Remote sensing airport element extraction method based on scene semantics and key point features
Through consistent regularization SAM and global adaptive dynamic attention module, combined with key point detection branches, the segmentation and detection problems of airport feature extraction in remote sensing images are solved, and high-precision airport feature extraction is achieved.
Patent Information
- Application Number
- CN202510592628.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-09
- Publication Date
- 2025-08-12
AI Technical Summary
The existing remote sensing image processing technology has difficulty in segmenting scene elements, blurred boundaries, and incorrect detection in airport factor extraction. The detection of the core element aircraft is difficult to adapt to multi-scale changes and occlusion conditions, resulting in low detection accuracy and low efficiency.
Using a method based on scene semantics and key point features, the airport scene elements are segmented through consistent regularization SAM, combined with the global adaptive dynamic attention module and key point detection branches, feature extraction and target positioning are optimized, and the model's adaptability in complex scenarios is enhanced.
It significantly improves the segmentation accuracy of airport scene elements and the detection accuracy of aircraft targets, and improves the practicality and efficiency of remote sensing image processing, especially under complex occlusion conditions.
Smart Images
Figure CN120472328A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to a remote sensing airport element extraction method based on scene semantics and key point features. Background Art
[0002] As key hubs in the global transportation system, the accurate extraction of airport surface elements (such as runways, aprons, and taxiways) and core elements (aircraft) is crucial for aviation safety, operational efficiency, and national security. High-resolution remote sensing technology provides rich data support for airport feature extraction, but existing methods have numerous shortcomings in segmenting airport surface elements and detecting core elements (aircraft). Traditional methods for airport surface element segmentation rely primarily on handcrafted features and predefined models, making it difficult to effectively distinguish spectrally similar regions, such as aprons and buildings, leading to false detections. Furthermore, the similarity in spectral and textural characteristics between different airport surface elements makes boundary processing difficult and blurry, easily leading to feature confusion and missegmentation. While deep learning methods have achieved breakthroughs in feature extraction and context modeling, existing methods for airport surface element segmentation still face data annotation dependencies and cost bottlenecks. They require large amounts of pixel-level annotated data, and the similarity in spectral and textural characteristics between different airport surface elements can easily lead to feature confusion. For detecting core elements (aircraft), existing technologies face challenges such as diverse object scales, feature distortion in complex environments, and feature loss due to occlusion. Aircraft targets vary widely in scale and have a polygonal structure, making a single convolution kernel inadequate for multi-scale feature extraction. Furthermore, complex environmental conditions such as changing lighting and cloud cover further complicate feature extraction. Under occlusion, key aircraft discriminant features cannot be fully rendered, interfering with the detection model's ability to accurately detect aircraft targets and increasing the risk of false and missed detections.
[0003] In summary, existing airport feature extraction techniques from remote sensing images suffer from numerous flaws in key steps such as candidate region generation, feature extraction, and target classification. These flaws limit the practicality and efficiency of target detection, especially when processing large-scale, complex remote sensing imagery. Therefore, it is necessary to develop new airport feature extraction methods to overcome these flaws and improve the accuracy and real-time performance of airport feature extraction from remote sensing images. Summary of the Invention
[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide a remote sensing airport feature extraction method based on scene semantics and key point features.
[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:
[0006] A remote sensing airport feature extraction method based on scene semantics and key point features includes the following steps:
[0007] Step 1: Build an airport surface element segmentation method based on the consistent regularized Segment Anything Model (SAM). Through dual-branch strong and weak enhancement and CutMix data enhancement, combined with an adaptive pseudo-label optimization mechanism, this method achieves accurate segmentation of airport surface elements (such as aprons, runways, buildings, and lawns). It effectively eliminates areas unrelated to aircraft targets (such as buildings and lawns), providing preprocessing support for subsequent aircraft detection tasks.
[0008] Step 2: After pre-processing and segmentation, a global adaptive dynamic attention (GADA) module is designed. This module dynamically adjusts the receptive field distribution through deformable convolution to cope with the multi-scale changes and irregular geometric shapes of aircraft targets. Combining GCNet global context modeling and ECA channel attention, it synergistically optimizes local details and global semantic information, significantly enhancing feature extraction capabilities.
[0009] In step three, a key point detection branch is introduced to accurately locate the key parts of the aircraft (such as the nose and wings), and to optimize the positioning accuracy of occluded targets using a geometric constraint mechanism to ensure that aircraft targets can still be reliably detected under complex occlusion conditions.
[0010] It should be noted that the remote sensing airport feature extraction method based on scene semantics and key point features includes an airport scene feature segmentation method based on consistency regularized SAM, a global adaptive dynamic attention (GADA) module and a key point detection branch.
