the weighted pseudo-label loss to obtain the adaptive weight adjustment result.

[0034] Furthermore, the expression for the weighted pseudo-label loss is: , in, For weighted pseudo-label loss, Let JS divergence be the metric. Let exp be the standard classification loss, and exp be the exponential function.

[0035] Specifically, due to the difference in feature representation capabilities between the two branches, they often produce different prediction results in uncertain regions. This is especially true for unlabeled target domain samples. When making predictions, the main branch It receives finer, deeper features, while the auxiliary branches... Receiving shallower features, the inconsistency between the predictions of the two prediction branches is regarded as a manifestation of the model's uncertainty about the prediction, and is further addressed through... Quantitative measurement is performed. This paper uses the Jensen-Shannon Divergence (JSD) to measure the inconsistency between two predicted distributions, thus serving as a quantitative indicator of uncertainty. The JSD is calculated based on the KL divergence between the two distributions and their mean distribution M. The mean distribution is determined based on the first and second prediction results, and its expression is: , Where M is the average distribution, This is the first prediction result. This is the second prediction result; The JS divergence is determined based on the average distribution, and the expression for the JS divergence is: , in, Let JS divergence be the metric. This is the first prediction result. This is the second prediction result, where M is the average distribution. Let KL divergence be a metric. The value range is [0,1]. The larger the value, the higher the prediction uncertainty of the model at that position.

[0036] Based on this uncertainty measure, an adaptive weight is introduced. This weight is negatively correlated with uncertainty, allowing the model to dynamically focus on high-confidence predictions during training, effectively suppressing noise generated in uncertain regions. The weighted pseudo-label loss is determined based on the JS divergence, and the expression for the weighted pseudo-label loss is: , in, For weighted pseudo-label loss, Let JS divergence be the metric. Let be the standard classification loss, and exp be the exponential function. Based on the weighted pseudo-label loss, adaptive weight adjustment is performed on uncertain regions to obtain the adaptive weight adjustment result. In the refinement stage, a probabilistic weighting mechanism based on the inconsistency of the two prediction branches is introduced to adaptively adjust the weights of uncertain prediction regions, thereby suppressing noise propagation.

[0037] S207: Based on the overlap rate determination, the initial pseudo-labels are geometrically corrected to obtain the geometric correction result, and a refined pseudo-label is generated based on the geometric correction result and the adaptive weight adjustment result; In a specific implementation of this invention, the step of geometrically correcting the initial pseudo-label based on the overlap rate determination to obtain a geometric correction result includes: generating several fine boundary masks based on the trained segmentation network, calculating the intersection-union ratio (IU) of each fine boundary mask with the initial pseudo-label; using the IU of each fine boundary mask with the initial pseudo-label as the overlap rate, comparing the overlap rate with a preset threshold to obtain a comparison result; and geometrically correcting the initial pseudo-label based on the comparison result to obtain a geometric correction result.

[0038] Specifically, based on the trained segmentation network, several fine-grained boundary masks are generated. The intersection-union ratio (IoU) of each fine-grained boundary mask with the initial pseudo-label is calculated. This IoU is used as the overlap rate, which is compared with a preset threshold to obtain the comparison result. Based on the comparison result, the initial pseudo-label is geometrically corrected to obtain the geometrically corrected result. The algorithm receives the initial pseudo-label predicted by the model and the semantically empty but precisely boundary-defined mask generated by the model. Subsequently, the algorithm traverses each mask and calculates its IoU with the corresponding initial pseudo-label. Based on the preset threshold, if the IoU is higher than the threshold, the mask is used to replace the initial pseudo-label as the correction result; otherwise, the original pseudo-label is retained, which is the geometrically corrected result. Based on the geometrically corrected result and the adaptive weight adjustment result, refined pseudo-labels are generated. By constructing an uncertainty modeling structure for the main prediction branch and the auxiliary prediction branch, the target domain pseudo-labels are probabilistically weighted and filtered. Furthermore, the geometric prior generated by the visual basic model is introduced to structurally correct the pseudo-label boundaries, thereby generating the final refined pseudo-labels.

[0039] S208: Based on the refined pseudo-labels, the trained segmentation network is iteratively self-trained to obtain the segmentation network after iterative self-training, the target remote sensing image is acquired, and the target remote sensing image is extracted across scales based on the segmentation network after iterative self-training to obtain the cross-scale building extraction result.

