Lightweight deformable strip feature extraction method for remote sensing image building extraction

Abstract

The invention discloses a lightweight deformable strip feature extraction method for remote sensing image building extraction, and belongs to the technical field of remote sensing. The method comprises the steps that a deformable strip attention backbone network is constructed, and ImageNet1K is used for pre-training for 300 rounds; the method comprises the following steps: cutting a building-containing data set into 1024 * 1024 image blocks and dividing a training / verification / test set; embedding the backbone network into a remote sensing task framework; loading a pre-training weight training network; and the test model outputs a result. The backbone network is divided into four stages, down-sampling is carried out in each stage, then deformable strip attention modules are stacked, the number of the deformable strip attention modules in the first second stage and the fourth stage is two, the number of the deformable strip attention modules in the third stage is four, the modules are subjected to 1 * 1 convolution dimensionality reduction and GELU activation, and branches are subjected to depth separable convolution and strip convolution K value 7 / 9 / 11 / 13 processing and then multiplied with original features to be output. The method gives consideration to precision and efficiency, and is suitable for high-resolution remote sensing image building extraction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing, and particularly relates to a lightweight deformable strip feature extraction method for building extraction of remote sensing images. BACKGROUND

[0002] In the field of remote sensing, building extraction is a fundamental but critical task, which is widely applied in urban planning, land use monitoring, and post-disaster assessment. Although deep learning methods have significantly promoted the development of this field, the complex structure of buildings, large scale differences, and texture similarity between background regions (such as roads, trees, parking lots, etc.) and buildings still pose challenges to the feature extraction process.

[0003] Traditional feature extraction methods mainly rely on hand-designed feature extraction techniques, such as SIFT (Scale-Invariant Feature Transform), HOG (Histogram of Oriented Gradients), or Haar-like features. These methods extract edge, texture, or shape information through predefined rules, and perform well in specific simple scenarios (such as single background or fixed scale targets). However, with the improvement of remote sensing image resolution and complexity, traditional methods face serious limitations: their feature representation ability is limited by human design, and cannot effectively adapt to diverse target scales, complex backgrounds (such as forests, urban mixed areas, fog interference), and dense distribution and irregular shapes of targets. In addition, traditional methods have high computational cost when processing high-resolution images, and are difficult to capture the details of small targets or occluded targets, resulting in insufficient performance in post-processing.

[0004] With the rise of deep learning technology, especially the widespread application of convolutional neural networks (CNN), the field of remote sensing image processing has ushered in a technological revolution. Deep learning models can automatically learn high-level semantic features from raw pixel data through multiple levels of nonlinear transformation, thereby significantly improving the performance of various tasks. For example, Faster R-CNN uses a region proposal network (RPN) combined with a CNN backbone network (such as ResNet, VGG) to achieve target detection; U-Net optimizes semantic segmentation through an encoding-decoding structure; and DehazeNet and other models design specialized feature extraction layers for dehazing tasks. These models have achieved remarkable results in their respective tasks and have gradually been introduced into the field of remote sensing.

[0005] Traditional convolutional neural network (CNN) is difficult to balance fine-grained local edge and large-scale structure information, especially in small building detection and dense scene, which is prone to miss detection and false detection. In order to solve this problem, in recent years, researchers try to introduce multi-scale convolution structure, attention mechanism and spatial modeling method to enhance the modeling ability of network to scale change and spatial relationship. However, the existing method is often designed insufficiently in the selection of convolution kernel scale and the fusion of spatial weight, which is difficult to realize efficient and accurate feature enhancement.

[0006] In addition, the semantic boundary of the building is often confused with the background, especially in the area where the roof color is similar to the road material, and the scale span of the building in the remote sensing image is large, from small shed to large factory. The existing model lacks dynamic perception ability, and it is difficult to automatically adjust the convolution receptive field and spatial attention area according to the image content, resulting in insufficient generalization of feature extraction. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a lightweight deformable strip feature extraction method for building extraction of remote sensing images, which can realize efficient feature extraction and calculation optimization, adaptive strip convolution design and residual connection to improve stability.

[0008] To solve the above technical problems, the technical scheme adopted by the present application is: A lightweight deformable strip feature extraction method for building extraction of remote sensing images, comprising the following steps: Step1, constructing a deformable strip attention modeling building feature extraction backbone network, and pre-training using ImgaNet1k; Step2, cutting the public dataset or private dataset containing building categories to generate pictures of a specified size, and the number of overlapping pixels of the cut pictures meets the set value; Step3, embedding the deformable strip attention modeling building feature extraction backbone network into the corresponding remote sensing task framework respectively; Step4, loading the pre-training weight of Step1, and training the network of Step3 using the dataset cut in Step2; Step5, testing the model trained in Step4 to detect the performance of the model and generate the final result.