[0011] It should be noted that the described airport surface element segmentation method based on consistency regularized SAM preserves the main structure and semantic information of the image through dual-branch strong and weak enhancement and CutMix data augmentation operations. An entropy-aware dynamic weighting mechanism quantifies pixel-level uncertainty, suppresses noise propagation, and improves the reliability of pseudo-labeling. It achieves precise segmentation of surface elements such as aprons and runways with limited annotated data, effectively eliminating surface element areas containing non-aircraft objects such as buildings and lawns. This provides preprocessing support for subsequent core element aircraft detection tasks.
[0012] It should be noted that step 2 also includes deformable convolution, GCNet global context modeling and ECA channel attention.
[0013] It should be noted that the deformable convolution in step 2 can dynamically adjust the convolution kernel sampling position and adaptively capture the geometric features of the local features of the aircraft target.
[0014] It should be noted that the GCNet global context modeling and ECA channel attention are introduced in the step 2 to enhance the spatial distinction ability and key channel feature weights, solve the feature distortion problem in strong reflective lighting and low contrast scenes, and enhance the network's adaptability to complex scenes.
[0015] It should be noted that step three includes a key point detection branch, which defines rigid geometric key points such as the nose tip and wing end points. By accurately locating these key parts, the positioning accuracy of the occluded target is optimized to ensure that the aircraft target can still be reliably detected under complex occlusion conditions.
[0016] Compared with the prior art, the present invention has the following beneficial effects:
[0017] 1. In order to enhance the adaptability of the model to complex airport scenes, the present invention constructs an airport scene element segmentation method based on consistency regularized SAM. Through the consistency regularized SAM method, the problem of blurred boundaries of airport scene element segmentation in complex scenes is effectively solved.
[0018] 2. Targeting aircraft target detection, a core element of airports after preprocessing, faced challenges such as multi-scale target variation, difficulty extracting features in complex environments, and feature loss due to occlusion. This led to the construction of a Global Adaptive Dynamic Attention (GADA) module. This module significantly enhances feature extraction capabilities through deformable convolution and attention mechanisms, significantly strengthening the network's ability to extract features in complex scenarios.
[0019] 3. To address the problem of aircraft targets losing key features and reducing detection accuracy in remote sensing images due to occlusion (such as clouds, buildings, etc.), a method based on key point detection is proposed. The key point detection branch optimizes the positioning accuracy of occluded targets through a geometric constraint mechanism, significantly improving the robustness and accuracy of detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 Schematic diagram of the overall structure of the airport surface element segmentation preprocessing method of the consistency regularized SAM of the present invention;
[0021] Figure 2 This is a schematic diagram of the overall structure of the aircraft detection network for dynamic attention and key point positioning of the present invention;
[0022] Figure 3 This is a schematic diagram of the structure of the global adaptive dynamic attention module of the present invention;
[0023] Figure 4 Schematic diagram comparing the traditional standard convolution and deformable convolution of the present invention;
[0024] Figure 5This is a schematic diagram of the GCNet global context modeling structure of the present invention;
[0025] Figure 6 Schematic diagram of the ECA attention structure of the present invention;
[0026] Figure 7 It is a schematic diagram of the key point branch structure of the present invention. DETAILED DESCRIPTION
[0027] The present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the following embodiments are based on the present technical solution and provide detailed implementation methods and specific operating processes, but the protection scope of the present invention is not limited to these embodiments.
[0028] like Figures 1 to 7 As shown, the present invention is a remote sensing airport feature extraction method based on scene semantics and key point features, comprising the following steps:
[0029] Step 1: Construct an airport surface element segmentation method based on consistency regularized SAM. This method addresses the problem of blurred segmentation boundaries caused by complex scene interference in airport surface element segmentation. This method uses dual-branch strong and weak enhancement and CutMix data enhancement, combined with an adaptive pseudo-label optimization mechanism, to achieve accurate segmentation of surface elements such as aprons and runways with limited annotated data. It also effectively eliminates surface element areas containing non-aircraft targets such as buildings and lawns, providing preprocessing support for the subsequent core element aircraft target detection task.
[0030] The consistency regularization theory approach, based on the smoothness and clustering assumptions, perturbs the input data to ensure consistency in the model's predictions before and after these perturbations. In the dual-branch strong and weak augmentation and CutMix data augmentation methods, the weak augmentation operation consists solely of geometric transformations, aiming to preserve the image's primary structure and semantic information. The strong augmentation operation combines photometric distortion with structural perturbations to simulate the noise, illumination variations, and local occlusions common in remote sensing imagery. For labeled data, the model uses a standard cross-entropy loss to supervise the learning of category boundaries. For unlabeled data, the predictions generated by the weak augmentation branch serve as pseudo-labels, constraining the output consistency of the strong augmentation branch. Furthermore, CutMix data augmentation is introduced to provide diverse perturbations, improving the model's robustness and generalization capabilities. In the adaptive pseudo-label optimization mechanism, an entropy-aware dynamic weighting mechanism is designed to address the inherent uncertainty of pseudo-labels in class boundary regions. Based on the entropy map predicted by the strong augmentation branch, this mechanism quantifies pixel-level uncertainty and suppresses noise propagation, improving the reliability of pseudo-labels.