[0040] In the specific implementation of this invention, the step of acquiring target remote sensing images and extracting buildings across scales from the target remote sensing images based on an iteratively self-trained segmentation network to obtain cross-scale building extraction results includes: acquiring target remote sensing images from different source domains; preprocessing the target remote sensing images from different source domains to obtain preprocessed target remote sensing images; and inputting the preprocessed target remote sensing images from different source domains into the iteratively self-trained segmentation network to extract buildings across scales to obtain cross-scale building extraction results.

[0041] Specifically, the trained segmentation network is iteratively self-trained based on the refined pseudo-labels to obtain a segmentation network after iterative self-training. Guided by the refined pseudo-labels, iterative self-training achieves synergistic optimization of semantic probability information and geometric structure information, thereby significantly improving the accuracy and robustness of cross-scale building footprint extraction under conditions without target domain annotation. Remote sensing images of targets from different source domains are acquired, representing remote sensing influences of the target domain at different resolutions or imaging conditions. These images are preprocessed to obtain preprocessed target remote sensing images, including noise reduction and image enhancement. The preprocessed target remote sensing influences from different source domains are input into the iteratively self-trained segmentation network for cross-scale building extraction, yielding cross-scale building extraction results. The contour footprints of buildings are extracted from images at different resolutions or imaging conditions, which are the extraction results.

[0042] To verify the effectiveness and universality of the proposed technical solution, numerous experiments were conducted on remote sensing datasets, and a comprehensive comparison was made with various existing semantic segmentation model domain adaptation methods. Dataset: Source Domain Inria Buildings Dataset: This dataset contains 180 high-resolution images from several cities, each 5000×5000 pixels in size, with a resolution of 0.3 meters. These images were also cropped into 1008 sub-images of 1024×1024 pixels for experimental use.

[0043] The target domain dataset uses drone aerial photography dataset with a spatial resolution of 0.03m and an image size of 1024×1024. Data annotation is performed manually using ArcGIS Pro's deep learning image annotation function, with annotations at the pixel level. A total of 824 training images and 412 test images are used.

[0044] All models were run on a workstation equipped with an NVIDIA RTX 6000 GPU and implemented using the PyTorch deep learning framework. Training was performed using the AdamW optimizer with an initial learning rate of 0.00001, momentum of 0.9, weight decay of 0.005, batch size of 4, and 20,000 iterations.

[0045] All models were trained on the Inria dataset (source domain) and evaluated on the UAV dataset (target domain). Traditional self-training methods such as PyCDA, CBST, Iast, and DCA all demonstrated some adaptability, with CBST achieving the best performance among the comparison methods at 72.08% IoU. The proposed CSG-Net achieved the best performance among all comparison methods at 73.05% IoU and 84.43% F1 score. Compared to the equally strong self-trained CBST, the cross-scale geometrically refined building footprint extraction framework CSG-Net still outperformed it. This fully demonstrates that after addressing the basic semantic noise of pseudo-labels through probabilistic weighting and other strategies, the geometric constraint algorithm also played a role. By performing hard correction of pseudo-labels at the boundary level, the method of this invention effectively overcomes the problems of building shape misalignment and contour blurring caused by drastic scale changes, thus achieving a final advantage in pixel-level classification accuracy. Furthermore, to comprehensively evaluate the actual computational cost of the models, following a similar evaluation method by Li, the training and testing times of each model were measured under a unified experimental setup. Training time was defined as the total time (h) required to complete 20,000 iterations, while testing time was defined as the total time (s) required for the model to complete one full prediction and evaluation on the entire validation set. CSG-Net's total training time (10.19 h) was slightly higher than other self-training methods, which is the necessary overhead for its more complex refinement and regularization steps. However, in the more crucial testing time, CSG-Net took 546 seconds, the shortest among all compared methods, significantly outperforming DCA (591 seconds) and PyCDA (706 seconds). This demonstrates that CSG-Net achieves high accuracy while also possessing a high level of inference efficiency.