[0009] In the above Step1, the following steps are included: Download the ImageNet1K dataset from the official website, which is 160G in size, embed the deformable strip attention modeling building feature extraction backbone network into the official pre-training framework, and pre-train the network for 300 rounds using the official pre-training framework.

[0010] In Step 2 described above, the following steps are included: Step 2.1, download the dataset from the official website of the public remote sensing dataset and decompress it, and the decompressed dataset is divided into a training set, a validation set and a test set; Step 2.2, the decompressed dataset is cropped, and the prepared processing script is used for cutting, the number of overlapping pixels of the picture is 500 during cutting, and the cut dataset is divided into a training validation set, a validation set and a test set.

[0011] In Step 3 described above, the building feature extraction backbone network constructed by the deformable strip attention modeling is divided into the following four stages: Step 3.1, in the first stage, for an input arbitrary RGB three-color image, first, a down-sampling block is used to obtain a feature map 102, then the feature map 102 is taken as input, two proposed deformable strip attention modules are stacked to obtain an output feature map 105; The network architecture of this stage is as follows: input three-color RGB image→down-sampling block 101→feature map 102; feature map 102→deformable strip attention module 103→deformable strip attention module 104→feature map 105; Step 3.2, in the second stage, the construction process is the same as the first stage, the final output feature map 105 of the first stage is taken as input, first, a down-sampling block is used to obtain a feature map 107, then the feature map 107 is taken as input, two proposed deformable strip attention modules are stacked to obtain an output feature map 110; The network architecture of this stage is as follows: feature map 105→down-sampling block 106→feature map 107; feature map 107→deformable strip attention module 108→deformable strip attention module 109→feature map 110; Step 3.3, in the third stage, the overall process is the same as the first and second stages, the deformable strip attention module is stacked four times in the third stage, the feature map 110 output by the second stage is taken as input, a down-sampling block is used to obtain a feature map 112, then the feature map 112 is taken as input, four proposed deformable strip attention modules are stacked to obtain an output feature map 117; The network architecture of this stage is as follows: feature map 110→down-sampling block 111→feature map 112; feature map 112→deformable strip attention module 113→deformable strip attention module 114→deformable strip attention module 115→deformable strip attention module 116→feature map 117; Step3.4, the fourth stage, the construction process is the same as the first stage, using the final output feature map 117 of the third stage as input, first using the down-sampling block to obtain feature map 1119, then taking feature map 119 as input, stacking two proposed deformable strip attention modules to obtain the output feature map 122; The network architecture of this stage is as follows: Feature map 117→down-sampling block 118→feature map 119; Feature map 119→deformable strip attention module 120→deformable strip attention module 121→feature map 122.

[0012] In the above Step3, in the four stages of the deformable strip attention modeling building feature extraction backbone network of the building, the deformable strip attention module is stacked, and the network structure construction sub-step of the deformable strip attention modeling is as follows: The input feature 200 is subjected to channel dimension reduction by a convolution D1 with a size of 1*1, and then the reduced feature map is activated by a GELU function D2 to obtain an activated feature map 201; The activated feature will be operated in two branches: one branch is reserved, and the other branch will be operated as follows; The activated feature map is subjected to a depth separable convolution D3 with a size of 5*5 to obtain a feature map with local information; The obtained feature map is subjected to a linear transformation by a convolution D4 with a size of 1*1 to obtain a feature map 202; The obtained feature map 202 is input in the x-axis direction through a strip convolution D5 with a size of 1*K; The obtained feature map is input in the y-axis direction through a strip convolution D6 with a size of K*1; The obtained feature map is re-used for channel dimension operation by a 1*1 convolution to obtain a feature map 203; The obtained feature map 203 is multiplied by the activated feature map 201 to obtain the final output feature map 204.

[0013] In the above network structure construction process of the deformable strip attention modeling, the value of the strip convolution K will keep increasing with the depth of the network. In the shallow layer of the network, a set strip convolution is used to obtain texture features, and with the deepening of the network, the size of the strip convolution is gradually expanded to obtain semantic features of the building.

[0014] In the above four stages, the value of K is 7, 9, 11, and 13, respectively.

[0015] In the above Step2, the data set is cut into 1024*1024 pictures.

[0016] In Step 2 described above, the number of pixels of the cut picture overlapping is 500.