[0031] Step 2: Design a global adaptive dynamic attention (GADA) module to dynamically adjust the receptive field distribution through deformable convolution to effectively cope with the multi-scale changes and irregular geometric shapes of aircraft targets; combine GCNet global context modeling and ECA channel attention to collaboratively optimize local details and global semantic information, significantly enhancing the network's feature extraction capabilities in complex scenes. Deformable convolution can better adapt to the multi-scale changes and irregular geometric shapes of aircraft targets by dynamically adjusting the sampling position of the convolution kernel. By establishing global pixel dependencies, GCNet enables the network to focus on the target area from the perspective of the entire image, significantly enhancing spatial differentiation capabilities. The ECA attention mechanism enhances the feature representation of important channels by adaptively weighting each channel, while suppressing the influence of irrelevant channels, thereby improving network performance.
[0032] Step 3: In remote sensing images, aircraft targets can be obscured by occlusion (e.g., clouds, buildings, etc.), leading to the loss of key features and reduced detection accuracy. This characteristic often results in the loss of discriminative features during target detection, increasing the difficulty of detection. Keypoint detection plays a crucial role in the research of deep learning-based target detection algorithms. Its primary function is to achieve refined aircraft detection by accurately locating these key components. In this method, the keypoint detection branch precisely locates key components of the aircraft (e.g., nose, wings), and utilizes geometric constraints (e.g., symmetry and angle constraints) to optimize the localization accuracy of occluded targets. For example, an aircraft has a bilaterally symmetrical, rigid structure. This characteristic can be exploited to infer the location of a keypoint on one side when it is occluded by using the positions of the visible keypoints on the other side and symmetry constraints. Furthermore, the spatial distribution of keypoints can be constrained by using angular relationships, such as the angle between the line connecting the wing endpoints and the central axis, further improving localization accuracy.
[0033] (1) Consistency regularization constraint: Schematic diagram as shown Figure 1 As shown. Where L cons is the consistency regularization constraint loss, y pl is a binary pseudo label. W and H represent the length and width respectively.
[0034] The calculation formula of the consistency regularization constraint is as follows:
[0035]
[0036] in:
[0037]
[0038] (2) Key point symmetry constraint: The calculation formula is:
[0039] y=kx+b
[0040] in:
[0041]
[0042] (3) Key point angle constraint: The calculation formula is:
[0043]
[0044] in:
[0045] k is the slope of the central axis, θ plane ∈[85°,95°]
[0046] (4) Key point branch loss function design: The calculation formula is:
[0047]
[0048] Specifically, if Figure 1 As shown in the figure, the overall structure of the consistency regularized SAM airport surface element segmentation preprocessing method includes the input remote sensing image, the preprocessing module, the SAM segmentation module, the dual-branch enhancement module, the adaptive pseudo-label optimization module, and the output segmentation result. The preprocessing module performs preliminary processing on the input remote sensing image, the SAM segmentation module segments the airport surface elements based on consistency regularization, the dual-branch enhancement module improves the robustness of the model through strong and weak enhancement strategies, and the adaptive pseudo-label optimization module optimizes the segmentation results, ultimately outputting accurate airport surface element segmentation results.
[0049] Specifically, if Figure 2 As shown in Figure 2, the overall architecture of the aircraft detection network using dynamic attention and keypoint localization includes a preprocessed image input, a feature extraction module, a GADA module, a keypoint detection branch, and output detection results. The feature extraction module extracts image features, the GADA module enhances feature extraction, and the keypoint detection branch locates the key parts of the aircraft, ultimately outputting the detection results of the aircraft target.
[0050] like Figure 3 As shown in Figure 1, the global adaptive dynamic attention module architecture includes an input feature map, a deformable convolutional layer, a GCNet global context modeling layer, an ECA channel attention layer, and an output enhanced feature map. The deformable convolutional layer dynamically adjusts the receptive field, the GCNet layer performs global context modeling, and the ECA layer optimizes channel attention, ultimately outputting an enhanced feature map.