[0046] To more intuitively compare the extraction performance of different networks in different scenarios, five representative scenarios were selected from the test set for visualization, such as... Figure 5 As shown in the figure, by comparing the extraction results of each method (columns c to g) with the true labels (column b), the performance of different networks in the face of huge domain differences can be intuitively reflected.

[0047] In densely built and complex areas like Scene 1 and Scene 3, most comparison methods exhibit severe adhesion and fragmentation, making it difficult to clearly separate individual buildings. For example, in the results of CBST and Iast, multiple buildings are incorrectly merged into a large patch. In contrast, the CSG-Net proposed in this invention better preserves the independence and structural integrity of buildings, achieving the highest discriminative power in dense areas. In areas with relatively regular building outlines like Scene 2 and Scene 5, CSG-Net demonstrates a significant advantage in boundary sharpness and geometric shape. Other methods, such as PyCda and DCA, produce predictions with blurry edges, while the building footprint outlines generated by CSG-Net are smoother, more complete, and highly consistent with the geometry of the real labels. In areas containing large industrial buildings like Scene 4, CSG-Net also performs excellently, effectively avoiding omissions within large targets and maintaining their regular rectangular outlines.

[0048] In this embodiment of the invention, a segmentation network is trained based on labeled source domain data; initial pseudo-labels are generated using unlabeled target domain images based on the trained segmentation network; adaptive weight adjustment of uncertain regions is performed on the initial pseudo-labels based on dual prediction branches; geometric correction is performed on the initial pseudo-labels based on overlap rate determination; refined pseudo-labels are generated based on the geometric correction results and adaptive weight adjustment results; the trained segmentation network is iteratively self-trained based on the refined pseudo-labels; and cross-scale building extraction is performed on the target remote sensing images based on the iteratively self-trained segmentation network. This achieves synergistic optimization of semantic probability information and geometric structure information, effectively improving the model's building extraction accuracy and cross-domain adaptability in cross-platform and cross-resolution remote sensing images under unlabeled target domain conditions.

[0049] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of the structural composition of a cross-scale building extraction device based on pseudo-label refinement in an embodiment of the present invention. The device includes: Network training module 31: used to acquire labeled source domain data and train the segmentation network based on the source domain data to obtain a trained segmentation network; Initial pseudo-label module 32: used to generate initial pseudo-labels based on the trained segmentation network using unlabeled target domain images; Weight adjustment module 33: used to perform adaptive weight adjustment on the uncertain region of the initial pseudo-label based on the main prediction branch and the auxiliary prediction branch, and obtain the adaptive weight adjustment result; Refined pseudo-label module 34: used to perform geometric correction on the initial pseudo-label based on the overlap rate determination, obtain the geometric correction result, and generate refined pseudo-label based on the geometric correction result and the adaptive weight adjustment result; Building extraction module 35: used to iteratively self-train the trained segmentation network based on the refined pseudo-labels to obtain the segmentation network after iterative self-training, acquire the target remote sensing image, and extract buildings across scales from the target remote sensing image based on the segmentation network after iterative self-training to obtain the building extraction results.

[0050] In the specific implementation of this invention, the specific implementation of the device item can be referred to the implementation of the method item above, and will not be repeated here.

[0051] In this embodiment of the invention, a segmentation network is trained based on labeled source domain data; initial pseudo-labels are generated using unlabeled target domain images based on the trained segmentation network; adaptive weight adjustment of uncertain regions is performed on the initial pseudo-labels based on dual prediction branches; geometric correction is performed on the initial pseudo-labels based on overlap rate determination; refined pseudo-labels are generated based on the geometric correction results and adaptive weight adjustment results; the trained segmentation network is iteratively self-trained based on the refined pseudo-labels; and cross-scale building extraction is performed on the target remote sensing images based on the iteratively self-trained segmentation network. This achieves synergistic optimization of semantic probability information and geometric structure information, effectively improving the model's building extraction accuracy and cross-domain adaptability in cross-platform and cross-resolution remote sensing images under unlabeled target domain conditions.

[0052] This invention provides a computer-readable storage medium storing a computer program. When executed by a processor, this program implements the cross-scale building extraction method based on pseudo-label refinement, as described in any of the above embodiments. The computer-readable storage medium includes, but is not limited to, any type of disk (including floppy disks, hard disks, optical disks, CD-ROMs, and magneto-optical disks), ROM (Read-Only Memory), RAM (Random Access Memory), EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, magnetic cards, or optical cards. In other words, the storage device includes any medium that stores or transmits information in a readable form by a device (e.g., a computer, a mobile phone), and can be a read-only memory, a disk, or an optical disk, etc.