[0017] In the network training process of Step 4 described above, an AdamW optimizer is used, the network momentum is set to 0.9, the weight decay factor is set to 0.05, the learning rate is preheated through a warm-up method, the learning rate is updated using a cosine annealing strategy, the number of training rounds is set to 24 rounds, the initial learning rate is 0.0001, and random flip, rotation and scaling data augmentation operations are applied to the data set during training to improve the model generalization ability.

[0018] The lightweight deformable strip feature extraction method for building extraction of remote sensing images has the following beneficial effects: 1. A lightweight deformable strip feature extraction method for buildings in remote sensing images is proposed. In view of the challenges of fuzzy building boundaries, adjacent buildings easy to stick together and difficult to separate in remote sensing images, the invention uses the directional perception characteristics of strip convolution and the high efficiency of depth separable convolution to improve the modeling ability of local edges and long strip structures.

[0019] 2. The deformable mechanism and residual connection design are introduced on the basis of strip convolution, so that the network can dynamically adjust the convolution receptive field and spatial attention area according to the image content, effectively enhancing the robustness and feature expression ability of the model in scenes with large scale difference and complex background interference.

[0020] 3. Efficient trade-off between computational efficiency and performance accuracy. The invention greatly reduces the model parameter quantity and computational overhead through lightweight design and convolution structure optimization, meets the strict requirements of remote sensing image processing on computing resources, and improves the deployment efficiency and scalability of the model while maintaining high accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0021] The present application will be further described below in conjunction with the drawings and examples: Figure 1 The deformable strip attention modeling building feature extraction backbone network framework proposed by the present application; Figure 2 The deformable strip attention module proposed by the present application. DETAILED DESCRIPTION

[0022] The technical solutions of the present application will be described in detail below in conjunction with the drawings and examples.

[0023] A lightweight deformable strip feature extraction method for building extraction of remote sensing images comprises the following steps: Step1, construct a deformable strip attention modeling building feature extraction backbone network, pre-train using ImgaNet1k; Step2, cut the public dataset containing building categories (such as Potsdam and Vaihingen) or private dataset (such as the building dataset of the Three Gorges Reservoir area) to generate pictures of a specified size, and the number of overlapping pixels of the cut pictures meets the set value; Step3, embed the deformable strip attention modeling building feature extraction backbone network into the corresponding remote sensing task framework respectively; Step4, load the pre-training weights of Step1, and train the network of Step3 using the dataset cut in Step2; Step5, test the model trained in Step4 to detect the performance of the model and generate the final result.

[0024] In Step1 above, the following steps are included: Download the ImageNet1K dataset from the official website, which is 160G in size, embed the deformable strip attention modeling building feature extraction backbone network into the official pre-training framework, and pre-train for 300 rounds using the official pre-training framework.

[0025] In Step2 above, the following steps are included: Step2.1, download the dataset from the official website of some public remote sensing datasets and decompress it. For example, Potsdam, the decompressed dataset is divided into training set, validation set and test set; Step2.2, the original dataset pictures are large, and direct training requires a lot of training resources, so the decompressed dataset needs to be cropped. Use the written processing script to cut, and the number of overlapping pixels when cutting is 500. The cut dataset is divided into training validation set, validation set and test set.

[0026] In Step3 above, the theme architecture of the deformable strip attention modeling building feature extraction backbone network constructed is as shown in Figure 1 The deformable strip attention modeling building feature extraction backbone network is mainly divided into the following four stages: Step3.1, in the first stage, for an input arbitrary RGB three-color image, first adopt a downsampling block to obtain a feature map 102, then take the feature map 102 as input, stack two proposed deformable strip attention modules to obtain an output feature map 105; The network architecture of this stage is as follows: Input trichromatic RGB image -> down-sampling block 101 -> feature map 102; Feature map 102 -> deformable strip attention module 103 -> deformable strip attention module 104 -> feature map 105; Step3.2, the second stage, the construction process is the same as the first stage, using the final output feature map 105 of the first stage as input, first using the down-sampling block to obtain the feature map 107, then taking the feature map 107 as input, stacking two proposed deformable strip attention modules to obtain the output feature map 110; The network architecture of this stage is as follows: Feature map 105 -> down-sampling block 106 -> feature map 107; Feature map 107 -> deformable strip attention module 108 -> deformable strip attention module 109 -> feature map 110; Step3.3, the third stage, the overall process is the same as the first two stages, the third stage stacks four deformable strip attention modules, takes the feature map 110 output by the second stage as input, uses a down-sampling block to obtain a feature map 112, then takes the feature map 112 as input, stacks four proposed deformable strip attention modules to obtain an output feature map 117; The network architecture of this stage is as follows: Feature map 110 -> down-sampling block 111 -> feature map 112; Feature map 112 -> deformable strip attention module 113 -> deformable strip attention module 114 -> deformable strip attention module 115 -> deformable strip attention module 116 -> feature map 117; Step3.4, the fourth stage, the construction process is the same as the first stage, using the final output feature map 117 of the third stage as input, first using the down-sampling block to obtain the feature map 1119, then taking the feature map 119 as input, stacking two proposed deformable strip attention modules to obtain the output feature map 122; The network architecture of this stage is as follows: Feature map 117 -> down-sampling block 118 -> feature map 119; Feature map 119 -> deformable strip attention module 120 -> deformable strip attention module 121 -> feature map 122.