[0051] like Figure 4As shown in the figure, in the comparison between traditional standard convolution and deformable convolution, traditional convolution uses a fixed convolution kernel for feature extraction, while deformable convolution dynamically adjusts the position of the convolution kernel by learning the offset to better adapt to the shape and scale changes of the target.
[0052] like Figure 5 As shown in Figure 2, the GCNet global context modeling structure includes an input feature map, a global context extraction layer, a feature fusion layer, and an output global context feature map. The global context extraction layer captures global information, and the feature fusion layer fuses global information with local features to enhance the expressiveness of features.
[0053] like Figure 6 As shown in Figure 2, the ECA attention structure consists of an input feature map, a channel attention layer, and an output channel enhanced feature map. The channel attention layer enhances the features of important channels and suppresses unimportant channels by learning the correlation between channels.
[0054] Finally, if Figure 7 As shown in Figure 1, a keypoint branch is connected in parallel to the detection network. The structure includes an input feature map, a keypoint localization layer, a geometric constraint layer, and an output keypoint localization result. The keypoint localization layer accurately locates the key parts of the aircraft, and the geometric constraint layer optimizes the localization results using geometric relationships, ultimately outputting the optimized keypoint positions.
[0055] Those skilled in the art can make various corresponding changes and modifications based on the above technical solutions and concepts, and all of these changes and modifications should be included in the scope of protection of the claims of the present invention.
Claims
1. A remote sensing airport feature extraction method based on scene semantics and key point features, characterized by: The method comprises the following steps: Step 1: Construct a preprocessing method for airport surface element segmentation based on consistency regularization. Through dual-branch strong and weak enhancement and CutMix data enhancement, combined with an adaptive pseudo-label optimization mechanism, this method achieves accurate segmentation of airport surface elements and effectively eliminates surface element areas where non-aircraft targets exist. Step 2: For the scene elements that contain the core element aircraft target after segmentation preprocessing, an aircraft detection network based on dynamic attention and key point positioning is constructed. A global adaptive dynamic attention module is designed, which dynamically adjusts the receptive field distribution through deformable convolution. This is combined with GCNet global context modeling and ECA channel attention to enhance feature extraction capabilities. Step three: introduce a key point detection branch into the aircraft detection network, accurately locate the key parts of the aircraft, and use the geometric constraint mechanism to optimize the positioning accuracy of the occluded aircraft target.
2. The remote sensing airport element extraction method based on scene semantics and key point features according to claim 1 is characterized in that: In step 1, the airport scene element segmentation method based on consistency regularized SAM, through dual-branch strong and weak enhancement and CutMix data enhancement, combined with an adaptive pseudo-label optimization mechanism, achieves accurate segmentation of scene elements such as aprons and runways under limited labeled data, and effectively eliminates scene element areas where non-aircraft targets such as buildings and lawns exist.
3. The remote sensing airport element extraction method based on scene semantics and key point features according to claim 2 is characterized in that: The dual-branch strong and weak enhancement includes a weak enhancement operation and a strong enhancement operation. The weak enhancement operation retains the main structure and semantic information of the image, and the strong enhancement operation combines photometric distortion and structural perturbation.
4. The remote sensing airport element extraction method based on scene semantics and key point features according to claim 1 is characterized in that: In step 1, the adaptive pseudo-label optimization mechanism quantifies pixel-level uncertainty and suppresses noise propagation through an entropy-aware dynamic weighting mechanism, thereby improving the reliability of pseudo-labels.
5. The remote sensing airport element extraction method based on scene semantics and key point features according to claim 1 is characterized in that: In step 2, the aircraft detection network includes designing a global adaptive dynamic attention module, dynamically adjusting the receptive field distribution through deformable convolution, and combining GCNet global context modeling and ECA channel attention to significantly enhance feature extraction capabilities.
6. The remote sensing airport element extraction method based on scene semantics and key point features according to claim 5 is characterized in that: The global adaptive dynamic attention module dynamically adjusts the convolution kernel sampling position through deformable convolution to adaptively capture the geometric features of the local features of the aircraft target; Combining GCNet global context modeling and ECA channel attention, the spatial discrimination ability and key channel feature weights are enhanced respectively, solving the feature distortion problem in strong reflective lighting and low contrast scenes.
7. The remote sensing airport element extraction method based on scene semantics and key point features according to claim 1 is characterized in that: In step three, by accurately locating the key parts of the aircraft and optimizing the positioning accuracy of the occluded targets using a geometric constraint mechanism, the robustness and accuracy of the detection are significantly improved.
8. The remote sensing airport element extraction method based on scene semantics and key point features according to claim 7 is characterized in that: The key point detection branch defines multiple rigid geometric key points at the nose tip and wing endpoints, and optimizes the positioning accuracy of the occluded target using the geometric constraints of the key points such as the nose and wings.