[0053] Example 4 Please see Figure 4 , Figure 4 This is a schematic diagram of the structural composition of the electronic device in an embodiment of the present invention.

[0054] This invention also provides an electronic device, such as... Figure 4 As shown, the electronic device includes a memory 41, a processor 43, and a computer program 42 stored in the memory 41 and executable on the processor 43. Those skilled in the art will understand that... Figure 4 The illustrated electronic device does not constitute a limitation on all devices and may include more or fewer components than illustrated, or combine certain components. Memory 41 can be used to store computer program 42 and various functional modules. Processor 43 runs the computer program 42 stored in memory 41, thereby performing various functional applications and data processing of the device. Memory can be internal memory or external memory, or both. Internal memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, or random access memory. External memory may include hard disks, floppy disks, ZIP disks, USB flash drives, magnetic tapes, etc. Processor 43 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, a single-chip microcomputer, or a processor 43, or any conventional processor, etc. The processors and memories disclosed in this invention include, but are not limited to, these types of processors and memories. The processors and memories disclosed in this invention are merely examples and not intended to be limiting.

[0055] As one embodiment, the electronic device includes: one or more processors 43, a memory 41, and one or more computer programs 42, wherein the one or more computer programs 42 are stored in the memory 41 and configured to be executed by the one or more processors 43, and the one or more computer programs 42 are configured to perform the cross-scale building extraction method based on pseudo-label refinement in any of the above embodiments. For specific implementation processes, please refer to the above embodiments, which will not be repeated here.

[0056] In this embodiment of the invention, a segmentation network is trained based on labeled source domain data; initial pseudo-labels are generated using unlabeled target domain images based on the trained segmentation network; adaptive weight adjustment of uncertain regions is performed on the initial pseudo-labels based on dual prediction branches; geometric correction is performed on the initial pseudo-labels based on overlap rate determination; refined pseudo-labels are generated based on the geometric correction results and adaptive weight adjustment results; the trained segmentation network is iteratively self-trained based on the refined pseudo-labels; and cross-scale building extraction is performed on the target remote sensing images based on the iteratively self-trained segmentation network. This achieves synergistic optimization of semantic probability information and geometric structure information, effectively improving the model's building extraction accuracy and cross-domain adaptability in cross-platform and cross-resolution remote sensing images under unlabeled target domain conditions.

[0057] Furthermore, the above provides a detailed description of a cross-scale building extraction method and related apparatus based on pseudo-label refinement provided by the embodiments of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A cross-scale building extraction method based on pseudo-label refinement, characterized in that, The method includes: Acquire labeled source domain data and train the segmentation network based on the source domain data to obtain a trained segmentation network; Initial pseudo-labels are generated from unlabeled target domain images based on a trained segmentation network. Based on the main prediction branch and the auxiliary prediction branch, the initial pseudo-label is adaptively weighted in the uncertain region to obtain the adaptive weighting result. The initial pseudo-labels are geometrically corrected based on the overlap rate determination to obtain the geometric correction result, and refined pseudo-labels are generated based on the geometric correction result and the adaptive weight adjustment result. The trained segmentation network is iteratively self-trained based on the refined pseudo-labels to obtain the segmentation network after iterative self-training. The target remote sensing image is then acquired, and cross-scale building extraction is performed on the target remote sensing image based on the segmentation network after iterative self-training to obtain the cross-scale building extraction result.

2. The method for cross-scale building extraction based on pseudo-label refinement according to claim 1, characterized in that, The step of training the segmentation network based on the source domain data to obtain a trained segmentation network includes: The source domain data is subjected to noise reduction processing to obtain the noise-reduced source domain data; Image enhancement processing is performed on the preprocessed source domain data to obtain image-enhanced source domain data. The segmentation network is then trained based on the image-enhanced source domain data to obtain a trained segmentation network.