[0027] In the above Step3, in the four stages of the building feature extraction backbone network constructed by the deformable strip attention modeling, the deformable strip attention module is mainly stacked, and the network structure construction sub-step of the deformable strip attention modeling is as follows: The input feature 200 is subjected to channel dimension reduction by a convolution D1 with a size of 1*1, and then the reduced feature map is activated by a GELU function D2 to obtain an activated feature map 201. The activated feature will be subjected to two branch operations: one branch is retained, and the other branch will be subjected to the following operations. The activated feature map is subjected to a depth separable convolution D3 with a size of 5*5 to obtain a feature map with local information. The obtained feature map is subjected to linear transformation by a convolution D4 with a size of 1*1 to obtain a feature map 202. The obtained feature map 202 is input to a bar convolution D5 with a size of 1*K in the x-axis direction. The obtained feature map is input to a bar convolution D6 with a size of K*1 in the y-axis direction. The obtained feature map is subjected to channel dimension increasing operation by a 1*1 convolution to obtain a feature map 203. The obtained feature map 203 is multiplied by the activated feature map 201 to obtain a final output feature map 204.

[0028] In the network structure construction process of the deformable strip attention modeling described above, the value of the bar convolution K will keep increasing with the depth of the network. In the shallow layer of the network, a smaller bar convolution is used to obtain texture features, and with the deepening of the network, the size of the bar convolution is gradually expanded to obtain semantic features of the building.

[0029] In the four stages described above, the value of K is 7, 9, 11, and 13, respectively.

[0030] In Step 2 described above, the data set is cut into 1024*1024 pictures.

[0031] In Step 2 described above, the number of overlapping pixels of the cut pictures is 500.

[0032] In the network training process of Step 4 described above, the AdamW optimizer is used, the network momentum is set to 0.9, the weight decay factor is set to 0.05, the learning rate is preheated by the warm-up method, the learning rate is updated by the cosine annealing strategy, the training number is set to 24 rounds, the initial learning rate is 0.0001, and the data set is subjected to random flip, rotation and scaling data augmentation operations during training to improve the generalization ability of the model.

[0033] Embodiment 1: A building feature extraction backbone network based on deformable strip attention modeling includes the following steps: Step S0: data preprocessing; cut the downloaded public dataset such as DOTA, Vaihingen to generate pictures with a size of 1024x1024. The number of overlapping pixels of the cut pictures is 500 to ensure the continuity of information between adjacent image blocks, so as to better capture the features of the target in different scales and contexts.

[0034] Step S1: model construction and pre-training; construct the proposed deformable strip attention modeling building feature extraction backbone network, the model design considers multi-level feature fusion to enhance the recognition ability of different size targets, and uses ImgaNet1k for pre-training.

[0035] Step S2: load the pre-training weights generated in step S1, embed the proposed feature extraction algorithm into the corresponding framework, and use the cut remote sensing dataset for training. In the training process, data enhancement techniques such as random flipping, rotation and scaling are used to improve the generalization ability of the model, and finally the remote sensing target detection model is obtained.

[0036] Step S3: use the weights trained in S2 to generate test results; use the trained algorithm model to detect or segment the remote sensing image to be detected or segmented, and generate the final detection or segmentation image. In the detection or segmentation process, the model analyzes the multi-scale features of the image, generates the bounding box and class label of the target, and visualizes the detection or segmentation results for further analysis and application.

[0037] The construction of the deformable strip attention feature extraction backbone network is completed through the above steps.

[0038] As shown in Figure 1 , when constructing the deformable strip attention feature extraction backbone network, the deformable strip attention module needs to be constructed first, and the detailed construction steps are as follows: The size of the input feature 200 is CxHxW. First, use a 1x1 convolution D1 to perform channel dimension reduction operation to obtain a reduced feature map; use GELU function D2 to activate the feature map to obtain an activated feature map 201.

[0039] The GELU function is approximately in the following form: ; The activated feature map 201 will be operated in two branches: one branch is directly retained, and the other branch is processed as follows: Use a 5x5 depth separable convolution D3 on the activated feature map to obtain a feature map with local information; then use a 1x1 convolution D4 to perform linear transformation to obtain a feature map 202.

[0040] The feature map 202 is respectively input into a strip convolution D5 with a size of 1xK in the x-axis direction and a strip convolution D6 with a size of Kx1 in the y-axis direction to obtain a new feature representation; and then a channel dimension increasing operation is performed through a 1x1 convolution D7 to obtain a feature map 203.

[0041] Finally, the feature map 203 is multiplied element by element with the activated feature map 201 to obtain a final output feature map 204.

[0042] In the above construction process, the value of the strip convolution K gradually increases with the network depth. A smaller K value is used in the shallow stage to obtain texture features, and the K value is gradually expanded to obtain semantic features as the network deepens. In the four stages, the values of K are 7, 9, 11 and 13 respectively.

[0043] The construction of the deformable strip attention module is beneficial to simultaneously obtaining multi-scale local and global information, enhancing the feature expression ability through layer-by-layer expanding strip convolution, and realizing feature screening and enhancement through element-by-element multiplication, thereby improving the modeling and representation capabilities of the network for image features.

[0044] The overall network construction process is as follows: Input three-color RGB image → down-sampling block 101 → feature map 102; Feature map 102 → deformable strip attention module 103 → deformable strip attention module 104 → feature map 105; Feature map 105 → down-sampling block 106 → feature map 107; Feature map 107 → deformable strip attention module 108 → deformable strip attention module 109 → feature map 110; Feature map 110 → down-sampling block 111 → feature map 112; Feature map 112 → deformable strip attention module 113 → deformable strip attention module 114 → deformable strip attention module 115 → deformable strip attention module 116 → feature map 117; Feature map 117 → down-sampling block 118 → feature map 119; Feature map 119 → deformable strip attention module 120 → deformable strip attention module 121 → feature map 122; When constructing the deformable strip attention module, the process is as follows: Input feature 200 (CxHxW) → 1x1 convolution D1 for channel dimension reduction → feature map → GELU function D2 activation → activated feature map 201.

[0045] Activation feature map 201 is divided into two paths: Branch 1: Keep it directly; Branch 2: Activate feature map 201 → 5×5 depthwise separable convolution D3 → Feature map → 1×1 convolution D4 → Feature map 202 → 1×K strip convolution in the x-axis direction D5 → K×1 strip convolution in the y-axis direction D6 → 1×1 convolution D7 channel dimensionality upgrade → Feature map 203.

[0046] Finally: Feature map 203 × Activate feature map 201 → Output feature map 204.

[0047] This invention demonstrates significant algorithmic innovation and performance improvement in the field of remote sensing image processing, particularly in building extraction tasks. Its core contributions are as follows: First, by introducing a deformable strip attention modeling module, combined with multi-level feature fusion and dynamic weight allocation mechanisms, the model can efficiently capture the features of buildings in remote sensing images at different scales (e.g., high-rise and low-rise buildings) and spatial structures (e.g., boundaries and context), thereby significantly enhancing the accuracy and robustness of building detection, segmentation, and related tasks. Second, in the data preprocessing and model training stages, high-resolution image segmentation (1024×1024 pixels, 500-pixel overlap) and data augmentation techniques (e.g., random rotation, brightness adjustment) are employed, and depthwise separable convolutional layers are used for feature extraction, effectively improving the model's generalization ability and adaptability to complex backgrounds such as occlusion and material diversity. Furthermore, through the network's adaptive dynamic kernel-space adjustment and attention mechanism, the ability to express key building features such as roof edges and multi-layer structures has been enhanced, significantly improving the practicality and cross-task transfer potential of the backbone network, and providing solid technical support for the application of building extraction in urban planning and disaster monitoring.

[0048] Example 2: To explain the feasibility of this scheme, further explanation will be provided below with experimental data.

[0049] (1) Experimental setup: To improve efficiency and save training time, the building feature extraction backbone network based on deformable strip attention modeling proposed in the application is first pre-trained on the ImageNet1K dataset for 300 epochs to obtain high-quality pre-training weights. This experiment designs a comparative experiment for two downstream tasks of building classes in remote sensing images, covering two public segmentation datasets (Potsdam and Vaihingen) and a private building detection dataset. Specifically, in the target detection task, Oriented-RCNN (O-RCNN) is used as the overall framework, and the building feature extraction backbone network based on deformable strip attention modeling proposed in the application is embedded as the backbone network. In the image segmentation task, UNetformer is used as the overall framework, and the backbone network is embedded as the feature extraction backbone network. In the change detection task, Changer is used as the overall framework, and the backbone network is also embedded as the backbone network.