3. The method for cross-scale building extraction based on pseudo-label refinement according to claim 1, characterized in that, The adaptive weight adjustment of the initial pseudo-label for uncertain regions based on the main prediction branch and the auxiliary prediction branch, to obtain the adaptive weight adjustment result, includes: Based on the trained segmentation network, feature extraction is performed on unlabeled target domain images to obtain deep semantic features and shallow semantic features; The main prediction branch predicts uncertain regions based on the deep semantic features to obtain a first prediction result; The auxiliary prediction branch predicts uncertain regions based on the shallow semantic features to obtain a second prediction result; Based on the first and second prediction results, adaptive weight adjustment is performed on the uncertain region to obtain the adaptive weight adjustment result.

4. The method for cross-scale building extraction based on pseudo-label refinement according to claim 3, characterized in that, The adaptive weight adjustment of the uncertain region based on the first and second prediction results to obtain the adaptive weight adjustment result includes: The average distribution is determined based on the first and second prediction results, and the expression for the average distribution is: ， Where M is the average distribution, This is the first prediction result. This is the second prediction result; The JS divergence is determined based on the average distribution, and the expression for the JS divergence is: ， in, Let JS divergence be the metric. This is the first prediction result. This is the second prediction result, where M is the average distribution. Let KL divergence be a metric. The weighted pseudo-label loss is determined based on the JS divergence, and the adaptive weight adjustment is performed on the uncertain region based on the weighted pseudo-label loss to obtain the adaptive weight adjustment result.

5. The method for cross-scale building extraction based on pseudo-label refinement according to claim 4, characterized in that, The expression for the weighted pseudo-label loss is: ， in, For weighted pseudo-label loss, Let JS divergence be the metric. Let exp be the standard classification loss, and exp be the exponential function.

6. The method for cross-scale building extraction based on pseudo-label refinement according to claim 1, characterized in that, The geometric correction of the initial pseudo-label based on the overlap rate determination to obtain the geometric correction result includes: Several fine boundary masks are generated based on the trained segmentation network, and the intersection-union ratio of each fine boundary mask with the initial pseudo-label is calculated. The intersection-union ratio of each fine boundary mask with the initial pseudo label is used as the overlap rate. The overlap rate is compared with a preset threshold to obtain the comparison result. Based on the comparison results, the initial pseudo-labels are geometrically corrected to obtain the geometrically corrected results.

7. The method for cross-scale building extraction based on pseudo-label refinement according to claim 1, characterized in that, The process of acquiring the target remote sensing image and extracting buildings across scales from the target remote sensing image based on an iteratively self-trained segmentation network to obtain the cross-scale building extraction results includes: Acquire remote sensing images of targets from different source domains; Preprocessing is performed on target remote sensing images from different source domains to obtain preprocessed target remote sensing images; The preprocessed remote sensing effects of targets from different source domains are input into the iteratively self-trained segmentation network to extract buildings across scales, thus obtaining the results of the building extraction.

8. A cross-scale building extraction device based on pseudo-label refinement, characterized in that, The device includes: Network training module: used to acquire labeled source domain data and train the segmentation network based on the source domain data to obtain a trained segmentation network; Initial pseudo-label module: used to generate initial pseudo-labels based on the trained segmentation network using unlabeled target domain images; Weight adjustment module: used to adaptively adjust the weights of the initial pseudo-labels in uncertain regions based on the main prediction branch and the auxiliary prediction branch, and obtain the adaptive weight adjustment result; The refined pseudo-label module is used to perform geometric correction on the initial pseudo-labels based on the overlap rate determination, obtain the geometric correction result, and generate refined pseudo-labels based on the geometric correction result and the adaptive weight adjustment result. Building extraction module: used to iteratively self-train the trained segmentation network based on the refined pseudo-labels to obtain the segmentation network after iterative self-training, acquire the target remote sensing image, and extract buildings across scales from the target remote sensing image based on the segmentation network after iterative self-training to obtain the cross-scale building extraction results.

9. An electronic device, the electronic device comprising a processor and a memory, characterized in that, The memory is used to store instructions, and the processor is used to invoke the instructions in the memory to cause the electronic device to execute the cross-scale building extraction method based on pseudo-label refinement as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed on an electronic device, cause the electronic device to perform the cross-scale building extraction method based on pseudo-label refinement as described in any one of claims 1 to 7.

Images