[0050] (2) Dataset setup: To verify the building nature of the method of the application, training and verification are performed on three mainstream datasets: In the target detection task, to prove the effectiveness of our invention, we propose a remote sensing image dataset including two classes of unfinished buildings and completed buildings, named YCBD (Yi Chang Building Dataset). These images were taken by a drone in the rural area of Yichang, China, with a spatial resolution of 0.2 to 0.4 m and a shooting height of 500-1000 m. By dividing the original image into 1000 × 1000 pixel segments, to ensure roughly balanced distribution, the training set, validation set and test set are randomly divided into approximately 6:2:2 ratio.

[0051] In the image segmentation task, the Vaihingen and Potsdam datasets are selected for experimental verification. The Vaihingen dataset contains 33 high-resolution aerial images covering the suburban area of Vaihingen, Germany, with a resolution of about 9 centimeters per pixel, composed of red, green, blue (RGB) and infrared (IR) channels, with a total coverage area of about 1.4 square kilometers, providing detailed pixel-level annotations, divided into five categories: buildings, vegetation, low vegetation, trees, and ground. The images contain shadows, occlusions, and lighting changes, making them suitable for testing the model's robustness in complex urban environments and building edge localization. The Potsdam dataset contains 38 high-resolution aerial images covering the city center and its surrounding areas in Potsdam, Germany, with a resolution of about 5 centimeters per pixel, including RGB and IR channels, with a total coverage area of about 3.4 square kilometers, providing detailed pixel-level annotations, covering six categories: buildings, vegetation, roads, ground, low vegetation, and vehicles. The data has strong diversity, including high-resolution details and spectral overlap, making it suitable for evaluating the model's adaptability in fine segmentation and multi-class scenarios.

[0052] In the change detection task, the LEVIR-CD dataset is selected for experimental verification. This dataset contains 637 pairs of high-resolution remote sensing images (resolution about 0.5m / pixel), covering multiple cities in China, with a time span from 2002 to 2018, and annotated building change areas, with a total of about 31,333 samples, covering urban expansion, building demolition, and other change types. Its high resolution and multi-time characteristics make it an ideal test platform for change detection tasks.

[0053] (3) Parameter settings: All the above methods use the AdamW optimizer for optimization, with a network momentum of 0.9 and a weight decay factor of 0.05. The warm-up method is used for preheating, and cosine annealing is used to update the learning rate. The training rounds are 24 rounds, and the initial learning rate is 0.0001.

[0054] (4) Evaluation indicators: To comprehensively evaluate the performance of the building feature extraction backbone network based on deformable strip attention modeling proposed in the present application in different remote sensing tasks, we selected evaluation indicators highly relevant to the characteristics of each task. In the target detection task, AP50 was used as the main evaluation indicator. AP50 refers to using 0.5 as the threshold of IoU, Intersection over Union, when calculating the average precision (AP). This indicator can effectively reflect the accuracy of the detector, especially for evaluating the accuracy of target positioning and classification. In the semantic segmentation task, mIoU, Mean Intersection over Union, was used as the main evaluation indicator. mIoU calculates the intersection over union of each class and takes the average value, comprehensively measuring the segmentation accuracy and boundary consistency of the model in pixel-level classification, especially suitable for multi-class segmentation tasks in complex scenes. In the change detection task, F1 score was used as the main evaluation indicator. F1 score considers the harmonic mean of precision (Precision) and recall (Recall) to evaluate the balanced performance of the model in detecting changed areas, especially suitable for handling unbalanced data sets and the robustness evaluation of multi-temporal image changes. Through the comprehensive use of the above indicators, the superior performance of the backbone network of the present application in multi-task scenarios is systematically verified.

[0055] Final experimental results: Table 1: Performance comparison and evaluation on YCBD

[0056] Table 2: Quantitative evaluation of building class and state-of-the-art method on Vaihingen and Potsdam

[0057] Table 3: Quantitative evaluation on Levir-CD and state-of-the-art method

[0058] The experimental results on multiple remote sensing data sets show that the proposed HAKSNet network significantly outperforms many existing state-of-the-art methods in terms of performance, demonstrating its strong feature extraction capability.

[0059] 1) In the building extraction target detection task, the performance evaluation based on the YCBD dataset shows that the HAKS method has achieved significant advantages in building detection accuracy, with the unfinished building UB and completed building CB categories reaching AP values of 81.20 and 80.90, significantly better than other methods such as PKINet (80.25 and 73.62) and S2ANet (78.11 and 69.13), which highlights the outstanding ability of HAKS in extracting the features of unfinished and completed buildings and accurately positioning the boundaries, especially in complex scenes. In addition, HAKS achieved an overall AP50 index of 80.20, surpassing LSKNet (77.49) and O-RCNN (75.81), further verifying its superior performance in dealing with the diversity of buildings (including differences in the construction stage) and occlusion problems in high-resolution remote sensing images.

[0060] 2) In the semantic segmentation task, the performance evaluation of the building class based on the Vaihingen and Potsdam datasets in the semantic segmentation task shows that the HAKSNet-S method performs outstandingly in building extraction. On the Vaihingen dataset, HAKSNet-S achieved mF1 of 91.79%, OA of 93.45%, and mIoU of 85.10%, significantly better than other methods, further verifying its robust adaptability to the multi-layer structure and material diversity of buildings under high-resolution conditions.

[0061] 3) In the change detection task, the performance evaluation based on the LEVIR-CD dataset shows that the HAKSNet-S method performs outstandingly in building change detection, achieving a Precision of 93.71%, a Recall of 91.75%, an F1 score of 92.41%, and an IoU of 86.45%, significantly better than other methods such as LSKNet-S (93.34%, 91.23%, 92.27%, 85.65%) and DiFormer (93.75%, 90.59%, 92.15%, 85.44%). This result highlights the superior ability of HAKSNet-S in accurately identifying building change areas (such as new construction or demolition) and balancing Precision and Recall, especially in handling complex scenarios with high resolution (0.5m / pixel) and large time span (2002-2018) in the LEVIR-CD dataset. The superior performance of HAKSNet-S further verifies the significant advantages of its deformable strip attention modeling mechanism in capturing subtle features of building changes and enhancing spatial context adaptability, providing an efficient solution for remote sensing change detection tasks.

[0062] In summary, the lightweight deformable strip feature extraction method proposed in the present application exhibits excellent performance in building detection, semantic segmentation, and change detection tasks. This method not only has innovative value in multi-task adaptation ability, spatial perception modeling precision, and parameter efficiency optimization, but also effectively addresses the challenges of building multi-scale distribution, boundary ambiguity, and complex background through adaptive strip convolution and residual feature fusion, providing an efficient and practical technical path for building extraction in high-resolution remote sensing scenes, laying a solid foundation for the development of remote sensing image intelligent processing. The present application solves the following technical problems: 1) Efficient feature extraction and computational optimization: Using pure convolution operation combined with depth separable convolution, about 70% of floating-point operations (FLOPs) are reduced compared to traditional convolution, significantly reducing parameter quantity and computational complexity. Efficient computation is achieved in high-resolution remote sensing image building feature extraction, adapting to the hardware acceleration needs of large strip convolution machines; 2) Adaptive strip convolution design: Using variable-size strip convolution at different network layers, the method flexibly adapts to the diversified scale and geometric characteristics of building targets such as edges and textures, improving feature extraction accuracy and robustness, especially for remote sensing image processing in complex backgrounds; 3) Residual connection improves stability: Through residual connection to fuse input features and convolution output attention results, the model focuses on key features, suppresses background noise, avoids gradient vanishing problems in deep networks, and ensures the training stability of high-resolution image deep networks. 4. Hardware-friendly and engineering application: The pure convolutional structure and variable strip convolutional design are highly adaptive to parallel computing devices such as GPUs / TPUs, supporting large-scale remote sensing image real-time processing, and suitable for smart city, environmental monitoring and other scenarios; 5. The network structure is simple, efficient and highly portable, and is easy to integrate into existing detection and segmentation frameworks.

Claims

1. A lightweight deformable strip feature extraction method for building extraction in remote sensing images, characterized in that, Includes the following steps: Step 1: Construct a deformable strip attention modeling backbone network for building feature extraction, and pre-train it using ImgaNet1k; Step 2: Cut the public or private dataset containing building categories into images of a specified size, ensuring that the number of overlapping pixels in the cut images meets the set value. Step 3: Embed the deformable strip attention modeling building feature extraction backbone network into the corresponding remote sensing task framework; Step 4: Load the pre-trained weights from Step 1 and train the network from Step 3 using the dataset segmented in Step 2. Step 5: Test the model trained in Step 4 to check its performance and generate the final results.

2. The lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 1, characterized in that, Step 1 includes the following steps: Download the ImageNet1K dataset from the official website (160GB in size), embed the deformable strip attention modeling building feature extraction backbone network into the official pre-training framework, and perform 300 rounds of pre-training using the official pre-training framework.

3. The lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Download the dataset from the official website of the publicly available remote sensing dataset and decompress it. The decompressed dataset is divided into training set, validation set and test set. Step 2.2: Crop the decompressed dataset using the written processing script. The number of overlapping pixels between images is 500. The cropped dataset is divided into training and validation sets, validation set, and test set.

4. A lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 1, characterized in that, In Step 3, the constructed deformable strip attention modeling backbone network for building feature extraction is divided into the following four stages: Step 3.1, First stage: For any input RGB three-color image, firstly, a downsampling block is used to obtain feature map 102. Then, feature map 102 is used as input, and two deformable strip attention modules are stacked to obtain output feature map 105. The network architecture at this stage is as follows: Input a three-color RGB image → downsampling block 101 → feature map 102; Feature map 102 → Deformable strip attention module 103 → Deformable strip attention module 104 → Feature map 105; Step 3.2, the second stage, the construction process is the same as the first stage. The final output feature map 105 of the first stage is used as input. First, a downsampling block is used to obtain feature map 107. Then, feature map 107 is used as input, and two deformable strip attention modules are stacked to obtain the output feature map 110. The network architecture at this stage is as follows: Feature map 105 → downsampling block 106 → feature map 107; Feature map 107 → Deformable strip attention module 108 → Deformable strip attention module 109 → Feature map 110; Step 3.3, the third stage, the overall process is the same as the first and second stages. In the third stage, the deformable strip attention modules are stacked four times. The feature map 110 output from the second stage is used as input, and a downsampling block is used to obtain the feature map 112. Then, the feature map 112 is used as input, and the four deformable strip attention modules proposed and constructed are stacked to obtain the output feature map 117. The network architecture at this stage is as follows: Feature map 110 → downsampling block 111 → feature map 112; Feature map 112 → Deformable strip attention module 113 → Deformable strip attention module 114 → Deformable strip attention module 115 → Deformable strip attention module 116 → Feature map 117; Step 3.4, the fourth stage, the construction process is the same as the first stage. The final output feature map 117 of the third stage is used as input. First, a downsampling block is used to obtain feature map 1119. Then, feature map 119 is used as input, and two deformable strip attention modules are stacked to obtain the output feature map 122. The network architecture at this stage is as follows: Feature map 117 → downsampling block 118 → feature map 119; Feature map 119 → Deformable strip attention module 120 → Deformable strip attention module 121 → Feature map 122.

5. A lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 4, characterized in that, In Step 3, the deformable strip attention module is stacked in the four stages of the deformable strip attention modeling building feature extraction backbone network. The sub-steps for constructing the deformable strip attention modeling network structure are as follows: The input feature 200 is subjected to channel dimensionality reduction through a 1*1 convolution D1, and then the dimensionality-reduced feature map is activated using the GELU function D2 to obtain the activated feature map 201. The activated feature will undergo two branch operations: one branch will retain it, and the other branch will perform the following operations; The activated feature map is then subjected to a 5x5 depthwise separable convolution D3 to obtain a feature map with local information. The obtained feature map is linearly transformed using a 1*1 convolution D4 to obtain feature map 202; The obtained feature map 202 is used as input and passed through a bar convolution D5 of size 1*K in the x-axis direction; The obtained feature map is used as input and passed through a bar convolution of size K*1 D6 in the y-axis direction; The obtained feature map is then subjected to channel upsizing operation using a 1*1 roll-to-roll machine to obtain feature map 203; The obtained feature map 203 is multiplied by the activation feature map 201 to obtain the final output feature map 204.

6. A lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 5, characterized in that, In the process of constructing the network structure for deformable strip attention modeling, the value of the strip convolution K increases continuously with the depth of the network. The set strip convolution is used in the shallow layers of the network to obtain texture features, while as the network deepens, the size of the strip convolution is gradually increased to obtain semantic features of buildings.

7. A lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 6, characterized in that, In the four stages, K takes values ​​of 7, 9, 11, and 13, respectively.

8. A lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 1, characterized in that, In Step 2, the dataset is cut into 1024×1024 images.

9. A lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 1, characterized in that, In Step 2, the number of overlapping pixels in the cut images is 500.

10. A lightweight deformable strip feature extraction method for building extraction in remote sensing images according to claim 1, characterized in that, In the network training process described in Step 4, the AdamW optimizer was used, the network momentum was set to 0.9 and the weight decay factor was set to 0.

05. The learning rate was warmed up using the warm-up method and updated using the cosine annealing strategy. The number of training rounds was set to 24 rounds, the initial learning rate was 0.0001, and data augmentation operations such as random flipping, rotation and scaling were applied to the dataset during training to improve the model's generalization ability.

Images